{
 "S1sh2::signage": {
  "fp": "96001d125bd436a5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S1sh2": {
  "input_fingerprint": "9311f60884b3d76e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 작은 물고기의 주둥이가 낡은 로봇 손가락 끝에 톡 맞닿은 찰나의 클로즈업.\n\nLOCATION (lock): On the open seabed, at a robot finger protruding from a submerged rubbish heap in blue water. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Old robot finger (Protruding from submerged rubbish and touching the fish's snout) — The fingertip points obliquely toward the fish from the right side; used as Small contact anchor beside the fish's face; Submerged rubbish pile (Resting on the seabed around the protruding finger); used as Lower-frame context establishing where the finger emerges.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Blue underwater daylight with restrained contrast preserves the quiet curiosity of the contact and the worn finger's readable contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the blue underwater daylight, an old robot remains buried in seabed rubbish with a finger protruding. The robot's chest logo is already worn, although the buried chest need not be exposed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 작은 물고기 (작은 머리, 작은 입, 머리 양옆의 눈) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 작은 물고기의 주둥이가 낡은 로봇 손가락 끝에 톡 맞닿은 찰나의 클로즈업.\n\nLOCATION (lock): On the open seabed, at a robot finger protruding from a submerged rubbish heap in blue water. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Old robot finger (Protruding from submerged rubbish and touching the fish's snout) — The fingertip points obliquely toward the fish from the right side; used as Small contact anchor beside the fish's face; Submerged rubbish pile (Resting on the seabed around the protruding finger); used as Lower-frame context establishing where the finger emerges.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Blue underwater daylight with restrained contrast preserves the quiet curiosity of the contact and the worn finger's readable contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the blue underwater daylight, an old robot remains buried in seabed rubbish with a finger protruding. The robot's chest logo is already worn, although the buried chest need not be exposed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 작은 물고기 (작은 머리, 작은 입, 머리 양옆의 눈) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 작은 물고기의 주둥이가 낡은 로봇 손가락 끝에 톡 맞닿은 찰나의 클로즈업.\n\nLOCATION (lock): On the open seabed, at a robot finger protruding from a submerged rubbish heap in blue water. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Old robot finger (Protruding from submerged rubbish and touching the fish's snout) — The fingertip points obliquely toward the fish from the right side; used as Small contact anchor beside the fish's face; Submerged rubbish pile (Resting on the seabed around the protruding finger); used as Lower-frame context establishing where the finger emerges.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Blue underwater daylight with restrained contrast preserves the quiet curiosity of the contact and the worn finger's readable contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the blue underwater daylight, an old robot remains buried in seabed rubbish with a finger protruding. The robot's chest logo is already worn, although the buried chest need not be exposed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 작은 물고기 (작은 머리, 작은 입, 머리 양옆의 눈) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "로봇 손가락이 화면 오른쪽에서 비스듬히 뻗어 나와 왼쪽의 작은 물고기 주둥이와 정확히 맞닿아 있음.",
    "built_space": "푸른 물속 해저면에 다양한 형태의 낡은 쓰레기 더미가 하단부를 채우며 자연스럽게 깔려 있음.",
    "entities": "작은 물고기, 녹슬고 낡은 마디가 있는 로봇 손가락, 해저 쓰레기 더미 모두 지시된 대로 나타남.",
    "hard_violations": [],
    "physics": "물고기는 물속에서 유영하며 부력을 유지하고 있고, 손가락은 하단의 쓰레기 더미에 안정적으로 고정되어 있음."
   },
   {
    "label": "B",
    "direction": "로봇 손가락이 화면 왼쪽에서 비스듬히 뻗어 나와 오른쪽의 물고기 주둥이와 맞닿아 있음 (지시된 방향과 반대임).",
    "built_space": "해저면에 얽힌 전선과 고철 등의 쓰레기 더미가 넓게 분포되어 있음.",
    "entities": "작은 물고기, 금속 질감의 낡은 로봇 손가락, 해저 쓰레기 더미가 묘사됨.",
    "hard_violations": [],
    "physics": "물고기는 수중에 자연스럽게 떠 있고, 로봇 손가락은 쓰레기 더미 사이에 단단히 박혀 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽에서 뻗어 나온 로봇 손가락과 물고기가 맞닿는 클로즈업 구도 및 수중 환경을 지시문대로 충실히 구현함."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "물고기와 로봇 손가락이 닿는 찰나는 잘 표현했으나, 손가락이 지시된 오른쪽이 아닌 왼쪽에서 뻗어 나와 구도 지시를 어김."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇 손가락이 화면 오른쪽에서 비스듬히 뻗어 나와 왼쪽의 작은 물고기 주둥이와 정확히 맞닿아 있음.",
        "built_space": "푸른 물속 해저면에 다양한 형태의 낡은 쓰레기 더미가 하단부를 채우며 자연스럽게 깔려 있음.",
        "entities": "작은 물고기, 녹슬고 낡은 마디가 있는 로봇 손가락, 해저 쓰레기 더미 모두 지시된 대로 나타남.",
        "hard_violations": [],
        "physics": "물고기는 물속에서 유영하며 부력을 유지하고 있고, 손가락은 하단의 쓰레기 더미에 안정적으로 고정되어 있음."
       },
       {
        "label": "B",
        "direction": "로봇 손가락이 화면 왼쪽에서 비스듬히 뻗어 나와 오른쪽의 물고기 주둥이와 맞닿아 있음 (지시된 방향과 반대임).",
        "built_space": "해저면에 얽힌 전선과 고철 등의 쓰레기 더미가 넓게 분포되어 있음.",
        "entities": "작은 물고기, 금속 질감의 낡은 로봇 손가락, 해저 쓰레기 더미가 묘사됨.",
        "hard_violations": [],
        "physics": "물고기는 수중에 자연스럽게 떠 있고, 로봇 손가락은 쓰레기 더미 사이에 단단히 박혀 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽에서 뻗어 나온 로봇 손가락과 물고기가 맞닿는 클로즈업 구도 및 수중 환경을 지시문대로 충실히 구현함."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "물고기와 로봇 손가락이 닿는 찰나는 잘 표현했으나, 손가락이 지시된 오른쪽이 아닌 왼쪽에서 뻗어 나와 구도 지시를 어김."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "로봇 손가락이 화면 오른쪽에서 비스듬히 뻗어 나와 왼쪽의 작은 물고기 주둥이와 정확히 맞닿아 있음.",
        "built_space": "푸른 물속 해저면에 다양한 형태의 낡은 쓰레기 더미가 하단부를 채우며 자연스럽게 깔려 있음.",
        "entities": "작은 물고기, 녹슬고 낡은 마디가 있는 로봇 손가락, 해저 쓰레기 더미 모두 지시된 대로 나타남.",
        "hard_violations": [],
        "physics": "물고기는 물속에서 유영하며 부력을 유지하고 있고, 손가락은 하단의 쓰레기 더미에 안정적으로 고정되어 있음."
       },
       {
        "label": "B",
        "direction": "로봇 손가락이 화면 왼쪽에서 비스듬히 뻗어 나와 오른쪽의 물고기 주둥이와 맞닿아 있음 (지시된 방향과 반대임).",
        "built_space": "해저면에 얽힌 전선과 고철 등의 쓰레기 더미가 넓게 분포되어 있음.",
        "entities": "작은 물고기, 금속 질감의 낡은 로봇 손가락, 해저 쓰레기 더미가 묘사됨.",
        "hard_violations": [],
        "physics": "물고기는 수중에 자연스럽게 떠 있고, 로봇 손가락은 쓰레기 더미 사이에 단단히 박혀 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "주둥이와 손끝의 접촉은 선명하지만, 손가락이 왼쪽 아래에서 뻗어 나와 명시된 오른쪽 배치를 뒤집었고 물고기도 손가락에 비해 크게 보입니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "오른쪽 쓰레기더미에서 비스듬히 나온 손끝과 왼쪽 작은 물고기의 주둥이가 만나는 클로즈업으로, 지정된 방향·배치·수중 환경을 가장 충실히 구현합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "물고기는 오른쪽에서 왼쪽 아래를 향하며 작은 입을 금속 손끝에 대고 있습니다. 손가락은 왼쪽 아래에서 오른쪽 위의 물고기 주둥이를 향합니다. 접촉 대상은 맞지만, 오른쪽에서 물고기를 향해야 하는 손가락의 화면 방향은 반대입니다.",
        "built_space": "개방된 푸른 해저이며 건축물이나 고정 설비는 없습니다. 아래쪽과 왼쪽에 금속 잔해와 케이블로 된 쓰레기더미가 있고, 그 안에서 관절이 있는 손가락 하나가 나옵니다. 쓰레기더미는 하단 맥락을 제공하지만 손가락의 출현 위치가 지정과 다릅니다.",
        "entities": "은회색 물고기 한 마리와 부식된 로봇 손가락 하나가 보입니다. 물고기는 작은 입과 머리 옆의 정상적인 눈을 갖지만 몸통이 오른쪽 경계 밖으로 잘리고, 손가락에 비해 상당히 크게 보입니다. 로봇의 가슴은 매몰되어 보이지 않으며, 사람이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "물고기는 물속에서 지느러미와 부력으로 자세를 유지하는 자연스러운 수영 상태입니다. 금속 손가락의 밑부분은 해저 쓰레기더미에 연결되어 지지되고, 주변 잔해도 바닥에 쌓여 있습니다. 주둥이와 손끝의 가벼운 접촉은 물리적으로 가능합니다."
       },
       {
        "label": "B",
        "direction": "왼쪽 물고기는 오른쪽 손끝을 향하고, 오른쪽 아래에서 솟은 손가락은 왼쪽 위의 물고기 주둥이를 향합니다. 주둥이 바로 앞에 손끝이 닿는 배치로, 지정된 접근 방향과 접촉 대상이 맞습니다.",
        "built_space": "푸른 물속의 개방된 해저가 배경이며 건축물이나 고정 설비는 없습니다. 하단과 오른쪽의 쓰레기더미에서 로봇 손가락 하나가 비스듬히 돌출됩니다. 손가락의 기부가 잔해 속에 묻혀 있어 요청한 장소와 배치가 성립합니다.",
        "entities": "작은 회색 물고기 한 마리가 작은 머리와 입, 머리 옆의 눈을 갖고 있습니다. 오른쪽에는 마모·부식된 금속 관절 손가락이 있으며, 로봇의 나머지 몸과 가슴은 노출되지 않습니다. 손가락 전체의 화면 비중은 작은 접촉 기준점이라는 설명보다 다소 크지만, 필요한 대상은 모두 식별됩니다. 사람이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "물고기는 꼬리와 지느러미를 펼친 채 수중에서 수평 자세를 유지하므로 부력과 유영으로 설명됩니다. 손가락은 오른쪽 아래 잔해 속 기부에 연결되어 지지되고, 쓰레기는 해저에 놓여 있습니다. 접촉 자세에 관통이나 무지지 부유는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "주둥이와 손끝의 접촉은 선명하지만, 손가락이 왼쪽 아래에서 뻗어 나와 명시된 오른쪽 배치를 뒤집었고 물고기도 손가락에 비해 크게 보입니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "오른쪽 쓰레기더미에서 비스듬히 나온 손끝과 왼쪽 작은 물고기의 주둥이가 만나는 클로즈업으로, 지정된 방향·배치·수중 환경을 가장 충실히 구현합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "물고기는 오른쪽에서 왼쪽 아래를 향하며 작은 입을 금속 손끝에 대고 있습니다. 손가락은 왼쪽 아래에서 오른쪽 위의 물고기 주둥이를 향합니다. 접촉 대상은 맞지만, 오른쪽에서 물고기를 향해야 하는 손가락의 화면 방향은 반대입니다.",
        "built_space": "개방된 푸른 해저이며 건축물이나 고정 설비는 없습니다. 아래쪽과 왼쪽에 금속 잔해와 케이블로 된 쓰레기더미가 있고, 그 안에서 관절이 있는 손가락 하나가 나옵니다. 쓰레기더미는 하단 맥락을 제공하지만 손가락의 출현 위치가 지정과 다릅니다.",
        "entities": "은회색 물고기 한 마리와 부식된 로봇 손가락 하나가 보입니다. 물고기는 작은 입과 머리 옆의 정상적인 눈을 갖지만 몸통이 오른쪽 경계 밖으로 잘리고, 손가락에 비해 상당히 크게 보입니다. 로봇의 가슴은 매몰되어 보이지 않으며, 사람이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "물고기는 물속에서 지느러미와 부력으로 자세를 유지하는 자연스러운 수영 상태입니다. 금속 손가락의 밑부분은 해저 쓰레기더미에 연결되어 지지되고, 주변 잔해도 바닥에 쌓여 있습니다. 주둥이와 손끝의 가벼운 접촉은 물리적으로 가능합니다."
       },
       {
        "label": "A",
        "direction": "왼쪽 물고기는 오른쪽 손끝을 향하고, 오른쪽 아래에서 솟은 손가락은 왼쪽 위의 물고기 주둥이를 향합니다. 주둥이 바로 앞에 손끝이 닿는 배치로, 지정된 접근 방향과 접촉 대상이 맞습니다.",
        "built_space": "푸른 물속의 개방된 해저가 배경이며 건축물이나 고정 설비는 없습니다. 하단과 오른쪽의 쓰레기더미에서 로봇 손가락 하나가 비스듬히 돌출됩니다. 손가락의 기부가 잔해 속에 묻혀 있어 요청한 장소와 배치가 성립합니다.",
        "entities": "작은 회색 물고기 한 마리가 작은 머리와 입, 머리 옆의 눈을 갖고 있습니다. 오른쪽에는 마모·부식된 금속 관절 손가락이 있으며, 로봇의 나머지 몸과 가슴은 노출되지 않습니다. 손가락 전체의 화면 비중은 작은 접촉 기준점이라는 설명보다 다소 크지만, 필요한 대상은 모두 식별됩니다. 사람이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "물고기는 꼬리와 지느러미를 펼친 채 수중에서 수평 자세를 유지하므로 부력과 유영으로 설명됩니다. 손가락은 오른쪽 아래 잔해 속 기부에 연결되어 지지되고, 쓰레기는 해저에 놓여 있습니다. 접촉 자세에 관통이나 무지지 부유는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.292
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.292
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1292
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "오른쪽에서 뻗어 나온 로봇 손가락과 물고기가 맞닿는 클로즈업 구도 및 수중 환경을 지시문대로 충실히 구현함."
   },
   {
    "label": "B",
    "score": 1292,
    "verdict_ko": "물고기와 로봇 손가락이 닿는 찰나는 잘 표현했으나, 손가락이 지시된 오른쪽이 아닌 왼쪽에서 뻗어 나와 구도 지시를 어김."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd70-1310-780b-b337-0259d18c9738",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S1sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:15:29.200343+00:00",
  "fingerprint": "1236dd4f5c048ec001f66071cd1afa069220333b4231f037857909818fa0099a",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S1sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S1sh2_sel.png",
  "source_sha256": "200ab420423edfb30961b9ec7d329cde1909124d0ddc9986230f18fd846cd88a",
  "file": "S1sh2_cine.png",
  "staged_sha256": "f4d14be9ad1a2ed330e44a5a2131f239622d5be69705b88572d53c2396c2235e",
  "latency_ms": 10512
 },
 "S1sh4::signage": {
  "fp": "0e5ee861cdb82ffc",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S1sh4": {
  "input_fingerprint": "c7040acec7d51cc4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수면 위에서부터 깊은 바다 속으로 뻗어 내려온 거대한 금속 크레인의 집게가 쓰레기 더미 위를 우악스럽게 덮치며 내려오는 도중, 물살을 밀어내고 있는 찰나의 프레임.\n\nLOCATION (lock): In open water immediately above a rubbish-covered seabed, beneath a crane grab descending from the surface. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Descending crane claw in the upper-center of the frame, midground, moves toward Rubbish pile along the lower edge; Rubbish pile along the lower edge in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Crane claw (Descending from above toward the rubbish and displacing water) — Its underside and separated gripping members face the low, laterally offset camera; used as Overhead threat spanning the upper composition without becoming a solid frame-filling mass; Descending bubbles (Moving down from above around the approaching claw); used as Visible directional cue linking the overhead disturbance to the approaching mechanism; Submerged rubbish pile (Below the approaching claw); used as Narrow lower-edge spatial reference preserving continuity with the contact shot.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Blue underwater daylight separates the descending metal claw and bubbles from the surrounding water with forceful but readable contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old robot remains buried in the seabed rubbish, its finger protruding and its chest logo worn. Bubbles descend through the blue water as the large crane approaches from the surface.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 작은 물고기 (작은 머리, 작은 입, 머리 양옆의 눈) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is.. That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수면 위에서부터 깊은 바다 속으로 뻗어 내려온 거대한 금속 크레인의 집게가 쓰레기 더미 위를 우악스럽게 덮치며 내려오는 도중, 물살을 밀어내고 있는 찰나의 프레임.\n\nLOCATION (lock): In open water immediately above a rubbish-covered seabed, beneath a crane grab descending from the surface. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Descending crane claw in the upper-center of the frame, midground, moves toward Rubbish pile along the lower edge; Rubbish pile along the lower edge in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Crane claw (Descending from above toward the rubbish and displacing water) — Its underside and separated gripping members face the low, laterally offset camera; used as Overhead threat spanning the upper composition without becoming a solid frame-filling mass; Descending bubbles (Moving down from above around the approaching claw); used as Visible directional cue linking the overhead disturbance to the approaching mechanism; Submerged rubbish pile (Below the approaching claw); used as Narrow lower-edge spatial reference preserving continuity with the contact shot.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Blue underwater daylight separates the descending metal claw and bubbles from the surrounding water with forceful but readable contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old robot remains buried in the seabed rubbish, its finger protruding and its chest logo worn. Bubbles descend through the blue water as the large crane approaches from the surface.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 작은 물고기 (작은 머리, 작은 입, 머리 양옆의 눈) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is.. That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수면 위에서부터 깊은 바다 속으로 뻗어 내려온 거대한 금속 크레인의 집게가 쓰레기 더미 위를 우악스럽게 덮치며 내려오는 도중, 물살을 밀어내고 있는 찰나의 프레임.\n\nLOCATION (lock): In open water immediately above a rubbish-covered seabed, beneath a crane grab descending from the surface. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Descending crane claw in the upper-center of the frame, midground, moves toward Rubbish pile along the lower edge; Rubbish pile along the lower edge in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Crane claw (Descending from above toward the rubbish and displacing water) — Its underside and separated gripping members face the low, laterally offset camera; used as Overhead threat spanning the upper composition without becoming a solid frame-filling mass; Descending bubbles (Moving down from above around the approaching claw); used as Visible directional cue linking the overhead disturbance to the approaching mechanism; Submerged rubbish pile (Below the approaching claw); used as Narrow lower-edge spatial reference preserving continuity with the contact shot.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Blue underwater daylight separates the descending metal claw and bubbles from the surrounding water with forceful but readable contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old robot remains buried in the seabed rubbish, its finger protruding and its chest logo worn. Bubbles descend through the blue water as the large crane approaches from the surface.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 작은 물고기 (작은 머리, 작은 입, 머리 양옆의 눈) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is.. That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "크레인 집게가 아래쪽 쓰레기 더미와 로봇 손가락을 향해 내려오고 있으며, 집게 중앙부에서 기포와 물살이 수직으로 강하게 밀려 내려감.",
    "built_space": "하단 전경에 쓰레기가 깔린 해저면이 위치하고 그 위로 수중 공간이 열려 있음. 카메라는 낮은 위치에서 크레인 집게의 아랫부분을 올려다보는 구도임.",
    "entities": "벌어진 상태의 거대한 금속 크레인 집게, 낡은 로봇 손가락 1개, 사각형 로고 파편, 하강하는 기포 무리. 샷 텍스트에 언급되지 않은 물고기는 올바르게 제외됨.",
    "hard_violations": [],
    "physics": "크레인 집게는 상단에서 매달려 하강하며 물리적인 수압을 발생시키고 있음. 로봇 손가락은 해저 쓰레기 더미에 단단히 박혀 지탱됨."
   },
   {
    "label": "B",
    "direction": "크레인 집게가 아래쪽 로봇 손가락을 향하고 있으며, 기포가 집게 주변에서 위아래로 퍼지고 있음.",
    "built_space": "하단 전경에 쓰레기 더미와 로봇 손가락이 배치되었고, 카메라는 낮은 위치에서 크레인을 올려다봄.",
    "entities": "여러 갈래의 발톱을 가진 크레인 집게, 낡은 로봇 손가락, 부서진 쓰레기들. 물고기는 정상적으로 생략됨.",
    "hard_violations": [
     "[gemini-pro] 크레인 집게의 중심부에 용도를 알 수 없는 기형적인 덩어리(작은 집게 형태의 구조물)가 결합되어 있어 기계적으로 불가능한 형태를 띰."
    ],
    "physics": "크레인은 하강 중이고 로봇 손가락은 바닥에 고정되어 있으나, 물살을 강하게 밀어내는 물리적 힘의 표현은 다소 약함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 쓰레기 더미(특히 우측의 사각형 파편)와 로봇 손가락의 각도를 정확히 유지했으며, 하강하는 크레인이 물살을 밀어내는 역동적인 찰나를 매우 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "레퍼런스에 존재하던 바닥 파편 디테일이 상당히 누락되었고, 크레인 집게의 구조가 기계적으로 불가능한 형태로 왜곡되어 완성도를 떨어뜨렸습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "크레인 집게가 아래쪽 쓰레기 더미와 로봇 손가락을 향해 내려오고 있으며, 집게 중앙부에서 기포와 물살이 수직으로 강하게 밀려 내려감.",
        "built_space": "하단 전경에 쓰레기가 깔린 해저면이 위치하고 그 위로 수중 공간이 열려 있음. 카메라는 낮은 위치에서 크레인 집게의 아랫부분을 올려다보는 구도임.",
        "entities": "벌어진 상태의 거대한 금속 크레인 집게, 낡은 로봇 손가락 1개, 사각형 로고 파편, 하강하는 기포 무리. 샷 텍스트에 언급되지 않은 물고기는 올바르게 제외됨.",
        "hard_violations": [],
        "physics": "크레인 집게는 상단에서 매달려 하강하며 물리적인 수압을 발생시키고 있음. 로봇 손가락은 해저 쓰레기 더미에 단단히 박혀 지탱됨."
       },
       {
        "label": "B",
        "direction": "크레인 집게가 아래쪽 로봇 손가락을 향하고 있으며, 기포가 집게 주변에서 위아래로 퍼지고 있음.",
        "built_space": "하단 전경에 쓰레기 더미와 로봇 손가락이 배치되었고, 카메라는 낮은 위치에서 크레인을 올려다봄.",
        "entities": "여러 갈래의 발톱을 가진 크레인 집게, 낡은 로봇 손가락, 부서진 쓰레기들. 물고기는 정상적으로 생략됨.",
        "hard_violations": [
         "크레인 집게의 중심부에 용도를 알 수 없는 기형적인 덩어리(작은 집게 형태의 구조물)가 결합되어 있어 기계적으로 불가능한 형태를 띰."
        ],
        "physics": "크레인은 하강 중이고 로봇 손가락은 바닥에 고정되어 있으나, 물살을 강하게 밀어내는 물리적 힘의 표현은 다소 약함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 쓰레기 더미(특히 우측의 사각형 파편)와 로봇 손가락의 각도를 정확히 유지했으며, 하강하는 크레인이 물살을 밀어내는 역동적인 찰나를 매우 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "레퍼런스에 존재하던 바닥 파편 디테일이 상당히 누락되었고, 크레인 집게의 구조가 기계적으로 불가능한 형태로 왜곡되어 완성도를 떨어뜨렸습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "크레인 집게가 아래쪽 쓰레기 더미와 로봇 손가락을 향해 내려오고 있으며, 집게 중앙부에서 기포와 물살이 수직으로 강하게 밀려 내려감.",
        "built_space": "하단 전경에 쓰레기가 깔린 해저면이 위치하고 그 위로 수중 공간이 열려 있음. 카메라는 낮은 위치에서 크레인 집게의 아랫부분을 올려다보는 구도임.",
        "entities": "벌어진 상태의 거대한 금속 크레인 집게, 낡은 로봇 손가락 1개, 사각형 로고 파편, 하강하는 기포 무리. 샷 텍스트에 언급되지 않은 물고기는 올바르게 제외됨.",
        "hard_violations": [],
        "physics": "크레인 집게는 상단에서 매달려 하강하며 물리적인 수압을 발생시키고 있음. 로봇 손가락은 해저 쓰레기 더미에 단단히 박혀 지탱됨."
       },
       {
        "label": "B",
        "direction": "크레인 집게가 아래쪽 로봇 손가락을 향하고 있으며, 기포가 집게 주변에서 위아래로 퍼지고 있음.",
        "built_space": "하단 전경에 쓰레기 더미와 로봇 손가락이 배치되었고, 카메라는 낮은 위치에서 크레인을 올려다봄.",
        "entities": "여러 갈래의 발톱을 가진 크레인 집게, 낡은 로봇 손가락, 부서진 쓰레기들. 물고기는 정상적으로 생략됨.",
        "hard_violations": [
         "크레인 집게의 중심부에 용도를 알 수 없는 기형적인 덩어리(작은 집게 형태의 구조물)가 결합되어 있어 기계적으로 불가능한 형태를 띰."
        ],
        "physics": "크레인은 하강 중이고 로봇 손가락은 바닥에 고정되어 있으나, 물살을 강하게 밀어내는 물리적 힘의 표현은 다소 약함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "벌어진 집게의 밑면과 쓰레기를 향한 접근은 잘 보이지만, 집게와 전경 더미가 커서 중경의 집게와 좁은 하단 쓰레기 띠라는 지정 구도에서 B보다 벗어난다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "상단 중앙의 분리된 집게, 그 아래 열린 물 공간, 좁은 하단 쓰레기 더미가 지정된 와이드 구도에 더 충실하며 기포와 부식된 손가락의 연속성도 유지한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "벌어진 집게 끝들이 아래쪽 쓰레기 더미를 향하고, 집게 중심 아래에 로봇 손가락이 돌출되어 있다. 기포는 상단 연결부 주변에서 집게 아래 손가락 근처까지 이어진다. 접근 경로는 맞지만 정지된 기포만으로 하강 방향을 확정할 수는 없다. 사람의 시선이나 무기는 없다.",
        "built_space": "건축물 없는 개방된 수중 공간이다. 상단 중앙에 집게 장치 하나가 있으며, 서로 떨어진 여러 집게날의 안쪽과 밑면이 보인다. 밝은 수면이 위에, 해저가 아래에 놓인다. 쓰레기는 하단 전체와 오른쪽 전경에 쌓여 있어 지정된 좁은 하단 띠보다 넓고, 집게도 비교적 가깝고 크게 보인다. 중복된 장치나 문제 되는 반사는 없다.",
        "entities": "부식된 금속 크레인 집게, 기포, 푸른 주간 수중광, 해저 폐기물이 보인다. 하단 중앙 오른쪽의 굵고 마디진 로봇 손가락은 참고 이미지처럼 폐기물에서 왼쪽 위로 돌출되며 부식 질감도 유사하다. 로봇 가슴과 마모된 표식은 판별되지 않는다. 사람과 물고기는 없으며 이 샷에서 요구되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "집게는 상단 프레임 밖으로 이어지는 금속 연결 구조에 매달려 있고, 벌어진 날들은 중앙 기구에 연결되어 있다. 로봇 손가락의 기부는 해저 쓰레기에 묻혀 지지된다. 폐기물도 바닥에 놓여 있다. 집게 주변 기포와 교란된 물은 접근 중 물을 밀어내는 상황에 부합하며, 지지 없이 떠 있는 고형물은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "집게의 네 개 주요 날 끝이 하단 중앙 쓰레기 더미 쪽으로 향한다. 집게 아래의 기포 기둥은 돌출된 로봇 손가락 뒤쪽 더미까지 이어져 접근 대상과 연결된다. 기포의 실제 이동 방향은 정지 화면만으로 확정하기 어렵지만, 위에서 아래로 이어지는 교란 경로는 명확하다. 사람의 시선이나 무기는 없다.",
        "built_space": "개방된 바닷속에 집게 장치 하나가 상단 중앙에 배치되고, 그 아래로 상당한 물 공간이 남는다. 벌어진 네 주요 날 사이로 물이 보이며 안쪽 면도 드러난다. 낮은 수중 시점이지만 좌우 비대칭은 약해 측면으로 비껴난 정도는 크지 않다. 쓰레기 더미는 주로 하단 가장자리에 제한되어 지정된 공간 배치에 가깝다. 건축 설비나 중복 장치, 불가능한 반사는 없다.",
        "entities": "거대한 부식 금속 집게, 다량의 기포, 푸른 주간 바닷물, 침전물로 덮인 금속 폐기물이 보인다. 로봇의 마디진 손가락이 하단 중앙에서 왼쪽 위로 돌출되어 참고 이미지의 방향과 매몰 상태를 유지한다. 오른쪽 판재에 사각 요철이 있으나 로봇 가슴이나 표식인지는 확정할 수 없다. 사람이나 물고기는 없고 읽을 수 있는 글자도 없다.",
        "hard_violations": [],
        "physics": "집게 상부 연결 구조가 화면 위로 이어지고 중앙 기구가 각 날을 지지한다. 손가락은 폐기물에 묻힌 기부로 지지되며 주변 판재와 막대도 더미에 걸치거나 박혀 있다. 아래로 길게 이어진 기포와 물의 교란은 하강하는 장치가 만드는 흐름으로 해석 가능하다. 지지 없이 공중이나 수중에 정지한 고형물은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "벌어진 집게의 밑면과 쓰레기를 향한 접근은 잘 보이지만, 집게와 전경 더미가 커서 중경의 집게와 좁은 하단 쓰레기 띠라는 지정 구도에서 B보다 벗어난다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "상단 중앙의 분리된 집게, 그 아래 열린 물 공간, 좁은 하단 쓰레기 더미가 지정된 와이드 구도에 더 충실하며 기포와 부식된 손가락의 연속성도 유지한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "벌어진 집게 끝들이 아래쪽 쓰레기 더미를 향하고, 집게 중심 아래에 로봇 손가락이 돌출되어 있다. 기포는 상단 연결부 주변에서 집게 아래 손가락 근처까지 이어진다. 접근 경로는 맞지만 정지된 기포만으로 하강 방향을 확정할 수는 없다. 사람의 시선이나 무기는 없다.",
        "built_space": "건축물 없는 개방된 수중 공간이다. 상단 중앙에 집게 장치 하나가 있으며, 서로 떨어진 여러 집게날의 안쪽과 밑면이 보인다. 밝은 수면이 위에, 해저가 아래에 놓인다. 쓰레기는 하단 전체와 오른쪽 전경에 쌓여 있어 지정된 좁은 하단 띠보다 넓고, 집게도 비교적 가깝고 크게 보인다. 중복된 장치나 문제 되는 반사는 없다.",
        "entities": "부식된 금속 크레인 집게, 기포, 푸른 주간 수중광, 해저 폐기물이 보인다. 하단 중앙 오른쪽의 굵고 마디진 로봇 손가락은 참고 이미지처럼 폐기물에서 왼쪽 위로 돌출되며 부식 질감도 유사하다. 로봇 가슴과 마모된 표식은 판별되지 않는다. 사람과 물고기는 없으며 이 샷에서 요구되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "집게는 상단 프레임 밖으로 이어지는 금속 연결 구조에 매달려 있고, 벌어진 날들은 중앙 기구에 연결되어 있다. 로봇 손가락의 기부는 해저 쓰레기에 묻혀 지지된다. 폐기물도 바닥에 놓여 있다. 집게 주변 기포와 교란된 물은 접근 중 물을 밀어내는 상황에 부합하며, 지지 없이 떠 있는 고형물은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "집게의 네 개 주요 날 끝이 하단 중앙 쓰레기 더미 쪽으로 향한다. 집게 아래의 기포 기둥은 돌출된 로봇 손가락 뒤쪽 더미까지 이어져 접근 대상과 연결된다. 기포의 실제 이동 방향은 정지 화면만으로 확정하기 어렵지만, 위에서 아래로 이어지는 교란 경로는 명확하다. 사람의 시선이나 무기는 없다.",
        "built_space": "개방된 바닷속에 집게 장치 하나가 상단 중앙에 배치되고, 그 아래로 상당한 물 공간이 남는다. 벌어진 네 주요 날 사이로 물이 보이며 안쪽 면도 드러난다. 낮은 수중 시점이지만 좌우 비대칭은 약해 측면으로 비껴난 정도는 크지 않다. 쓰레기 더미는 주로 하단 가장자리에 제한되어 지정된 공간 배치에 가깝다. 건축 설비나 중복 장치, 불가능한 반사는 없다.",
        "entities": "거대한 부식 금속 집게, 다량의 기포, 푸른 주간 바닷물, 침전물로 덮인 금속 폐기물이 보인다. 로봇의 마디진 손가락이 하단 중앙에서 왼쪽 위로 돌출되어 참고 이미지의 방향과 매몰 상태를 유지한다. 오른쪽 판재에 사각 요철이 있으나 로봇 가슴이나 표식인지는 확정할 수 없다. 사람이나 물고기는 없고 읽을 수 있는 글자도 없다.",
        "hard_violations": [],
        "physics": "집게 상부 연결 구조가 화면 위로 이어지고 중앙 기구가 각 날을 지지한다. 손가락은 폐기물에 묻힌 기부로 지지되며 주변 판재와 막대도 더미에 걸치거나 박혀 있다. 아래로 길게 이어진 기포와 물의 교란은 하강하는 장치가 만드는 흐름으로 해석 가능하다. 지지 없이 공중이나 수중에 정지한 고형물은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.46
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.21
   },
   "violations": {
    "B": [
     "[gemini-pro] 크레인 집게의 중심부에 용도를 알 수 없는 기형적인 덩어리(작은 집게 형태의 구조물)가 결합되어 있어 기계적으로 불가능한 형태를 띰."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1210
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스의 쓰레기 더미(특히 우측의 사각형 파편)와 로봇 손가락의 각도를 정확히 유지했으며, 하강하는 크레인이 물살을 밀어내는 역동적인 찰나를 매우 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1210,
    "verdict_ko": "레퍼런스에 존재하던 바닥 파편 디테일이 상당히 누락되었고, 크레인 집게의 구조가 기계적으로 불가능한 형태로 왜곡되어 완성도를 떨어뜨렸습니다.  ★위반: [gemini-pro] 크레인 집게의 중심부에 용도를 알 수 없는 기형적인 덩어리(작은 집게 형태의 구조물)가 결합되어 있어 기계적으로 불가능한 형태를 띰."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S1sh2_sel.png",
    "asset_id": "1acff503-1f22-4233-9b51-160cc0cca376",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd70-c07c-7b1c-97da-25630d1cd468",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S1sh2"
  },
  "lane_policy": "ab_select_bypass:bg_only:share_plan_prev_bgonly"
 },
 "S1sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:15:40.036918+00:00",
  "fingerprint": "c73d872ddcc371da6b9c8127dccee7f0d22b6076dd8f399aa232e8756ce34626",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S1sh4_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S1sh4_sel.png",
  "source_sha256": "2b0aea46c98e60a3b555a9bc497b6361e750885a6bf86a45b3124a28d84d19d4",
  "file": "S1sh4_cine.png",
  "staged_sha256": "ebd2bdc7b96365aea1d08749bcc3fc0cfd79e281146c02d246254af4f471c659",
  "latency_ms": 11912
 },
 "S2sh3::signage": {
  "fp": "d1cb5033d4281af7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S2sh3": {
  "input_fingerprint": "526a754538fe77c2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갑판 위 쏟아진 쓰레기 더미 틈에 파묻힌 채, 굵은 그물에 몸이 칭칭 감겨 있는 낡고 부서진 고철 로봇 찰리의 정지된 외관.\n\nLOCATION (lock): Among freshly dumped rubbish on the exposed deck of an ocean-going waste collection ship. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Coarse netting (Tightly wrapped around Charlie's body); used as Crossing lines that interrupt the body silhouette and communicate restraint; Rubbish on the collection ship (Deposited on the deck and partially burying Charlie); used as Uneven foreground and lateral occlusion around the robot; Collection ship deck (Supporting the deposited rubbish) — A small portion of its upper surface is visible beneath the pile; used as Lower-edge grounding that locates the discovery aboard the ship.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and measured contrast reveals the old robot's damage and net bindings without assigning him an unsupported active glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rubbish lies heaped on the collection ship's deck, with the old scrap robot Charlie entangled in a net among it. Charlie retains the worn UBIC logo on his chest.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갑판 위 쏟아진 쓰레기 더미 틈에 파묻힌 채, 굵은 그물에 몸이 칭칭 감겨 있는 낡고 부서진 고철 로봇 찰리의 정지된 외관.\n\nLOCATION (lock): Among freshly dumped rubbish on the exposed deck of an ocean-going waste collection ship. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Coarse netting (Tightly wrapped around Charlie's body); used as Crossing lines that interrupt the body silhouette and communicate restraint; Rubbish on the collection ship (Deposited on the deck and partially burying Charlie); used as Uneven foreground and lateral occlusion around the robot; Collection ship deck (Supporting the deposited rubbish) — A small portion of its upper surface is visible beneath the pile; used as Lower-edge grounding that locates the discovery aboard the ship.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and measured contrast reveals the old robot's damage and net bindings without assigning him an unsupported active glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rubbish lies heaped on the collection ship's deck, with the old scrap robot Charlie entangled in a net among it. Charlie retains the worn UBIC logo on his chest.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갑판 위 쏟아진 쓰레기 더미 틈에 파묻힌 채, 굵은 그물에 몸이 칭칭 감겨 있는 낡고 부서진 고철 로봇 찰리의 정지된 외관.\n\nLOCATION (lock): Among freshly dumped rubbish on the exposed deck of an ocean-going waste collection ship. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Coarse netting (Tightly wrapped around Charlie's body); used as Crossing lines that interrupt the body silhouette and communicate restraint; Rubbish on the collection ship (Deposited on the deck and partially burying Charlie); used as Uneven foreground and lateral occlusion around the robot; Collection ship deck (Supporting the deposited rubbish) — A small portion of its upper surface is visible beneath the pile; used as Lower-edge grounding that locates the discovery aboard the ship.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and measured contrast reveals the old robot's damage and net bindings without assigning him an unsupported active glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rubbish lies heaped on the collection ship's deck, with the old scrap robot Charlie entangled in a net among it. Charlie retains the worn UBIC logo on his chest.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "로봇의 시선은 화면 우측 전방을 향함.",
    "built_space": "금속 재질의 평평한 배 갑판 위에 쓰레기 더미가 흩어져 있음.",
    "entities": "각진 머리 외장과 흰 얼굴판을 가진 고릴라형 로봇, 점 형태의 눈과 선 형태의 입, 가슴의 UBIC 로고, 몸을 조이는 굵은 그물, 쓰레기 더미 모두 지시사항과 일치함.",
    "hard_violations": [
     "[gpt-high] 가슴의 UBIC 문자가 판독 가능하여, 모든 표기를 판독 불가능하게 하라는 최종 지시를 위반한다."
    ],
    "physics": "로봇의 하반신과 등은 쓰레기 더미와 갑판 바닥에 의해 안정적으로 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "로봇의 시선은 정면 좌측을 향함.",
    "built_space": "난간과 바다가 보이는 배의 갑판 위로 쓰레기들이 넓게 퍼져 있음.",
    "entities": "고릴라형 로봇이나 머리 외장이 각지지 않고 둥글며 얼굴판과 같이 흰색임. 가슴의 UBIC 로고와 쓰레기 더미는 존재하나, 그물이 칭칭 감기지 않고 느슨하게 덮여 있음.",
    "hard_violations": [
     "[gpt-high] 가슴의 UBIC 문자가 선명하게 읽혀, 모든 표기를 판독 불가능하게 하라는 최종 지시를 위반한다."
    ],
    "physics": "로봇은 갑판과 주변 쓰레기 더미의 지지를 받으며 파묻혀 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "각진 머리 외장과 흰 얼굴판의 대비, 그리고 몸을 단단히 옭아맨 그물의 구속감을 텍스트에 맞게 정확히 묘사했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "머리 외장이 둥글고 전체가 흰색이라 얼굴판과 구분되지 않으며, 그물이 칭칭 감기지 않고 망토처럼 느슨하게 걸쳐져 있어 지시와 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 시선은 화면 우측 전방을 향함.",
        "built_space": "금속 재질의 평평한 배 갑판 위에 쓰레기 더미가 흩어져 있음.",
        "entities": "각진 머리 외장과 흰 얼굴판을 가진 고릴라형 로봇, 점 형태의 눈과 선 형태의 입, 가슴의 UBIC 로고, 몸을 조이는 굵은 그물, 쓰레기 더미 모두 지시사항과 일치함.",
        "hard_violations": [],
        "physics": "로봇의 하반신과 등은 쓰레기 더미와 갑판 바닥에 의해 안정적으로 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "로봇의 시선은 정면 좌측을 향함.",
        "built_space": "난간과 바다가 보이는 배의 갑판 위로 쓰레기들이 넓게 퍼져 있음.",
        "entities": "고릴라형 로봇이나 머리 외장이 각지지 않고 둥글며 얼굴판과 같이 흰색임. 가슴의 UBIC 로고와 쓰레기 더미는 존재하나, 그물이 칭칭 감기지 않고 느슨하게 덮여 있음.",
        "hard_violations": [],
        "physics": "로봇은 갑판과 주변 쓰레기 더미의 지지를 받으며 파묻혀 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "각진 머리 외장과 흰 얼굴판의 대비, 그리고 몸을 단단히 옭아맨 그물의 구속감을 텍스트에 맞게 정확히 묘사했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "머리 외장이 둥글고 전체가 흰색이라 얼굴판과 구분되지 않으며, 그물이 칭칭 감기지 않고 망토처럼 느슨하게 걸쳐져 있어 지시와 다릅니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 시선은 화면 우측 전방을 향함.",
        "built_space": "금속 재질의 평평한 배 갑판 위에 쓰레기 더미가 흩어져 있음.",
        "entities": "각진 머리 외장과 흰 얼굴판을 가진 고릴라형 로봇, 점 형태의 눈과 선 형태의 입, 가슴의 UBIC 로고, 몸을 조이는 굵은 그물, 쓰레기 더미 모두 지시사항과 일치함.",
        "hard_violations": [],
        "physics": "로봇의 하반신과 등은 쓰레기 더미와 갑판 바닥에 의해 안정적으로 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "로봇의 시선은 정면 좌측을 향함.",
        "built_space": "난간과 바다가 보이는 배의 갑판 위로 쓰레기들이 넓게 퍼져 있음.",
        "entities": "고릴라형 로봇이나 머리 외장이 각지지 않고 둥글며 얼굴판과 같이 흰색임. 가슴의 UBIC 로고와 쓰레기 더미는 존재하나, 그물이 칭칭 감기지 않고 느슨하게 덮여 있음.",
        "hard_violations": [],
        "physics": "로봇은 갑판과 주변 쓰레기 더미의 지지를 받으며 파묻혀 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "쓰레기에 묻힌 정지 상태와 갑판 하단은 잘 보이지만, 그물이 몸을 칭칭 구속하기보다 걸쳐져 있고 가슴의 선명한 UBIC 문자가 가독성 금지를 위반한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "각진 고릴라형 머리와 몸통을 거듭 가로지르는 굵은 결박이 요청에 더 충실하지만, 읽히는 UBIC 문자 때문에 이 후보 역시 최종 사용에는 부적합하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴은 카메라 쪽을 향하면서 화면 왼쪽으로 조금 돌아가 있다. 점 형태의 눈은 특정 대상을 응시한다고 단정하기 어렵다. 조준물이나 이동하는 물체는 없으며, 별도의 시선 목표를 요구하지 않는 정지 외관 장면과 모순되지 않는다.",
        "built_space": "뒤쪽과 왼쪽에 보강재가 있는 선체 측벽이 이어지고, 왼쪽 너머로 바다가 보인다. 뒤쪽에는 떨어져 있는 쓰레기 더미 두 곳과 넓게 노출된 갑판이 있으며, 화면 하단에도 녹슨 철판 표면이 보인다. 찰리는 전경 쓰레기 더미 안에 상체를 세운 채 묻혀 있다. 요청한 하단의 갑판 단서는 있으나, 배경 갑판까지 상당히 많이 노출된다. 반사상이나 중복된 인물은 없다.",
        "entities": "등장 개체는 로봇 한 대뿐이다. 고릴라형 기계 얼굴, 흰 얼굴판, 작은 점 눈, 선 형태의 입이 있으며 머리 외장은 각진 면보다 둥근 윤곽이 두드러진다. 금속 몸에는 녹과 마모가 보이고, 주변에는 병·용기·밧줄·폐금속 등의 쓰레기가 쌓여 있다. 굵은 그물은 어깨와 팔을 덮지만 가슴 중앙이 넓게 드러나 결박보다는 걸친 상태에 가깝다. 가슴의 UBIC 표기는 명확히 읽힌다. 낮의 자연광이며 눈에 능동적인 발광은 보이지 않는다.",
        "hard_violations": [
         "가슴의 UBIC 문자가 선명하게 읽혀, 모든 표기를 판독 불가능하게 하라는 최종 지시를 위반한다."
        ],
        "physics": "몸의 하부는 쓰레기에 가려져 있고 상체와 팔은 주변 더미에 기대어 지지된다. 화면 왼쪽의 기계 손도 폐기물 위에 놓여 있다. 그물은 어깨와 팔에 걸려 아래로 처지며, 쓰레기는 갑판과 서로의 표면에 받쳐져 있다. 지지 없이 떠 있는 개체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 머리는 약간 기울어져 있으며 얼굴은 카메라 쪽에서 화면 오른쪽으로 조금 돌아가 있다. 눈이 향하는 명확한 대상은 없다. 무기나 진행 방향이 있는 물체는 없고, 행동 중이 아니라 정지된 로봇 외관으로 읽힌다.",
        "built_space": "상단과 왼쪽에 금속 선체 측벽이 있고, 왼쪽 뒤에는 밧줄이 감긴 짧은 기둥 하나가 보인다. 배경 중앙은 수직 앞면을 가진 넓은 철판 단차 또는 덮개 같은 구조로 채워져 있다. 찰리는 그 앞의 쓰레기 더미 안에 놓여 있으며 하단과 양옆이 폐기물로 가려진다. 상체 중심의 미디엄 숏은 맞지만, 요청한 화면 하단의 작은 갑판 노출은 뚜렷하지 않고 대신 뒤쪽 철판 면적이 크다. 반사상이나 인물 중복은 없다.",
        "entities": "등장 개체는 낡은 로봇 한 대이다. 각진 기계 머리 외장, 고릴라형 흰 얼굴판, 작은 점 형태의 눈과 선 형태의 입이 요청과 잘 맞는다. 몸통과 팔에는 녹슨 금속판과 노출된 기계 부위가 보인다. 굵은 그물이 양어깨와 팔을 감싸고, 여러 겹의 굵은 밧줄이 가슴과 허리를 교차해 구속감을 만든다. 병·파손 용기·폐금속·폐그물이 몸 주변과 하단을 묻고 있다. 가슴의 작은 UBIC 문자도 읽을 수 있다. 낮의 절제된 색조이며 눈의 발광은 없다.",
        "hard_violations": [
         "가슴의 UBIC 문자가 판독 가능하여, 모든 표기를 판독 불가능하게 하라는 최종 지시를 위반한다."
        ],
        "physics": "로봇의 몸은 쓰레기 더미에 기대어 있고, 드러난 팔과 손은 전경 폐기물에 받쳐져 있다. 밧줄과 그물은 어깨·몸통·팔의 표면에 밀착하거나 그 사이로 늘어져 있어 지지와 장력이 납득된다. 주변 폐기물도 서로 포개져 놓여 있으며, 지지 없이 공중에 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "쓰레기에 묻힌 정지 상태와 갑판 하단은 잘 보이지만, 그물이 몸을 칭칭 구속하기보다 걸쳐져 있고 가슴의 선명한 UBIC 문자가 가독성 금지를 위반한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "각진 고릴라형 머리와 몸통을 거듭 가로지르는 굵은 결박이 요청에 더 충실하지만, 읽히는 UBIC 문자 때문에 이 후보 역시 최종 사용에는 부적합하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴은 카메라 쪽을 향하면서 화면 왼쪽으로 조금 돌아가 있다. 점 형태의 눈은 특정 대상을 응시한다고 단정하기 어렵다. 조준물이나 이동하는 물체는 없으며, 별도의 시선 목표를 요구하지 않는 정지 외관 장면과 모순되지 않는다.",
        "built_space": "뒤쪽과 왼쪽에 보강재가 있는 선체 측벽이 이어지고, 왼쪽 너머로 바다가 보인다. 뒤쪽에는 떨어져 있는 쓰레기 더미 두 곳과 넓게 노출된 갑판이 있으며, 화면 하단에도 녹슨 철판 표면이 보인다. 찰리는 전경 쓰레기 더미 안에 상체를 세운 채 묻혀 있다. 요청한 하단의 갑판 단서는 있으나, 배경 갑판까지 상당히 많이 노출된다. 반사상이나 중복된 인물은 없다.",
        "entities": "등장 개체는 로봇 한 대뿐이다. 고릴라형 기계 얼굴, 흰 얼굴판, 작은 점 눈, 선 형태의 입이 있으며 머리 외장은 각진 면보다 둥근 윤곽이 두드러진다. 금속 몸에는 녹과 마모가 보이고, 주변에는 병·용기·밧줄·폐금속 등의 쓰레기가 쌓여 있다. 굵은 그물은 어깨와 팔을 덮지만 가슴 중앙이 넓게 드러나 결박보다는 걸친 상태에 가깝다. 가슴의 UBIC 표기는 명확히 읽힌다. 낮의 자연광이며 눈에 능동적인 발광은 보이지 않는다.",
        "hard_violations": [
         "가슴의 UBIC 문자가 선명하게 읽혀, 모든 표기를 판독 불가능하게 하라는 최종 지시를 위반한다."
        ],
        "physics": "몸의 하부는 쓰레기에 가려져 있고 상체와 팔은 주변 더미에 기대어 지지된다. 화면 왼쪽의 기계 손도 폐기물 위에 놓여 있다. 그물은 어깨와 팔에 걸려 아래로 처지며, 쓰레기는 갑판과 서로의 표면에 받쳐져 있다. 지지 없이 떠 있는 개체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 머리는 약간 기울어져 있으며 얼굴은 카메라 쪽에서 화면 오른쪽으로 조금 돌아가 있다. 눈이 향하는 명확한 대상은 없다. 무기나 진행 방향이 있는 물체는 없고, 행동 중이 아니라 정지된 로봇 외관으로 읽힌다.",
        "built_space": "상단과 왼쪽에 금속 선체 측벽이 있고, 왼쪽 뒤에는 밧줄이 감긴 짧은 기둥 하나가 보인다. 배경 중앙은 수직 앞면을 가진 넓은 철판 단차 또는 덮개 같은 구조로 채워져 있다. 찰리는 그 앞의 쓰레기 더미 안에 놓여 있으며 하단과 양옆이 폐기물로 가려진다. 상체 중심의 미디엄 숏은 맞지만, 요청한 화면 하단의 작은 갑판 노출은 뚜렷하지 않고 대신 뒤쪽 철판 면적이 크다. 반사상이나 인물 중복은 없다.",
        "entities": "등장 개체는 낡은 로봇 한 대이다. 각진 기계 머리 외장, 고릴라형 흰 얼굴판, 작은 점 형태의 눈과 선 형태의 입이 요청과 잘 맞는다. 몸통과 팔에는 녹슨 금속판과 노출된 기계 부위가 보인다. 굵은 그물이 양어깨와 팔을 감싸고, 여러 겹의 굵은 밧줄이 가슴과 허리를 교차해 구속감을 만든다. 병·파손 용기·폐금속·폐그물이 몸 주변과 하단을 묻고 있다. 가슴의 작은 UBIC 문자도 읽을 수 있다. 낮의 절제된 색조이며 눈의 발광은 없다.",
        "hard_violations": [
         "가슴의 UBIC 문자가 판독 가능하여, 모든 표기를 판독 불가능하게 하라는 최종 지시를 위반한다."
        ],
        "physics": "로봇의 몸은 쓰레기 더미에 기대어 있고, 드러난 팔과 손은 전경 폐기물에 받쳐져 있다. 밧줄과 그물은 어깨·몸통·팔의 표면에 밀착하거나 그 사이로 늘어져 있어 지지와 장력이 납득된다. 주변 폐기물도 서로 포개져 놓여 있으며, 지지 없이 공중에 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.321
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.071
   },
   "violations": {
    "B": [
     "[gpt-high] 가슴의 UBIC 문자가 선명하게 읽혀, 모든 표기를 판독 불가능하게 하라는 최종 지시를 위반한다."
    ],
    "A": [
     "[gpt-high] 가슴의 UBIC 문자가 판독 가능하여, 모든 표기를 판독 불가능하게 하라는 최종 지시를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1071
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "각진 머리 외장과 흰 얼굴판의 대비, 그리고 몸을 단단히 옭아맨 그물의 구속감을 텍스트에 맞게 정확히 묘사했습니다.  ★위반: [gpt-high] 가슴의 UBIC 문자가 판독 가능하여, 모든 표기를 판독 불가능하게 하라는 최종 지시를 위반한다."
   },
   {
    "label": "B",
    "score": 1071,
    "verdict_ko": "머리 외장이 둥글고 전체가 흰색이라 얼굴판과 구분되지 않으며, 그물이 칭칭 감기지 않고 망토처럼 느슨하게 걸쳐져 있어 지시와 다릅니다.  ★위반: [gpt-high] 가슴의 UBIC 문자가 선명하게 읽혀, 모든 표기를 판독 불가능하게 하라는 최종 지시를 위반한다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd71-83d6-79c2-96b6-8b77dcfc4c2d",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S2sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:15:52.248713+00:00",
  "fingerprint": "f18967a10f2e753617b5a16b465011c92ca887f128df40eba1843e54161486b1",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S2sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S2sh3_sel.png",
  "source_sha256": "2387a2d79332c60f38df8a5fef3e6cc2312e14e085c471755fd5791af2dc00c4",
  "file": "S2sh3_cine.png",
  "staged_sha256": "a84bcbc3a5f6602673224dad16517db582d9789f826a8bfb5ac562d0ad09b959",
  "latency_ms": 12327
 },
 "S3sh2::confined_fp_apt": {
  "applies": true,
  "reason_ko": "이 샷은 덤프트럭의 운전석 내부를 배경으로 하며, 운전자가 조향 장치(핸들) 앞에 앉아 있는 정확한 위치와 자세가 중요합니다. 운전석과 조작부의 공간적 배치가 틀릴 경우 매우 어색하고 명백한 오류가 되므로 평면도 레이아웃 보조가 필요합니다.",
  "input_fingerprint": "42ec10f78a183e5d"
 },
 "S3sh2::signage": {
  "fp": "81e8a43aeac10fa3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::562ac6794aee": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/confinedfp_base_562ac6794aee.png",
  "place_text": "Inside the compact driver's cab of a self-driving dump truck carrying scrap, with the steering wheel directly before the driver's seat and daylight through the windshield.",
  "input_fingerprint": "af85c7b8fad563d9"
 },
 "S3sh2::confined_fp": {
  "reads": {
   "controls": "One steering wheel located directly in front of the left-hand seat.",
   "mirrors": "No mirrors are depicted in the diagram.",
   "camera": "Positioned in the center of the cabin between the two seats, pointing diagonally towards the front-left to frame the driver.",
   "occupants": "One occupant seated in the left-hand driver's seat."
  },
  "mismatches": [],
  "scene_description_en": "The camera is stationed in the center of the cabin, angled obliquely left toward the driver's station. The Asian Korean driver occupies the left-hand seat, appearing prominently in the near center-left of the shot and facing towards the right edge of the frame. Situated slightly further away, the steering wheel appears in the lower-right midground, facing back toward the driver. A section of the broad front windshield spans the far right background, admitting daylight into the cab. The right-hand passenger seat is located behind the camera's field of view and remains off-screen. No reflective surfaces or mirrors are visible from this vantage point.",
  "fixed": true,
  "input_fingerprint": "78d0742dbd944a43"
 },
 "era_assess::c9fb1c3110e40cd2": {
  "subjects": [],
  "subject_text": "이현우 일행과 수빈의 트럭 운전석·조수석\n지붕이 없는 낡은 운전 공간. 운전대와 계기판 앞에 운전석·조수석이 나란히 놓이고 전면 유리와 양옆 문이 둘러싼다.",
  "identity": "canonical",
  "scope_id": "L94",
  "scope_role": "location_exterior",
  "scope_sha": "7c2ba56be0030e86"
 },
 "S3sh2": {
  "input_fingerprint": "9e1b413d5b8d2eb8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자율주행 중인 덤프트럭 운전석에 앉은 아시아계 한국인 운전사의 고개가 핸들 쪽으로 푹 꺾여 잠든 찰나.\n\nLOCATION (lock): Inside the compact driver's cab of a self-driving dump truck carrying scrap, with the steering wheel directly before the driver's seat and daylight through the windshield. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Steering wheel (In front of the sleeping driver while the truck operates autonomously) — Seen obliquely from the passenger side beneath the driver's lowered head; used as Lower-right reference revealing the direction of the head drop; Driver's seat (Occupied by the sleeping driver) — Its side and a narrow section behind his shoulder are visible; used as Limited background context supporting the slumped seated posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight appropriate to the cab gives the sleeping face gentle tonal separation while maintaining an uneasy, unembellished stillness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The dump-truck driver is asleep in the driver's seat, with his torso seated and his head drooping forward toward the steering wheel. His arm and leg positions and any contact between his head and the wheel are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The self-driving dump truck carries a full load of scrap, including the old, net-entangled Charlie with his worn chest logo. No unloading has occurred yet.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed in the center of the cabin, angled obliquely left toward the driver's station. The Asian Korean driver occupies the left-hand seat, appearing prominently in the near center-left of the shot and facing towards the right edge of the frame. Situated slightly further away, the steering wheel appears in the lower-right midground, facing back toward the driver. A section of the broad front windshield spans the far right background, admitting daylight into the cab. The right-hand passenger seat is located behind the camera's field of view and remains off-screen. No reflective surfaces or mirrors are visible from this vantage point.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자율주행 중인 덤프트럭 운전석에 앉은 아시아계 한국인 운전사의 고개가 핸들 쪽으로 푹 꺾여 잠든 찰나.\n\nLOCATION (lock): Inside the compact driver's cab of a self-driving dump truck carrying scrap, with the steering wheel directly before the driver's seat and daylight through the windshield. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight appropriate to the cab gives the sleeping face gentle tonal separation while maintaining an uneasy, unembellished stillness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The dump-truck driver is asleep in the driver's seat, with his torso seated and his head drooping forward toward the steering wheel. His arm and leg positions and any contact between his head and the wheel are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The self-driving dump truck carries a full load of scrap, including the old, net-entangled Charlie with his worn chest logo. No unloading has occurred yet.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed in the center of the cabin, angled obliquely left toward the driver's station. The Asian Korean driver occupies the left-hand seat, appearing prominently in the near center-left of the shot and facing towards the right edge of the frame. Situated slightly further away, the steering wheel appears in the lower-right midground, facing back toward the driver. A section of the broad front windshield spans the far right background, admitting daylight into the cab. The right-hand passenger seat is located behind the camera's field of view and remains off-screen. No reflective surfaces or mirrors are visible from this vantage point.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자율주행 중인 덤프트럭 운전석에 앉은 아시아계 한국인 운전사의 고개가 핸들 쪽으로 푹 꺾여 잠든 찰나.\n\nLOCATION (lock): Inside the compact driver's cab of a self-driving dump truck carrying scrap, with the steering wheel directly before the driver's seat and daylight through the windshield. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight appropriate to the cab gives the sleeping face gentle tonal separation while maintaining an uneasy, unembellished stillness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The dump-truck driver is asleep in the driver's seat, with his torso seated and his head drooping forward toward the steering wheel. His arm and leg positions and any contact between his head and the wheel are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The self-driving dump truck carries a full load of scrap, including the old, net-entangled Charlie with his worn chest logo. No unloading has occurred yet.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S3sh2_confinedfp.png",
     "asset_id": null,
     "role": null
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S3sh2_confinedfp.png",
     "asset_id": null,
     "role": null
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "조수석에서 운전석을 향해 카메라가 위치하며, 운전자의 고개는 앞쪽 핸들 방향으로 숙여짐.",
    "built_space": "트럭 운전석 내부. 운전대와 대시보드, 시트가 배치된 좌측 운전석 구조임.",
    "entities": "잠든 아시아계 운전자가 있고 창밖 짐칸에 그물에 얽힌 찰리가 서 있으나, 대시보드 화면에 읽을 수 있는 한국어 텍스트가 존재함.",
    "hard_violations": [
     "[gemini-pro] 대시보드 화면에 읽을 수 있는 텍스트 노출",
     "[gpt-high] 대시보드 중앙 화면에 판독 가능한 한글 상태 문구를 표시하여 이미지 어디에도 읽을 수 있는 글자를 넣지 말라는 조건을 위반한다."
    ],
    "physics": "운전자의 상체는 시트에 기대어 앞으로 숙여져 있으며, 팔은 아래로 늘어져 있음."
   },
   {
    "label": "B",
    "direction": "조수석에서 운전석을 바라보는 구도이며, 운전자의 고개는 우측 하단에 위치한 핸들 쪽으로 꺾여 있음.",
    "built_space": "트럭 운전석 내부. 우측 하단의 운전대와 대시보드, 운전석 시트의 측면 및 어깨 뒤편이 명확히 보임.",
    "entities": "잠든 아시아계 운전자와 창밖의 고철, 그물망이 확인되나 가슴 로고가 있는 찰리의 구체적인 모습은 보이지 않음.",
    "hard_violations": [],
    "physics": "운전자의 몸은 시트에 안정적으로 지지되어 있고, 오른손은 허벅지 위에 자연스럽게 놓여 중력 법칙에 부합함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "대시보드 디스플레이에 읽을 수 있는 텍스트가 명확하게 노출되어 텍스트 삽입 금지 지침을 위반했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 조수석 카메라 앵글과 운전자의 수면 자세, 하단 운전대 배치를 잘 따랐으나 창밖 짐칸의 찰리 디테일이 생략되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "조수석에서 운전석을 향해 카메라가 위치하며, 운전자의 고개는 앞쪽 핸들 방향으로 숙여짐.",
        "built_space": "트럭 운전석 내부. 운전대와 대시보드, 시트가 배치된 좌측 운전석 구조임.",
        "entities": "잠든 아시아계 운전자가 있고 창밖 짐칸에 그물에 얽힌 찰리가 서 있으나, 대시보드 화면에 읽을 수 있는 한국어 텍스트가 존재함.",
        "hard_violations": [
         "대시보드 화면에 읽을 수 있는 텍스트 노출"
        ],
        "physics": "운전자의 상체는 시트에 기대어 앞으로 숙여져 있으며, 팔은 아래로 늘어져 있음."
       },
       {
        "label": "B",
        "direction": "조수석에서 운전석을 바라보는 구도이며, 운전자의 고개는 우측 하단에 위치한 핸들 쪽으로 꺾여 있음.",
        "built_space": "트럭 운전석 내부. 우측 하단의 운전대와 대시보드, 운전석 시트의 측면 및 어깨 뒤편이 명확히 보임.",
        "entities": "잠든 아시아계 운전자와 창밖의 고철, 그물망이 확인되나 가슴 로고가 있는 찰리의 구체적인 모습은 보이지 않음.",
        "hard_violations": [],
        "physics": "운전자의 몸은 시트에 안정적으로 지지되어 있고, 오른손은 허벅지 위에 자연스럽게 놓여 중력 법칙에 부합함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "대시보드 디스플레이에 읽을 수 있는 텍스트가 명확하게 노출되어 텍스트 삽입 금지 지침을 위반했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 조수석 카메라 앵글과 운전자의 수면 자세, 하단 운전대 배치를 잘 따랐으나 창밖 짐칸의 찰리 디테일이 생략되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "조수석에서 운전석을 향해 카메라가 위치하며, 운전자의 고개는 앞쪽 핸들 방향으로 숙여짐.",
        "built_space": "트럭 운전석 내부. 운전대와 대시보드, 시트가 배치된 좌측 운전석 구조임.",
        "entities": "잠든 아시아계 운전자가 있고 창밖 짐칸에 그물에 얽힌 찰리가 서 있으나, 대시보드 화면에 읽을 수 있는 한국어 텍스트가 존재함.",
        "hard_violations": [
         "대시보드 화면에 읽을 수 있는 텍스트 노출"
        ],
        "physics": "운전자의 상체는 시트에 기대어 앞으로 숙여져 있으며, 팔은 아래로 늘어져 있음."
       },
       {
        "label": "B",
        "direction": "조수석에서 운전석을 바라보는 구도이며, 운전자의 고개는 우측 하단에 위치한 핸들 쪽으로 꺾여 있음.",
        "built_space": "트럭 운전석 내부. 우측 하단의 운전대와 대시보드, 운전석 시트의 측면 및 어깨 뒤편이 명확히 보임.",
        "entities": "잠든 아시아계 운전자와 창밖의 고철, 그물망이 확인되나 가슴 로고가 있는 찰리의 구체적인 모습은 보이지 않음.",
        "hard_violations": [],
        "physics": "운전자의 몸은 시트에 안정적으로 지지되어 있고, 오른손은 허벅지 위에 자연스럽게 놓여 중력 법칙에 부합함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "읽을 수 있는 문구 없이 착석한 수면 자세를 보여주지만, 요구된 클로즈업보다 훨씬 넓고 고개가 핸들 쪽으로 푹 꺾인 정도도 부족하다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "핸들로 깊이 떨어진 고개는 더 정확하지만, 클로즈업을 지키지 않았고 화면에 판독 가능한 한글 문구를 넣어 명시적 금지 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "운전사는 눈을 감고 턱을 아래로 내렸으며 얼굴은 오른쪽 아래의 핸들 방향을 향한다. 다만 고개가 가슴 쪽으로 조금 숙여진 정도라 핸들을 향해 푹 꺾인 순간은 약하다. 핸들은 운전자 앞에 있고 카메라는 조수석 쪽에서 비스듬히 바라본다.",
        "built_space": "운전석 하나와 왼쪽 전경의 조수석 일부, 핸들 하나, 계기판과 중앙 조작부가 보인다. 운전자는 핸들 바로 뒤 운전석에 앉아 있어 기본 배치는 도면과 맞는다. 그러나 머리받침부터 허벅지까지, 좌석 등받침 대부분과 넓은 대시보드까지 보여 지정된 클로즈업 및 좁은 좌석 배경을 벗어난다. 측면 창밖의 고철 적재물은 이 차량 자체의 뒤쪽 적재함인지 확인되지 않는다.",
        "entities": "중년으로 보이는 동아시아계 남성 한 명이 작업복을 입고 잠들어 있다. 한국인이라는 국적은 외형만으로 확인할 수 없지만 지정된 인물상과 충돌하지 않는다. 운전석과 핸들, 낮빛이 들어오는 창은 확인된다. 창밖에는 고철과 그물이 보이지만 찰리는 식별되지 않으며, 원래 클로즈업 밖에 있을 적재물의 비노출 자체는 결함이 아니다. 자율주행 여부는 정지 이미지로 확정할 수 없다. 판독 가능한 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "몸통과 골반은 좌석에 지지되고 등은 등받침 쪽에 놓여 있다. 두 손은 허벅지와 무릎 부근에 내려놓았으며 팔도 아래로 처져 있다. 머리는 목을 축으로 앞으로 숙여져 있고 공중에 떠 있는 신체나 물체는 없다. 핸들은 조향축에, 열쇠는 시동 장치에 연결되어 있다."
       },
       {
        "label": "B",
        "direction": "운전사는 눈을 감고 상체와 목을 앞으로 깊이 굽혔다. 얼굴이 오른쪽 아래 핸들 중앙 가까이 떨어져 있어 고개가 향하는 대상이 핸들이라는 점은 A보다 명확하다. 조수석 쪽에서 보는 사선 방향도 맞는다.",
        "built_space": "운전석 하나, 왼쪽 전경의 조수석 일부, 핸들 하나와 계기판, 중앙 화면이 보인다. 운전자는 자신의 좌석에 앉아 앞쪽 핸들로 숙이고 있으며 기본 좌석 관계는 맞는다. 다만 허벅지와 넓은 등받침, 천장, 대시보드 및 외부 고철장까지 포함한 넓은 구도로 클로즈업 조건을 어긴다. 측면 거울에 고철장이 비치는 것은 가능한 배치지만, 측면 창 너머 적재물이 이 트럭의 적재함에 실려 있는지는 확인되지 않는다.",
        "entities": "중년으로 보이는 동아시아계 남성 한 명이 낡은 작업복을 입고 잠들어 있다. 한국 국적은 시각적으로 확정할 수 없다. 창밖에는 그물에 얽힌 낡은 인간형 기계와 가슴 표식이 보여 찰리를 표현한 것으로 읽힌다. 추가 사람은 보이지 않는다. 낮빛과 핸들, 운전석은 맞지만 중앙 화면에 읽을 수 있는 한글 상태 문구가 있어 무문자 조건에 어긋난다.",
        "hard_violations": [
         "대시보드 중앙 화면에 판독 가능한 한글 상태 문구를 표시하여 이미지 어디에도 읽을 수 있는 글자를 넣지 말라는 조건을 위반한다."
        ],
        "physics": "골반은 좌석에 놓여 있고 몸통은 앞으로 접혀 있다. 머리 옆면은 핸들 중앙부에 기대거나 맞닿은 것으로 보이며, 깊이 숙인 머리의 지지가 성립한다. 안쪽 손은 허벅지 사이에 내려앉고 가까운 팔은 무릎 쪽으로 처져 프레임 아래로 이어진다. 보이는 범위에서 억지로 들어 올린 팔다리는 없다. 창밖 인간형 기계의 하체 지지는 고철에 가려져 있어 부유한다고 단정할 근거는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "읽을 수 있는 문구 없이 착석한 수면 자세를 보여주지만, 요구된 클로즈업보다 훨씬 넓고 고개가 핸들 쪽으로 푹 꺾인 정도도 부족하다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "핸들로 깊이 떨어진 고개는 더 정확하지만, 클로즈업을 지키지 않았고 화면에 판독 가능한 한글 문구를 넣어 명시적 금지 조건을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "운전사는 눈을 감고 턱을 아래로 내렸으며 얼굴은 오른쪽 아래의 핸들 방향을 향한다. 다만 고개가 가슴 쪽으로 조금 숙여진 정도라 핸들을 향해 푹 꺾인 순간은 약하다. 핸들은 운전자 앞에 있고 카메라는 조수석 쪽에서 비스듬히 바라본다.",
        "built_space": "운전석 하나와 왼쪽 전경의 조수석 일부, 핸들 하나, 계기판과 중앙 조작부가 보인다. 운전자는 핸들 바로 뒤 운전석에 앉아 있어 기본 배치는 도면과 맞는다. 그러나 머리받침부터 허벅지까지, 좌석 등받침 대부분과 넓은 대시보드까지 보여 지정된 클로즈업 및 좁은 좌석 배경을 벗어난다. 측면 창밖의 고철 적재물은 이 차량 자체의 뒤쪽 적재함인지 확인되지 않는다.",
        "entities": "중년으로 보이는 동아시아계 남성 한 명이 작업복을 입고 잠들어 있다. 한국인이라는 국적은 외형만으로 확인할 수 없지만 지정된 인물상과 충돌하지 않는다. 운전석과 핸들, 낮빛이 들어오는 창은 확인된다. 창밖에는 고철과 그물이 보이지만 찰리는 식별되지 않으며, 원래 클로즈업 밖에 있을 적재물의 비노출 자체는 결함이 아니다. 자율주행 여부는 정지 이미지로 확정할 수 없다. 판독 가능한 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "몸통과 골반은 좌석에 지지되고 등은 등받침 쪽에 놓여 있다. 두 손은 허벅지와 무릎 부근에 내려놓았으며 팔도 아래로 처져 있다. 머리는 목을 축으로 앞으로 숙여져 있고 공중에 떠 있는 신체나 물체는 없다. 핸들은 조향축에, 열쇠는 시동 장치에 연결되어 있다."
       },
       {
        "label": "A",
        "direction": "운전사는 눈을 감고 상체와 목을 앞으로 깊이 굽혔다. 얼굴이 오른쪽 아래 핸들 중앙 가까이 떨어져 있어 고개가 향하는 대상이 핸들이라는 점은 A보다 명확하다. 조수석 쪽에서 보는 사선 방향도 맞는다.",
        "built_space": "운전석 하나, 왼쪽 전경의 조수석 일부, 핸들 하나와 계기판, 중앙 화면이 보인다. 운전자는 자신의 좌석에 앉아 앞쪽 핸들로 숙이고 있으며 기본 좌석 관계는 맞는다. 다만 허벅지와 넓은 등받침, 천장, 대시보드 및 외부 고철장까지 포함한 넓은 구도로 클로즈업 조건을 어긴다. 측면 거울에 고철장이 비치는 것은 가능한 배치지만, 측면 창 너머 적재물이 이 트럭의 적재함에 실려 있는지는 확인되지 않는다.",
        "entities": "중년으로 보이는 동아시아계 남성 한 명이 낡은 작업복을 입고 잠들어 있다. 한국 국적은 시각적으로 확정할 수 없다. 창밖에는 그물에 얽힌 낡은 인간형 기계와 가슴 표식이 보여 찰리를 표현한 것으로 읽힌다. 추가 사람은 보이지 않는다. 낮빛과 핸들, 운전석은 맞지만 중앙 화면에 읽을 수 있는 한글 상태 문구가 있어 무문자 조건에 어긋난다.",
        "hard_violations": [
         "대시보드 중앙 화면에 판독 가능한 한글 상태 문구를 표시하여 이미지 어디에도 읽을 수 있는 글자를 넣지 말라는 조건을 위반한다."
        ],
        "physics": "골반은 좌석에 놓여 있고 몸통은 앞으로 접혀 있다. 머리 옆면은 핸들 중앙부에 기대거나 맞닿은 것으로 보이며, 깊이 숙인 머리의 지지가 성립한다. 안쪽 손은 허벅지 사이에 내려앉고 가까운 팔은 무릎 쪽으로 처져 프레임 아래로 이어진다. 보이는 범위에서 억지로 들어 올린 팔다리는 없다. 창밖 인간형 기계의 하체 지지는 고철에 가려져 있어 부유한다고 단정할 근거는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.929,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.679,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 대시보드 화면에 읽을 수 있는 텍스트 노출",
     "[gpt-high] 대시보드 중앙 화면에 판독 가능한 한글 상태 문구를 표시하여 이미지 어디에도 읽을 수 있는 글자를 넣지 말라는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 679,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 679,
    "verdict_ko": "대시보드 디스플레이에 읽을 수 있는 텍스트가 명확하게 노출되어 텍스트 삽입 금지 지침을 위반했습니다.  ★위반: [gemini-pro] 대시보드 화면에 읽을 수 있는 텍스트 노출 / [gpt-high] 대시보드 중앙 화면에 판독 가능한 한글 상태 문구를 표시하여 이미지 어디에도 읽을 수 있는 글자를 넣지 말라는 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 조수석 카메라 앵글과 운전자의 수면 자세, 하단 운전대 배치를 잘 따랐으나 창밖 짐칸의 찰리 디테일이 생략되었습니다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S3sh2_confinedfp.png",
    "asset_id": null,
    "role": null
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd72-4e2e-76c8-8ef1-52bc7f7baa1c",
  "confined_fp": {
   "base_key": "confinedfp::562ac6794aee",
   "apt_reason": "이 샷은 덤프트럭의 운전석 내부를 배경으로 하며, 운전자가 조향 장치(핸들) 앞에 앉아 있는 정확한 위치와 자세가 중요합니다. 운전석과 조작부의 공간적 배치가 틀릴 경우 매우 어색하고 명백한 오류가 되므로 평면도 레이아웃 보조가 필요합니다.",
   "fixed": true,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S3sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:16:04.968654+00:00",
  "fingerprint": "8a16d2675a4fb23248f9b696816a45e5252f8885e1809d006846f10c4fb1f238",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S3sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S3sh2_sel.png",
  "source_sha256": "314d32d8782cafe89a5a03556372f0025275f0cd85fd1d50f311d3882415b691",
  "file": "S3sh2_cine.png",
  "staged_sha256": "8f41e389559333a5fb5af0dc7be94cfa90f66fb6cb05208b83d755b3f54b95cc",
  "latency_ms": 15198
 },
 "S4sh3::confined_fp_apt": {
  "applies": true,
  "reason_ko": "차량 내부의 운전석에 앉은 인물이 조수석 창문 밖으로 특정 방향(건너편 주택)을 가리키는 장면이므로, 좌석 위치와 가리키는 방향 간의 정확한 공간적 배치가 필수적입니다. 방향이나 좌석이 틀리면 상황 전달에 큰 오류가 발생합니다.",
  "input_fingerprint": "62792926154f6942"
 },
 "S4sh3::signage": {
  "fp": "bb99f78c83578e7b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::c19457c11668": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/confinedfp_base_c19457c11668.png",
  "place_text": "Inside the front cabin of a car parked in an affluent residential street, beside the passenger-side window facing a house across the road in daylight.",
  "input_fingerprint": "d310fa303effaebb"
 },
 "S4sh3::confined_fp": {
  "reads": {
   "controls": "A steering wheel is located at the front left station. A center console sits between the two front seats.",
   "mirrors": "No mirrors are drawn or indicated in the diagram.",
   "camera": "The camera is positioned outside the vehicle, forward of the passenger side, pointing backward and inward towards the passenger window.",
   "occupants": "Pedro occupies the driver's seat on the front left. The passenger seat and both rear seats are empty."
  },
  "mismatches": [],
  "scene_description_en": "The camera is stationed outside the car, forward of the passenger side, looking backward and inward. In the foreground on the left, the open passenger window frames the scene, facing outward. Pedro's arm extends from deep inside the cabin through this window, pointing towards the right side of the screen. In the far background on the left, Pedro occupies the driver's seat, facing towards the right of the frame with an earpiece on his visible ear. The steering wheel sits in the distant center directly in front of him, while the center console lies in the deep background between the seats. No mirrors are present in the diagram.",
  "fixed": false,
  "input_fingerprint": "d3772165248677dd"
 },
 "era_assess::f1d12b223d77fa83": {
  "subjects": [],
  "subject_text": "고급주택가에 주차된 페드로의 자동차 내부\n운전석과 조수석이 나란한 승용차 실내. 운전대와 대시보드가 있고, 창문 너머로 고급주택가가 보인다.",
  "identity": "canonical",
  "scope_id": "L05",
  "scope_role": "location_interior",
  "scope_sha": "5dd290ed9e6c0fff"
 },
 "S4sh3": {
  "input_fingerprint": "1930506110bbf4e9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 페드로가 조수석 창밖으로 팔을 뻗어 길 건너편의 고급주택 쪽을 정확히 가리키는 손끝.\n\nLOCATION (lock): Inside the front cabin of a car parked in an affluent residential street, beside the passenger-side window facing a house across the road in daylight. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Luxury house indicated by Pedro's fingertip in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Passenger-side window opening (Open enough for Pedro's arm to extend outside) — Seen obliquely from outside and forward of the opening; used as Slim edge framing the arm without exposing the rest of the cabin; Luxury house across the street (The house Pedro is identifying) — Its street-facing exterior is visible beyond the pointing hand; used as Upper-right destination of the pointing vector, kept smaller than the foreground hand grouping; Street between the car and house (Separating the stopped car from the indicated house); used as Depth interval preventing the house from appearing adjacent to the window.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and clear tonal separation keeps the pointing hand and the house simultaneously readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked in the upscale residential neighborhood, across the road from the target house. The house plan is crumpled and poorly drawn. 페드로: Pedro remains in the driver's seat and has an earpiece in his ear. He points toward the target house.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed outside the car, forward of the passenger side, looking backward and inward. In the foreground on the left, the open passenger window frames the scene, facing outward. Pedro's arm extends from deep inside the cabin through this window, pointing towards the right side of the screen. In the far background on the left, Pedro occupies the driver's seat, facing towards the right of the frame with an earpiece on his visible ear. The steering wheel sits in the distant center directly in front of him, while the center console lies in the deep background between the seats. No mirrors are present in the diagram.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 페드로가 조수석 창밖으로 팔을 뻗어 길 건너편의 고급주택 쪽을 정확히 가리키는 손끝.\n\nLOCATION (lock): Inside the front cabin of a car parked in an affluent residential street, beside the passenger-side window facing a house across the road in daylight. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and clear tonal separation keeps the pointing hand and the house simultaneously readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked in the upscale residential neighborhood, across the road from the target house. The house plan is crumpled and poorly drawn. 페드로: Pedro remains in the driver's seat and has an earpiece in his ear. He points toward the target house.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed outside the car, forward of the passenger side, looking backward and inward. In the foreground on the left, the open passenger window frames the scene, facing outward. Pedro's arm extends from deep inside the cabin through this window, pointing towards the right side of the screen. In the far background on the left, Pedro occupies the driver's seat, facing towards the right of the frame with an earpiece on his visible ear. The steering wheel sits in the distant center directly in front of him, while the center console lies in the deep background between the seats. No mirrors are present in the diagram.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 페드로가 조수석 창밖으로 팔을 뻗어 길 건너편의 고급주택 쪽을 정확히 가리키는 손끝.\n\nLOCATION (lock): Inside the front cabin of a car parked in an affluent residential street, beside the passenger-side window facing a house across the road in daylight. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and clear tonal separation keeps the pointing hand and the house simultaneously readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked in the upscale residential neighborhood, across the road from the target house. The house plan is crumpled and poorly drawn. 페드로: Pedro remains in the driver's seat and has an earpiece in his ear. He points toward the target house.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S4sh3_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "페드로",
     "path": "<bytes:826925>",
     "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S4sh3_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "페드로",
     "path": "<bytes:826925>",
     "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 조수석 쪽을 향하고 있으며, 오른팔을 차 안에서 조수석 창문 방향으로 뻗어 창밖의 집을 가리킴.",
    "built_space": "좌핸들 차량의 실내. 카메라는 운전석(왼쪽) 창문 바깥에 위치하여 차량 내부 전체(운전대, 대시보드, 조수석 등)를 횡단하여 비추고 있음.",
    "entities": "페드로(비니, 가죽 재킷, 왼쪽 귀에 무선 이어폰 착용)가 운전석에 앉아 있으며, 조수석 창문 너머로 고급 주택이 보임.",
    "hard_violations": [
     "[gemini-pro] 카메라가 조수석 창문 바깥(외부 전방)이 아닌 운전석 창문 바깥에 위치하여, '차량 내부가 노출되지 않도록 슬림하게 프레이밍'하라는 명시적 지시를 완전히 위반함."
    ],
    "physics": "운전석에 앉아 체중을 지탱하고 있으며, 뻗은 팔의 자세와 위치가 물리적으로 자연스러움."
   },
   {
    "label": "B",
    "direction": "시선은 앞쪽을 향하고, 왼팔을 운전석 창밖으로 뻗어 배경의 집을 가리킴.",
    "built_space": "좌핸들 차량의 측면. 카메라는 운전석(왼쪽) 창문 바깥 전방에 위치하여 차체 외부와 운전석 내부 일부를 비춤.",
    "entities": "페드로(가죽 재킷, 오른쪽 귀에 이어피스)가 운전석에 앉아 있으며, 배경에 고급 주택이 위치함.",
    "hard_violations": [
     "[gemini-pro] 카메라가 조수석 창문 바깥이 아닌 운전석 창문 바깥에 위치함.",
     "[gemini-pro] 조수석 창밖으로 팔을 뻗어야 한다는 명시적 액션 지시를 어기고 운전석 창밖으로 왼팔을 뻗음.",
     "[gpt-high] 팔을 내민 곳이 지정된 조수석 창이 아니라 운전석 바로 옆 창으로 나타나, 배치도와 핵심 창밖 지목 동선을 위반한다."
    ],
    "physics": "운전석에 앉은 상태에서 왼팔을 창틀 위로 자연스럽게 뻗어 지탱하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 카메라 위치(조수석 창문 밖)를 무시하고 운전석 측면에서 촬영하여 실내를 완전히 노출한 치명적인 연출 위반입니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "카메라 위치 위반에 더해, 조수석이 아닌 운전석 창밖으로 팔을 뻗어 핵심 행동 지시까지 완전히 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 조수석 쪽을 향하고 있으며, 오른팔을 차 안에서 조수석 창문 방향으로 뻗어 창밖의 집을 가리킴.",
        "built_space": "좌핸들 차량의 실내. 카메라는 운전석(왼쪽) 창문 바깥에 위치하여 차량 내부 전체(운전대, 대시보드, 조수석 등)를 횡단하여 비추고 있음.",
        "entities": "페드로(비니, 가죽 재킷, 왼쪽 귀에 무선 이어폰 착용)가 운전석에 앉아 있으며, 조수석 창문 너머로 고급 주택이 보임.",
        "hard_violations": [
         "카메라가 조수석 창문 바깥(외부 전방)이 아닌 운전석 창문 바깥에 위치하여, '차량 내부가 노출되지 않도록 슬림하게 프레이밍'하라는 명시적 지시를 완전히 위반함."
        ],
        "physics": "운전석에 앉아 체중을 지탱하고 있으며, 뻗은 팔의 자세와 위치가 물리적으로 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선은 앞쪽을 향하고, 왼팔을 운전석 창밖으로 뻗어 배경의 집을 가리킴.",
        "built_space": "좌핸들 차량의 측면. 카메라는 운전석(왼쪽) 창문 바깥 전방에 위치하여 차체 외부와 운전석 내부 일부를 비춤.",
        "entities": "페드로(가죽 재킷, 오른쪽 귀에 이어피스)가 운전석에 앉아 있으며, 배경에 고급 주택이 위치함.",
        "hard_violations": [
         "카메라가 조수석 창문 바깥이 아닌 운전석 창문 바깥에 위치함.",
         "조수석 창밖으로 팔을 뻗어야 한다는 명시적 액션 지시를 어기고 운전석 창밖으로 왼팔을 뻗음."
        ],
        "physics": "운전석에 앉은 상태에서 왼팔을 창틀 위로 자연스럽게 뻗어 지탱하고 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 카메라 위치(조수석 창문 밖)를 무시하고 운전석 측면에서 촬영하여 실내를 완전히 노출한 치명적인 연출 위반입니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "카메라 위치 위반에 더해, 조수석이 아닌 운전석 창밖으로 팔을 뻗어 핵심 행동 지시까지 완전히 어겼습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 조수석 쪽을 향하고 있으며, 오른팔을 차 안에서 조수석 창문 방향으로 뻗어 창밖의 집을 가리킴.",
        "built_space": "좌핸들 차량의 실내. 카메라는 운전석(왼쪽) 창문 바깥에 위치하여 차량 내부 전체(운전대, 대시보드, 조수석 등)를 횡단하여 비추고 있음.",
        "entities": "페드로(비니, 가죽 재킷, 왼쪽 귀에 무선 이어폰 착용)가 운전석에 앉아 있으며, 조수석 창문 너머로 고급 주택이 보임.",
        "hard_violations": [
         "카메라가 조수석 창문 바깥(외부 전방)이 아닌 운전석 창문 바깥에 위치하여, '차량 내부가 노출되지 않도록 슬림하게 프레이밍'하라는 명시적 지시를 완전히 위반함."
        ],
        "physics": "운전석에 앉아 체중을 지탱하고 있으며, 뻗은 팔의 자세와 위치가 물리적으로 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선은 앞쪽을 향하고, 왼팔을 운전석 창밖으로 뻗어 배경의 집을 가리킴.",
        "built_space": "좌핸들 차량의 측면. 카메라는 운전석(왼쪽) 창문 바깥 전방에 위치하여 차체 외부와 운전석 내부 일부를 비춤.",
        "entities": "페드로(가죽 재킷, 오른쪽 귀에 이어피스)가 운전석에 앉아 있으며, 배경에 고급 주택이 위치함.",
        "hard_violations": [
         "카메라가 조수석 창문 바깥이 아닌 운전석 창문 바깥에 위치함.",
         "조수석 창밖으로 팔을 뻗어야 한다는 명시적 액션 지시를 어기고 운전석 창밖으로 왼팔을 뻗음."
        ],
        "physics": "운전석에 앉은 상태에서 왼팔을 창틀 위로 자연스럽게 뻗어 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "손끝은 길 건너 주택을 향하지만 조수석이 아닌 운전석 창으로 팔을 내밀었고, 손끝 클로즈업 대신 얼굴·실내·차체를 크게 노출했다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "운전석에서 조수석 창 쪽으로 뻗는 관계는 맞지만, 손끝이 주택보다 아래를 향하고 넓게 드러난 실내가 지정된 손끝 클로즈업을 훼손한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "검지는 화면 오른쪽으로 거의 수평으로 뻗어 길 건너 주택의 어두운 차고 입구 쪽을 가리킨다. 얼굴과 시선도 오른쪽을 향한다. 주택을 지목하는 관계는 읽히지만 오른쪽 위를 향하는 상승 벡터는 약하다.",
        "built_space": "열린 측면 창 하나, 사이드미러 하나, 운전대 하나, 실내 후사경 하나가 보인다. 페드로가 창 바로 안쪽 운전대 뒤에 앉아 가까운 창으로 팔을 내밀어 운전석 창으로 읽힌다. 배치도의 반대편 조수석 창을 가로질러 뻗는 구성이 아니다. 창틀과 문짝이 화면 왼쪽과 아래를 크게 차지하고 실내도 드러나므로 얇은 창 가장자리만 남기는 구도와 다르다. 도로는 차와 주택 사이를 분명하게 분리한다.",
        "entities": "젊은 남성 한 명과 그의 손·팔이 보인다. 얼굴은 참고의 라틴계 혼혈 청소년 설정과 대체로 부합하며 낡은 검은 가죽 재킷과 귀의 이어피스가 있다. 다만 참고의 검은 비니는 없다. 오른쪽 위에는 유리와 콘크리트 외장의 고급주택이 있고 낮 장면이다. 구겨진 주택 도면은 보이지 않으며 이 구도에서 필수 노출 대상은 아니다. 읽을 수 있는 글자나 도식은 없다.",
        "hard_violations": [
         "팔을 내민 곳이 지정된 조수석 창이 아니라 운전석 바로 옆 창으로 나타나, 배치도와 핵심 창밖 지목 동선을 위반한다."
        ],
        "physics": "팔은 보이는 어깨에서 소매와 손목을 거쳐 손으로 연결되며, 열린 창을 통과한다. 검지는 펴지고 나머지 손가락은 접혀 있어 가능한 지목 자세다. 몸은 좌석에 앉아 있으며 차는 도로에 놓여 있다. 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "검지는 화면 오른쪽으로 거의 수평으로 향한다. 그 연장선은 배경 주택의 외벽보다 아래인 앞유리 하단·대시보드 높이로 지나가므로 오른쪽 위 주택을 정확히 지목하지 않는다. 페드로의 얼굴과 시선은 차량 전방 쪽으로 향하며 손끝의 목표 주택을 명확히 확인하는 시선은 아니다.",
        "built_space": "가까운 조수석 창 하나와 부분적으로 올라온 유리, 빈 조수석 하나, 운전석 하나, 운전대 하나, 중앙 콘솔 하나, 실내 후사경 하나가 보인다. 페드로는 먼 쪽 운전석에 앉아 가까운 조수석 쪽으로 팔을 뻗어 기본 좌석 관계는 배치도에 맞는다. 그러나 카메라는 창 앞쪽에서 손만 비스듬히 잡기보다 넓은 창을 통해 실내 대부분을 보여준다. 주택은 주로 앞유리 너머에 나타나며 차와 목표 주택 사이 도로의 깊이 간격도 불명확하다. 유리의 반사와 투과에서 명백히 불가능한 중복은 보이지 않는다.",
        "entities": "젊은 남성 한 명이 보이며 검은 비니, 낡은 검은 가죽 재킷, 손목 장식과 귀의 이어피스는 참고와 대체로 부합한다. 보이는 얼굴과 피부색도 설정에 크게 어긋나지 않지만 정확한 나이와 민족성은 이 이미지로 단정하기 어렵다. 큰 전경 손과 배경의 고급주택이 있고 낮이다. 주택 도면은 보이지 않는다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "한 손은 운전대 윗부분을 잡고, 다른 팔은 어깨에서 조수석 창 쪽으로 이어진다. 뻗은 손은 내려간 유리의 윗선보다 높아 창 개구부를 통과할 여지가 있다. 손가락과 손목의 지목 자세는 물리적으로 가능하며 몸통은 운전석에 지지된다. 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "손끝은 길 건너 주택을 향하지만 조수석이 아닌 운전석 창으로 팔을 내밀었고, 손끝 클로즈업 대신 얼굴·실내·차체를 크게 노출했다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "운전석에서 조수석 창 쪽으로 뻗는 관계는 맞지만, 손끝이 주택보다 아래를 향하고 넓게 드러난 실내가 지정된 손끝 클로즈업을 훼손한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "검지는 화면 오른쪽으로 거의 수평으로 뻗어 길 건너 주택의 어두운 차고 입구 쪽을 가리킨다. 얼굴과 시선도 오른쪽을 향한다. 주택을 지목하는 관계는 읽히지만 오른쪽 위를 향하는 상승 벡터는 약하다.",
        "built_space": "열린 측면 창 하나, 사이드미러 하나, 운전대 하나, 실내 후사경 하나가 보인다. 페드로가 창 바로 안쪽 운전대 뒤에 앉아 가까운 창으로 팔을 내밀어 운전석 창으로 읽힌다. 배치도의 반대편 조수석 창을 가로질러 뻗는 구성이 아니다. 창틀과 문짝이 화면 왼쪽과 아래를 크게 차지하고 실내도 드러나므로 얇은 창 가장자리만 남기는 구도와 다르다. 도로는 차와 주택 사이를 분명하게 분리한다.",
        "entities": "젊은 남성 한 명과 그의 손·팔이 보인다. 얼굴은 참고의 라틴계 혼혈 청소년 설정과 대체로 부합하며 낡은 검은 가죽 재킷과 귀의 이어피스가 있다. 다만 참고의 검은 비니는 없다. 오른쪽 위에는 유리와 콘크리트 외장의 고급주택이 있고 낮 장면이다. 구겨진 주택 도면은 보이지 않으며 이 구도에서 필수 노출 대상은 아니다. 읽을 수 있는 글자나 도식은 없다.",
        "hard_violations": [
         "팔을 내민 곳이 지정된 조수석 창이 아니라 운전석 바로 옆 창으로 나타나, 배치도와 핵심 창밖 지목 동선을 위반한다."
        ],
        "physics": "팔은 보이는 어깨에서 소매와 손목을 거쳐 손으로 연결되며, 열린 창을 통과한다. 검지는 펴지고 나머지 손가락은 접혀 있어 가능한 지목 자세다. 몸은 좌석에 앉아 있으며 차는 도로에 놓여 있다. 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "검지는 화면 오른쪽으로 거의 수평으로 향한다. 그 연장선은 배경 주택의 외벽보다 아래인 앞유리 하단·대시보드 높이로 지나가므로 오른쪽 위 주택을 정확히 지목하지 않는다. 페드로의 얼굴과 시선은 차량 전방 쪽으로 향하며 손끝의 목표 주택을 명확히 확인하는 시선은 아니다.",
        "built_space": "가까운 조수석 창 하나와 부분적으로 올라온 유리, 빈 조수석 하나, 운전석 하나, 운전대 하나, 중앙 콘솔 하나, 실내 후사경 하나가 보인다. 페드로는 먼 쪽 운전석에 앉아 가까운 조수석 쪽으로 팔을 뻗어 기본 좌석 관계는 배치도에 맞는다. 그러나 카메라는 창 앞쪽에서 손만 비스듬히 잡기보다 넓은 창을 통해 실내 대부분을 보여준다. 주택은 주로 앞유리 너머에 나타나며 차와 목표 주택 사이 도로의 깊이 간격도 불명확하다. 유리의 반사와 투과에서 명백히 불가능한 중복은 보이지 않는다.",
        "entities": "젊은 남성 한 명이 보이며 검은 비니, 낡은 검은 가죽 재킷, 손목 장식과 귀의 이어피스는 참고와 대체로 부합한다. 보이는 얼굴과 피부색도 설정에 크게 어긋나지 않지만 정확한 나이와 민족성은 이 이미지로 단정하기 어렵다. 큰 전경 손과 배경의 고급주택이 있고 낮이다. 주택 도면은 보이지 않는다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "한 손은 운전대 윗부분을 잡고, 다른 팔은 어깨에서 조수석 창 쪽으로 이어진다. 뻗은 손은 내려간 유리의 윗선보다 높아 창 개구부를 통과할 여지가 있다. 손가락과 손목의 지목 자세는 물리적으로 가능하며 몸통은 운전석에 지지된다. 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.0
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 카메라가 조수석 창문 바깥(외부 전방)이 아닌 운전석 창문 바깥에 위치하여, '차량 내부가 노출되지 않도록 슬림하게 프레이밍'하라는 명시적 지시를 완전히 위반함."
    ],
    "B": [
     "[gemini-pro] 카메라가 조수석 창문 바깥이 아닌 운전석 창문 바깥에 위치함.",
     "[gemini-pro] 조수석 창밖으로 팔을 뻗어야 한다는 명시적 액션 지시를 어기고 운전석 창밖으로 왼팔을 뻗음.",
     "[gpt-high] 팔을 내민 곳이 지정된 조수석 창이 아니라 운전석 바로 옆 창으로 나타나, 배치도와 핵심 창밖 지목 동선을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지정된 카메라 위치(조수석 창문 밖)를 무시하고 운전석 측면에서 촬영하여 실내를 완전히 노출한 치명적인 연출 위반입니다.  ★위반: [gemini-pro] 카메라가 조수석 창문 바깥(외부 전방)이 아닌 운전석 창문 바깥에 위치하여, '차량 내부가 노출되지 않도록 슬림하게 프레이밍'하라는 명시적 지시를 완전히 위반함."
   },
   {
    "label": "B",
    "score": 750,
    "verdict_ko": "카메라 위치 위반에 더해, 조수석이 아닌 운전석 창밖으로 팔을 뻗어 핵심 행동 지시까지 완전히 어겼습니다.  ★위반: [gemini-pro] 카메라가 조수석 창문 바깥이 아닌 운전석 창문 바깥에 위치함. / [gemini-pro] 조수석 창밖으로 팔을 뻗어야 한다는 명시적 액션 지시를 어기고 운전석 창밖으로 왼팔을 뻗음. / [gpt-high] 팔을 내민 곳이 지정된 조수석 창이 아니라 운전석 바로 옆 창으로 나타나, 배치도와 핵심 창밖 지목 동선을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S4sh3_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "페드로",
    "path": "<bytes:826925>",
    "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd73-4776-721f-acbf-dd7a6a2101c4",
  "confined_fp": {
   "base_key": "confinedfp::c19457c11668",
   "apt_reason": "차량 내부의 운전석에 앉은 인물이 조수석 창문 밖으로 특정 방향(건너편 주택)을 가리키는 장면이므로, 좌석 위치와 가리키는 방향 간의 정확한 공간적 배치가 필수적입니다. 방향이나 좌석이 틀리면 상황 전달에 큰 오류가 발생합니다.",
   "fixed": false,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S4sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:16:20.546233+00:00",
  "fingerprint": "2bb5ea09da172e9301ce8b6f7228cef525f0d909a0e4fdc31e1198a8b5caa9a0",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S4sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S4sh3_sel.png",
  "source_sha256": "775e7ab8732ee192dfb288d804b642725444d34bcb1f2e0ccd0c1a4ea9a2a359",
  "file": "S4sh3_cine.png",
  "staged_sha256": "d17d88875c697493d3225c65a602a95039f127be42076ab2cbd5d525dce94880",
  "latency_ms": 11019
 },
 "S4sh5::signage": {
  "fp": "cb4071b76941d40f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S4sh5": {
  "input_fingerprint": "d4c9d07f62669973",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고급주택 근처에 멈춰 선 이현우가 검지 손가락을 들어 자신의 귓바퀴에 꽂힌 소형 무선 이어폰 바깥쪽을 꾹 누르는 순간.\n\nLOCATION (lock): On the residential street beside the upscale house across from the parked car, outdoors in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Small wireless earpiece (Worn in Hyunwoo's ear and being pressed) — Its exposed outer side is visible between the fingertip and ear; used as Small focal detail within the human close-up; Luxury house exterior (Near Hyunwoo after his approach across the street) — Only an oblique exterior fragment is retained behind him; used as Soft left-edge location cue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight and restrained contrast retain natural facial detail and the small finger-to-ear contact without introducing a stylized device glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked across the road from the upscale house. 이현우: Hyunwoo is outside the car, approaching the house across the road. He checks the communication setup at his ear.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고급주택 근처에 멈춰 선 이현우가 검지 손가락을 들어 자신의 귓바퀴에 꽂힌 소형 무선 이어폰 바깥쪽을 꾹 누르는 순간.\n\nLOCATION (lock): On the residential street beside the upscale house across from the parked car, outdoors in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Small wireless earpiece (Worn in Hyunwoo's ear and being pressed) — Its exposed outer side is visible between the fingertip and ear; used as Small focal detail within the human close-up; Luxury house exterior (Near Hyunwoo after his approach across the street) — Only an oblique exterior fragment is retained behind him; used as Soft left-edge location cue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight and restrained contrast retain natural facial detail and the small finger-to-ear contact without introducing a stylized device glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked across the road from the upscale house. 이현우: Hyunwoo is outside the car, approaching the house across the road. He checks the communication setup at his ear.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고급주택 근처에 멈춰 선 이현우가 검지 손가락을 들어 자신의 귓바퀴에 꽂힌 소형 무선 이어폰 바깥쪽을 꾹 누르는 순간.\n\nLOCATION (lock): On the residential street beside the upscale house across from the parked car, outdoors in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Small wireless earpiece (Worn in Hyunwoo's ear and being pressed) — Its exposed outer side is visible between the fingertip and ear; used as Small focal detail within the human close-up; Luxury house exterior (Near Hyunwoo after his approach across the street) — Only an oblique exterior fragment is retained behind him; used as Soft left-edge location cue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight and restrained contrast retain natural facial detail and the small finger-to-ear contact without introducing a stylized device glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked across the road from the upscale house. 이현우: Hyunwoo is outside the car, approaching the house across the road. He checks the communication setup at his ear.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선은 오른쪽 아래를 향하고 있으며, 오른쪽 검지 손가락이 오른쪽 귀에 꽂힌 작은 검은색 무선 이어폰의 바깥쪽을 정확히 누르고 있음.",
    "built_space": "화면 왼쪽 배경으로 고급 주택의 외관이 부드럽게 아웃포커싱되어 있으며, 오른쪽 하단에는 도로 위에 정차된 차량의 일부가 보임.",
    "entities": "이현우의 얼굴, 헝클어진 머리, 낡은 후드티 모두 캐릭터 레퍼런스와 일치함. 이어폰은 지시된 대로 작고 바깥면이 잘 보이는 무선 형태임.",
    "hard_violations": [],
    "physics": "손가락이 이어폰을 누르는 동작에 맞춰 손과 팔의 위치가 물리적으로 자연스럽게 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "이현우의 시선은 오른쪽 아래를 향하고 있으며, 오른쪽 검지 손가락이 오른쪽 귀에 꽂힌 기둥 형태의 회색 무선 이어폰을 짚고 있음.",
    "built_space": "화면 왼쪽 배경으로 고급 주택이 보이고, 그 옆으로 이어지는 도로 중간에 은색 차량이 위치해 있음.",
    "entities": "이현우의 인상과 복장은 레퍼런스와 잘 맞으나, 귀에 꽂힌 이어폰이 소형이라기보다는 크기가 큰 기둥형 이어폰임.",
    "hard_violations": [],
    "physics": "손가락이 이어폰에 닿아 있으며, 손과 팔이 자연스럽게 중력과 인체 구조에 맞게 유지되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지시된 소형 무선 이어폰의 형태와 손가락으로 바깥쪽을 누르는 동작을 매우 사실적으로 묘사했으며, 배경의 고급 주택 컷과 캐릭터의 외형이 레퍼런스와 완벽하게 일치합니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "캐릭터와 배경의 전반적인 분위기는 잘 맞으나, 무선 이어폰의 크기가 다소 커서 '소형'이라는 조건에 부합하지 않으며 백그라운드의 차량 배치가 프롬프트의 정차된 묘사와 약간 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 오른쪽 아래를 향하고 있으며, 오른쪽 검지 손가락이 오른쪽 귀에 꽂힌 작은 검은색 무선 이어폰의 바깥쪽을 정확히 누르고 있음.",
        "built_space": "화면 왼쪽 배경으로 고급 주택의 외관이 부드럽게 아웃포커싱되어 있으며, 오른쪽 하단에는 도로 위에 정차된 차량의 일부가 보임.",
        "entities": "이현우의 얼굴, 헝클어진 머리, 낡은 후드티 모두 캐릭터 레퍼런스와 일치함. 이어폰은 지시된 대로 작고 바깥면이 잘 보이는 무선 형태임.",
        "hard_violations": [],
        "physics": "손가락이 이어폰을 누르는 동작에 맞춰 손과 팔의 위치가 물리적으로 자연스럽게 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 오른쪽 아래를 향하고 있으며, 오른쪽 검지 손가락이 오른쪽 귀에 꽂힌 기둥 형태의 회색 무선 이어폰을 짚고 있음.",
        "built_space": "화면 왼쪽 배경으로 고급 주택이 보이고, 그 옆으로 이어지는 도로 중간에 은색 차량이 위치해 있음.",
        "entities": "이현우의 인상과 복장은 레퍼런스와 잘 맞으나, 귀에 꽂힌 이어폰이 소형이라기보다는 크기가 큰 기둥형 이어폰임.",
        "hard_violations": [],
        "physics": "손가락이 이어폰에 닿아 있으며, 손과 팔이 자연스럽게 중력과 인체 구조에 맞게 유지되고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지시된 소형 무선 이어폰의 형태와 손가락으로 바깥쪽을 누르는 동작을 매우 사실적으로 묘사했으며, 배경의 고급 주택 컷과 캐릭터의 외형이 레퍼런스와 완벽하게 일치합니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "캐릭터와 배경의 전반적인 분위기는 잘 맞으나, 무선 이어폰의 크기가 다소 커서 '소형'이라는 조건에 부합하지 않으며 백그라운드의 차량 배치가 프롬프트의 정차된 묘사와 약간 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 오른쪽 아래를 향하고 있으며, 오른쪽 검지 손가락이 오른쪽 귀에 꽂힌 작은 검은색 무선 이어폰의 바깥쪽을 정확히 누르고 있음.",
        "built_space": "화면 왼쪽 배경으로 고급 주택의 외관이 부드럽게 아웃포커싱되어 있으며, 오른쪽 하단에는 도로 위에 정차된 차량의 일부가 보임.",
        "entities": "이현우의 얼굴, 헝클어진 머리, 낡은 후드티 모두 캐릭터 레퍼런스와 일치함. 이어폰은 지시된 대로 작고 바깥면이 잘 보이는 무선 형태임.",
        "hard_violations": [],
        "physics": "손가락이 이어폰을 누르는 동작에 맞춰 손과 팔의 위치가 물리적으로 자연스럽게 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 오른쪽 아래를 향하고 있으며, 오른쪽 검지 손가락이 오른쪽 귀에 꽂힌 기둥 형태의 회색 무선 이어폰을 짚고 있음.",
        "built_space": "화면 왼쪽 배경으로 고급 주택이 보이고, 그 옆으로 이어지는 도로 중간에 은색 차량이 위치해 있음.",
        "entities": "이현우의 인상과 복장은 레퍼런스와 잘 맞으나, 귀에 꽂힌 이어폰이 소형이라기보다는 크기가 큰 기둥형 이어폰임.",
        "hard_violations": [],
        "physics": "손가락이 이어폰에 닿아 있으며, 손과 팔이 자연스럽게 중력과 인체 구조에 맞게 유지되고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "검지로 이어폰 외측을 누르는 클로즈업과 왼쪽의 비스듬한 주택 단서가 더 충실하며, 다만 이어폰은 요구한 소형 장치보다 다소 길고 크다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "작은 이어폰의 접촉 동작은 정확하지만, 주택 정면이 배경을 넓게 차지하고 자동차가 인물 바로 곁의 전경에 들어와 지정된 공간 관계가 덜 명확하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 화면 오른쪽 바깥을 약간 아래로 바라보며 렌즈를 응시하지 않는다. 올린 검지 끝은 보이는 귀에 착용한 회색 이어폰의 외측 상단에 직접 닿아 있다. 손가락이 향하는 대상은 귀 주변 피부가 아니라 이어폰이다.",
        "built_space": "야외 주택가의 낮 장면이다. 왼쪽에 밝은 석재 주택의 비스듬한 외벽, 현관 한 곳, 그 앞의 계단 한 구간, 아래층 큰 창 한 곳과 위층 창 부분들이 보인다. 낮은 검은 울타리와 석재 기둥, 조경수가 있으며 참고 장소의 주요 재료와 형태에 부합한다. 인물은 주택 앞 보행 공간 가까이에 있고, 자동차 한 대는 오른쪽 도로의 후경에 있다. 차가 도로 맞은편에 있는지는 이 크롭만으로 확정하기 어렵다. 불가능한 반사나 중복 설비는 보이지 않는다.",
        "entities": "한 명의 젊은 한국계 남성으로 보이며, 검은 헝클어진 머리와 얼굴 윤곽, 낡은 짙은 회색 후드가 인물 참고와 대체로 일치한다. 손의 피부색과 젊은 외관도 인물과 연결된다. 귀에는 회색 무선 이어폰 한 개가 있으나 아래로 길게 내려오는 형태여서 소형이라는 요구에는 다소 덜 맞는다. 고급주택과 자동차 한 대가 보이고 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "손목과 소매가 화면 아래까지 이어지며, 접은 나머지 손가락과 편 검지가 귀를 누르는 동작을 자연스럽게 만든다. 이어폰은 귀에 끼워져 있고 검지 끝이 바깥 면을 받쳐 누른다. 하체는 클로즈업 밖이므로 발의 지지는 확인할 수 없지만 부유를 나타내는 단서는 없다. 자동차는 도로 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 화면 오른쪽 아래의 화면 밖을 향한다. 검지는 귀에 끼운 짙은색 이어폰 외측을 정확히 누르며, 손톱 위쪽으로 장치의 둥근 외측 가장자리가 드러난다. 카메라를 향한 응시는 없다.",
        "built_space": "낮의 주택 외부이며 왼쪽 배경에 현관 한 곳, 앞 계단 한 구간, 아래층 큰 창 한 곳과 위층 창 부분들이 보인다. 왼쪽 가장자리와 오른쪽 후경에는 석재 경계 기둥이 보이고 잔디와 조경수가 이어진다. 재료는 참고 주택과 유사하지만 외벽이 비교적 정면에 가깝고 크게 남아, 작은 비스듬한 왼쪽 단서라는 요구보다 존재감이 강하다. 자동차 한 대의 앞유리 주변과 측면 거울이 오른쪽 아래 전경에 보여 인물 바로 곁에 있는 듯하며, 도로 건너 주차된 차라는 관계는 확인되지 않는다.",
        "entities": "젊은 한국계 남성 한 명이며 검은 헝클어진 머리, 얼굴과 체격, 해진 짙은 회색 후드가 참고 인물과 대체로 맞는다. 손과 소매도 같은 인물의 것으로 자연스럽게 연결된다. 귀 안의 작은 짙은색 무선 이어폰은 요구한 크기에 잘 맞고 외측에 인위적인 발광은 없다. 주택과 자동차 일부 외에 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "손은 소매에서 이어지는 손목으로 지지되고 검지가 이어폰 외측에 접촉한다. 이어폰은 귓바퀴 안에 안정적으로 끼워져 있어 떠 있지 않다. 접힌 손가락과 손목의 자세는 실제로 귀의 장치를 누를 수 있는 형태다. 발과 자동차 바퀴는 프레임 밖이므로 직접적인 지면 접촉은 확인할 수 없으나, 보이는 부분에 물리적으로 불가능한 자세나 부유는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "검지로 이어폰 외측을 누르는 클로즈업과 왼쪽의 비스듬한 주택 단서가 더 충실하며, 다만 이어폰은 요구한 소형 장치보다 다소 길고 크다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "작은 이어폰의 접촉 동작은 정확하지만, 주택 정면이 배경을 넓게 차지하고 자동차가 인물 바로 곁의 전경에 들어와 지정된 공간 관계가 덜 명확하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "인물은 화면 오른쪽 바깥을 약간 아래로 바라보며 렌즈를 응시하지 않는다. 올린 검지 끝은 보이는 귀에 착용한 회색 이어폰의 외측 상단에 직접 닿아 있다. 손가락이 향하는 대상은 귀 주변 피부가 아니라 이어폰이다.",
        "built_space": "야외 주택가의 낮 장면이다. 왼쪽에 밝은 석재 주택의 비스듬한 외벽, 현관 한 곳, 그 앞의 계단 한 구간, 아래층 큰 창 한 곳과 위층 창 부분들이 보인다. 낮은 검은 울타리와 석재 기둥, 조경수가 있으며 참고 장소의 주요 재료와 형태에 부합한다. 인물은 주택 앞 보행 공간 가까이에 있고, 자동차 한 대는 오른쪽 도로의 후경에 있다. 차가 도로 맞은편에 있는지는 이 크롭만으로 확정하기 어렵다. 불가능한 반사나 중복 설비는 보이지 않는다.",
        "entities": "한 명의 젊은 한국계 남성으로 보이며, 검은 헝클어진 머리와 얼굴 윤곽, 낡은 짙은 회색 후드가 인물 참고와 대체로 일치한다. 손의 피부색과 젊은 외관도 인물과 연결된다. 귀에는 회색 무선 이어폰 한 개가 있으나 아래로 길게 내려오는 형태여서 소형이라는 요구에는 다소 덜 맞는다. 고급주택과 자동차 한 대가 보이고 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "손목과 소매가 화면 아래까지 이어지며, 접은 나머지 손가락과 편 검지가 귀를 누르는 동작을 자연스럽게 만든다. 이어폰은 귀에 끼워져 있고 검지 끝이 바깥 면을 받쳐 누른다. 하체는 클로즈업 밖이므로 발의 지지는 확인할 수 없지만 부유를 나타내는 단서는 없다. 자동차는 도로 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "인물의 시선은 화면 오른쪽 아래의 화면 밖을 향한다. 검지는 귀에 끼운 짙은색 이어폰 외측을 정확히 누르며, 손톱 위쪽으로 장치의 둥근 외측 가장자리가 드러난다. 카메라를 향한 응시는 없다.",
        "built_space": "낮의 주택 외부이며 왼쪽 배경에 현관 한 곳, 앞 계단 한 구간, 아래층 큰 창 한 곳과 위층 창 부분들이 보인다. 왼쪽 가장자리와 오른쪽 후경에는 석재 경계 기둥이 보이고 잔디와 조경수가 이어진다. 재료는 참고 주택과 유사하지만 외벽이 비교적 정면에 가깝고 크게 남아, 작은 비스듬한 왼쪽 단서라는 요구보다 존재감이 강하다. 자동차 한 대의 앞유리 주변과 측면 거울이 오른쪽 아래 전경에 보여 인물 바로 곁에 있는 듯하며, 도로 건너 주차된 차라는 관계는 확인되지 않는다.",
        "entities": "젊은 한국계 남성 한 명이며 검은 헝클어진 머리, 얼굴과 체격, 해진 짙은 회색 후드가 참고 인물과 대체로 맞는다. 손과 소매도 같은 인물의 것으로 자연스럽게 연결된다. 귀 안의 작은 짙은색 무선 이어폰은 요구한 크기에 잘 맞고 외측에 인위적인 발광은 없다. 주택과 자동차 일부 외에 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "손은 소매에서 이어지는 손목으로 지지되고 검지가 이어폰 외측에 접촉한다. 이어폰은 귓바퀴 안에 안정적으로 끼워져 있어 떠 있지 않다. 접힌 손가락과 손목의 자세는 실제로 귀의 장치를 누를 수 있는 형태다. 발과 자동차 바퀴는 프레임 밖이므로 직접적인 지면 접촉은 확인할 수 없으나, 보이는 부분에 물리적으로 불가능한 자세나 부유는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.8
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.8
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1875,
   "B": 1800
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "지시된 소형 무선 이어폰의 형태와 손가락으로 바깥쪽을 누르는 동작을 매우 사실적으로 묘사했으며, 배경의 고급 주택 컷과 캐릭터의 외형이 레퍼런스와 완벽하게 일치합니다."
   },
   {
    "label": "B",
    "score": 1800,
    "verdict_ko": "캐릭터와 배경의 전반적인 분위기는 잘 맞으나, 무선 이어폰의 크기가 다소 커서 '소형'이라는 조건에 부합하지 않으며 백그라운드의 차량 배치가 프롬프트의 정차된 묘사와 약간 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S4sh3_sel.png",
    "asset_id": "30d579bb-6717-4a54-b01c-981ebf0cd05a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd73-fe53-7d10-946b-5fce9eca32f1",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S4sh3"
  }
 },
 "S4sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:16:31.905590+00:00",
  "fingerprint": "c33b73dd4ee160d8543851a1d1e35ee9e652a2539586d9b6b916719289d2c6fa",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S4sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S4sh5_sel.png",
  "source_sha256": "b3f3220de3c7779f1c6bea94dc534ba5eb9ea2ba246f57416abbe584191080c4",
  "file": "S4sh5_cine.png",
  "staged_sha256": "baccc735e914bc8d5fdeb48d3a590c87d65d05664f57ea6d69f7f140b46f2d1a",
  "latency_ms": 10170
 },
 "S5sh10::signage": {
  "fp": "58f8afb30627cda5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::b267a8a73d07716f": {
  "subjects": [],
  "subject_text": "고급주택 담장·현관 앞·정원\n높은 담장 안에 정원과 현관 앞 공간이 있는 주택 외부. 현관에는 구식 열쇠 잠금장치가 있고 거실 쪽에는 큰 통유리창이 있다.",
  "identity": "canonical",
  "scope_id": "L07",
  "scope_role": "location_exterior",
  "scope_sha": "be1b5fd21ec6d382"
 },
 "S5sh10::bgfirst_bg": {
  "input_fingerprint": "f3f65a9fc81938b0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 열린 창문 밖으로 힘없이 꺾인 백인 여성형 집사 로봇의 상체를 벽 모퉁이의 이현우가 두 팔로 덥석 받아 안는 순간.\n\nLOCATION (lock): Outside an upscale house, at the open living-room window beside an exterior wall corner.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open living-room window behind the collapsing robot in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Living-room window opening (Open, with the robot collapsing through it) — Seen from outside at an oblique angle; the opening exposes a limited portion of the living room; used as Shared spatial anchor connecting the interior body to Hyunwoo's exterior catch; Adjacent exterior wall corner (Beside the window where Hyunwoo had hidden) — Its outer edge lies toward the right side of the composition; used as Lateral boundary that explains Hyunwoo's short interception route.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight appropriate to the open window preserves humanlike facial detail and clear arm-to-body separation without revealing the robot's still-concealed metal interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 열린 창문 밖으로 힘없이 꺾인 백인 여성형 집사 로봇의 상체를 벽 모퉁이의 이현우가 두 팔로 덥석 받아 안는 순간.\n\nLOCATION (lock): Outside an upscale house, at the open living-room window beside an exterior wall corner.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open living-room window behind the collapsing robot in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Living-room window opening (Open, with the robot collapsing through it) — Seen from outside at an oblique angle; the opening exposes a limited portion of the living room; used as Shared spatial anchor connecting the interior body to Hyunwoo's exterior catch; Adjacent exterior wall corner (Beside the window where Hyunwoo had hidden) — Its outer edge lies toward the right side of the composition; used as Lateral boundary that explains Hyunwoo's short interception route.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight appropriate to the open window preserves humanlike facial detail and clear arm-to-body separation without revealing the robot's still-concealed metal interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S5sh10__bgfirst_bg.png",
  "asset_id": "ce41f054-4da2-4f2d-91e9-68adc725b197",
  "input_asset_ids": [
   "1dd7fd01-6688-4035-ac05-5f0a0c3704b3",
   "f93bd100-ada4-44ab-b370-a7e1748ac239",
   "93e38602-1247-4591-8e5c-e38083816dd4"
  ]
 },
 "S5sh10": {
  "input_fingerprint": "b87d2cb807365d76",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 창문 밖으로 힘없이 꺾인 백인 여성형 집사 로봇의 상체를 벽 모퉁이의 이현우가 두 팔로 덥석 받아 안는 순간.\n\nLOCATION (lock): Outside an upscale house, at the open living-room window beside an exterior wall corner. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open living-room window behind the collapsing robot in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Living-room window opening (Open, with the robot collapsing through it) — Seen from outside at an oblique angle; the opening exposes a limited portion of the living room; used as Shared spatial anchor connecting the interior body to Hyunwoo's exterior catch; Adjacent exterior wall corner (Beside the window where Hyunwoo had hidden) — Its outer edge lies toward the right side of the composition; used as Lateral boundary that explains Hyunwoo's short interception route.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight appropriate to the open window preserves humanlike facial detail and clear arm-to-body separation without revealing the robot's still-concealed metal interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The disabled female butler robot is collapsing limply outside the open window, with her upper body caught and supported in both of Hyunwoo's arms, her mouth wide open, and her eyes rolled back. Her exact head angle, arm positions, leg positions, and lower-body support are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room glass window is open, while the front door retains its old key-operated lock. The female-form butler robot is incapacitated but still externally intact; the answered phone is playing the noisy digital recording. 이현우: Hyunwoo is beside the open living-room window, catching the collapsing butler robot in his arms. His ear communication setup remains in place.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 백인 여성형 집사 로봇 (백인 여성형 얼굴, 젊은 성인 외형, 인간 같은 얼굴 표면) — wearing: 부유한 주택에 어울리는 클래식한 흑백 톤의 메이드 앞치마와 단정한 원피스. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 창문 밖으로 힘없이 꺾인 백인 여성형 집사 로봇의 상체를 벽 모퉁이의 이현우가 두 팔로 덥석 받아 안는 순간.\n\nLOCATION (lock): Outside an upscale house, at the open living-room window beside an exterior wall corner. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open living-room window behind the collapsing robot in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Living-room window opening (Open, with the robot collapsing through it) — Seen from outside at an oblique angle; the opening exposes a limited portion of the living room; used as Shared spatial anchor connecting the interior body to Hyunwoo's exterior catch; Adjacent exterior wall corner (Beside the window where Hyunwoo had hidden) — Its outer edge lies toward the right side of the composition; used as Lateral boundary that explains Hyunwoo's short interception route.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight appropriate to the open window preserves humanlike facial detail and clear arm-to-body separation without revealing the robot's still-concealed metal interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The disabled female butler robot is collapsing limply outside the open window, with her upper body caught and supported in both of Hyunwoo's arms, her mouth wide open, and her eyes rolled back. Her exact head angle, arm positions, leg positions, and lower-body support are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room glass window is open, while the front door retains its old key-operated lock. The female-form butler robot is incapacitated but still externally intact; the answered phone is playing the noisy digital recording. 이현우: Hyunwoo is beside the open living-room window, catching the collapsing butler robot in his arms. His ear communication setup remains in place.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 백인 여성형 집사 로봇 (백인 여성형 얼굴, 젊은 성인 외형, 인간 같은 얼굴 표면) — wearing: 부유한 주택에 어울리는 클래식한 흑백 톤의 메이드 앞치마와 단정한 원피스. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 창문 밖으로 힘없이 꺾인 백인 여성형 집사 로봇의 상체를 벽 모퉁이의 이현우가 두 팔로 덥석 받아 안는 순간.\n\nLOCATION (lock): Outside an upscale house, at the open living-room window beside an exterior wall corner. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open living-room window behind the collapsing robot in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Living-room window opening (Open, with the robot collapsing through it) — Seen from outside at an oblique angle; the opening exposes a limited portion of the living room; used as Shared spatial anchor connecting the interior body to Hyunwoo's exterior catch; Adjacent exterior wall corner (Beside the window where Hyunwoo had hidden) — Its outer edge lies toward the right side of the composition; used as Lateral boundary that explains Hyunwoo's short interception route.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight appropriate to the open window preserves humanlike facial detail and clear arm-to-body separation without revealing the robot's still-concealed metal interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The disabled female butler robot is collapsing limply outside the open window, with her upper body caught and supported in both of Hyunwoo's arms, her mouth wide open, and her eyes rolled back. Her exact head angle, arm positions, leg positions, and lower-body support are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room glass window is open, while the front door retains its old key-operated lock. The female-form butler robot is incapacitated but still externally intact; the answered phone is playing the noisy digital recording. 이현우: Hyunwoo is beside the open living-room window, catching the collapsing butler robot in his arms. His ear communication setup remains in place.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 백인 여성형 집사 로봇 (백인 여성형 얼굴, 젊은 성인 외형, 인간 같은 얼굴 표면) — wearing: 부유한 주택에 어울리는 클래식한 흑백 톤의 메이드 앞치마와 단정한 원피스. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S5sh10__bgfirst_bg.png",
     "asset_id": "ce41f054-4da2-4f2d-91e9-68adc725b197",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S5sh10.png",
     "asset_id": "1dd7fd01-6688-4035-ac05-5f0a0c3704b3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L07B01.png",
     "asset_id": "f93bd100-ada4-44ab-b370-a7e1748ac239",
     "role": "location_plate"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_luxury_house_sel.png",
     "asset_id": "93e38602-1247-4591-8e5c-e38083816dd4",
     "role": "structure_seed_look"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선은 로봇을 향하고, 로봇은 위를 쳐다봄.",
    "built_space": "좌측 상단에 창문, 우측에 모퉁이가 위치해 구도와 일치함. 건물 형태는 Structure 대신 Location 사진을 따름.",
    "entities": "이현우는 흰색 티셔츠를 입어 의상 오류. 로봇은 지시와 달리 기계 목 부위가 노출됨.",
    "hard_violations": [],
    "physics": "이현우가 두 팔로 로봇을 안정적으로 지탱하며, 몸이 중력에 맞게 늘어짐."
   },
   {
    "label": "B",
    "direction": "이현우는 로봇을, 로봇은 허공을 향함.",
    "built_space": "우측에 창문, 좌측에 모퉁이가 있어 요구 구도와 반대됨. 건물 형태는 Structure 사진과 일치함.",
    "entities": "이현우는 지정된 후드를 착용함. 로봇은 지나치게 인공적인 마네킹 형태로 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 잉여 신체 부위: 이현우의 우측 어깨 위에 정체불명의 흰색 팔과 손이 얹혀 있음."
    ],
    "physics": "이현우가 몸을 지탱 중이나, 어깨에 놓인 잉여 팔은 늘어진 상태라는 설정과 중력을 거스름."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지정된 화면 구도(좌측 상단 창문, 우측 모퉁이)를 잘 따랐으나, 건물 구조와 캐릭터 의상 레퍼런스 반영에 실패함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "의상과 건물 레퍼런스는 일치하나 구도가 반전되었고 잉여 신체가 발생하여 실격됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 로봇을 향하고, 로봇은 위를 쳐다봄.",
        "built_space": "좌측 상단에 창문, 우측에 모퉁이가 위치해 구도와 일치함. 건물 형태는 Structure 대신 Location 사진을 따름.",
        "entities": "이현우는 흰색 티셔츠를 입어 의상 오류. 로봇은 지시와 달리 기계 목 부위가 노출됨.",
        "hard_violations": [],
        "physics": "이현우가 두 팔로 로봇을 안정적으로 지탱하며, 몸이 중력에 맞게 늘어짐."
       },
       {
        "label": "B",
        "direction": "이현우는 로봇을, 로봇은 허공을 향함.",
        "built_space": "우측에 창문, 좌측에 모퉁이가 있어 요구 구도와 반대됨. 건물 형태는 Structure 사진과 일치함.",
        "entities": "이현우는 지정된 후드를 착용함. 로봇은 지나치게 인공적인 마네킹 형태로 묘사됨.",
        "hard_violations": [
         "잉여 신체 부위: 이현우의 우측 어깨 위에 정체불명의 흰색 팔과 손이 얹혀 있음."
        ],
        "physics": "이현우가 몸을 지탱 중이나, 어깨에 놓인 잉여 팔은 늘어진 상태라는 설정과 중력을 거스름."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지정된 화면 구도(좌측 상단 창문, 우측 모퉁이)를 잘 따랐으나, 건물 구조와 캐릭터 의상 레퍼런스 반영에 실패함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "의상과 건물 레퍼런스는 일치하나 구도가 반전되었고 잉여 신체가 발생하여 실격됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 로봇을 향하고, 로봇은 위를 쳐다봄.",
        "built_space": "좌측 상단에 창문, 우측에 모퉁이가 위치해 구도와 일치함. 건물 형태는 Structure 대신 Location 사진을 따름.",
        "entities": "이현우는 흰색 티셔츠를 입어 의상 오류. 로봇은 지시와 달리 기계 목 부위가 노출됨.",
        "hard_violations": [],
        "physics": "이현우가 두 팔로 로봇을 안정적으로 지탱하며, 몸이 중력에 맞게 늘어짐."
       },
       {
        "label": "B",
        "direction": "이현우는 로봇을, 로봇은 허공을 향함.",
        "built_space": "우측에 창문, 좌측에 모퉁이가 있어 요구 구도와 반대됨. 건물 형태는 Structure 사진과 일치함.",
        "entities": "이현우는 지정된 후드를 착용함. 로봇은 지나치게 인공적인 마네킹 형태로 묘사됨.",
        "hard_violations": [
         "잉여 신체 부위: 이현우의 우측 어깨 위에 정체불명의 흰색 팔과 손이 얹혀 있음."
        ],
        "physics": "이현우가 몸을 지탱 중이나, 어깨에 놓인 잉여 팔은 늘어진 상태라는 설정과 중력을 거스름."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "기와·벽돌과 현우의 후드 의상은 참조에 가깝지만, 창이 오른쪽 배경에 놓여 지정 구도를 뒤집고 로봇의 얼굴도 인간 피부보다 인형 표면에 가깝다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "왼쪽 위 열린 창과 오른쪽 벽 모퉁이 사이에서 두 팔로 받아내는 배치는 더 정확하지만, 구조 참조와 다른 창호·외벽, 바뀐 의상과 노출된 목 기계부가 충실도를 낮춘다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 오른쪽에 있는 로봇의 얼굴을 바라본다. 로봇의 눈은 위로 돌아가 있고 입은 열려 있다. 로봇은 창 앞에서 머리가 오른쪽, 하체가 왼쪽 아래로 기울어져 있다. 열린 창의 주된 개구부가 로봇 뒤 오른쪽에 있어, 창을 왼쪽 위에 두라는 방향 배치와 다르다.",
        "built_space": "밝은 프레임의 열린 창 하나, 오른쪽 외벽 모서리 하나, 왼쪽 현관문 하나와 벽돌 기둥 두 개, 기와 처마가 보인다. 창 안에는 계단 일부가 보이고 현관 앞에는 목제 탁자와 의자가 있다. 기와·밝은 벽·벽돌은 구조 참조와 닮았지만, 참조의 바닥까지 내려오는 거실 유리 개구부가 높은 창턱의 여닫이창으로 바뀌었다. 현우는 창 바깥에 있으나 화면의 오른쪽 모퉁이보다 왼쪽 현관 쪽에 치우쳐 있다. 명백히 불가능한 반사는 보이지 않는다.",
        "entities": "인물은 현우와 여성형 로봇 두 명뿐이다. 현우는 젊은 동아시아계 남성 외형, 헝클어진 검은 머리, 낡은 짙은 후드와 귀 통신장치를 갖춰 인물 참조에 가깝다. 로봇은 젊은 백인 여성형 외모와 흑백 메이드 원피스·앞치마를 갖췄지만, 얼굴과 팔이 매우 균일한 백색이고 관절선이 있어 인간 같은 피부보다는 마네킹에 가깝다. 금속 내부는 드러나지 않는다. 창턱에는 전화기로 보이는 검은 기기가 있으나 녹음 재생 여부는 정지영상으로 확인할 수 없다. 읽을 수 있는 문자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "현우의 보이는 손은 로봇의 복부를 감싸며, 다른 팔은 몸통 뒤로 가려져 지지 접점이 분명하지 않다. 로봇의 한 팔은 아래로 늘어지고 다른 손은 현우의 어깨에 얹혀 지지된다. 뒤로 젖혀진 머리와 기울어진 몸통은 받아 안는 동작으로 설명 가능하며, 몸 전체가 무지지 상태로 떠 있다고 볼 근거는 없다. 두 사람의 발과 로봇 하체의 최종 지지는 화면 밖이다."
       },
       {
        "label": "B",
        "direction": "현우의 시선은 자신이 붙잡는 로봇의 몸통과 얼굴 쪽으로 내려간다. 로봇은 눈을 위로 돌리고 입을 크게 벌린 채 머리를 뒤로 젖혔다. 왼쪽 위 열린 창에서 바깥쪽 아래로 무너져 현우에게 안기는 방향이 읽혀 지정된 공간 관계에 더 가깝다.",
        "built_space": "왼쪽 위에 열린 큰 창 하나, 위쪽 벽에 작은 창 하나, 오른쪽에 외벽 모퉁이 하나가 보인다. 창 안에는 소파 일부와 식탁 쪽 조명이 제한적으로 드러난다. 현우는 창 바깥과 오른쪽 모퉁이 사이에 있어 요청한 배치에 맞는다. 다만 짙은 창틀과 높게 솟은 창턱, 현대식 외벽은 구조 참조의 밝은 프레임과 바닥까지 내려오는 거실 개구부를 재현하지 못한다. 유리의 수목 반사는 외부 촬영 위치에서 가능해 보인다.",
        "entities": "현우와 여성형 로봇만 보인다. 현우는 젊은 동아시아계 남성 외형, 검은 헝클어진 머리와 귀 통신장치를 갖췄지만, 참조의 낡은 짙은 후드 대신 흰 반소매 티셔츠를 입었다. 로봇은 젊은 백인 여성형 얼굴, 인간 같은 피부, 흑백 메이드 원피스와 앞치마를 갖췄다. 그러나 목에 검은 기계 연결부가 노출되어 외관이 온전하고 금속 내부가 감춰져 있어야 한다는 조건에 어긋난다. 전화기와 현관 잠금장치는 화면에서 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "현우의 한 손과 팔은 로봇의 갈비뼈 아래 몸통을 받치고, 다른 손과 팔은 허리·골반 쪽을 붙잡아 두 팔의 지지가 분명하다. 현우는 무릎을 굽히고 다리를 벌려 무게를 받아내며, 발은 화면 밖이다. 로봇의 보이는 팔과 손은 아래로 늘어지고 머리는 뒤로 처져 있어 무력화된 상태와 부합한다. 로봇 하체는 잘려 있지만 상체에는 명확한 지지가 있어 공중 부유로 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "기와·벽돌과 현우의 후드 의상은 참조에 가깝지만, 창이 오른쪽 배경에 놓여 지정 구도를 뒤집고 로봇의 얼굴도 인간 피부보다 인형 표면에 가깝다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "왼쪽 위 열린 창과 오른쪽 벽 모퉁이 사이에서 두 팔로 받아내는 배치는 더 정확하지만, 구조 참조와 다른 창호·외벽, 바뀐 의상과 노출된 목 기계부가 충실도를 낮춘다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 오른쪽에 있는 로봇의 얼굴을 바라본다. 로봇의 눈은 위로 돌아가 있고 입은 열려 있다. 로봇은 창 앞에서 머리가 오른쪽, 하체가 왼쪽 아래로 기울어져 있다. 열린 창의 주된 개구부가 로봇 뒤 오른쪽에 있어, 창을 왼쪽 위에 두라는 방향 배치와 다르다.",
        "built_space": "밝은 프레임의 열린 창 하나, 오른쪽 외벽 모서리 하나, 왼쪽 현관문 하나와 벽돌 기둥 두 개, 기와 처마가 보인다. 창 안에는 계단 일부가 보이고 현관 앞에는 목제 탁자와 의자가 있다. 기와·밝은 벽·벽돌은 구조 참조와 닮았지만, 참조의 바닥까지 내려오는 거실 유리 개구부가 높은 창턱의 여닫이창으로 바뀌었다. 현우는 창 바깥에 있으나 화면의 오른쪽 모퉁이보다 왼쪽 현관 쪽에 치우쳐 있다. 명백히 불가능한 반사는 보이지 않는다.",
        "entities": "인물은 현우와 여성형 로봇 두 명뿐이다. 현우는 젊은 동아시아계 남성 외형, 헝클어진 검은 머리, 낡은 짙은 후드와 귀 통신장치를 갖춰 인물 참조에 가깝다. 로봇은 젊은 백인 여성형 외모와 흑백 메이드 원피스·앞치마를 갖췄지만, 얼굴과 팔이 매우 균일한 백색이고 관절선이 있어 인간 같은 피부보다는 마네킹에 가깝다. 금속 내부는 드러나지 않는다. 창턱에는 전화기로 보이는 검은 기기가 있으나 녹음 재생 여부는 정지영상으로 확인할 수 없다. 읽을 수 있는 문자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "현우의 보이는 손은 로봇의 복부를 감싸며, 다른 팔은 몸통 뒤로 가려져 지지 접점이 분명하지 않다. 로봇의 한 팔은 아래로 늘어지고 다른 손은 현우의 어깨에 얹혀 지지된다. 뒤로 젖혀진 머리와 기울어진 몸통은 받아 안는 동작으로 설명 가능하며, 몸 전체가 무지지 상태로 떠 있다고 볼 근거는 없다. 두 사람의 발과 로봇 하체의 최종 지지는 화면 밖이다."
       },
       {
        "label": "A",
        "direction": "현우의 시선은 자신이 붙잡는 로봇의 몸통과 얼굴 쪽으로 내려간다. 로봇은 눈을 위로 돌리고 입을 크게 벌린 채 머리를 뒤로 젖혔다. 왼쪽 위 열린 창에서 바깥쪽 아래로 무너져 현우에게 안기는 방향이 읽혀 지정된 공간 관계에 더 가깝다.",
        "built_space": "왼쪽 위에 열린 큰 창 하나, 위쪽 벽에 작은 창 하나, 오른쪽에 외벽 모퉁이 하나가 보인다. 창 안에는 소파 일부와 식탁 쪽 조명이 제한적으로 드러난다. 현우는 창 바깥과 오른쪽 모퉁이 사이에 있어 요청한 배치에 맞는다. 다만 짙은 창틀과 높게 솟은 창턱, 현대식 외벽은 구조 참조의 밝은 프레임과 바닥까지 내려오는 거실 개구부를 재현하지 못한다. 유리의 수목 반사는 외부 촬영 위치에서 가능해 보인다.",
        "entities": "현우와 여성형 로봇만 보인다. 현우는 젊은 동아시아계 남성 외형, 검은 헝클어진 머리와 귀 통신장치를 갖췄지만, 참조의 낡은 짙은 후드 대신 흰 반소매 티셔츠를 입었다. 로봇은 젊은 백인 여성형 얼굴, 인간 같은 피부, 흑백 메이드 원피스와 앞치마를 갖췄다. 그러나 목에 검은 기계 연결부가 노출되어 외관이 온전하고 금속 내부가 감춰져 있어야 한다는 조건에 어긋난다. 전화기와 현관 잠금장치는 화면에서 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "현우의 한 손과 팔은 로봇의 갈비뼈 아래 몸통을 받치고, 다른 손과 팔은 허리·골반 쪽을 붙잡아 두 팔의 지지가 분명하다. 현우는 무릎을 굽히고 다리를 벌려 무게를 받아내며, 발은 화면 밖이다. 로봇의 보이는 팔과 손은 아래로 늘어지고 머리는 뒤로 처져 있어 무력화된 상태와 부합한다. 로봇 하체는 잘려 있지만 상체에는 명확한 지지가 있어 공중 부유로 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.433
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.183
   },
   "violations": {
    "B": [
     "[gemini-pro] 잉여 신체 부위: 이현우의 우측 어깨 위에 정체불명의 흰색 팔과 손이 얹혀 있음."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1183
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 화면 구도(좌측 상단 창문, 우측 모퉁이)를 잘 따랐으나, 건물 구조와 캐릭터 의상 레퍼런스 반영에 실패함."
   },
   {
    "label": "B",
    "score": 1183,
    "verdict_ko": "의상과 건물 레퍼런스는 일치하나 구도가 반전되었고 잉여 신체가 발생하여 실격됨.  ★위반: [gemini-pro] 잉여 신체 부위: 이현우의 우측 어깨 위에 정체불명의 흰색 팔과 손이 얹혀 있음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L07B01.png",
    "asset_id": "f93bd100-ada4-44ab-b370-a7e1748ac239",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_luxury_house_sel.png",
    "asset_id": "93e38602-1247-4591-8e5c-e38083816dd4",
    "role": "structure_seed_look"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd74-a692-773d-90e4-25d189446f00",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S5sh10__bgfirst_bg.png",
   "bg_asset_id": "ce41f054-4da2-4f2d-91e9-68adc725b197",
   "bg_record_key": "S5sh10::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S5sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:16:42.506865+00:00",
  "fingerprint": "2aec8a72f7ef21aa487e897b9f27595da24580a234744bbe3302b405bcaf0739",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S5sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S5sh10_sel.png",
  "source_sha256": "da30a49697c3f30e04bc7c2a689149492b50ac1949c1955491be1488419554fe",
  "file": "S5sh10_cine.png",
  "staged_sha256": "8aae377ac5932136d8b2bc1dbe1689eec28995a19d352a3ea8e9d8def8bc2879",
  "latency_ms": 11662
 },
 "S5sh20::signage": {
  "fp": "6538eb13a91f159d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S5sh20::bgfirst_bg": {
  "input_fingerprint": "d327bd804c77f27e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 2층 창문 틀을 밟고 밖으로 몸을 날리려는 이현우의 오른쪽 종아리 바지를 셰퍼드가 사나운 이빨로 콱 물고 늘어지는 순간.\n\nLOCATION (lock): At the inside edge of a second-floor window in the upscale house, with daylight entering through the open escape window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open second-floor window and supporting lower frame in the upper-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Second-floor window frame (Window open, with Hyunwoo's foot bearing on the frame) — Its interior side and supporting lower edge are visible on the right; used as Footing reference and upper-right escape destination; Second-floor interior (Behind Hyunwoo along the shepherd's approach); used as Lower-left depth context distinguishing the dog's interior approach from Hyunwoo's outward movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight from the open window and ambient interior illumination keep the trouser grip and footing readable without adding unsupported injury detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 2층 창문 틀을 밟고 밖으로 몸을 날리려는 이현우의 오른쪽 종아리 바지를 셰퍼드가 사나운 이빨로 콱 물고 늘어지는 순간.\n\nLOCATION (lock): At the inside edge of a second-floor window in the upscale house, with daylight entering through the open escape window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open second-floor window and supporting lower frame in the upper-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Second-floor window frame (Window open, with Hyunwoo's foot bearing on the frame) — Its interior side and supporting lower edge are visible on the right; used as Footing reference and upper-right escape destination; Second-floor interior (Behind Hyunwoo along the shepherd's approach); used as Lower-left depth context distinguishing the dog's interior approach from Hyunwoo's outward movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight from the open window and ambient interior illumination keep the trouser grip and footing readable without adding unsupported injury detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S5sh20__bgfirst_bg.png",
  "asset_id": "e85a3e9a-38f2-44d2-ae4f-f94391f1ced7",
  "input_asset_ids": [
   "737ad8ec-60cb-4d2d-8fed-3f17f28c46a9",
   "f93bd100-ada4-44ab-b370-a7e1748ac239",
   "93e38602-1247-4591-8e5c-e38083816dd4"
  ]
 },
 "S5sh20": {
  "input_fingerprint": "9e78312103616769",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 2층 창문 틀을 밟고 밖으로 몸을 날리려는 이현우의 오른쪽 종아리 바지를 셰퍼드가 사나운 이빨로 콱 물고 늘어지는 순간.\n\nLOCATION (lock): At the inside edge of a second-floor window in the upscale house, with daylight entering through the open escape window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open second-floor window and supporting lower frame in the upper-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Second-floor window frame (Window open, with Hyunwoo's foot bearing on the frame) — Its interior side and supporting lower edge are visible on the right; used as Footing reference and upper-right escape destination; Second-floor interior (Behind Hyunwoo along the shepherd's approach); used as Lower-left depth context distinguishing the dog's interior approach from Hyunwoo's outward movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight from the open window and ambient interior illumination keep the trouser grip and footing readable without adding unsupported injury detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room window remains open, and the disabled butler robot lies on the floor with torn head covering exposing metal underneath. The second-floor escape window is now open, and the shepherd has reached the upstairs window. 이현우: Hyunwoo is at the open second-floor window and sustains a dog bite to his leg. His ear communication setup has not been removed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 셰퍼드 (길쭉한 주둥이, 곧게 선 삼각형 귀, 촘촘한 얼굴 털) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 2층 창문 틀을 밟고 밖으로 몸을 날리려는 이현우의 오른쪽 종아리 바지를 셰퍼드가 사나운 이빨로 콱 물고 늘어지는 순간.\n\nLOCATION (lock): At the inside edge of a second-floor window in the upscale house, with daylight entering through the open escape window. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open second-floor window and supporting lower frame in the upper-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Second-floor window frame (Window open, with Hyunwoo's foot bearing on the frame) — Its interior side and supporting lower edge are visible on the right; used as Footing reference and upper-right escape destination; Second-floor interior (Behind Hyunwoo along the shepherd's approach); used as Lower-left depth context distinguishing the dog's interior approach from Hyunwoo's outward movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight from the open window and ambient interior illumination keep the trouser grip and footing readable without adding unsupported injury detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room window remains open, and the disabled butler robot lies on the floor with torn head covering exposing metal underneath. The second-floor escape window is now open, and the shepherd has reached the upstairs window. 이현우: Hyunwoo is at the open second-floor window and sustains a dog bite to his leg. His ear communication setup has not been removed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 셰퍼드 (길쭉한 주둥이, 곧게 선 삼각형 귀, 촘촘한 얼굴 털) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 2층 창문 틀을 밟고 밖으로 몸을 날리려는 이현우의 오른쪽 종아리 바지를 셰퍼드가 사나운 이빨로 콱 물고 늘어지는 순간.\n\nLOCATION (lock): At the inside edge of a second-floor window in the upscale house, with daylight entering through the open escape window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open second-floor window and supporting lower frame in the upper-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Second-floor window frame (Window open, with Hyunwoo's foot bearing on the frame) — Its interior side and supporting lower edge are visible on the right; used as Footing reference and upper-right escape destination; Second-floor interior (Behind Hyunwoo along the shepherd's approach); used as Lower-left depth context distinguishing the dog's interior approach from Hyunwoo's outward movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight from the open window and ambient interior illumination keep the trouser grip and footing readable without adding unsupported injury detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room window remains open, and the disabled butler robot lies on the floor with torn head covering exposing metal underneath. The second-floor escape window is now open, and the shepherd has reached the upstairs window. 이현우: Hyunwoo is at the open second-floor window and sustains a dog bite to his leg. His ear communication setup has not been removed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 셰퍼드 (길쭉한 주둥이, 곧게 선 삼각형 귀, 촘촘한 얼굴 털) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S5sh20__bgfirst_bg.png",
     "asset_id": "e85a3e9a-38f2-44d2-ae4f-f94391f1ced7",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S5sh20.png",
     "asset_id": "737ad8ec-60cb-4d2d-8fed-3f17f28c46a9",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L07B01.png",
     "asset_id": "f93bd100-ada4-44ab-b370-a7e1748ac239",
     "role": "location_plate"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_luxury_house_sel.png",
     "asset_id": "93e38602-1247-4591-8e5c-e38083816dd4",
     "role": "structure_seed_look"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 창문 밖으로 몸을 내밀려 하고 있으며, 셰퍼드는 그의 바닥에 있는 왼쪽 다리를 향해 주둥이를 뻗어 바지를 물고 있음 (지정된 오른쪽 종아리가 아님).",
    "built_space": "고급 주택의 2층 내부로 추정됨. 우측에 바깥으로 열린 큰 창문이 있고, 좌측 뒤쪽으로 TV와 거실 가구들이 배치됨. 좌측 바닥에 로봇이 누워 있음.",
    "entities": "이현우: 무선 이어폰과 검은 머리는 일치하나, 레퍼런스와 전혀 다른 회색 반팔 티셔츠와 바지를 착용함. 셰퍼드: 개. 로봇: 바닥에 쓰러진 형태.",
    "hard_violations": [],
    "physics": "이현우는 왼쪽 발을 바닥에 디디고 오른쪽 발을 창틀에 올린 채 몸을 지탱함. 셰퍼드는 네 발을 바닥에 딛고 서서 다리를 물고 있음. 지지 구조에 물리적 오류는 없음."
   },
   {
    "label": "B",
    "direction": "이현우는 우측의 열린 창문 밖으로 몸을 날리려 하고 있으며, 셰퍼드는 뒤로 뻗어 있는 그의 오른쪽 종아리 바지를 정확히 물고 늘어지고 있음.",
    "built_space": "2층 복도 또는 방 내부. 우측 상단에 안으로 열린 창문이 있으며, 좌측 하단으로 깊이감 있는 실내 복도 공간이 이어짐. 카메라 구도와 배경 배치가 지시사항과 일치함.",
    "entities": "이현우: 검은 헝클어진 머리, 무선 이어폰 착용, 레퍼런스와 정확히 일치하는 낡은 짙은 회색 후드티와 무릎이 찢어진 청바지를 착용함. 셰퍼드: 개. 로봇: 좌측 하단 바닥에 머리 덮개가 파손되어 내부가 드러난 채 쓰러져 있음.",
    "hard_violations": [],
    "physics": "이현우는 왼쪽 발을 창틀에 올리고 오른손으로 창틀을 잡아 체중을 싣고 있으며, 오른쪽 다리는 허공에 뜬 채 셰퍼드에게 물려 당겨지고 있음. 셰퍼드는 바닥에 앞발을 딛고 체중을 실어 바지를 끌어당기고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "캐릭터 레퍼런스의 지정된 의상을 완전히 무시하였으며, 프롬프트에 명시된 '오른쪽 종아리'가 아닌 바닥에 딛고 있는 왼쪽 다리를 개가 물고 있어 핵심 연출 지시를 크게 위반했습니다."
       },
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "캐릭터의 얼굴과 지정된 낡은 후드티, 찢어진 청바지 의상을 정확히 재현했으며, 오른쪽 종아리를 셰퍼드가 물고 늘어지는 상황과 지시된 화면 구도를 완벽하게 구현한 훌륭한 결과물입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 창문 밖으로 몸을 내밀려 하고 있으며, 셰퍼드는 그의 바닥에 있는 왼쪽 다리를 향해 주둥이를 뻗어 바지를 물고 있음 (지정된 오른쪽 종아리가 아님).",
        "built_space": "고급 주택의 2층 내부로 추정됨. 우측에 바깥으로 열린 큰 창문이 있고, 좌측 뒤쪽으로 TV와 거실 가구들이 배치됨. 좌측 바닥에 로봇이 누워 있음.",
        "entities": "이현우: 무선 이어폰과 검은 머리는 일치하나, 레퍼런스와 전혀 다른 회색 반팔 티셔츠와 바지를 착용함. 셰퍼드: 개. 로봇: 바닥에 쓰러진 형태.",
        "hard_violations": [],
        "physics": "이현우는 왼쪽 발을 바닥에 디디고 오른쪽 발을 창틀에 올린 채 몸을 지탱함. 셰퍼드는 네 발을 바닥에 딛고 서서 다리를 물고 있음. 지지 구조에 물리적 오류는 없음."
       },
       {
        "label": "B",
        "direction": "이현우는 우측의 열린 창문 밖으로 몸을 날리려 하고 있으며, 셰퍼드는 뒤로 뻗어 있는 그의 오른쪽 종아리 바지를 정확히 물고 늘어지고 있음.",
        "built_space": "2층 복도 또는 방 내부. 우측 상단에 안으로 열린 창문이 있으며, 좌측 하단으로 깊이감 있는 실내 복도 공간이 이어짐. 카메라 구도와 배경 배치가 지시사항과 일치함.",
        "entities": "이현우: 검은 헝클어진 머리, 무선 이어폰 착용, 레퍼런스와 정확히 일치하는 낡은 짙은 회색 후드티와 무릎이 찢어진 청바지를 착용함. 셰퍼드: 개. 로봇: 좌측 하단 바닥에 머리 덮개가 파손되어 내부가 드러난 채 쓰러져 있음.",
        "hard_violations": [],
        "physics": "이현우는 왼쪽 발을 창틀에 올리고 오른손으로 창틀을 잡아 체중을 싣고 있으며, 오른쪽 다리는 허공에 뜬 채 셰퍼드에게 물려 당겨지고 있음. 셰퍼드는 바닥에 앞발을 딛고 체중을 실어 바지를 끌어당기고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "캐릭터 레퍼런스의 지정된 의상을 완전히 무시하였으며, 프롬프트에 명시된 '오른쪽 종아리'가 아닌 바닥에 딛고 있는 왼쪽 다리를 개가 물고 있어 핵심 연출 지시를 크게 위반했습니다."
       },
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "캐릭터의 얼굴과 지정된 낡은 후드티, 찢어진 청바지 의상을 정확히 재현했으며, 오른쪽 종아리를 셰퍼드가 물고 늘어지는 상황과 지시된 화면 구도를 완벽하게 구현한 훌륭한 결과물입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 창문 밖으로 몸을 내밀려 하고 있으며, 셰퍼드는 그의 바닥에 있는 왼쪽 다리를 향해 주둥이를 뻗어 바지를 물고 있음 (지정된 오른쪽 종아리가 아님).",
        "built_space": "고급 주택의 2층 내부로 추정됨. 우측에 바깥으로 열린 큰 창문이 있고, 좌측 뒤쪽으로 TV와 거실 가구들이 배치됨. 좌측 바닥에 로봇이 누워 있음.",
        "entities": "이현우: 무선 이어폰과 검은 머리는 일치하나, 레퍼런스와 전혀 다른 회색 반팔 티셔츠와 바지를 착용함. 셰퍼드: 개. 로봇: 바닥에 쓰러진 형태.",
        "hard_violations": [],
        "physics": "이현우는 왼쪽 발을 바닥에 디디고 오른쪽 발을 창틀에 올린 채 몸을 지탱함. 셰퍼드는 네 발을 바닥에 딛고 서서 다리를 물고 있음. 지지 구조에 물리적 오류는 없음."
       },
       {
        "label": "B",
        "direction": "이현우는 우측의 열린 창문 밖으로 몸을 날리려 하고 있으며, 셰퍼드는 뒤로 뻗어 있는 그의 오른쪽 종아리 바지를 정확히 물고 늘어지고 있음.",
        "built_space": "2층 복도 또는 방 내부. 우측 상단에 안으로 열린 창문이 있으며, 좌측 하단으로 깊이감 있는 실내 복도 공간이 이어짐. 카메라 구도와 배경 배치가 지시사항과 일치함.",
        "entities": "이현우: 검은 헝클어진 머리, 무선 이어폰 착용, 레퍼런스와 정확히 일치하는 낡은 짙은 회색 후드티와 무릎이 찢어진 청바지를 착용함. 셰퍼드: 개. 로봇: 좌측 하단 바닥에 머리 덮개가 파손되어 내부가 드러난 채 쓰러져 있음.",
        "hard_violations": [],
        "physics": "이현우는 왼쪽 발을 창틀에 올리고 오른손으로 창틀을 잡아 체중을 싣고 있으며, 오른쪽 다리는 허공에 뜬 채 셰퍼드에게 물려 당겨지고 있음. 셰퍼드는 바닥에 앞발을 딛고 체중을 실어 바지를 끌어당기고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "창틀에 실린 발과 오른쪽 종아리의 바지를 당기는 물림이 명확하고 인물 의상도 맞지만, 지정한 미디엄 숏보다 넓으며 창호는 구조 참조와 다릅니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "탈출 방향과 실내 접근 구도는 맞지만 셰퍼드가 왼쪽 종아리를 겨냥하고, 창틀을 밟는 접촉이 불명확하며 의상도 인물 참조와 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우의 얼굴과 상체는 오른쪽 위 열린 창밖을 향한다. 셰퍼드는 왼쪽 실내에서 오른쪽 위로 주둥이를 뻗어 뒤에 남은 종아리의 바지 천을 실제로 물고 당긴다. 옆모습의 골반과 다리 연결상 창틀에 접어 올린 다리는 왼쪽, 물린 뒤쪽 다리는 오른쪽으로 읽힌다.",
        "built_space": "오른쪽에 큰 창 개구부 하나, 안쪽으로 열린 창짝 하나, 하부 창턱이 보인다. 현우의 신발 하나가 하부 창턱에 놓이고 셰퍼드는 왼쪽 복도 바닥에 있다. 왼쪽에는 여러 문틀과 바닥에 쓰러진 로봇 한 대가 보인다. 창밖 하단에는 가로 난간 두 줄도 있다. 탈출구와 실내 접근로의 배치는 맞지만, 검은 여닫이 창호는 구조 참조의 밝은 색 창호와 다르다. 사람과 개의 몸 전체 및 긴 복도까지 포함해 미디엄 숏보다 넓다.",
        "entities": "젊은 동아시아계 남성 한 명, 검은색 헝클어진 머리, 낡은 짙은 후드티, 해진 청바지와 운동화가 보이며 현우 참조에 가깝다. 귀에는 검은 통신 장치가 보인다. 개 한 마리는 긴 주둥이, 선 귀, 조밀한 검정·황갈색 털을 가진 셰퍼드다. 뒤에는 흰 외피가 손상되고 내부가 드러난 인간형 로봇 한 대가 누워 있다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "창턱 위 신발의 밑창이 수평 하부 프레임에 닿아 현우를 지지하고, 접힌 무릎과 앞으로 기운 몸통은 도약 직전 자세로 성립한다. 반대쪽 다리는 뒤로 내려와 있으며 셰퍼드의 턱에 잡힌 바지 천이 팽팽하다. 셰퍼드는 보이는 발들로 마룻바닥을 딛고 버틴다. 로봇은 바닥에 누워 지지된다. 근거 없이 공중에 떠 있는 몸은 없다."
       },
       {
        "label": "B",
        "direction": "등을 보이는 현우는 오른쪽 위 창밖으로 상체와 오른팔을 뻗는다. 오른쪽 무릎은 창 쪽으로 올라가 있고 왼쪽 다리는 실내에 남아 있다. 따라서 왼쪽 아래 셰퍼드의 주둥이가 닿는 대상은 요청한 오른쪽 종아리가 아니라 왼쪽 종아리다. 입은 크게 열려 있고 바지에 닿지만 천을 단단히 움켜 당기는 표현은 A보다 약하다.",
        "built_space": "오른쪽 위에 큰 창 개구부 하나와 열린 창짝 하나, 하부 창턱이 있다. 왼쪽 실내에는 텔레비전과 낮은 장, 조명, 문 개구부가 보이고 오른쪽 아래에는 소파 일부가 들어온다. 로봇 한 대는 뒤쪽 바닥에 누워 있다. 개의 접근 방향과 탈출구 위치는 요청에 맞고 인물은 화면을 크게 차지한다. 다만 이 창 역시 구조 참조의 밝은 창틀 대신 짙은 여닫이 창호이며, 밖에 보이는 기와집만으로 현재 건물의 구조 일치를 확인할 수 없다.",
        "entities": "검은 헝클어진 머리의 젊은 남성 한 명이 보이지만 얼굴 대부분이 가려져 참조 얼굴 일치는 확인하기 어렵다. 검은 반소매 티셔츠, 회갈색 바지, 검은 운동화는 참조의 후드티와 해진 청바지 차림과 다르다. 귀 부근에는 작은 통신 장치로 보이는 것이 있다. 셰퍼드 한 마리는 긴 주둥이와 선 삼각형 귀, 검정·황갈색 털을 갖췄다. 배경에는 머리 외피가 손상된 인간형 로봇 한 대가 보인다. 읽히는 글자는 없다.",
        "hard_violations": [],
        "physics": "현우의 왼손은 창턱을 짚고 왼발 앞부분은 실내 바닥에 닿아 있어 몸의 지지는 성립한다. 그러나 올린 오른발은 밑창 전체가 창턱에 놓이기보다 뒤꿈치가 실내 쪽 아래로 처지고 앞부분만 창턱 가장자리에 걸친 모습이어서, 창틀을 확실히 밟고 도약하려는 지지 관계가 약하다. 셰퍼드의 하체 일부는 잘렸지만 보이는 앞다리는 바닥으로 이어져 접근 자세가 가능하다. 로봇은 바닥에 지지된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "창틀에 실린 발과 오른쪽 종아리의 바지를 당기는 물림이 명확하고 인물 의상도 맞지만, 지정한 미디엄 숏보다 넓으며 창호는 구조 참조와 다릅니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "탈출 방향과 실내 접근 구도는 맞지만 셰퍼드가 왼쪽 종아리를 겨냥하고, 창틀을 밟는 접촉이 불명확하며 의상도 인물 참조와 다릅니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우의 얼굴과 상체는 오른쪽 위 열린 창밖을 향한다. 셰퍼드는 왼쪽 실내에서 오른쪽 위로 주둥이를 뻗어 뒤에 남은 종아리의 바지 천을 실제로 물고 당긴다. 옆모습의 골반과 다리 연결상 창틀에 접어 올린 다리는 왼쪽, 물린 뒤쪽 다리는 오른쪽으로 읽힌다.",
        "built_space": "오른쪽에 큰 창 개구부 하나, 안쪽으로 열린 창짝 하나, 하부 창턱이 보인다. 현우의 신발 하나가 하부 창턱에 놓이고 셰퍼드는 왼쪽 복도 바닥에 있다. 왼쪽에는 여러 문틀과 바닥에 쓰러진 로봇 한 대가 보인다. 창밖 하단에는 가로 난간 두 줄도 있다. 탈출구와 실내 접근로의 배치는 맞지만, 검은 여닫이 창호는 구조 참조의 밝은 색 창호와 다르다. 사람과 개의 몸 전체 및 긴 복도까지 포함해 미디엄 숏보다 넓다.",
        "entities": "젊은 동아시아계 남성 한 명, 검은색 헝클어진 머리, 낡은 짙은 후드티, 해진 청바지와 운동화가 보이며 현우 참조에 가깝다. 귀에는 검은 통신 장치가 보인다. 개 한 마리는 긴 주둥이, 선 귀, 조밀한 검정·황갈색 털을 가진 셰퍼드다. 뒤에는 흰 외피가 손상되고 내부가 드러난 인간형 로봇 한 대가 누워 있다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "창턱 위 신발의 밑창이 수평 하부 프레임에 닿아 현우를 지지하고, 접힌 무릎과 앞으로 기운 몸통은 도약 직전 자세로 성립한다. 반대쪽 다리는 뒤로 내려와 있으며 셰퍼드의 턱에 잡힌 바지 천이 팽팽하다. 셰퍼드는 보이는 발들로 마룻바닥을 딛고 버틴다. 로봇은 바닥에 누워 지지된다. 근거 없이 공중에 떠 있는 몸은 없다."
       },
       {
        "label": "A",
        "direction": "등을 보이는 현우는 오른쪽 위 창밖으로 상체와 오른팔을 뻗는다. 오른쪽 무릎은 창 쪽으로 올라가 있고 왼쪽 다리는 실내에 남아 있다. 따라서 왼쪽 아래 셰퍼드의 주둥이가 닿는 대상은 요청한 오른쪽 종아리가 아니라 왼쪽 종아리다. 입은 크게 열려 있고 바지에 닿지만 천을 단단히 움켜 당기는 표현은 A보다 약하다.",
        "built_space": "오른쪽 위에 큰 창 개구부 하나와 열린 창짝 하나, 하부 창턱이 있다. 왼쪽 실내에는 텔레비전과 낮은 장, 조명, 문 개구부가 보이고 오른쪽 아래에는 소파 일부가 들어온다. 로봇 한 대는 뒤쪽 바닥에 누워 있다. 개의 접근 방향과 탈출구 위치는 요청에 맞고 인물은 화면을 크게 차지한다. 다만 이 창 역시 구조 참조의 밝은 창틀 대신 짙은 여닫이 창호이며, 밖에 보이는 기와집만으로 현재 건물의 구조 일치를 확인할 수 없다.",
        "entities": "검은 헝클어진 머리의 젊은 남성 한 명이 보이지만 얼굴 대부분이 가려져 참조 얼굴 일치는 확인하기 어렵다. 검은 반소매 티셔츠, 회갈색 바지, 검은 운동화는 참조의 후드티와 해진 청바지 차림과 다르다. 귀 부근에는 작은 통신 장치로 보이는 것이 있다. 셰퍼드 한 마리는 긴 주둥이와 선 삼각형 귀, 검정·황갈색 털을 갖췄다. 배경에는 머리 외피가 손상된 인간형 로봇 한 대가 보인다. 읽히는 글자는 없다.",
        "hard_violations": [],
        "physics": "현우의 왼손은 창턱을 짚고 왼발 앞부분은 실내 바닥에 닿아 있어 몸의 지지는 성립한다. 그러나 올린 오른발은 밑창 전체가 창턱에 놓이기보다 뒤꿈치가 실내 쪽 아래로 처지고 앞부분만 창턱 가장자리에 걸친 모습이어서, 창틀을 확실히 밟고 도약하려는 지지 관계가 약하다. 셰퍼드의 하체 일부는 잘렸지만 보이는 앞다리는 바닥으로 이어져 접근 자세가 가능하다. 로봇은 바닥에 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.871,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.871,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 871,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 871,
    "verdict_ko": "캐릭터 레퍼런스의 지정된 의상을 완전히 무시하였으며, 프롬프트에 명시된 '오른쪽 종아리'가 아닌 바닥에 딛고 있는 왼쪽 다리를 개가 물고 있어 핵심 연출 지시를 크게 위반했습니다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "캐릭터의 얼굴과 지정된 낡은 후드티, 찢어진 청바지 의상을 정확히 재현했으며, 오른쪽 종아리를 셰퍼드가 물고 늘어지는 상황과 지시된 화면 구도를 완벽하게 구현한 훌륭한 결과물입니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L07B01.png",
    "asset_id": "f93bd100-ada4-44ab-b370-a7e1748ac239",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_luxury_house_sel.png",
    "asset_id": "93e38602-1247-4591-8e5c-e38083816dd4",
    "role": "structure_seed_look"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd75-67fe-7b6b-8913-d8f89d98d2bc",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S5sh20__bgfirst_bg.png",
   "bg_asset_id": "e85a3e9a-38f2-44d2-ae4f-f94391f1ced7",
   "bg_record_key": "S5sh20::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S5sh20::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:16:54.590621+00:00",
  "fingerprint": "c88830b01c168b35cbf5ed63856b81d16db2375374a4bd02c979c5bf800d5e71",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S5sh20_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S5sh20_sel.png",
  "source_sha256": "69e385bfa045b4b0b657a3c313e8545e128d9d0db60ed9180443ba907c8f28f6",
  "file": "S5sh20_cine.png",
  "staged_sha256": "aa1c618d5b8db7ca87bb3977587b84abb978025ff6c495d7b9741d3ec6b8ac89",
  "latency_ms": 12315
 },
 "S5sh24::signage": {
  "fp": "6dba5f18508595b2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S5sh24::bgfirst_bg": {
  "input_fingerprint": "f22d7a78ece3791e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 잔디밭에 꼿꼿이 선 이현우의 몸을 빙 둘러싼 채, 짙은 남색 제복을 입은 아시아계 한국인 경찰 여러 명이 권총을 일제히 겨눈 위압적인 구도.\n\nLOCATION (lock): On the lawn outside the upscale house, directly below the second-floor escape window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Lawn (Occupied by Hyunwoo and the surrounding police) — Its ground plane is seen obliquely from above; used as Continuous shared surface making the near-side and far-side encirclement legible; Police handguns (Raised and aimed at Hyunwoo) — Their barrels converge inward from different positions around him, not toward the camera; used as Multiple small directional lines enclosing the central figure without oversized foreground weapons.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation retains the dark navy uniforms and separates the encircling figures clearly against the lawn.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 잔디밭에 꼿꼿이 선 이현우의 몸을 빙 둘러싼 채, 짙은 남색 제복을 입은 아시아계 한국인 경찰 여러 명이 권총을 일제히 겨눈 위압적인 구도.\n\nLOCATION (lock): On the lawn outside the upscale house, directly below the second-floor escape window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Lawn (Occupied by Hyunwoo and the surrounding police) — Its ground plane is seen obliquely from above; used as Continuous shared surface making the near-side and far-side encirclement legible; Police handguns (Raised and aimed at Hyunwoo) — Their barrels converge inward from different positions around him, not toward the camera; used as Multiple small directional lines enclosing the central figure without oversized foreground weapons.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation retains the dark navy uniforms and separates the encircling figures clearly against the lawn.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S5sh24__bgfirst_bg.png",
  "asset_id": "c0bf2b7e-d98e-4a61-87e7-b483e5281967",
  "input_asset_ids": [
   "795bd908-c628-42a4-8d1f-e085c55947c3",
   "2c87ee92-f9b1-40d2-a384-6e45ad3dd181"
  ]
 },
 "S5sh24": {
  "input_fingerprint": "5648f56e40bb6816",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 잔디밭에 꼿꼿이 선 이현우의 몸을 빙 둘러싼 채, 짙은 남색 제복을 입은 아시아계 한국인 경찰 여러 명이 권총을 일제히 겨눈 위압적인 구도.\n\nLOCATION (lock): On the lawn outside the upscale house, directly below the second-floor escape window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Lawn (Occupied by Hyunwoo and the surrounding police) — Its ground plane is seen obliquely from above; used as Continuous shared surface making the near-side and far-side encirclement legible; Police handguns (Raised and aimed at Hyunwoo) — Their barrels converge inward from different positions around him, not toward the camera; used as Multiple small directional lines enclosing the central figure without oversized foreground weapons.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation retains the dark navy uniforms and separates the encircling figures clearly against the lawn.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The second-floor escape window and the living-room window remain open. The incapacitated butler robot remains on the floor with exposed metal in its damaged head. 이현우: Hyunwoo has landed outside and stood up, retaining the fresh dog-bite injury to his leg.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 아시아계 한국인 경찰 여러 명 right now, so 아시아계 한국인 경찰 여러 명's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 아시아계 한국인 경찰 여러 명: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 잔디밭에 꼿꼿이 선 이현우의 몸을 빙 둘러싼 채, 짙은 남색 제복을 입은 아시아계 한국인 경찰 여러 명이 권총을 일제히 겨눈 위압적인 구도.\n\nLOCATION (lock): On the lawn outside the upscale house, directly below the second-floor escape window. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Lawn (Occupied by Hyunwoo and the surrounding police) — Its ground plane is seen obliquely from above; used as Continuous shared surface making the near-side and far-side encirclement legible; Police handguns (Raised and aimed at Hyunwoo) — Their barrels converge inward from different positions around him, not toward the camera; used as Multiple small directional lines enclosing the central figure without oversized foreground weapons.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation retains the dark navy uniforms and separates the encircling figures clearly against the lawn.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The second-floor escape window and the living-room window remain open. The incapacitated butler robot remains on the floor with exposed metal in its damaged head. 이현우: Hyunwoo has landed outside and stood up, retaining the fresh dog-bite injury to his leg.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 아시아계 한국인 경찰 여러 명 right now, so 아시아계 한국인 경찰 여러 명's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 아시아계 한국인 경찰 여러 명: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 잔디밭에 꼿꼿이 선 이현우의 몸을 빙 둘러싼 채, 짙은 남색 제복을 입은 아시아계 한국인 경찰 여러 명이 권총을 일제히 겨눈 위압적인 구도.\n\nLOCATION (lock): On the lawn outside the upscale house, directly below the second-floor escape window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Lawn (Occupied by Hyunwoo and the surrounding police) — Its ground plane is seen obliquely from above; used as Continuous shared surface making the near-side and far-side encirclement legible; Police handguns (Raised and aimed at Hyunwoo) — Their barrels converge inward from different positions around him, not toward the camera; used as Multiple small directional lines enclosing the central figure without oversized foreground weapons.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation retains the dark navy uniforms and separates the encircling figures clearly against the lawn.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The second-floor escape window and the living-room window remain open. The incapacitated butler robot remains on the floor with exposed metal in its damaged head. 이현우: Hyunwoo has landed outside and stood up, retaining the fresh dog-bite injury to his leg.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 아시아계 한국인 경찰 여러 명 right now, so 아시아계 한국인 경찰 여러 명's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 아시아계 한국인 경찰 여러 명: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S5sh24__bgfirst_bg.png",
     "asset_id": "c0bf2b7e-d98e-4a61-87e7-b483e5281967",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S5sh24.png",
     "asset_id": "795bd908-c628-42a4-8d1f-e085c55947c3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L07B02.png",
     "asset_id": "2c87ee92-f9b1-40d2-a384-6e45ad3dd181",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "경찰관들과 화면 하단의 총구가 모두 중앙의 이현우를 향해 정확히 겨냥되고 있음.",
    "built_space": "참조 사진의 주택 외관과 잔디밭이 일치함. 2층 창문이 열려 있고, 1층 거실 바닥에 쓰러진 로봇이 보임.",
    "entities": "경찰들은 남색 제복을 입음. 이현우는 얼굴과 바지가 참조와 일치하나 후드티 대신 반팔을 입고 있음. 로봇이 존재함.",
    "hard_violations": [],
    "physics": "모든 인물이 잔디밭에 안정적으로 서 있으며, 로봇은 바닥에 누워 있음. 하단의 손은 화면 밖 신체에 의해 지지됨."
   },
   {
    "label": "B",
    "direction": "경찰관들의 총구가 중앙의 이현우를 향해 겨냥되고 있음.",
    "built_space": "참조 사진의 주택과 잔디밭이 구현됨. 2층 창문이 열려 있으나 1층 거실 내부에 쓰러진 로봇은 없음.",
    "entities": "이현우의 얼굴과 후드티, 바지가 참조 이미지와 정확히 일치함. 경찰들은 남색 제복을 착용함.",
    "hard_violations": [],
    "physics": "모든 인물이 잔디밭 위에 안정적인 자세로 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "부감 샷과 와이드 구도, 거실 내부의 로봇 등 우선순위가 높은 촬영 지시를 훌륭히 따랐으나, 이현우의 후드티 의상이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이현우의 의상은 정확하게 재현했으나, 지시된 부감 샷을 무시하고 전경에 거대한 무기와 인물을 배치하여 카메라 및 구도 지시를 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "경찰관들과 화면 하단의 총구가 모두 중앙의 이현우를 향해 정확히 겨냥되고 있음.",
        "built_space": "참조 사진의 주택 외관과 잔디밭이 일치함. 2층 창문이 열려 있고, 1층 거실 바닥에 쓰러진 로봇이 보임.",
        "entities": "경찰들은 남색 제복을 입음. 이현우는 얼굴과 바지가 참조와 일치하나 후드티 대신 반팔을 입고 있음. 로봇이 존재함.",
        "hard_violations": [],
        "physics": "모든 인물이 잔디밭에 안정적으로 서 있으며, 로봇은 바닥에 누워 있음. 하단의 손은 화면 밖 신체에 의해 지지됨."
       },
       {
        "label": "B",
        "direction": "경찰관들의 총구가 중앙의 이현우를 향해 겨냥되고 있음.",
        "built_space": "참조 사진의 주택과 잔디밭이 구현됨. 2층 창문이 열려 있으나 1층 거실 내부에 쓰러진 로봇은 없음.",
        "entities": "이현우의 얼굴과 후드티, 바지가 참조 이미지와 정확히 일치함. 경찰들은 남색 제복을 착용함.",
        "hard_violations": [],
        "physics": "모든 인물이 잔디밭 위에 안정적인 자세로 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "부감 샷과 와이드 구도, 거실 내부의 로봇 등 우선순위가 높은 촬영 지시를 훌륭히 따랐으나, 이현우의 후드티 의상이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이현우의 의상은 정확하게 재현했으나, 지시된 부감 샷을 무시하고 전경에 거대한 무기와 인물을 배치하여 카메라 및 구도 지시를 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "경찰관들과 화면 하단의 총구가 모두 중앙의 이현우를 향해 정확히 겨냥되고 있음.",
        "built_space": "참조 사진의 주택 외관과 잔디밭이 일치함. 2층 창문이 열려 있고, 1층 거실 바닥에 쓰러진 로봇이 보임.",
        "entities": "경찰들은 남색 제복을 입음. 이현우는 얼굴과 바지가 참조와 일치하나 후드티 대신 반팔을 입고 있음. 로봇이 존재함.",
        "hard_violations": [],
        "physics": "모든 인물이 잔디밭에 안정적으로 서 있으며, 로봇은 바닥에 누워 있음. 하단의 손은 화면 밖 신체에 의해 지지됨."
       },
       {
        "label": "B",
        "direction": "경찰관들의 총구가 중앙의 이현우를 향해 겨냥되고 있음.",
        "built_space": "참조 사진의 주택과 잔디밭이 구현됨. 2층 창문이 열려 있으나 1층 거실 내부에 쓰러진 로봇은 없음.",
        "entities": "이현우의 얼굴과 후드티, 바지가 참조 이미지와 정확히 일치함. 경찰들은 남색 제복을 착용함.",
        "hard_violations": [],
        "physics": "모든 인물이 잔디밭 위에 안정적인 자세로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "현우의 얼굴·후드 의상과 직립 자세는 잘 맞지만, 눈높이 구도와 크게 잘린 전경 경찰·권총 때문에 위에서 내려다보는 잔디의 연속면과 앞뒤 포위 배치가 제대로 드러나지 않는다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "비스듬한 하이앵글 와이드숏으로 잔디 위 앞뒤 포위 관계를 가장 충실하게 구현했지만, 하단의 큰 전경 권총과 현우의 반팔 의상·불명확한 물린 상처는 요구와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "좌우 및 뒤쪽 경찰의 시선과 권총은 대체로 중앙 현우의 상체를 향한다. 아래쪽 두 경찰도 팔을 안쪽으로 뻗어 현우를 조준하며, 관객을 향해 총구를 들이대지는 않는다. 다만 겹쳐진 전경 팔과 큰 권총 때문에 개별 사선과 원형 포위 관계가 일부 가려진다. 현우는 정면을 바라본다.",
        "built_space": "밝은 외벽에 아래층 대형 유리 개구부 두 곳, 오른쪽 현관 한 곳, 위층 큰 창 한 곳과 작은 창 한 곳이 보인다. 큰 위층 창과 작은 창은 열려 있지만 아래층 거실창의 열린 부분은 명확하지 않다. 참고의 주요 입면 배치는 대체로 유지되지만, 현관 차양은 전경 인물에 가려 확인하기 어렵다. 중앙 현우와 경찰 일곱 명이 잔디 위에 배치되어 있으나 카메라가 거의 눈높이라 가까운 쪽과 먼 쪽을 연결하는 지면이 충분히 펼쳐지지 않는다. 유리의 수목 반사는 가능한 배치다.",
        "entities": "현우는 검은 헝클어진 머리의 젊은 동아시아계 남성으로, 참고의 얼굴과 마른 체형, 낡은 짙은 후드, 찢어진 청바지, 운동화에 가깝다. 무릎의 찢어진 부분은 보이나 신선한 개 물림 상처는 식별되지 않는다. 경찰 일곱 명은 성인 동아시아계 남성으로 보이고 짙은 남색 제복과 권총을 갖췄다. 국적은 외모만으로 확정할 수 없다. 실내 바닥의 로봇은 보이지 않으며, 실내가 가려져 그 상태를 판단할 수 없다. 명확히 읽히는 문구나 화면 오버레이는 없다.",
        "hard_violations": [],
        "physics": "현우의 두 발은 잔디에 닿고 몸은 꼿꼿하게 서 있다. 전신이 보이는 경찰들은 벌린 발로 지면을 딛고, 권총은 각 경찰의 손에 잡혀 있다. 전경 경찰의 하체는 화면 밖이지만 상체가 공중에 떠 있다는 증거는 없다. 지지 없이 떠 있는 물체나 명백히 불가능한 관절 자세는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽·뒤쪽·오른쪽·앞쪽 경찰들이 중앙 현우를 향해 몸과 시선을 돌리고 권총을 안쪽으로 겨눈다. 왼쪽 앞 경찰의 총구는 오른쪽 위의 현우 쪽, 오른쪽 앞 경찰의 총구는 왼쪽 위의 현우 쪽을 향한다. 오른쪽 끝 경찰은 현우의 몸통 또는 허리 높이로 조준한다. 하단 중앙의 별도 손에 잡힌 권총도 장면 안쪽을 향하지만 단축되어 정확한 조준점은 확인하기 어렵다. 현우의 얼굴은 화면 오른쪽 경찰 쪽으로 돌아가 있다.",
        "built_space": "위에서 비스듬히 본 잔디가 하나의 연속면으로 펼쳐지고, 전신 경찰 일곱 명과 하단의 부분적으로 보이는 경찰 한 명이 현우의 앞뒤와 좌우를 둘러싼다. 외벽에는 아래층 대형 유리 개구부 두 곳, 오른쪽 현관과 차양 각 한 곳, 위층 큰 창과 작은 창 각 한 곳, 아래층 벽등 두 개가 보인다. 위층 큰 창과 아래층 거실창은 열린 상태로 읽힌다. 참고의 밝은 외벽·검은 창틀·담장·잔디라는 장소 구성이 유지된다. 현관 주변 화분은 두 개로 참고의 한 개와 다르고 창 세부도 완전히 같지는 않다. 부자연스러운 거울상은 보이지 않는다.",
        "entities": "현우는 검은 헝클어진 머리와 마른 체형의 젊은 동아시아계 남성이지만, 참고의 긴팔 후드 대신 짙은 반팔 티셔츠를 입었다. 낡고 찢어진 청바지는 유지되지만 신선한 물린 상처는 뚜렷하지 않다. 경찰들은 남색 제복·모자·방호조끼를 착용한 성인들로, 얼굴이 보이는 인물들은 동아시아계로 보인다. 권총은 각 경찰의 손에 들려 있다. 오른쪽 실내 바닥에는 검고 흰 몸체와 금속성 머리 손상이 보이는 로봇이 누워 있어 지속 상태와 부합한다. 읽을 수 있는 글자나 오버레이는 없다.",
        "hard_violations": [],
        "physics": "현우는 두 발로 잔디를 딛고 서 있으며 팔을 내리고 있다. 전신 경찰들은 모두 지면에 발을 붙인 넓은 사격 자세로 체중을 지탱한다. 권총은 손으로 잡고 있으며 하단 전경 권총에도 손과 제복 소매가 이어진다. 실내 로봇은 바닥에 누워 바닥의 지지를 받는다. 지지 없는 부유나 불가능한 신체 배치는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "현우의 얼굴·후드 의상과 직립 자세는 잘 맞지만, 눈높이 구도와 크게 잘린 전경 경찰·권총 때문에 위에서 내려다보는 잔디의 연속면과 앞뒤 포위 배치가 제대로 드러나지 않는다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "비스듬한 하이앵글 와이드숏으로 잔디 위 앞뒤 포위 관계를 가장 충실하게 구현했지만, 하단의 큰 전경 권총과 현우의 반팔 의상·불명확한 물린 상처는 요구와 다르다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "좌우 및 뒤쪽 경찰의 시선과 권총은 대체로 중앙 현우의 상체를 향한다. 아래쪽 두 경찰도 팔을 안쪽으로 뻗어 현우를 조준하며, 관객을 향해 총구를 들이대지는 않는다. 다만 겹쳐진 전경 팔과 큰 권총 때문에 개별 사선과 원형 포위 관계가 일부 가려진다. 현우는 정면을 바라본다.",
        "built_space": "밝은 외벽에 아래층 대형 유리 개구부 두 곳, 오른쪽 현관 한 곳, 위층 큰 창 한 곳과 작은 창 한 곳이 보인다. 큰 위층 창과 작은 창은 열려 있지만 아래층 거실창의 열린 부분은 명확하지 않다. 참고의 주요 입면 배치는 대체로 유지되지만, 현관 차양은 전경 인물에 가려 확인하기 어렵다. 중앙 현우와 경찰 일곱 명이 잔디 위에 배치되어 있으나 카메라가 거의 눈높이라 가까운 쪽과 먼 쪽을 연결하는 지면이 충분히 펼쳐지지 않는다. 유리의 수목 반사는 가능한 배치다.",
        "entities": "현우는 검은 헝클어진 머리의 젊은 동아시아계 남성으로, 참고의 얼굴과 마른 체형, 낡은 짙은 후드, 찢어진 청바지, 운동화에 가깝다. 무릎의 찢어진 부분은 보이나 신선한 개 물림 상처는 식별되지 않는다. 경찰 일곱 명은 성인 동아시아계 남성으로 보이고 짙은 남색 제복과 권총을 갖췄다. 국적은 외모만으로 확정할 수 없다. 실내 바닥의 로봇은 보이지 않으며, 실내가 가려져 그 상태를 판단할 수 없다. 명확히 읽히는 문구나 화면 오버레이는 없다.",
        "hard_violations": [],
        "physics": "현우의 두 발은 잔디에 닿고 몸은 꼿꼿하게 서 있다. 전신이 보이는 경찰들은 벌린 발로 지면을 딛고, 권총은 각 경찰의 손에 잡혀 있다. 전경 경찰의 하체는 화면 밖이지만 상체가 공중에 떠 있다는 증거는 없다. 지지 없이 떠 있는 물체나 명백히 불가능한 관절 자세는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽·뒤쪽·오른쪽·앞쪽 경찰들이 중앙 현우를 향해 몸과 시선을 돌리고 권총을 안쪽으로 겨눈다. 왼쪽 앞 경찰의 총구는 오른쪽 위의 현우 쪽, 오른쪽 앞 경찰의 총구는 왼쪽 위의 현우 쪽을 향한다. 오른쪽 끝 경찰은 현우의 몸통 또는 허리 높이로 조준한다. 하단 중앙의 별도 손에 잡힌 권총도 장면 안쪽을 향하지만 단축되어 정확한 조준점은 확인하기 어렵다. 현우의 얼굴은 화면 오른쪽 경찰 쪽으로 돌아가 있다.",
        "built_space": "위에서 비스듬히 본 잔디가 하나의 연속면으로 펼쳐지고, 전신 경찰 일곱 명과 하단의 부분적으로 보이는 경찰 한 명이 현우의 앞뒤와 좌우를 둘러싼다. 외벽에는 아래층 대형 유리 개구부 두 곳, 오른쪽 현관과 차양 각 한 곳, 위층 큰 창과 작은 창 각 한 곳, 아래층 벽등 두 개가 보인다. 위층 큰 창과 아래층 거실창은 열린 상태로 읽힌다. 참고의 밝은 외벽·검은 창틀·담장·잔디라는 장소 구성이 유지된다. 현관 주변 화분은 두 개로 참고의 한 개와 다르고 창 세부도 완전히 같지는 않다. 부자연스러운 거울상은 보이지 않는다.",
        "entities": "현우는 검은 헝클어진 머리와 마른 체형의 젊은 동아시아계 남성이지만, 참고의 긴팔 후드 대신 짙은 반팔 티셔츠를 입었다. 낡고 찢어진 청바지는 유지되지만 신선한 물린 상처는 뚜렷하지 않다. 경찰들은 남색 제복·모자·방호조끼를 착용한 성인들로, 얼굴이 보이는 인물들은 동아시아계로 보인다. 권총은 각 경찰의 손에 들려 있다. 오른쪽 실내 바닥에는 검고 흰 몸체와 금속성 머리 손상이 보이는 로봇이 누워 있어 지속 상태와 부합한다. 읽을 수 있는 글자나 오버레이는 없다.",
        "hard_violations": [],
        "physics": "현우는 두 발로 잔디를 딛고 서 있으며 팔을 내리고 있다. 전신 경찰들은 모두 지면에 발을 붙인 넓은 사격 자세로 체중을 지탱한다. 권총은 손으로 잡고 있으며 하단 전경 권총에도 손과 제복 소매가 이어진다. 실내 로봇은 바닥에 누워 바닥의 지지를 받는다. 지지 없는 부유나 불가능한 신체 배치는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.238
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.238
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1238
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "부감 샷과 와이드 구도, 거실 내부의 로봇 등 우선순위가 높은 촬영 지시를 훌륭히 따랐으나, 이현우의 후드티 의상이 누락되었습니다."
   },
   {
    "label": "B",
    "score": 1238,
    "verdict_ko": "이현우의 의상은 정확하게 재현했으나, 지시된 부감 샷을 무시하고 전경에 거대한 무기와 인물을 배치하여 카메라 및 구도 지시를 심각하게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L07B02.png",
    "asset_id": "2c87ee92-f9b1-40d2-a384-6e45ad3dd181",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd76-33a1-7911-8be2-db98793a5f75",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S5sh24__bgfirst_bg.png",
   "bg_asset_id": "c0bf2b7e-d98e-4a61-87e7-b483e5281967",
   "bg_record_key": "S5sh24::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S5sh24::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:17:07.314884+00:00",
  "fingerprint": "da7dd297737f9a9819ac7e195fbb3da42e9a9661e3e89758ce4a977f5c2e9446",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S5sh24_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S5sh24_sel.png",
  "source_sha256": "5880d6e46002b13d47f3a554f241e455df852e4d2064a2be7898b11d8eef9324",
  "file": "S5sh24_cine.png",
  "staged_sha256": "cdaa3b94b7257f156c12e556864fef714df08771f8b6507f4705348d7e13cea1",
  "latency_ms": 12993
 },
 "S6sh21::signage": {
  "fp": "bdca3540763f1cf8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S6sh21": {
  "input_fingerprint": "a40ce1af2dd4f2de",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낡은 금속 고릴라 로봇 찰리가 쓰레기 더미 속에서 상체를 무겁게 일으켜 세운 웅장한 전신.\n\nLOCATION (lock): On a towering rubbish heap in an open-air refugee-settlement dump, at the scrap pile from which the robot emerges. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Rubbish mound around Charlie (Broken through by the emerging robot); used as Uneven base around the lower body that makes the torso lift physically legible; Surrounding rubbish heaps (Extending around the children and the emerging robot); used as Layered background and shared ground context, leaving clear space around the robot's full silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight gives the old metal body solemn weight while the explicitly blue light in Charlie's eyes supplies the shot's selective vivid accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rubbish mound has been disturbed by Charlie's emergence; the old gorilla-shaped robot retains his worn UBIC chest logo and the previously established net entanglement, with no removal described. The repaired jukebox has its horn speaker connected and an LP mounted, with tools and other records scattered around the blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낡은 금속 고릴라 로봇 찰리가 쓰레기 더미 속에서 상체를 무겁게 일으켜 세운 웅장한 전신.\n\nLOCATION (lock): On a towering rubbish heap in an open-air refugee-settlement dump, at the scrap pile from which the robot emerges. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Rubbish mound around Charlie (Broken through by the emerging robot); used as Uneven base around the lower body that makes the torso lift physically legible; Surrounding rubbish heaps (Extending around the children and the emerging robot); used as Layered background and shared ground context, leaving clear space around the robot's full silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight gives the old metal body solemn weight while the explicitly blue light in Charlie's eyes supplies the shot's selective vivid accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rubbish mound has been disturbed by Charlie's emergence; the old gorilla-shaped robot retains his worn UBIC chest logo and the previously established net entanglement, with no removal described. The repaired jukebox has its horn speaker connected and an LP mounted, with tools and other records scattered around the blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낡은 금속 고릴라 로봇 찰리가 쓰레기 더미 속에서 상체를 무겁게 일으켜 세운 웅장한 전신.\n\nLOCATION (lock): On a towering rubbish heap in an open-air refugee-settlement dump, at the scrap pile from which the robot emerges. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Rubbish mound around Charlie (Broken through by the emerging robot); used as Uneven base around the lower body that makes the torso lift physically legible; Surrounding rubbish heaps (Extending around the children and the emerging robot); used as Layered background and shared ground context, leaving clear space around the robot's full silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight gives the old metal body solemn weight while the explicitly blue light in Charlie's eyes supplies the shot's selective vivid accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rubbish mound has been disturbed by Charlie's emergence; the old gorilla-shaped robot retains his worn UBIC chest logo and the previously established net entanglement, with no removal described. The repaired jukebox has its horn speaker connected and an LP mounted, with tools and other records scattered around the blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "로봇은 정면을 향하고 군중은 로봇을 주시함.",
    "built_space": "참조 사진과 유사한 매립지. 담요 위에 나팔 스피커가 달린 주크박스 존재.",
    "entities": "흰 얼굴판과 각진 외장이 없는 고릴라 로봇. 텍스트에 없는 다수의 군중 등장.",
    "hard_violations": [
     "[gemini-pro] 숏 텍스트에 없는 인물 임의 추가",
     "[gpt-high] 찰리만 허용된 장면에 다수의 성인과 아이들을 임의로 추가했다.",
     "[gpt-high] 찰리 가슴의 영문 로고가 판독 가능하게 보여 무문자·무로고 지시를 위반한다."
    ],
    "physics": "쓰레기 잔해를 밀어내며 지면을 짚고 상체를 지탱함."
   },
   {
    "label": "B",
    "direction": "로봇이 정면을 응시함.",
    "built_space": "참조 사진과 매우 일치하는 매립지. 바닥 담요 주변과 상자 안에 스피커 존재.",
    "entities": "흰 얼굴판, 각진 외장, 점 형태의 눈 등 찰리의 특징이 완벽히 일치함.",
    "hard_violations": [
     "[gpt-high] 가슴의 영문 로고가 읽힐 정도로 노출되어, 어떤 글자나 로고도 판독 가능하게 보여서는 안 된다는 최종 지시를 위반한다."
    ],
    "physics": "바닥의 구멍에서 나와 양손으로 지면을 안정적으로 짚고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 로봇의 구체적인 얼굴 디자인을 완벽히 재현했고, 불필요한 인물 추가 없이 지침을 잘 준수했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "숏 텍스트에 없는 군중을 임의로 추가한 심각한 위반이 있으며, 로봇의 지정된 외형 묘사도 무시되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇은 정면을 향하고 군중은 로봇을 주시함.",
        "built_space": "참조 사진과 유사한 매립지. 담요 위에 나팔 스피커가 달린 주크박스 존재.",
        "entities": "흰 얼굴판과 각진 외장이 없는 고릴라 로봇. 텍스트에 없는 다수의 군중 등장.",
        "hard_violations": [
         "숏 텍스트에 없는 인물 임의 추가"
        ],
        "physics": "쓰레기 잔해를 밀어내며 지면을 짚고 상체를 지탱함."
       },
       {
        "label": "B",
        "direction": "로봇이 정면을 응시함.",
        "built_space": "참조 사진과 매우 일치하는 매립지. 바닥 담요 주변과 상자 안에 스피커 존재.",
        "entities": "흰 얼굴판, 각진 외장, 점 형태의 눈 등 찰리의 특징이 완벽히 일치함.",
        "hard_violations": [],
        "physics": "바닥의 구멍에서 나와 양손으로 지면을 안정적으로 짚고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 로봇의 구체적인 얼굴 디자인을 완벽히 재현했고, 불필요한 인물 추가 없이 지침을 잘 준수했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "숏 텍스트에 없는 군중을 임의로 추가한 심각한 위반이 있으며, 로봇의 지정된 외형 묘사도 무시되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "로봇은 정면을 향하고 군중은 로봇을 주시함.",
        "built_space": "참조 사진과 유사한 매립지. 담요 위에 나팔 스피커가 달린 주크박스 존재.",
        "entities": "흰 얼굴판과 각진 외장이 없는 고릴라 로봇. 텍스트에 없는 다수의 군중 등장.",
        "hard_violations": [
         "숏 텍스트에 없는 인물 임의 추가"
        ],
        "physics": "쓰레기 잔해를 밀어내며 지면을 짚고 상체를 지탱함."
       },
       {
        "label": "B",
        "direction": "로봇이 정면을 응시함.",
        "built_space": "참조 사진과 매우 일치하는 매립지. 바닥 담요 주변과 상자 안에 스피커 존재.",
        "entities": "흰 얼굴판, 각진 외장, 점 형태의 눈 등 찰리의 특징이 완벽히 일치함.",
        "hard_violations": [],
        "physics": "바닥의 구멍에서 나와 양손으로 지면을 안정적으로 짚고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리만 등장하고 각진 머리·흰 얼굴판·그물과 상체를 밀어 올리는 동작은 잘 맞지만, 읽히는 가슴 문자와 하체가 묻힌 구도 때문에 무문자·웅장한 전신 요구를 충족하지 못한다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "양팔로 잔해를 밀어 올리는 동작과 주크박스 소품은 드러나지만, 허용되지 않은 다수의 구경꾼과 읽히는 가슴 문자가 결정적인 위반이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴과 파란 눈은 카메라 쪽 정면을 향한다. 양팔은 아래·바깥쪽으로 뻗어 몸 양옆의 잔해를 짚는다. 특정 응시 대상은 지정되지 않았으므로 시선 충돌은 없다. 왼쪽 혼 스피커의 입구는 카메라 쪽을 향한다.",
        "built_space": "뒤와 왼쪽에 큰 고철 더미가 있고 오른쪽에는 울타리와 뒤로 이어지는 통로가 있다. 앞에는 담요 한 장과 상자 두 개, 오른쪽 아래에는 맨홀 뚜껑 한 개가 보여 참조 장소의 배치를 상당히 유지한다. 다만 찰리는 높은 쓰레기 산 위라기보다 평지 앞쪽의 잔해 구덩이에서 솟는 모습이며, 하체가 가려져 전신 윤곽이 확보되지 않는다.",
        "entities": "등장 인물은 비인간 로봇 찰리 한 기뿐이다. 낡고 녹슨 금속 몸체, 각진 머리 외장, 흰 얼굴판, 작은 파란 눈과 선 형태의 입이 보인다. 그물이 어깨부터 가슴과 허리까지 걸려 있다. 가슴에는 큰 영문 문자가 읽힌다. 담요에는 공구와 여러 레코드, 전자기기가 있고 왼쪽에 혼 스피커가 있지만, 수리된 주크박스에 혼이 연결되고 LP가 장착된 상태는 확인되지 않는다. 낮 장면은 맞는다.",
        "hard_violations": [
         "가슴의 영문 로고가 읽힐 정도로 노출되어, 어떤 글자나 로고도 판독 가능하게 보여서는 안 된다는 최종 지시를 위반한다."
        ],
        "physics": "찰리는 양손을 좌우의 고철과 지면 가장자리에 짚어 상체를 지탱한다. 팔로 몸을 밀어 올리는 동작은 물리적으로 가능하며, 하체는 잔해 아래에 가려져 있다. 그물은 어깨와 몸통에 걸려 늘어지고, 레코드와 공구는 담요나 지면 위에 놓여 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리는 얼굴을 약간 숙인 채 카메라 쪽을 바라보고, 두 팔은 앞쪽 아래의 잔해를 누른다. 주변 성인과 아이들은 대체로 중앙의 찰리를 향해 서서 바라본다. 혼 스피커의 입구는 앞쪽과 화면 오른쪽을 향한다. 군중의 시선 자체는 자연스럽지만 군중의 등장은 허용되지 않는다.",
        "built_space": "참조와 비슷한 후면 고철 산, 오른쪽 울타리와 통로, 앞쪽 담요 한 장과 상자 두 개가 보인다. 담요에는 혼을 단 주크박스 한 대와 별도의 라디오가 놓여 있다. 다수의 사람이 찰리 둘레와 전경을 둘러싸 참조의 빈 작업 공간을 군중 장면으로 바꾼다. 찰리는 높은 더미 위보다는 마당 높이의 잔해 속에 있고 다리와 발은 보이지 않는다.",
        "entities": "낡은 고릴라형 금속 로봇 한 기에 파란 눈, 선 형태의 입, 몸통을 가로지르는 그물과 읽히는 가슴 로고가 있다. 머리 외장은 요구된 각진 형태보다 둥글고 얼굴판도 흰색보다 어두운 회색이다. 화면에는 여러 피부색의 성인 남녀와 아이들이 다수 추가되었으며 이들 모두 허용된 등장 대상이 아니다. 주크박스 위에는 혼과 원반이 보이고 담요에 공구와 다른 레코드가 흩어져 있어 소품 상태는 A보다 분명하다. 낮 장면은 맞는다.",
        "hard_violations": [
         "찰리만 허용된 장면에 다수의 성인과 아이들을 임의로 추가했다.",
         "찰리 가슴의 영문 로고가 판독 가능하게 보여 무문자·무로고 지시를 위반한다."
        ],
        "physics": "찰리의 양손은 앞쪽 잔해에 닿아 상체를 받치며, 굽힌 팔과 전방으로 기울인 몸통은 무거운 몸을 일으키는 동작으로 읽힌다. 몸 아래에서 튀는 작은 파편은 로봇의 출현 충격으로 설명된다. 군중은 지면에 서 있고 전경 인물들의 발은 화면 밖에 있다. 주크박스는 담요 위에 놓여 있으며 혼은 그 위의 관에 연결되어 있다. 지지나 발사 원인 없이 떠 있는 대상은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "찰리만 등장하고 각진 머리·흰 얼굴판·그물과 상체를 밀어 올리는 동작은 잘 맞지만, 읽히는 가슴 문자와 하체가 묻힌 구도 때문에 무문자·웅장한 전신 요구를 충족하지 못한다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "양팔로 잔해를 밀어 올리는 동작과 주크박스 소품은 드러나지만, 허용되지 않은 다수의 구경꾼과 읽히는 가슴 문자가 결정적인 위반이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴과 파란 눈은 카메라 쪽 정면을 향한다. 양팔은 아래·바깥쪽으로 뻗어 몸 양옆의 잔해를 짚는다. 특정 응시 대상은 지정되지 않았으므로 시선 충돌은 없다. 왼쪽 혼 스피커의 입구는 카메라 쪽을 향한다.",
        "built_space": "뒤와 왼쪽에 큰 고철 더미가 있고 오른쪽에는 울타리와 뒤로 이어지는 통로가 있다. 앞에는 담요 한 장과 상자 두 개, 오른쪽 아래에는 맨홀 뚜껑 한 개가 보여 참조 장소의 배치를 상당히 유지한다. 다만 찰리는 높은 쓰레기 산 위라기보다 평지 앞쪽의 잔해 구덩이에서 솟는 모습이며, 하체가 가려져 전신 윤곽이 확보되지 않는다.",
        "entities": "등장 인물은 비인간 로봇 찰리 한 기뿐이다. 낡고 녹슨 금속 몸체, 각진 머리 외장, 흰 얼굴판, 작은 파란 눈과 선 형태의 입이 보인다. 그물이 어깨부터 가슴과 허리까지 걸려 있다. 가슴에는 큰 영문 문자가 읽힌다. 담요에는 공구와 여러 레코드, 전자기기가 있고 왼쪽에 혼 스피커가 있지만, 수리된 주크박스에 혼이 연결되고 LP가 장착된 상태는 확인되지 않는다. 낮 장면은 맞는다.",
        "hard_violations": [
         "가슴의 영문 로고가 읽힐 정도로 노출되어, 어떤 글자나 로고도 판독 가능하게 보여서는 안 된다는 최종 지시를 위반한다."
        ],
        "physics": "찰리는 양손을 좌우의 고철과 지면 가장자리에 짚어 상체를 지탱한다. 팔로 몸을 밀어 올리는 동작은 물리적으로 가능하며, 하체는 잔해 아래에 가려져 있다. 그물은 어깨와 몸통에 걸려 늘어지고, 레코드와 공구는 담요나 지면 위에 놓여 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리는 얼굴을 약간 숙인 채 카메라 쪽을 바라보고, 두 팔은 앞쪽 아래의 잔해를 누른다. 주변 성인과 아이들은 대체로 중앙의 찰리를 향해 서서 바라본다. 혼 스피커의 입구는 앞쪽과 화면 오른쪽을 향한다. 군중의 시선 자체는 자연스럽지만 군중의 등장은 허용되지 않는다.",
        "built_space": "참조와 비슷한 후면 고철 산, 오른쪽 울타리와 통로, 앞쪽 담요 한 장과 상자 두 개가 보인다. 담요에는 혼을 단 주크박스 한 대와 별도의 라디오가 놓여 있다. 다수의 사람이 찰리 둘레와 전경을 둘러싸 참조의 빈 작업 공간을 군중 장면으로 바꾼다. 찰리는 높은 더미 위보다는 마당 높이의 잔해 속에 있고 다리와 발은 보이지 않는다.",
        "entities": "낡은 고릴라형 금속 로봇 한 기에 파란 눈, 선 형태의 입, 몸통을 가로지르는 그물과 읽히는 가슴 로고가 있다. 머리 외장은 요구된 각진 형태보다 둥글고 얼굴판도 흰색보다 어두운 회색이다. 화면에는 여러 피부색의 성인 남녀와 아이들이 다수 추가되었으며 이들 모두 허용된 등장 대상이 아니다. 주크박스 위에는 혼과 원반이 보이고 담요에 공구와 다른 레코드가 흩어져 있어 소품 상태는 A보다 분명하다. 낮 장면은 맞는다.",
        "hard_violations": [
         "찰리만 허용된 장면에 다수의 성인과 아이들을 임의로 추가했다.",
         "찰리 가슴의 영문 로고가 판독 가능하게 보여 무문자·무로고 지시를 위반한다."
        ],
        "physics": "찰리의 양손은 앞쪽 잔해에 닿아 상체를 받치며, 굽힌 팔과 전방으로 기울인 몸통은 무거운 몸을 일으키는 동작으로 읽힌다. 몸 아래에서 튀는 작은 파편은 로봇의 출현 충격으로 설명된다. 군중은 지면에 서 있고 전경 인물들의 발은 화면 밖에 있다. 주크박스는 담요 위에 놓여 있으며 혼은 그 위의 관에 연결되어 있다. 지지나 발사 원인 없이 떠 있는 대상은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.679,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.429,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 숏 텍스트에 없는 인물 임의 추가",
     "[gpt-high] 찰리만 허용된 장면에 다수의 성인과 아이들을 임의로 추가했다.",
     "[gpt-high] 찰리 가슴의 영문 로고가 판독 가능하게 보여 무문자·무로고 지시를 위반한다."
    ],
    "B": [
     "[gpt-high] 가슴의 영문 로고가 읽힐 정도로 노출되어, 어떤 글자나 로고도 판독 가능하게 보여서는 안 된다는 최종 지시를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 429
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지정된 로봇의 구체적인 얼굴 디자인을 완벽히 재현했고, 불필요한 인물 추가 없이 지침을 잘 준수했습니다.  ★위반: [gpt-high] 가슴의 영문 로고가 읽힐 정도로 노출되어, 어떤 글자나 로고도 판독 가능하게 보여서는 안 된다는 최종 지시를 위반한다."
   },
   {
    "label": "A",
    "score": 429,
    "verdict_ko": "숏 텍스트에 없는 군중을 임의로 추가한 심각한 위반이 있으며, 로봇의 지정된 외형 묘사도 무시되었습니다.  ★위반: [gemini-pro] 숏 텍스트에 없는 인물 임의 추가 / [gpt-high] 찰리만 허용된 장면에 다수의 성인과 아이들을 임의로 추가했다. / [gpt-high] 찰리 가슴의 영문 로고가 판독 가능하게 보여 무문자·무로고 지시를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L11B01.png",
    "asset_id": "e9e2a760-e506-4460-9818-2803bd774a80",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd77-09f4-7ce3-9b59-a953e66ed69e",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S6sh21::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:17:20.635233+00:00",
  "fingerprint": "43a888be1941f19015097064b152d3c3a5785ad2e3098dcbe1b0c3aac592c63f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S6sh21_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S6sh21_sel.png",
  "source_sha256": "69dea80435bd10333dc8c719fcb9ea3af106f5adfb251f0969b4f53749324dca",
  "file": "S6sh21_cine.png",
  "staged_sha256": "25240813a6abcce7dcafee54e26fa56fc1211abcc301549de5eab094adbeb2b9",
  "latency_ms": 10927
 },
 "S6sh30::signage": {
  "fp": "9034c006966d49ab",
  "inscriptions": [
   {
    "text_native": "NEGATIVE",
    "source": "scene_text_quoted",
    "reason_ko": "찰리의 시야 인터페이스 화면 위에 붉은색으로 직접 표기되는 인식 판정 결과 텍스트이다.",
    "source_quote": "NEGATIVE"
   }
  ],
  "cues": [],
  "dropped": []
 },
 "S6sh30": {
  "input_fingerprint": "536883f133900222",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 시야 인터페이스 속 지소영의 얼굴 사진 위로 'NEGATIVE'라는 붉은 글자가 선명하게 겹쳐진 주관 시점 찰나.\n\nLOCATION (lock): At the open-air refugee-settlement rubbish heap, with the nearby girl's face viewed through the robot's recognition display. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Stored identification photograph (Visible in the visual interface before disappearing) — The image-bearing face is fully legible to the viewer and displays 지소영's portrait; used as Central recognition subject, separated from the indistinct live view; NEGATIVE interface label (Superimposed across the stored photograph) — The readable interface layer faces the viewer and displays exactly 'NEGATIVE'; used as Makes the failed identification immediately intelligible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the daytime live-view exposure beneath the interface, using the explicitly red 'NEGATIVE' lettering as the sole emphatic color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's visual interface overlays “NEGATIVE” on Soyoung's image before leaving only Amber's face; his eyes are blue-lit, his chest logo remains worn, and no removal of the earlier net is described. The repaired jukebox, connected speaker, mounted LP, blanket and scattered tools remain at the rubbish heap.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"NEGATIVE\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 시야 인터페이스 속 지소영의 얼굴 사진 위로 'NEGATIVE'라는 붉은 글자가 선명하게 겹쳐진 주관 시점 찰나.\n\nLOCATION (lock): At the open-air refugee-settlement rubbish heap, with the nearby girl's face viewed through the robot's recognition display. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Stored identification photograph (Visible in the visual interface before disappearing) — The image-bearing face is fully legible to the viewer and displays 지소영's portrait; used as Central recognition subject, separated from the indistinct live view; NEGATIVE interface label (Superimposed across the stored photograph) — The readable interface layer faces the viewer and displays exactly 'NEGATIVE'; used as Makes the failed identification immediately intelligible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the daytime live-view exposure beneath the interface, using the explicitly red 'NEGATIVE' lettering as the sole emphatic color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's visual interface overlays “NEGATIVE” on Soyoung's image before leaving only Amber's face; his eyes are blue-lit, his chest logo remains worn, and no removal of the earlier net is described. The repaired jukebox, connected speaker, mounted LP, blanket and scattered tools remain at the rubbish heap.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"NEGATIVE\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 시야 인터페이스 속 지소영의 얼굴 사진 위로 'NEGATIVE'라는 붉은 글자가 선명하게 겹쳐진 주관 시점 찰나.\n\nLOCATION (lock): At the open-air refugee-settlement rubbish heap, with the nearby girl's face viewed through the robot's recognition display. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Stored identification photograph (Visible in the visual interface before disappearing) — The image-bearing face is fully legible to the viewer and displays 지소영's portrait; used as Central recognition subject, separated from the indistinct live view; NEGATIVE interface label (Superimposed across the stored photograph) — The readable interface layer faces the viewer and displays exactly 'NEGATIVE'; used as Makes the failed identification immediately intelligible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the daytime live-view exposure beneath the interface, using the explicitly red 'NEGATIVE' lettering as the sole emphatic color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's visual interface overlays “NEGATIVE” on Soyoung's image before leaving only Amber's face; his eyes are blue-lit, his chest logo remains worn, and no removal of the earlier net is described. The repaired jukebox, connected speaker, mounted LP, blanket and scattered tools remain at the rubbish heap.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"NEGATIVE\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "저장 사진 속 성인 여성은 관객을 정면으로 바라보고, 붉은 ‘NEGATIVE’는 그 얼굴을 가로지른다. 실경의 소녀는 카메라에 등을 보이며 뒤쪽 찰리를 향한다. 찰리도 카메라 쪽을 향해 있어, 찰리가 가까운 소녀의 얼굴을 보는 시선 관계가 아니다.",
    "built_space": "참조의 고철 더미, 오른쪽 통로와 판재 울타리, 오른쪽 아래 맨홀 하나가 유지된다. 바닥에는 담요 하나와 여러 음반·도구가 있고, 왼쪽에는 나팔형 장치 하나, 중앙 왼쪽에는 라디오형 장치 하나가 보인다. 그러나 참조의 외부 카메라 구도를 거의 유지하여 찰리가 고철 더미 중앙에 그대로 보이고, 화면에는 주변 공간과 소녀의 하체까지 넓게 들어온다. 요구된 얼굴 사진 중심의 주관 클로즈업과 다르다.",
    "entities": "중앙에는 검은 머리의 동아시아계 성인 여성 증명사진이 있으며 지소영의 저장 사진 역할은 명확하다. 지소영의 별도 외모 참조는 없어 정확한 신원 일치 여부는 확인할 수 없다. 실경 소녀는 짙은 머리와 뒷모습만 보여 금발의 10세 한국계·백인 혼혈 앰버라는 지정과 맞지 않으며 얼굴도 보이지 않는다. 배경 찰리의 각진 고릴라형 머리, 흰 얼굴판, 푸른 눈과 그물은 남아 있다. ‘NEGATIVE’ 외에 ‘지소영’과 다른 인터페이스 문자가 추가되어 있다.",
    "hard_violations": [
     "찰리의 주관 시점이어야 하는데 찰리의 머리와 몸을 정면에서 보는 외부 시점으로 배치했다.",
     "배경의 찰리와 별개로 전경에 큰 로봇 팔·손이 들어와, 단일 찰리의 시점으로 설명할 수 없는 중복 로봇 신체 구성이 된다.",
     "허용된 ‘NEGATIVE’ 외에 읽을 수 있는 ‘지소영’ 및 추가 인터페이스 문자를 넣었다."
    ],
    "physics": "소녀는 곧게 서 있으며 발은 프레임 밖이므로 공중 부양으로 볼 근거는 없다. 찰리의 몸은 고철 더미 뒤로 이어지고 손은 앞쪽 잔해에 닿아 있다. 전경 로봇 손은 화면 밖으로 이어지는 팔에 연결되어 있으나 배경 찰리와의 신체 관계가 성립하지 않는다. 저장 사진과 문자는 명시된 시야 인터페이스이므로 물리적 받침이 필요한 부유 물체로 판단하지 않는다."
   },
   {
    "label": "A",
    "direction": "저장 사진 속 여성은 관객을 정면으로 바라보며 붉은 ‘NEGATIVE’가 얼굴 중앙을 가로지른다. 배경 찰리 역시 카메라 쪽을 향한다. 찰리가 바라봐야 할 앰버의 얼굴은 실경에 없고, 오히려 찰리 자신이 인식 화면 뒤의 대상처럼 보인다.",
    "built_space": "고철 더미와 오른쪽 통로·울타리, 맨홀 하나, 바닥 담요 하나가 참조와 대응한다. 왼쪽 나팔형 장치 하나와 담요 위 음반·도구도 흐릿하게 남아 있으나 세부 수량이나 연결 상태는 판독하기 어렵다. 저장 사진은 A보다 크게 제시되고 실경은 흐려져 두 층이 더 잘 분리된다. 다만 배경의 공간 배치와 찰리의 위치는 여전히 참조를 외부에서 바라본 구도다.",
    "entities": "중앙의 저장 사진은 검은 단발머리의 동아시아계 성인 여성으로, 지소영의 인물사진이라는 역할을 충족한다. 정확한 외모를 대조할 별도 참조는 없다. 배경에는 흰 얼굴판과 푸른 눈의 찰리가 있지만 앰버는 보이지 않는다. 붉은 ‘NEGATIVE’는 정확하고 선명하나, 왼쪽의 ‘지소영’, 오른쪽의 ‘2069’ 및 추가 숫자·문자가 함께 읽힌다. 주관 시점에서 제외되어야 할 찰리를 보여주는 대신 필요한 실시간 소녀 얼굴이 빠졌다.",
    "hard_violations": [
     "찰리의 눈으로 보는 장면에 찰리 자신의 머리와 몸을 정면 배치하여 주관 시점의 공간 관계를 위반했다.",
     "유일하게 허용된 ‘NEGATIVE’ 외에 ‘지소영’, ‘2069’ 등 읽을 수 있는 추가 문자를 넣었다."
    ],
    "physics": "배경 찰리의 몸과 팔은 잔해 뒤로 자연스럽게 이어지며 공중에 떠 있는 신체는 보이지 않는다. 담요와 음반·도구는 바닥이나 잔해에 놓여 있다. 중앙 사진과 문자는 요청된 가상 인식 화면이므로 손으로 들거나 바닥에 지지될 필요는 없다. 물리적 지지보다 카메라와 찰리의 위치 관계가 문제다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "붉은 ‘NEGATIVE’는 사진 위에 정확히 겹치지만, 찰리를 외부에서 보는 넓은 구도와 등을 돌린 소녀, 추가 문자로 주관 시점과 얼굴 인식 장면을 위반한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "저장 인물사진을 더 크게 강조하고 실경과 분리해 A보다 가깝지만, 찰리 자신이 보이는 잘못된 시점과 추가 문자, 앰버 얼굴의 부재로 재촬영이 필요하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "저장 사진 속 성인 여성은 관객을 정면으로 바라보고, 붉은 ‘NEGATIVE’는 그 얼굴을 가로지른다. 실경의 소녀는 카메라에 등을 보이며 뒤쪽 찰리를 향한다. 찰리도 카메라 쪽을 향해 있어, 찰리가 가까운 소녀의 얼굴을 보는 시선 관계가 아니다.",
        "built_space": "참조의 고철 더미, 오른쪽 통로와 판재 울타리, 오른쪽 아래 맨홀 하나가 유지된다. 바닥에는 담요 하나와 여러 음반·도구가 있고, 왼쪽에는 나팔형 장치 하나, 중앙 왼쪽에는 라디오형 장치 하나가 보인다. 그러나 참조의 외부 카메라 구도를 거의 유지하여 찰리가 고철 더미 중앙에 그대로 보이고, 화면에는 주변 공간과 소녀의 하체까지 넓게 들어온다. 요구된 얼굴 사진 중심의 주관 클로즈업과 다르다.",
        "entities": "중앙에는 검은 머리의 동아시아계 성인 여성 증명사진이 있으며 지소영의 저장 사진 역할은 명확하다. 지소영의 별도 외모 참조는 없어 정확한 신원 일치 여부는 확인할 수 없다. 실경 소녀는 짙은 머리와 뒷모습만 보여 금발의 10세 한국계·백인 혼혈 앰버라는 지정과 맞지 않으며 얼굴도 보이지 않는다. 배경 찰리의 각진 고릴라형 머리, 흰 얼굴판, 푸른 눈과 그물은 남아 있다. ‘NEGATIVE’ 외에 ‘지소영’과 다른 인터페이스 문자가 추가되어 있다.",
        "hard_violations": [
         "찰리의 주관 시점이어야 하는데 찰리의 머리와 몸을 정면에서 보는 외부 시점으로 배치했다.",
         "배경의 찰리와 별개로 전경에 큰 로봇 팔·손이 들어와, 단일 찰리의 시점으로 설명할 수 없는 중복 로봇 신체 구성이 된다.",
         "허용된 ‘NEGATIVE’ 외에 읽을 수 있는 ‘지소영’ 및 추가 인터페이스 문자를 넣었다."
        ],
        "physics": "소녀는 곧게 서 있으며 발은 프레임 밖이므로 공중 부양으로 볼 근거는 없다. 찰리의 몸은 고철 더미 뒤로 이어지고 손은 앞쪽 잔해에 닿아 있다. 전경 로봇 손은 화면 밖으로 이어지는 팔에 연결되어 있으나 배경 찰리와의 신체 관계가 성립하지 않는다. 저장 사진과 문자는 명시된 시야 인터페이스이므로 물리적 받침이 필요한 부유 물체로 판단하지 않는다."
       },
       {
        "label": "B",
        "direction": "저장 사진 속 여성은 관객을 정면으로 바라보며 붉은 ‘NEGATIVE’가 얼굴 중앙을 가로지른다. 배경 찰리 역시 카메라 쪽을 향한다. 찰리가 바라봐야 할 앰버의 얼굴은 실경에 없고, 오히려 찰리 자신이 인식 화면 뒤의 대상처럼 보인다.",
        "built_space": "고철 더미와 오른쪽 통로·울타리, 맨홀 하나, 바닥 담요 하나가 참조와 대응한다. 왼쪽 나팔형 장치 하나와 담요 위 음반·도구도 흐릿하게 남아 있으나 세부 수량이나 연결 상태는 판독하기 어렵다. 저장 사진은 A보다 크게 제시되고 실경은 흐려져 두 층이 더 잘 분리된다. 다만 배경의 공간 배치와 찰리의 위치는 여전히 참조를 외부에서 바라본 구도다.",
        "entities": "중앙의 저장 사진은 검은 단발머리의 동아시아계 성인 여성으로, 지소영의 인물사진이라는 역할을 충족한다. 정확한 외모를 대조할 별도 참조는 없다. 배경에는 흰 얼굴판과 푸른 눈의 찰리가 있지만 앰버는 보이지 않는다. 붉은 ‘NEGATIVE’는 정확하고 선명하나, 왼쪽의 ‘지소영’, 오른쪽의 ‘2069’ 및 추가 숫자·문자가 함께 읽힌다. 주관 시점에서 제외되어야 할 찰리를 보여주는 대신 필요한 실시간 소녀 얼굴이 빠졌다.",
        "hard_violations": [
         "찰리의 눈으로 보는 장면에 찰리 자신의 머리와 몸을 정면 배치하여 주관 시점의 공간 관계를 위반했다.",
         "유일하게 허용된 ‘NEGATIVE’ 외에 ‘지소영’, ‘2069’ 등 읽을 수 있는 추가 문자를 넣었다."
        ],
        "physics": "배경 찰리의 몸과 팔은 잔해 뒤로 자연스럽게 이어지며 공중에 떠 있는 신체는 보이지 않는다. 담요와 음반·도구는 바닥이나 잔해에 놓여 있다. 중앙 사진과 문자는 요청된 가상 인식 화면이므로 손으로 들거나 바닥에 지지될 필요는 없다. 물리적 지지보다 카메라와 찰리의 위치 관계가 문제다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "붉은 ‘NEGATIVE’는 사진 위에 정확히 겹치지만, 찰리를 외부에서 보는 넓은 구도와 등을 돌린 소녀, 추가 문자로 주관 시점과 얼굴 인식 장면을 위반한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "저장 인물사진을 더 크게 강조하고 실경과 분리해 A보다 가깝지만, 찰리 자신이 보이는 잘못된 시점과 추가 문자, 앰버 얼굴의 부재로 재촬영이 필요하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "저장 사진 속 성인 여성은 관객을 정면으로 바라보고, 붉은 ‘NEGATIVE’는 그 얼굴을 가로지른다. 실경의 소녀는 카메라에 등을 보이며 뒤쪽 찰리를 향한다. 찰리도 카메라 쪽을 향해 있어, 찰리가 가까운 소녀의 얼굴을 보는 시선 관계가 아니다.",
        "built_space": "참조의 고철 더미, 오른쪽 통로와 판재 울타리, 오른쪽 아래 맨홀 하나가 유지된다. 바닥에는 담요 하나와 여러 음반·도구가 있고, 왼쪽에는 나팔형 장치 하나, 중앙 왼쪽에는 라디오형 장치 하나가 보인다. 그러나 참조의 외부 카메라 구도를 거의 유지하여 찰리가 고철 더미 중앙에 그대로 보이고, 화면에는 주변 공간과 소녀의 하체까지 넓게 들어온다. 요구된 얼굴 사진 중심의 주관 클로즈업과 다르다.",
        "entities": "중앙에는 검은 머리의 동아시아계 성인 여성 증명사진이 있으며 지소영의 저장 사진 역할은 명확하다. 지소영의 별도 외모 참조는 없어 정확한 신원 일치 여부는 확인할 수 없다. 실경 소녀는 짙은 머리와 뒷모습만 보여 금발의 10세 한국계·백인 혼혈 앰버라는 지정과 맞지 않으며 얼굴도 보이지 않는다. 배경 찰리의 각진 고릴라형 머리, 흰 얼굴판, 푸른 눈과 그물은 남아 있다. ‘NEGATIVE’ 외에 ‘지소영’과 다른 인터페이스 문자가 추가되어 있다.",
        "hard_violations": [
         "찰리의 주관 시점이어야 하는데 찰리의 머리와 몸을 정면에서 보는 외부 시점으로 배치했다.",
         "배경의 찰리와 별개로 전경에 큰 로봇 팔·손이 들어와, 단일 찰리의 시점으로 설명할 수 없는 중복 로봇 신체 구성이 된다.",
         "허용된 ‘NEGATIVE’ 외에 읽을 수 있는 ‘지소영’ 및 추가 인터페이스 문자를 넣었다."
        ],
        "physics": "소녀는 곧게 서 있으며 발은 프레임 밖이므로 공중 부양으로 볼 근거는 없다. 찰리의 몸은 고철 더미 뒤로 이어지고 손은 앞쪽 잔해에 닿아 있다. 전경 로봇 손은 화면 밖으로 이어지는 팔에 연결되어 있으나 배경 찰리와의 신체 관계가 성립하지 않는다. 저장 사진과 문자는 명시된 시야 인터페이스이므로 물리적 받침이 필요한 부유 물체로 판단하지 않는다."
       },
       {
        "label": "A",
        "direction": "저장 사진 속 여성은 관객을 정면으로 바라보며 붉은 ‘NEGATIVE’가 얼굴 중앙을 가로지른다. 배경 찰리 역시 카메라 쪽을 향한다. 찰리가 바라봐야 할 앰버의 얼굴은 실경에 없고, 오히려 찰리 자신이 인식 화면 뒤의 대상처럼 보인다.",
        "built_space": "고철 더미와 오른쪽 통로·울타리, 맨홀 하나, 바닥 담요 하나가 참조와 대응한다. 왼쪽 나팔형 장치 하나와 담요 위 음반·도구도 흐릿하게 남아 있으나 세부 수량이나 연결 상태는 판독하기 어렵다. 저장 사진은 A보다 크게 제시되고 실경은 흐려져 두 층이 더 잘 분리된다. 다만 배경의 공간 배치와 찰리의 위치는 여전히 참조를 외부에서 바라본 구도다.",
        "entities": "중앙의 저장 사진은 검은 단발머리의 동아시아계 성인 여성으로, 지소영의 인물사진이라는 역할을 충족한다. 정확한 외모를 대조할 별도 참조는 없다. 배경에는 흰 얼굴판과 푸른 눈의 찰리가 있지만 앰버는 보이지 않는다. 붉은 ‘NEGATIVE’는 정확하고 선명하나, 왼쪽의 ‘지소영’, 오른쪽의 ‘2069’ 및 추가 숫자·문자가 함께 읽힌다. 주관 시점에서 제외되어야 할 찰리를 보여주는 대신 필요한 실시간 소녀 얼굴이 빠졌다.",
        "hard_violations": [
         "찰리의 눈으로 보는 장면에 찰리 자신의 머리와 몸을 정면 배치하여 주관 시점의 공간 관계를 위반했다.",
         "유일하게 허용된 ‘NEGATIVE’ 외에 ‘지소영’, ‘2069’ 등 읽을 수 있는 추가 문자를 넣었다."
        ],
        "physics": "배경 찰리의 몸과 팔은 잔해 뒤로 자연스럽게 이어지며 공중에 떠 있는 신체는 보이지 않는다. 담요와 음반·도구는 바닥이나 잔해에 놓여 있다. 중앙 사진과 문자는 요청된 가상 인식 화면이므로 손으로 들거나 바닥에 지지될 필요는 없다. 물리적 지지보다 카메라와 찰리의 위치 관계가 문제다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "totals": {
   "B": 1,
   "A": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1,
    "verdict_ko": "붉은 ‘NEGATIVE’는 사진 위에 정확히 겹치지만, 찰리를 외부에서 보는 넓은 구도와 등을 돌린 소녀, 추가 문자로 주관 시점과 얼굴 인식 장면을 위반한다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "저장 인물사진을 더 크게 강조하고 실경과 분리해 A보다 가깝지만, 찰리 자신이 보이는 잘못된 시점과 추가 문자, 앰버 얼굴의 부재로 재촬영이 필요하다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S6sh21_sel.png",
    "asset_id": "d4faca1c-5912-4a07-a66f-a164deed0b1b",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd77-be39-7d5e-a6da-8e48f0bd2280",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S6sh21"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S6sh30::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:17:31.898761+00:00",
  "fingerprint": "7efa3beaf8e502134bfe9e4725726c522d69ae15cf6aa9e25a9fc48d8c984afa",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S6sh30_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S6sh30_sel.png",
  "source_sha256": "888afb663bd822d87d24e7b0dc58a261b8b2befe31780163c8900d53c399d47d",
  "file": "S6sh30_cine.png",
  "staged_sha256": "9663678c83d3213d87947c0894da4b7cb0d3d3f0a0794bc2f85fd441e1403c0f",
  "latency_ms": 11793
 },
 "S6sh46::signage": {
  "fp": "6d896a59539e83ce",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S6sh46": {
  "input_fingerprint": "9eac026a20ea69da",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 라울과 앰버가 코트 입은 찰리의 양옆에 바짝 붙어 한 발을 내딛은 도중, 서로 마주 본 채 환하게 웃는 찰나.\n\nLOCATION (lock): Along the outdoor route leaving the refugee-settlement dump toward the container homes. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee settlement road (The companions are departing along it toward 페드로's home) — The route extends diagonally beyond the walking group; used as Leaves walking space ahead and prepares the later upward tilt; 찰리's old coat and hat (Worn as a disguise) — Their side and oblique front contours are visible on 찰리; used as Preserves the disguised silhouette between the children without exaggerating its size.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight contrast that preserves the children’s expressive eyes and the worn disguise without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie now wears an old coat and hat over his scrap-robot exterior, with his worn chest logo and earlier net entanglement concealed rather than explicitly removed. The repaired jukebox has been switched off. 앰버: Amber is walking toward Pedro's home, laughing after the danger has passed. Her mask and waist tool pouch remain established. 라울: Raul is walking toward Pedro's home, laughing with relief.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 라울과 앰버가 코트 입은 찰리의 양옆에 바짝 붙어 한 발을 내딛은 도중, 서로 마주 본 채 환하게 웃는 찰나.\n\nLOCATION (lock): Along the outdoor route leaving the refugee-settlement dump toward the container homes. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee settlement road (The companions are departing along it toward 페드로's home) — The route extends diagonally beyond the walking group; used as Leaves walking space ahead and prepares the later upward tilt; 찰리's old coat and hat (Worn as a disguise) — Their side and oblique front contours are visible on 찰리; used as Preserves the disguised silhouette between the children without exaggerating its size.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight contrast that preserves the children’s expressive eyes and the worn disguise without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie now wears an old coat and hat over his scrap-robot exterior, with his worn chest logo and earlier net entanglement concealed rather than explicitly removed. The repaired jukebox has been switched off. 앰버: Amber is walking toward Pedro's home, laughing after the danger has passed. Her mask and waist tool pouch remain established. 라울: Raul is walking toward Pedro's home, laughing with relief.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 라울과 앰버가 코트 입은 찰리의 양옆에 바짝 붙어 한 발을 내딛은 도중, 서로 마주 본 채 환하게 웃는 찰나.\n\nLOCATION (lock): Along the outdoor route leaving the refugee-settlement dump toward the container homes. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee settlement road (The companions are departing along it toward 페드로's home) — The route extends diagonally beyond the walking group; used as Leaves walking space ahead and prepares the later upward tilt; 찰리's old coat and hat (Worn as a disguise) — Their side and oblique front contours are visible on 찰리; used as Preserves the disguised silhouette between the children without exaggerating its size.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight contrast that preserves the children’s expressive eyes and the worn disguise without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie now wears an old coat and hat over his scrap-robot exterior, with his worn chest logo and earlier net entanglement concealed rather than explicitly removed. The repaired jukebox has been switched off. 앰버: Amber is walking toward Pedro's home, laughing after the danger has passed. Her mask and waist tool pouch remain established. 라울: Raul is walking toward Pedro's home, laughing with relief.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라를 향해 걸어오며, 아이들은 프롬프트의 지시물(서로)이 아닌 찰리를 쳐다보고 있음.",
    "built_space": "배경에 고철 더미와 길이 있으며, 레퍼런스에 없던 대형 주크박스가 뜬금없이 왼쪽에 배치됨.",
    "entities": "찰리는 모자와 코트를 착용함. 앰버와 라울은 레퍼런스와 완전히 다른 옷(후드티와 긴바지 등)을 입고 있어 신원이 어긋남.",
    "hard_violations": [
     "[gemini-pro] 앰버와 라울의 복장이 지정된 레퍼런스와 전혀 일치하지 않음",
     "[gemini-pro] 찰리의 코트 양쪽 소매가 아래로 늘어져 있음에도 아이들 어깨 위에 별도의 로봇 손이 얹혀 있어 신체 구조상 물리적으로 불가능함(여분의 팔)"
    ],
    "physics": "세 명 모두 바닥을 딛고 걷는 자세를 취하고 있으나, 찰리의 팔과 손 연결이 구조적으로 성립하지 않음."
   },
   {
    "label": "B",
    "direction": "카메라를 등지고 컨테이너 방향으로 걸어가며, 앰버와 라울이 서로 마주 보고 있음.",
    "built_space": "왼쪽의 고철 더미와 오른쪽의 컨테이너 주택이 프롬프트에서 지시한 이동 경로 및 배경과 정확히 일치함.",
    "entities": "찰리는 뒷모습으로 코트와 모자를 착용함. 앰버(원피스, 재킷)와 라울(티셔츠, 반바지)의 복장이 레퍼런스와 일치함. 앰버의 마스크는 머리 뒤쪽에 다소 어색하게 묶여 있음.",
    "hard_violations": [],
    "physics": "세 인물 모두 자연스럽게 지면을 딛고 걷고 있으며, 서로 손을 잡거나 옷깃을 끄는 동작의 지탱 상태가 정상적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "아이들이 서로 마주 보며 걷는 핵심 연출과 캐릭터들의 복장 레퍼런스를 정확히 구현하여 프롬프트 충실도가 높습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "아이들이 서로를 보지 않고 복장 레퍼런스를 완전히 무시했으며, 찰리의 해부학적 구조(추가 팔)에 심각한 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라를 향해 걸어오며, 아이들은 프롬프트의 지시물(서로)이 아닌 찰리를 쳐다보고 있음.",
        "built_space": "배경에 고철 더미와 길이 있으며, 레퍼런스에 없던 대형 주크박스가 뜬금없이 왼쪽에 배치됨.",
        "entities": "찰리는 모자와 코트를 착용함. 앰버와 라울은 레퍼런스와 완전히 다른 옷(후드티와 긴바지 등)을 입고 있어 신원이 어긋남.",
        "hard_violations": [
         "앰버와 라울의 복장이 지정된 레퍼런스와 전혀 일치하지 않음",
         "찰리의 코트 양쪽 소매가 아래로 늘어져 있음에도 아이들 어깨 위에 별도의 로봇 손이 얹혀 있어 신체 구조상 물리적으로 불가능함(여분의 팔)"
        ],
        "physics": "세 명 모두 바닥을 딛고 걷는 자세를 취하고 있으나, 찰리의 팔과 손 연결이 구조적으로 성립하지 않음."
       },
       {
        "label": "B",
        "direction": "카메라를 등지고 컨테이너 방향으로 걸어가며, 앰버와 라울이 서로 마주 보고 있음.",
        "built_space": "왼쪽의 고철 더미와 오른쪽의 컨테이너 주택이 프롬프트에서 지시한 이동 경로 및 배경과 정확히 일치함.",
        "entities": "찰리는 뒷모습으로 코트와 모자를 착용함. 앰버(원피스, 재킷)와 라울(티셔츠, 반바지)의 복장이 레퍼런스와 일치함. 앰버의 마스크는 머리 뒤쪽에 다소 어색하게 묶여 있음.",
        "hard_violations": [],
        "physics": "세 인물 모두 자연스럽게 지면을 딛고 걷고 있으며, 서로 손을 잡거나 옷깃을 끄는 동작의 지탱 상태가 정상적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "아이들이 서로 마주 보며 걷는 핵심 연출과 캐릭터들의 복장 레퍼런스를 정확히 구현하여 프롬프트 충실도가 높습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "아이들이 서로를 보지 않고 복장 레퍼런스를 완전히 무시했으며, 찰리의 해부학적 구조(추가 팔)에 심각한 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라를 향해 걸어오며, 아이들은 프롬프트의 지시물(서로)이 아닌 찰리를 쳐다보고 있음.",
        "built_space": "배경에 고철 더미와 길이 있으며, 레퍼런스에 없던 대형 주크박스가 뜬금없이 왼쪽에 배치됨.",
        "entities": "찰리는 모자와 코트를 착용함. 앰버와 라울은 레퍼런스와 완전히 다른 옷(후드티와 긴바지 등)을 입고 있어 신원이 어긋남.",
        "hard_violations": [
         "앰버와 라울의 복장이 지정된 레퍼런스와 전혀 일치하지 않음",
         "찰리의 코트 양쪽 소매가 아래로 늘어져 있음에도 아이들 어깨 위에 별도의 로봇 손이 얹혀 있어 신체 구조상 물리적으로 불가능함(여분의 팔)"
        ],
        "physics": "세 명 모두 바닥을 딛고 걷는 자세를 취하고 있으나, 찰리의 팔과 손 연결이 구조적으로 성립하지 않음."
       },
       {
        "label": "B",
        "direction": "카메라를 등지고 컨테이너 방향으로 걸어가며, 앰버와 라울이 서로 마주 보고 있음.",
        "built_space": "왼쪽의 고철 더미와 오른쪽의 컨테이너 주택이 프롬프트에서 지시한 이동 경로 및 배경과 정확히 일치함.",
        "entities": "찰리는 뒷모습으로 코트와 모자를 착용함. 앰버(원피스, 재킷)와 라울(티셔츠, 반바지)의 복장이 레퍼런스와 일치함. 앰버의 마스크는 머리 뒤쪽에 다소 어색하게 묶여 있음.",
        "hard_violations": [],
        "physics": "세 인물 모두 자연스럽게 지면을 딛고 걷고 있으며, 서로 손을 잡거나 옷깃을 끄는 동작의 지탱 상태가 정상적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "찰리 양옆에 밀착해 웃으며 걷는 순간과 코트의 사선 정면 윤곽은 더 충실하지만, 집들이 있는 쪽과 반대로 걷고 아이들의 의상·마스크·도구 주머니를 유지하지 못했다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "아이들의 참고 의상과 자연스러운 보행은 잘 유지했으나, 찰리를 거의 뒤에서 보여 정면 윤곽을 놓치고 이동 방향도 주거지보다 고철 더미 쪽으로 읽힌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "라울은 오른쪽의 동료들을 향해 고개를 돌려 웃고, 앰버는 왼쪽 위의 찰리를 보며 웃는다. 찰리도 오른쪽 앰버 쪽으로 얼굴을 돌린다. 세 사람의 발과 몸통은 화면 안쪽 고철 더미를 향하므로, 쓰레기장을 떠나 컨테이너 집으로 향한다는 이동 방향은 명확하게 구현되지 않았다.",
        "built_space": "중앙 흙길 양쪽에 컨테이너 벽이 있고, 오른쪽에는 철망 울타리 한 구간과 그 앞의 드럼통 하나, 왼쪽에는 실외기 하나와 드럼통 하나가 보인다. 뒤쪽 대부분을 고철 더미가 차지한다. 세 사람은 모두 길 위에 서 있으며 추가 인물이나 불가능한 반사는 없다. 참고의 녹슨 고철·먼지·임시 구조물 재질은 이어지지만, 길이 일행 너머 대각선으로 주거지까지 이어지는 구성은 약하다.",
        "entities": "등장 주체는 찰리와 두 아이뿐이다. 찰리는 갈색 낡은 모자와 긴 코트, 청바지와 부츠를 착용하고 회백색 기계 얼굴의 옆면을 보인다. 가슴 로고와 그물은 코트에 가려져 있다. 라울은 어린 갈색 피부의 남자아이로 묶은 곱슬머리, 낡은 녹색 티셔츠와 데님 반바지가 참고와 부합한다. 앰버는 어린 금발 여자아이로 녹색 재킷, 회색 원피스와 부츠가 부합하며, 흰 마스크는 얼굴이 아니라 등 위쪽에 걸려 있다. 두 아이의 세부 혼혈 정체성은 이 각도만으로 확정하기 어렵다. 앰버의 도구 주머니는 뒤쪽 시점과 팔에 가려 확인되지 않는다. 주크박스는 보이지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "세 사람 모두 한쪽 발로 지면을 지지하면서 반대쪽 발을 옮기는 보행 자세다. 라울의 손은 찰리의 코트 위 팔에 닿고, 앰버의 안쪽 팔도 찰리 쪽으로 연결된다. 코트는 어깨에 걸려 아래로 처지고 모자는 머리에 얹혀 있다. 앰버의 마스크는 목 부근에서 등에 걸린 모습이다. 지지 없이 떠 있는 몸이나 물건은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "앰버는 오른쪽 위의 찰리 얼굴을, 라울은 왼쪽 위의 찰리 얼굴을 보며 입을 벌려 웃는다. 찰리의 얼굴은 약간 왼쪽 아래 앰버 쪽을 향한다. 세 사람은 카메라 쪽 전경으로 걸어오는데, 컨테이너 주거지는 그들 뒤 오른쪽에 있어 집으로 향하기보다 집에서 멀어지는 방향으로 읽힌다.",
        "built_space": "왼쪽에 큰 고철 더미와 축음기 나팔 하나, 전경의 검은 상자 하나가 있고, 더미 가장자리에 주크박스 한 대가 놓여 있다. 흙길은 일행 뒤 오른쪽으로 대각선으로 이어지며, 끝에는 여러 컨테이너 주택이 두 층 높이로 배치되어 있다. 세 사람은 길 중앙에 나란히 있고 아이들이 찰리 양옆에 밀착한다. 참고의 고철 더미와 폐품 재질은 비교적 잘 이어지지만, 참고에서 확립된 주거지 쪽 구조물 배치는 달라졌다. 중복된 인물이나 불가능한 반사는 없다.",
        "entities": "찰리는 각진 회백색 고릴라형 기계 얼굴, 작은 눈과 선형 입, 낡은 갈색 모자와 코트를 갖춘다. 코트 앞이 열려 금속 몸통은 보이지만 글자 로고나 그물은 드러나지 않는다. 앰버는 어린 금발 여자아이, 라울은 갈색 피부와 뒤로 묶은 머리의 어린 남자아이로 표현된다. 다만 앰버의 원피스·녹색 재킷은 바지·청회색 겉옷으로, 라울의 녹색 티셔츠·반바지는 갈색 겉옷·긴바지로 바뀌었다. 앰버의 마스크와 허리 도구 주머니도 보이지 않는다. 얼굴의 세부 혼혈 정체성은 확정하기 어렵다. 주크박스는 조명이 켜지지 않은 모습이며, 명확히 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "찰리와 앰버는 지면에 닿은 한쪽 발이 몸을 받치고 반대쪽 발을 앞으로 옮긴다. 라울도 두 발의 앞뒤 간격과 체중 이동이 보행으로 성립한다. 찰리의 기계 손은 아이들의 어깨 바깥쪽에 닿아 있고, 아이들은 자신의 발로 체중을 지탱한다. 모자는 머리에, 코트는 어깨에 지지되며 주크박스와 폐품은 지면이나 고철 더미에 놓여 있다. 부유하거나 지지 없이 매달린 대상은 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "찰리 양옆에 밀착해 웃으며 걷는 순간과 코트의 사선 정면 윤곽은 더 충실하지만, 집들이 있는 쪽과 반대로 걷고 아이들의 의상·마스크·도구 주머니를 유지하지 못했다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "아이들의 참고 의상과 자연스러운 보행은 잘 유지했으나, 찰리를 거의 뒤에서 보여 정면 윤곽을 놓치고 이동 방향도 주거지보다 고철 더미 쪽으로 읽힌다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "라울은 오른쪽의 동료들을 향해 고개를 돌려 웃고, 앰버는 왼쪽 위의 찰리를 보며 웃는다. 찰리도 오른쪽 앰버 쪽으로 얼굴을 돌린다. 세 사람의 발과 몸통은 화면 안쪽 고철 더미를 향하므로, 쓰레기장을 떠나 컨테이너 집으로 향한다는 이동 방향은 명확하게 구현되지 않았다.",
        "built_space": "중앙 흙길 양쪽에 컨테이너 벽이 있고, 오른쪽에는 철망 울타리 한 구간과 그 앞의 드럼통 하나, 왼쪽에는 실외기 하나와 드럼통 하나가 보인다. 뒤쪽 대부분을 고철 더미가 차지한다. 세 사람은 모두 길 위에 서 있으며 추가 인물이나 불가능한 반사는 없다. 참고의 녹슨 고철·먼지·임시 구조물 재질은 이어지지만, 길이 일행 너머 대각선으로 주거지까지 이어지는 구성은 약하다.",
        "entities": "등장 주체는 찰리와 두 아이뿐이다. 찰리는 갈색 낡은 모자와 긴 코트, 청바지와 부츠를 착용하고 회백색 기계 얼굴의 옆면을 보인다. 가슴 로고와 그물은 코트에 가려져 있다. 라울은 어린 갈색 피부의 남자아이로 묶은 곱슬머리, 낡은 녹색 티셔츠와 데님 반바지가 참고와 부합한다. 앰버는 어린 금발 여자아이로 녹색 재킷, 회색 원피스와 부츠가 부합하며, 흰 마스크는 얼굴이 아니라 등 위쪽에 걸려 있다. 두 아이의 세부 혼혈 정체성은 이 각도만으로 확정하기 어렵다. 앰버의 도구 주머니는 뒤쪽 시점과 팔에 가려 확인되지 않는다. 주크박스는 보이지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "세 사람 모두 한쪽 발로 지면을 지지하면서 반대쪽 발을 옮기는 보행 자세다. 라울의 손은 찰리의 코트 위 팔에 닿고, 앰버의 안쪽 팔도 찰리 쪽으로 연결된다. 코트는 어깨에 걸려 아래로 처지고 모자는 머리에 얹혀 있다. 앰버의 마스크는 목 부근에서 등에 걸린 모습이다. 지지 없이 떠 있는 몸이나 물건은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "앰버는 오른쪽 위의 찰리 얼굴을, 라울은 왼쪽 위의 찰리 얼굴을 보며 입을 벌려 웃는다. 찰리의 얼굴은 약간 왼쪽 아래 앰버 쪽을 향한다. 세 사람은 카메라 쪽 전경으로 걸어오는데, 컨테이너 주거지는 그들 뒤 오른쪽에 있어 집으로 향하기보다 집에서 멀어지는 방향으로 읽힌다.",
        "built_space": "왼쪽에 큰 고철 더미와 축음기 나팔 하나, 전경의 검은 상자 하나가 있고, 더미 가장자리에 주크박스 한 대가 놓여 있다. 흙길은 일행 뒤 오른쪽으로 대각선으로 이어지며, 끝에는 여러 컨테이너 주택이 두 층 높이로 배치되어 있다. 세 사람은 길 중앙에 나란히 있고 아이들이 찰리 양옆에 밀착한다. 참고의 고철 더미와 폐품 재질은 비교적 잘 이어지지만, 참고에서 확립된 주거지 쪽 구조물 배치는 달라졌다. 중복된 인물이나 불가능한 반사는 없다.",
        "entities": "찰리는 각진 회백색 고릴라형 기계 얼굴, 작은 눈과 선형 입, 낡은 갈색 모자와 코트를 갖춘다. 코트 앞이 열려 금속 몸통은 보이지만 글자 로고나 그물은 드러나지 않는다. 앰버는 어린 금발 여자아이, 라울은 갈색 피부와 뒤로 묶은 머리의 어린 남자아이로 표현된다. 다만 앰버의 원피스·녹색 재킷은 바지·청회색 겉옷으로, 라울의 녹색 티셔츠·반바지는 갈색 겉옷·긴바지로 바뀌었다. 앰버의 마스크와 허리 도구 주머니도 보이지 않는다. 얼굴의 세부 혼혈 정체성은 확정하기 어렵다. 주크박스는 조명이 켜지지 않은 모습이며, 명확히 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "찰리와 앰버는 지면에 닿은 한쪽 발이 몸을 받치고 반대쪽 발을 앞으로 옮긴다. 라울도 두 발의 앞뒤 간격과 체중 이동이 보행으로 성립한다. 찰리의 기계 손은 아이들의 어깨 바깥쪽에 닿아 있고, 아이들은 자신의 발로 체중을 지탱한다. 모자는 머리에, 코트는 어깨에 지지되며 주크박스와 폐품은 지면이나 고철 더미에 놓여 있다. 부유하거나 지지 없이 매달린 대상은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.833
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.833
   },
   "violations": {
    "A": [
     "[gemini-pro] 앰버와 라울의 복장이 지정된 레퍼런스와 전혀 일치하지 않음",
     "[gemini-pro] 찰리의 코트 양쪽 소매가 아래로 늘어져 있음에도 아이들 어깨 위에 별도의 로봇 손이 얹혀 있어 신체 구조상 물리적으로 불가능함(여분의 팔)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1833,
   "A": 1179
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1833,
    "verdict_ko": "아이들이 서로 마주 보며 걷는 핵심 연출과 캐릭터들의 복장 레퍼런스를 정확히 구현하여 프롬프트 충실도가 높습니다."
   },
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "아이들이 서로를 보지 않고 복장 레퍼런스를 완전히 무시했으며, 찰리의 해부학적 구조(추가 팔)에 심각한 오류가 있습니다.  ★위반: [gemini-pro] 앰버와 라울의 복장이 지정된 레퍼런스와 전혀 일치하지 않음 / [gemini-pro] 찰리의 코트 양쪽 소매가 아래로 늘어져 있음에도 아이들 어깨 위에 별도의 로봇 손이 얹혀 있어 신체 구조상 물리적으로 불가능함(여분의 팔)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S6sh21_sel.png",
    "asset_id": "d4faca1c-5912-4a07-a66f-a164deed0b1b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:907223>",
    "asset_id": "12a69d16-0db4-4925-9828-f0f1eae433e2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:943979>",
    "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd78-8085-75b8-92a0-7224eb5a37b3",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S6sh21"
  }
 },
 "S6sh46::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:17:44.046550+00:00",
  "fingerprint": "ac7f875ca9238c19d6d54cf0de864076f664d2f77ab228d005c8ccd098a0ac89",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S6sh46_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S6sh46_sel.png",
  "source_sha256": "29aa4a82a26e4b9f3d91dc472ca407cc89453772de1a0a303d2a82ec5dde107b",
  "file": "S6sh46_cine.png",
  "staged_sha256": "cd42ecff5a98f9801cc7dce7110bd0ca015c0a3e7ee2f74a77972048301e932c",
  "latency_ms": 11771
 },
 "S7sh20::signage": {
  "fp": "2f295fa359b89e7a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::7c9beae4fac35ffa": {
  "subjects": [],
  "subject_text": "인천 난민촌 인공제방 작업장과 제방 벽 앞\n거대한 인공제방 벽과 지지대 앞에 보수 자재와 돌이 놓인 작업장. 노후한 벽에는 균열과 물이 스며든 흔적이 있고 바닥에 웅덩이가 있다.",
  "identity": "canonical",
  "scope_id": "L14",
  "scope_role": "location_exterior",
  "scope_sha": "f56d1cd5e3214356"
 },
 "S7sh20::bgfirst_bg": {
  "input_fingerprint": "2f2d536615f01a5b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 미연이 관리소장 코앞에 서서 손가락으로 금이 간 제방 벽을 정확히 가리키는 모습.\n\nLOCATION (lock): On the open repair worksite beside the refugee settlement's cracked seawall, near the stopped military truck.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: cracked embankment wall beyond 미연's pointing finger in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Cracked embankment wall (Severely cracked, with leakage identified by 미연) — The damaged wall face is visible beyond the direction of her extended finger; used as Provides visible evidence for her accusation rather than an anonymous backdrop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daylight exposure consistent across both figures and preserve enough background contrast for the wall cracks to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 미연이 관리소장 코앞에 서서 손가락으로 금이 간 제방 벽을 정확히 가리키는 모습.\n\nLOCATION (lock): On the open repair worksite beside the refugee settlement's cracked seawall, near the stopped military truck.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: cracked embankment wall beyond 미연's pointing finger in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Cracked embankment wall (Severely cracked, with leakage identified by 미연) — The damaged wall face is visible beyond the direction of her extended finger; used as Provides visible evidence for her accusation rather than an anonymous backdrop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daylight exposure consistent across both figures and preserve enough background contrast for the wall cracks to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S7sh20__bgfirst_bg.png",
  "asset_id": "0d7de940-6616-47b7-9eeb-55b0e83f7e16",
  "input_asset_ids": [
   "057cea98-c7f4-4fb0-aa3d-7c8ee6a77306",
   "689999ad-8d31-4952-8354-722109bc1dd0"
  ]
 },
 "S7sh20": {
  "input_fingerprint": "6842ece8c2451f6f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 미연이 관리소장 코앞에 서서 손가락으로 금이 간 제방 벽을 정확히 가리키는 모습.\n\nLOCATION (lock): On the open repair worksite beside the refugee settlement's cracked seawall, near the stopped military truck. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: cracked embankment wall beyond 미연's pointing finger in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Cracked embankment wall (Severely cracked, with leakage identified by 미연) — The damaged wall face is visible beyond the direction of her extended finger; used as Provides visible evidence for her accusation rather than an anonymous backdrop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daylight exposure consistent across both figures and preserve enough background contrast for the wall cracks to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall has severe cracks and visible water leakage; repair materials and equipment remain at the halted worksite. The military truck remains stopped in front of the forklift, and the puddle at the disembarkation point remains. 미연: Miyeon is out of the forklift and has pushed forward through the crowd near the truck. 관리소장: The camp administrator is beside the truck, attempting to return to it. His shoes remain dirty from the puddle splash.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴); 관리소장 (50대 남성, 대머리, 중년의 얼굴) — wearing: 싸구려 광택이 도는 낡고 통이 넓은 회색 양복과 흙탕물이 튄 구두. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 미연이 관리소장 코앞에 서서 손가락으로 금이 간 제방 벽을 정확히 가리키는 모습.\n\nLOCATION (lock): On the open repair worksite beside the refugee settlement's cracked seawall, near the stopped military truck. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: cracked embankment wall beyond 미연's pointing finger in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Cracked embankment wall (Severely cracked, with leakage identified by 미연) — The damaged wall face is visible beyond the direction of her extended finger; used as Provides visible evidence for her accusation rather than an anonymous backdrop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daylight exposure consistent across both figures and preserve enough background contrast for the wall cracks to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall has severe cracks and visible water leakage; repair materials and equipment remain at the halted worksite. The military truck remains stopped in front of the forklift, and the puddle at the disembarkation point remains. 미연: Miyeon is out of the forklift and has pushed forward through the crowd near the truck. 관리소장: The camp administrator is beside the truck, attempting to return to it. His shoes remain dirty from the puddle splash.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴); 관리소장 (50대 남성, 대머리, 중년의 얼굴) — wearing: 싸구려 광택이 도는 낡고 통이 넓은 회색 양복과 흙탕물이 튄 구두. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 미연이 관리소장 코앞에 서서 손가락으로 금이 간 제방 벽을 정확히 가리키는 모습.\n\nLOCATION (lock): On the open repair worksite beside the refugee settlement's cracked seawall, near the stopped military truck. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: cracked embankment wall beyond 미연's pointing finger in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Cracked embankment wall (Severely cracked, with leakage identified by 미연) — The damaged wall face is visible beyond the direction of her extended finger; used as Provides visible evidence for her accusation rather than an anonymous backdrop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daylight exposure consistent across both figures and preserve enough background contrast for the wall cracks to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall has severe cracks and visible water leakage; repair materials and equipment remain at the halted worksite. The military truck remains stopped in front of the forklift, and the puddle at the disembarkation point remains. 미연: Miyeon is out of the forklift and has pushed forward through the crowd near the truck. 관리소장: The camp administrator is beside the truck, attempting to return to it. His shoes remain dirty from the puddle splash.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴); 관리소장 (50대 남성, 대머리, 중년의 얼굴) — wearing: 싸구려 광택이 도는 낡고 통이 넓은 회색 양복과 흙탕물이 튄 구두. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S7sh20__bgfirst_bg.png",
     "asset_id": "0d7de940-6616-47b7-9eeb-55b0e83f7e16",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S7sh20.png",
     "asset_id": "057cea98-c7f4-4fb0-aa3d-7c8ee6a77306",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:878964>",
     "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_artificial_seawall_sel.png",
     "asset_id": "689999ad-8d31-4952-8354-722109bc1dd0",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:878964>",
     "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "미연은 관리소장을 바라보며 손가락으로 화면 왼쪽의 제방 벽을 정확히 가리키고 있습니다.",
    "built_space": "참조 이미지와 동일한 사선형 지지대가 있는 방파제 구조물이 왼쪽에 배치되었으며, 배경에 트럭과 작업 현장이 보입니다.",
    "entities": "미연은 참조 이미지의 인상착의와 일치하지만, 관리소장은 지시된 대머리가 아닌 풍성한 검은 머리로 나타납니다.",
    "hard_violations": [
     "[gpt-high] 트럭과 지게차 부근 배경에 안전모를 쓴 제삼의 작업자가 보입니다. 미연과 관리소장 외에는 어떤 사람도 추가하지 말라는 명시적 제한을 위반합니다."
    ],
    "physics": "두 인물 모두 지면에 두 발로 안정적으로 서 있습니다."
   },
   {
    "label": "B",
    "direction": "미연이 제방 벽이 아닌 관리소장의 얼굴을 향해 손가락을 찌르듯 가리키고 있습니다.",
    "built_space": "참조 이미지의 방파제와 완전히 다른, 거대한 곡선형 콘크리트 벽면이 배경을 차지하고 있습니다.",
    "entities": "미연과 회색 양복을 입은 대머리 관리소장 모두 지시문 및 참조 이미지의 외형 조건과 잘 일치합니다.",
    "hard_violations": [],
    "physics": "두 인물 모두 흙바닥에 정상적으로 서 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 방파제 구조와 벽을 가리키는 핵심 동작을 잘 구현했으나, 관리소장이 대머리가 아닌 점은 우선순위 4에서 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "관리소장의 외형은 정확히 일치하지만, 미연이 벽이 아닌 인물을 가리키고 있으며 참조된 장소와 전혀 다른 방파제를 그려내 연출과 장소 설정에서 크게 실패했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 관리소장을 바라보며 손가락으로 화면 왼쪽의 제방 벽을 정확히 가리키고 있습니다.",
        "built_space": "참조 이미지와 동일한 사선형 지지대가 있는 방파제 구조물이 왼쪽에 배치되었으며, 배경에 트럭과 작업 현장이 보입니다.",
        "entities": "미연은 참조 이미지의 인상착의와 일치하지만, 관리소장은 지시된 대머리가 아닌 풍성한 검은 머리로 나타납니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 두 발로 안정적으로 서 있습니다."
       },
       {
        "label": "B",
        "direction": "미연이 제방 벽이 아닌 관리소장의 얼굴을 향해 손가락을 찌르듯 가리키고 있습니다.",
        "built_space": "참조 이미지의 방파제와 완전히 다른, 거대한 곡선형 콘크리트 벽면이 배경을 차지하고 있습니다.",
        "entities": "미연과 회색 양복을 입은 대머리 관리소장 모두 지시문 및 참조 이미지의 외형 조건과 잘 일치합니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 흙바닥에 정상적으로 서 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 방파제 구조와 벽을 가리키는 핵심 동작을 잘 구현했으나, 관리소장이 대머리가 아닌 점은 우선순위 4에서 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "관리소장의 외형은 정확히 일치하지만, 미연이 벽이 아닌 인물을 가리키고 있으며 참조된 장소와 전혀 다른 방파제를 그려내 연출과 장소 설정에서 크게 실패했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "미연은 관리소장을 바라보며 손가락으로 화면 왼쪽의 제방 벽을 정확히 가리키고 있습니다.",
        "built_space": "참조 이미지와 동일한 사선형 지지대가 있는 방파제 구조물이 왼쪽에 배치되었으며, 배경에 트럭과 작업 현장이 보입니다.",
        "entities": "미연은 참조 이미지의 인상착의와 일치하지만, 관리소장은 지시된 대머리가 아닌 풍성한 검은 머리로 나타납니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 두 발로 안정적으로 서 있습니다."
       },
       {
        "label": "B",
        "direction": "미연이 제방 벽이 아닌 관리소장의 얼굴을 향해 손가락을 찌르듯 가리키고 있습니다.",
        "built_space": "참조 이미지의 방파제와 완전히 다른, 거대한 곡선형 콘크리트 벽면이 배경을 차지하고 있습니다.",
        "entities": "미연과 회색 양복을 입은 대머리 관리소장 모두 지시문 및 참조 이미지의 외형 조건과 잘 일치합니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 흙바닥에 정상적으로 서 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "추가 인물이라는 실격 사유는 없지만, 손가락이 균열 아닌 소장의 얼굴을 향하고 미디엄 숏 대신 전신을 보여주며 제방 구조도 참조와 다릅니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "균열을 가리키는 방향과 미디엄 숏, 제방 구조는 더 정확하지만 금지된 배경 인물을 추가해 실격이며 소장의 대머리 설정도 어깁니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 소장을 바라보면서 검지를 그의 코와 입 부근으로 뻗고 있습니다. 손가락의 진행 방향을 소장의 얼굴이 가로막으므로, 왼쪽 배경의 제방 균열을 정확히 지목하는 동작으로 읽히지 않습니다. 소장도 미연을 바라봅니다. 군용 트럭은 카메라 쪽으로 비스듬히 향해 정차해 있습니다.",
        "built_space": "배경에는 높고 상부가 굽은 콘크리트 벽 하나가 있으며 참조의 반복되는 돌출 부벽은 보이지 않습니다. 참조의 낮고 곧은 제방과 해안 도로 대신 높은 벽 앞의 넓은 공터로 바뀌었습니다. 오른쪽에 군용 트럭 한 대, 왼쪽에 물웅덩이 하나, 벽 아래에 보수 자재와 장비가 있습니다. 지게차는 일부 노란 장비에 가려져 확실하게 식별되지 않습니다. 두 사람 모두 공터에 서 있지만 전신이 나오는 넓은 구도이며, 미연은 소장에게 팔을 뻗어야 닿을 거리입니다.",
        "entities": "보이는 사람은 미연과 소장에 해당하는 두 명뿐입니다. 미연은 중년 동아시아계 여성으로 검은 단발과 낡은 갈색 작업복이 참조와 대체로 맞지만 검은 모자가 없습니다. 소장은 중년 남성이며 정수리가 벗겨진 머리, 광택 있는 회색 양복, 더러운 구두를 갖췄습니다. 심한 벽 균열과 젖은 흔적, 물웅덩이, 군용 트럭 및 보수 자재가 보입니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "두 사람의 신발은 모두 지면에 닿아 있고 체중을 지탱합니다. 미연의 뻗은 팔과 검지는 어깨와 팔꿈치에서 자연스럽게 이어집니다. 트럭은 바퀴로 지면에 서 있고 자재와 장비는 바닥이나 팔레트에 놓여 있습니다. 물웅덩이는 낮은 지면에 고여 있으며 지지 없이 떠 있는 물체나 인물은 없습니다."
       },
       {
        "label": "B",
        "direction": "미연의 검지는 화면 왼쪽의 벌어진 균열과 물이 흘러나오는 벽면을 향합니다. 손끝 너머에 실제 손상 부위가 보여 지목 대상이 명확합니다. 미연의 눈은 바로 앞 소장을 향하고 소장도 몸과 머리를 미연 쪽으로 돌리고 있습니다. 오른쪽 트럭은 도로 방향으로 놓여 있습니다.",
        "built_space": "왼쪽의 낮고 곧은 콘크리트 제방, 바다, 해안 도로와 반복되는 사다리꼴 부벽이 참조 구조를 잘 따릅니다. 잘린 전경 부벽을 포함해 최소 일곱 개의 부벽이 식별됩니다. 오른쪽에는 잘린 군용 트럭 한 대와 그 뒤쪽 지게차의 마스트 한 쌍, 벽돌 및 포대 적재물이 있습니다. 두 주인공은 도로 위에서 가까이 마주 서 있으며 소장의 등 너머로 미연을 보는 미디엄 구도입니다. 균열은 손끝 너머 왼쪽 배경에 있고 벽 아래에는 누수로 젖은 바닥이 이어집니다.",
        "entities": "미연은 검은 단발의 중년 동아시아계 여성으로 작업 재킷과 회색 셔츠가 참조에 가깝지만 모자가 없습니다. 소장은 회색 양복을 입은 남성이지만 뒷머리와 정수리에 검은 머리카락이 풍성하여 명시된 대머리와 맞지 않습니다. 신발은 프레임 밖이므로 진흙 상태는 확인할 수 없습니다. 두 주인공 외에 지게차와 적재물 사이에 안전모를 쓴 작업자 한 명이 추가로 보입니다. 균열, 흐르는 물, 보수 자재, 군용 트럭과 지게차는 식별되며 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "트럭과 지게차 부근 배경에 안전모를 쓴 제삼의 작업자가 보입니다. 미연과 관리소장 외에는 어떤 사람도 추가하지 말라는 명시적 제한을 위반합니다."
        ],
        "physics": "주인공들의 발은 프레임 밖이지만 몸통과 하체는 도로에 서 있는 자연스러운 자세로 이어지며 공중에 떠 있다는 징후는 없습니다. 미연의 팔과 손가락은 정상적인 관절 연결로 지탱됩니다. 배경 작업자는 적재물 뒤에서 상체만 보입니다. 보수 자재는 지면에 쌓여 있고 누수는 균열에서 아래로 떨어져 바닥에 고입니다. 지지 없는 물체나 불가능한 신체 자세는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "추가 인물이라는 실격 사유는 없지만, 손가락이 균열 아닌 소장의 얼굴을 향하고 미디엄 숏 대신 전신을 보여주며 제방 구조도 참조와 다릅니다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "균열을 가리키는 방향과 미디엄 숏, 제방 구조는 더 정확하지만 금지된 배경 인물을 추가해 실격이며 소장의 대머리 설정도 어깁니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "미연은 소장을 바라보면서 검지를 그의 코와 입 부근으로 뻗고 있습니다. 손가락의 진행 방향을 소장의 얼굴이 가로막으므로, 왼쪽 배경의 제방 균열을 정확히 지목하는 동작으로 읽히지 않습니다. 소장도 미연을 바라봅니다. 군용 트럭은 카메라 쪽으로 비스듬히 향해 정차해 있습니다.",
        "built_space": "배경에는 높고 상부가 굽은 콘크리트 벽 하나가 있으며 참조의 반복되는 돌출 부벽은 보이지 않습니다. 참조의 낮고 곧은 제방과 해안 도로 대신 높은 벽 앞의 넓은 공터로 바뀌었습니다. 오른쪽에 군용 트럭 한 대, 왼쪽에 물웅덩이 하나, 벽 아래에 보수 자재와 장비가 있습니다. 지게차는 일부 노란 장비에 가려져 확실하게 식별되지 않습니다. 두 사람 모두 공터에 서 있지만 전신이 나오는 넓은 구도이며, 미연은 소장에게 팔을 뻗어야 닿을 거리입니다.",
        "entities": "보이는 사람은 미연과 소장에 해당하는 두 명뿐입니다. 미연은 중년 동아시아계 여성으로 검은 단발과 낡은 갈색 작업복이 참조와 대체로 맞지만 검은 모자가 없습니다. 소장은 중년 남성이며 정수리가 벗겨진 머리, 광택 있는 회색 양복, 더러운 구두를 갖췄습니다. 심한 벽 균열과 젖은 흔적, 물웅덩이, 군용 트럭 및 보수 자재가 보입니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "두 사람의 신발은 모두 지면에 닿아 있고 체중을 지탱합니다. 미연의 뻗은 팔과 검지는 어깨와 팔꿈치에서 자연스럽게 이어집니다. 트럭은 바퀴로 지면에 서 있고 자재와 장비는 바닥이나 팔레트에 놓여 있습니다. 물웅덩이는 낮은 지면에 고여 있으며 지지 없이 떠 있는 물체나 인물은 없습니다."
       },
       {
        "label": "A",
        "direction": "미연의 검지는 화면 왼쪽의 벌어진 균열과 물이 흘러나오는 벽면을 향합니다. 손끝 너머에 실제 손상 부위가 보여 지목 대상이 명확합니다. 미연의 눈은 바로 앞 소장을 향하고 소장도 몸과 머리를 미연 쪽으로 돌리고 있습니다. 오른쪽 트럭은 도로 방향으로 놓여 있습니다.",
        "built_space": "왼쪽의 낮고 곧은 콘크리트 제방, 바다, 해안 도로와 반복되는 사다리꼴 부벽이 참조 구조를 잘 따릅니다. 잘린 전경 부벽을 포함해 최소 일곱 개의 부벽이 식별됩니다. 오른쪽에는 잘린 군용 트럭 한 대와 그 뒤쪽 지게차의 마스트 한 쌍, 벽돌 및 포대 적재물이 있습니다. 두 주인공은 도로 위에서 가까이 마주 서 있으며 소장의 등 너머로 미연을 보는 미디엄 구도입니다. 균열은 손끝 너머 왼쪽 배경에 있고 벽 아래에는 누수로 젖은 바닥이 이어집니다.",
        "entities": "미연은 검은 단발의 중년 동아시아계 여성으로 작업 재킷과 회색 셔츠가 참조에 가깝지만 모자가 없습니다. 소장은 회색 양복을 입은 남성이지만 뒷머리와 정수리에 검은 머리카락이 풍성하여 명시된 대머리와 맞지 않습니다. 신발은 프레임 밖이므로 진흙 상태는 확인할 수 없습니다. 두 주인공 외에 지게차와 적재물 사이에 안전모를 쓴 작업자 한 명이 추가로 보입니다. 균열, 흐르는 물, 보수 자재, 군용 트럭과 지게차는 식별되며 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "트럭과 지게차 부근 배경에 안전모를 쓴 제삼의 작업자가 보입니다. 미연과 관리소장 외에는 어떤 사람도 추가하지 말라는 명시적 제한을 위반합니다."
        ],
        "physics": "주인공들의 발은 프레임 밖이지만 몸통과 하체는 도로에 서 있는 자연스러운 자세로 이어지며 공중에 떠 있다는 징후는 없습니다. 미연의 팔과 손가락은 정상적인 관절 연결로 지탱됩니다. 배경 작업자는 적재물 뒤에서 상체만 보입니다. 보수 자재는 지면에 쌓여 있고 누수는 균열에서 아래로 떨어져 바닥에 고입니다. 지지 없는 물체나 불가능한 신체 자세는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.571
   },
   "violations": {
    "A": [
     "[gpt-high] 트럭과 지게차 부근 배경에 안전모를 쓴 제삼의 작업자가 보입니다. 미연과 관리소장 외에는 어떤 사람도 추가하지 말라는 명시적 제한을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1083,
   "B": 1571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "지정된 방파제 구조와 벽을 가리키는 핵심 동작을 잘 구현했으나, 관리소장이 대머리가 아닌 점은 우선순위 4에서 감점 요인입니다.  ★위반: [gpt-high] 트럭과 지게차 부근 배경에 안전모를 쓴 제삼의 작업자가 보입니다. 미연과 관리소장 외에는 어떤 사람도 추가하지 말라는 명시적 제한을 위반합니다."
   },
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "관리소장의 외형은 정확히 일치하지만, 미연이 벽이 아닌 인물을 가리키고 있으며 참조된 장소와 전혀 다른 방파제를 그려내 연출과 장소 설정에서 크게 실패했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_artificial_seawall_sel.png",
    "asset_id": "689999ad-8d31-4952-8354-722109bc1dd0",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:878964>",
    "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd79-420f-710a-aafa-13060481ecf5",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S7sh20__bgfirst_bg.png",
   "bg_asset_id": "0d7de940-6616-47b7-9eeb-55b0e83f7e16",
   "bg_record_key": "S7sh20::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S7sh20::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:17:56.217300+00:00",
  "fingerprint": "bd83ce2615ee59a45e26f3c7749e0db22656aded057dcc3bba9889640a4b20ab",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S7sh20_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S7sh20_sel.png",
  "source_sha256": "6ea06acb239bcb3317da1655193e64c0c90b49f0124172e3d2a60569a0828a2a",
  "file": "S7sh20_cine.png",
  "staged_sha256": "a56bd088d9d2eebfc17ebe0bcabdb4662e30862bdd2c5275c2f99d7b84753c0b",
  "latency_ms": 11607
 },
 "S7sh22::signage": {
  "fp": "e085892bf67a6515",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S7sh22": {
  "input_fingerprint": "9b5f493c97d59980",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 경비병의 두꺼운 장갑 낀 양손이 아시아계 난민 남성의 가슴팍을 맹렬히 밀쳐내는 타격의 찰나.\n\nLOCATION (lock): On the muddy ground where residents confront guards beside a military truck at the seawall repair site. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Embankment wall (Cracked and under interrupted repair) — Only a soft, partial view of the wall face remains beyond the torsos; used as Retains location continuity without competing with the point of contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight exposure, using controlled contrast to distinguish the contacting hands from the refugee’s chest.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the cracked embankment wall, wet worksite ground, and daylight appearance from the reference. EXCLUDE the pointing gesture and the loudspeaker as a carried foreground prop.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cracked seawall continues leaking, with the halted repair equipment and materials still in place. The military truck remains beside the forklift, near the puddled ground.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 관리소장 (50대 남성, 대머리, 중년의 얼굴) — wearing: 싸구려 광택이 도는 낡고 통이 넓은 회색 양복과 흙탕물이 튄 구두.; 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 경비병의 두꺼운 장갑 낀 양손이 아시아계 난민 남성의 가슴팍을 맹렬히 밀쳐내는 타격의 찰나.\n\nLOCATION (lock): On the muddy ground where residents confront guards beside a military truck at the seawall repair site. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Embankment wall (Cracked and under interrupted repair) — Only a soft, partial view of the wall face remains beyond the torsos; used as Retains location continuity without competing with the point of contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight exposure, using controlled contrast to distinguish the contacting hands from the refugee’s chest.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the cracked embankment wall, wet worksite ground, and daylight appearance from the reference. EXCLUDE the pointing gesture and the loudspeaker as a carried foreground prop.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cracked seawall continues leaking, with the halted repair equipment and materials still in place. The military truck remains beside the forklift, near the puddled ground.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 관리소장 (50대 남성, 대머리, 중년의 얼굴) — wearing: 싸구려 광택이 도는 낡고 통이 넓은 회색 양복과 흙탕물이 튄 구두.; 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 경비병의 두꺼운 장갑 낀 양손이 아시아계 난민 남성의 가슴팍을 맹렬히 밀쳐내는 타격의 찰나.\n\nLOCATION (lock): On the muddy ground where residents confront guards beside a military truck at the seawall repair site. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Embankment wall (Cracked and under interrupted repair) — Only a soft, partial view of the wall face remains beyond the torsos; used as Retains location continuity without competing with the point of contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight exposure, using controlled contrast to distinguish the contacting hands from the refugee’s chest.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the cracked embankment wall, wet worksite ground, and daylight appearance from the reference. EXCLUDE the pointing gesture and the loudspeaker as a carried foreground prop.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cracked seawall continues leaking, with the halted repair equipment and materials still in place. The military truck remains beside the forklift, near the puddled ground.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 관리소장 (50대 남성, 대머리, 중년의 얼굴) — wearing: 싸구려 광택이 도는 낡고 통이 넓은 회색 양복과 흙탕물이 튄 구두.; 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "경비병의 두 손이 난민 남성의 가슴과 왼쪽 팔뚝을 향해 맹렬히 밀치고 있음.",
    "built_space": "금이 간 콘크리트 방조제 벽이 배경에 넓게 보이며, 하단에 진흙탕 바닥이 보임.",
    "entities": "아시아계 난민 남성. 전술 장갑과 군복을 입은 경비병의 팔. 지시문에 없는 배경 인물 2명(미연 등).",
    "hard_violations": [
     "[gemini-pro] 숏 텍스트에 명시되지 않은 인물(배경의 여성 및 다른 남성) 임의 추가",
     "[gpt-high] 숏 텍스트에 없는 배경 인물 두 명을 추가했습니다."
    ],
    "physics": "난민 남성은 타격을 받아 뒤로 기울어지는 체중 이동을 보이며 하체로 지탱 중임."
   },
   {
    "label": "B",
    "direction": "두꺼운 장갑을 낀 양손이 난민 남성의 가슴팍을 향해 직접적으로 밀치고 있음.",
    "built_space": "금이 간 콘크리트 방조제 벽이 보이고, 바닥의 파편과 진흙이 일부 보임.",
    "entities": "아시아계 난민 남성. 두꺼운 장갑을 꼈으나 군복이 아닌 회색 양복(관리소장 복장)을 입은 팔.",
    "hard_violations": [
     "[gemini-pro] 난민 남성의 왼쪽 어깨 뒤에 몸통 없이 허공에 떠 있는 여분의 손(물리적으로 불가능한 신체 구조)"
    ],
    "physics": "난민은 밀리는 힘에 의해 뒤로 젖혀져 있으나, 어깨 뒤에 떠 있는 손을 지지하는 신체나 몸통이 전혀 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "숏 텍스트에 없는 인물들을 배경에 추가하는 치명적 오류를 범했으며, 요구된 클로즈업 대신 너무 넓은 미디엄 숏으로 렌더링되었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "난민 어깨 뒤에 몸통 없는 여분의 손이 나타나는 치명적 물리적 오류가 발생했고, 경비병의 팔에 이전 장면의 회색 양복이 잘못 입혀졌습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "경비병의 두 손이 난민 남성의 가슴과 왼쪽 팔뚝을 향해 맹렬히 밀치고 있음.",
        "built_space": "금이 간 콘크리트 방조제 벽이 배경에 넓게 보이며, 하단에 진흙탕 바닥이 보임.",
        "entities": "아시아계 난민 남성. 전술 장갑과 군복을 입은 경비병의 팔. 지시문에 없는 배경 인물 2명(미연 등).",
        "hard_violations": [
         "숏 텍스트에 명시되지 않은 인물(배경의 여성 및 다른 남성) 임의 추가"
        ],
        "physics": "난민 남성은 타격을 받아 뒤로 기울어지는 체중 이동을 보이며 하체로 지탱 중임."
       },
       {
        "label": "B",
        "direction": "두꺼운 장갑을 낀 양손이 난민 남성의 가슴팍을 향해 직접적으로 밀치고 있음.",
        "built_space": "금이 간 콘크리트 방조제 벽이 보이고, 바닥의 파편과 진흙이 일부 보임.",
        "entities": "아시아계 난민 남성. 두꺼운 장갑을 꼈으나 군복이 아닌 회색 양복(관리소장 복장)을 입은 팔.",
        "hard_violations": [
         "난민 남성의 왼쪽 어깨 뒤에 몸통 없이 허공에 떠 있는 여분의 손(물리적으로 불가능한 신체 구조)"
        ],
        "physics": "난민은 밀리는 힘에 의해 뒤로 젖혀져 있으나, 어깨 뒤에 떠 있는 손을 지지하는 신체나 몸통이 전혀 없음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "숏 텍스트에 없는 인물들을 배경에 추가하는 치명적 오류를 범했으며, 요구된 클로즈업 대신 너무 넓은 미디엄 숏으로 렌더링되었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "난민 어깨 뒤에 몸통 없는 여분의 손이 나타나는 치명적 물리적 오류가 발생했고, 경비병의 팔에 이전 장면의 회색 양복이 잘못 입혀졌습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "경비병의 두 손이 난민 남성의 가슴과 왼쪽 팔뚝을 향해 맹렬히 밀치고 있음.",
        "built_space": "금이 간 콘크리트 방조제 벽이 배경에 넓게 보이며, 하단에 진흙탕 바닥이 보임.",
        "entities": "아시아계 난민 남성. 전술 장갑과 군복을 입은 경비병의 팔. 지시문에 없는 배경 인물 2명(미연 등).",
        "hard_violations": [
         "숏 텍스트에 명시되지 않은 인물(배경의 여성 및 다른 남성) 임의 추가"
        ],
        "physics": "난민 남성은 타격을 받아 뒤로 기울어지는 체중 이동을 보이며 하체로 지탱 중임."
       },
       {
        "label": "B",
        "direction": "두꺼운 장갑을 낀 양손이 난민 남성의 가슴팍을 향해 직접적으로 밀치고 있음.",
        "built_space": "금이 간 콘크리트 방조제 벽이 보이고, 바닥의 파편과 진흙이 일부 보임.",
        "entities": "아시아계 난민 남성. 두꺼운 장갑을 꼈으나 군복이 아닌 회색 양복(관리소장 복장)을 입은 팔.",
        "hard_violations": [
         "난민 남성의 왼쪽 어깨 뒤에 몸통 없이 허공에 떠 있는 여분의 손(물리적으로 불가능한 신체 구조)"
        ],
        "physics": "난민은 밀리는 힘에 의해 뒤로 젖혀져 있으나, 어깨 뒤에 떠 있는 손을 지지하는 신체나 몸통이 전혀 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두꺼운 장갑 두 손이 남성의 가슴팍에 닿고 상체가 밀려나는 순간을 구현했지만, 배경과 허리까지 보여 접촉점 중심의 클로즈업은 다소 느슨합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시되지 않은 배경 인물 두 명이 추가되었고, 아래쪽 장갑이 가슴 대신 난민의 팔을 누르므로 양손으로 가슴팍을 밀치는 핵심 동작도 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽에서 뻗은 두 팔의 장갑이 남성의 윗가슴과 가슴 중앙을 향해 실제로 접촉합니다. 남성은 화면 오른쪽의 밀치는 사람을 바라보며 상체를 왼쪽 뒤로 젖힙니다. 경비병의 눈은 프레임 밖이라 시선은 확인할 수 없습니다.",
        "built_space": "뒤에는 균열과 세로 누수 흔적이 있는 콘크리트 방벽 한 면이 있고, 왼쪽 아래에는 기대어 둔 판재와 배관, 바닥에는 흙탕물 웅덩이와 케이블이 보입니다. 참고 장소의 재료와 낮빛은 이어지지만, 방벽과 작업장 바닥이 비교적 넓고 선명하여 지정된 부드러운 부분 배경보다 존재감이 큽니다. 트럭과 지게차는 프레임 밖이며, 이를 보여주려고 더 넓힐 필요는 없습니다.",
        "entities": "검은 머리의 성인 아시아계 남성 한 명과 밀치는 사람의 두 팔 및 얼굴 극히 일부가 보입니다. 남성은 낡은 갈색 작업 재킷과 줄무늬 속옷을 입어 난민 역할에 부합합니다. 두 장갑은 두껍고 긴 보호용 장갑으로 읽힙니다. 다만 밀치는 사람의 회색 소매는 군복보다 정장에 가까워 경비병 신분 표지는 약합니다. 미연이나 관리소장으로 명확히 식별되는 추가 인물, 확성기, 손가락질, 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 장갑은 각각 소매와 팔에 이어져 있으며 남성의 옷을 누릅니다. 가슴 부근의 눌림과 뒤로 기울어진 상체, 반사적으로 올라간 맨손은 밀침의 순간으로 가능합니다. 하체와 발은 잘려 지면 접촉을 직접 확인할 수 없지만, 몸이 공중에 떠 있는 모습은 아니며 접촉한 팔들이 힘의 출처를 보여줍니다."
       },
       {
        "label": "B",
        "direction": "오른쪽 경비병의 위쪽 손은 남성의 윗가슴을 향하지만 손바닥 전체가 강하게 밀착한 모습은 불명확합니다. 아래쪽 손은 가슴 앞에 가로놓인 남성의 팔을 누릅니다. 따라서 두 손의 힘이 모두 가슴팍에 직접 전달되는 장면은 아닙니다. 남성의 시선은 오른쪽 경비병의 얼굴을 향하며 경비병의 눈은 잘려 있습니다.",
        "built_space": "균열과 누수 얼룩이 있는 방벽 한 면, 왼쪽의 판재와 배관, 붉은 드럼통, 오른쪽 뒤의 주황색 수리 장비, 바닥의 큰 웅덩이가 보입니다. 장소의 기본 재료와 낮빛은 참고와 이어집니다. 그러나 벽 상단과 넓은 바닥까지 드러나 몸통 너머로 벽 일부만 흐리게 남겨야 하는 클로즈업보다 넓습니다. 배경 인물은 왼쪽 가장자리와 난민 바로 뒤에 각각 서 있습니다.",
        "entities": "전경에는 성인 아시아계 남성과 위장복·팔꿈치 보호대·두꺼운 전술 장갑을 착용한 경비병이 있어 주요 역할은 읽힙니다. 남성은 낡고 해진 회갈색 재킷을 입습니다. 그 외에 왼쪽 가장자리의 단발머리 인물 한 명과 남성 뒤에서 머리·몸통 일부 및 다리가 드러나는 인물 한 명이 추가되어 있습니다. 이 두 사람은 해당 숏 텍스트가 요구하지 않습니다. 확성기와 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "숏 텍스트에 없는 배경 인물 두 명을 추가했습니다."
        ],
        "physics": "경비병의 장갑 두 개는 각각 팔과 연결되고 아래쪽 손은 난민의 굽힌 팔에 지지되어 있습니다. 난민은 팔을 가슴 앞에 모으고 상체를 뒤로 기울여 방어적으로 버티는 자세이며, 해부학적으로 불가능하거나 공중에 떠 있는 모습은 아닙니다. 전경 인물들의 발은 프레임 밖이므로 지면 접촉은 확인할 수 없습니다. 다만 옷의 압박과 손의 접촉은 양손 가슴 타격보다 팔을 막거나 제지하는 동작에 가깝습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두꺼운 장갑 두 손이 남성의 가슴팍에 닿고 상체가 밀려나는 순간을 구현했지만, 배경과 허리까지 보여 접촉점 중심의 클로즈업은 다소 느슨합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지시되지 않은 배경 인물 두 명이 추가되었고, 아래쪽 장갑이 가슴 대신 난민의 팔을 누르므로 양손으로 가슴팍을 밀치는 핵심 동작도 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽에서 뻗은 두 팔의 장갑이 남성의 윗가슴과 가슴 중앙을 향해 실제로 접촉합니다. 남성은 화면 오른쪽의 밀치는 사람을 바라보며 상체를 왼쪽 뒤로 젖힙니다. 경비병의 눈은 프레임 밖이라 시선은 확인할 수 없습니다.",
        "built_space": "뒤에는 균열과 세로 누수 흔적이 있는 콘크리트 방벽 한 면이 있고, 왼쪽 아래에는 기대어 둔 판재와 배관, 바닥에는 흙탕물 웅덩이와 케이블이 보입니다. 참고 장소의 재료와 낮빛은 이어지지만, 방벽과 작업장 바닥이 비교적 넓고 선명하여 지정된 부드러운 부분 배경보다 존재감이 큽니다. 트럭과 지게차는 프레임 밖이며, 이를 보여주려고 더 넓힐 필요는 없습니다.",
        "entities": "검은 머리의 성인 아시아계 남성 한 명과 밀치는 사람의 두 팔 및 얼굴 극히 일부가 보입니다. 남성은 낡은 갈색 작업 재킷과 줄무늬 속옷을 입어 난민 역할에 부합합니다. 두 장갑은 두껍고 긴 보호용 장갑으로 읽힙니다. 다만 밀치는 사람의 회색 소매는 군복보다 정장에 가까워 경비병 신분 표지는 약합니다. 미연이나 관리소장으로 명확히 식별되는 추가 인물, 확성기, 손가락질, 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 장갑은 각각 소매와 팔에 이어져 있으며 남성의 옷을 누릅니다. 가슴 부근의 눌림과 뒤로 기울어진 상체, 반사적으로 올라간 맨손은 밀침의 순간으로 가능합니다. 하체와 발은 잘려 지면 접촉을 직접 확인할 수 없지만, 몸이 공중에 떠 있는 모습은 아니며 접촉한 팔들이 힘의 출처를 보여줍니다."
       },
       {
        "label": "A",
        "direction": "오른쪽 경비병의 위쪽 손은 남성의 윗가슴을 향하지만 손바닥 전체가 강하게 밀착한 모습은 불명확합니다. 아래쪽 손은 가슴 앞에 가로놓인 남성의 팔을 누릅니다. 따라서 두 손의 힘이 모두 가슴팍에 직접 전달되는 장면은 아닙니다. 남성의 시선은 오른쪽 경비병의 얼굴을 향하며 경비병의 눈은 잘려 있습니다.",
        "built_space": "균열과 누수 얼룩이 있는 방벽 한 면, 왼쪽의 판재와 배관, 붉은 드럼통, 오른쪽 뒤의 주황색 수리 장비, 바닥의 큰 웅덩이가 보입니다. 장소의 기본 재료와 낮빛은 참고와 이어집니다. 그러나 벽 상단과 넓은 바닥까지 드러나 몸통 너머로 벽 일부만 흐리게 남겨야 하는 클로즈업보다 넓습니다. 배경 인물은 왼쪽 가장자리와 난민 바로 뒤에 각각 서 있습니다.",
        "entities": "전경에는 성인 아시아계 남성과 위장복·팔꿈치 보호대·두꺼운 전술 장갑을 착용한 경비병이 있어 주요 역할은 읽힙니다. 남성은 낡고 해진 회갈색 재킷을 입습니다. 그 외에 왼쪽 가장자리의 단발머리 인물 한 명과 남성 뒤에서 머리·몸통 일부 및 다리가 드러나는 인물 한 명이 추가되어 있습니다. 이 두 사람은 해당 숏 텍스트가 요구하지 않습니다. 확성기와 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "숏 텍스트에 없는 배경 인물 두 명을 추가했습니다."
        ],
        "physics": "경비병의 장갑 두 개는 각각 팔과 연결되고 아래쪽 손은 난민의 굽힌 팔에 지지되어 있습니다. 난민은 팔을 가슴 앞에 모으고 상체를 뒤로 기울여 방어적으로 버티는 자세이며, 해부학적으로 불가능하거나 공중에 떠 있는 모습은 아닙니다. 전경 인물들의 발은 프레임 밖이므로 지면 접촉은 확인할 수 없습니다. 다만 옷의 압박과 손의 접촉은 양손 가슴 타격보다 팔을 막거나 제지하는 동작에 가깝습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.375,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.125,
    "B": 1.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 숏 텍스트에 명시되지 않은 인물(배경의 여성 및 다른 남성) 임의 추가",
     "[gpt-high] 숏 텍스트에 없는 배경 인물 두 명을 추가했습니다."
    ],
    "B": [
     "[gemini-pro] 난민 남성의 왼쪽 어깨 뒤에 몸통 없이 허공에 떠 있는 여분의 손(물리적으로 불가능한 신체 구조)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1125,
   "B": 1500
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "숏 텍스트에 없는 인물들을 배경에 추가하는 치명적 오류를 범했으며, 요구된 클로즈업 대신 너무 넓은 미디엄 숏으로 렌더링되었습니다.  ★위반: [gemini-pro] 숏 텍스트에 명시되지 않은 인물(배경의 여성 및 다른 남성) 임의 추가 / [gpt-high] 숏 텍스트에 없는 배경 인물 두 명을 추가했습니다."
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "난민 어깨 뒤에 몸통 없는 여분의 손이 나타나는 치명적 물리적 오류가 발생했고, 경비병의 팔에 이전 장면의 회색 양복이 잘못 입혀졌습니다.  ★위반: [gemini-pro] 난민 남성의 왼쪽 어깨 뒤에 몸통 없이 허공에 떠 있는 여분의 손(물리적으로 불가능한 신체 구조)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S7sh20_sel.png",
    "asset_id": "713d7291-1da9-4402-a7ed-e91b1e2e3a4a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1368135>",
    "asset_id": "df58dc98-6a17-497c-8753-2337102be459",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd7a-026b-7e05-8d72-3b1e2edecd8e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S7sh20"
  }
 },
 "S7sh22::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:18:08.161247+00:00",
  "fingerprint": "be8141d9317cd21c13c9084f0ddc51d1aa270471add025343da02a5e3cfbf092",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S7sh22_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S7sh22_sel.png",
  "source_sha256": "c785cb5bf2fa6f352054579c34079774752b2156990c7d18a419833dbb253a84",
  "file": "S7sh22_cine.png",
  "staged_sha256": "e33e605550bedb1caa2c20702f6781b6e5e9f807102ffb8fc296131a4e0d0587",
  "latency_ms": 11292
 },
 "S7sh26::signage": {
  "fp": "38cce9be8ad8c94b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S7sh26": {
  "input_fingerprint": "b4af0b85a4a07878",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 군용 트럭이 타이어 뒤로 흙먼지를 거칠게 뿜어내며 뒤쪽 도로를 향해 질주하는 mid-action 후면 외관.\n\nLOCATION (lock): On the outdoor access road leading away from the refugee settlement's seawall repair site. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: roadway departure direction beyond the truck in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Military truck (Accelerating away with 관리소장 inside) — Its rear and a narrow portion of its side face the camera; used as Receding subject whose decreasing size communicates withdrawal; Departure road (Occupied by the departing truck) — Extends from the lower frame toward the upper-center distance; used as Defines the departure vector and leaves the camera visibly outside it; Tire-thrown earth (Thrown backward behind the accelerating tires); used as Local motion evidence immediately behind the vehicle, not generalized haze.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the established daylight treatment, keeping the truck’s retreat readable without adding a new atmospheric lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military truck is departing, leaving the forklift, construction equipment and repair materials at the halted worksite. The seawall's severe cracks and water leakage remain unresolved.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 관리소장 (50대 남성, 대머리, 중년의 얼굴) — wearing: 싸구려 광택이 도는 낡고 통이 넓은 회색 양복과 흙탕물이 튄 구두.; 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 군용 트럭이 타이어 뒤로 흙먼지를 거칠게 뿜어내며 뒤쪽 도로를 향해 질주하는 mid-action 후면 외관.\n\nLOCATION (lock): On the outdoor access road leading away from the refugee settlement's seawall repair site. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: roadway departure direction beyond the truck in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Military truck (Accelerating away with 관리소장 inside) — Its rear and a narrow portion of its side face the camera; used as Receding subject whose decreasing size communicates withdrawal; Departure road (Occupied by the departing truck) — Extends from the lower frame toward the upper-center distance; used as Defines the departure vector and leaves the camera visibly outside it; Tire-thrown earth (Thrown backward behind the accelerating tires); used as Local motion evidence immediately behind the vehicle, not generalized haze.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the established daylight treatment, keeping the truck’s retreat readable without adding a new atmospheric lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military truck is departing, leaving the forklift, construction equipment and repair materials at the halted worksite. The seawall's severe cracks and water leakage remain unresolved.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 관리소장 (50대 남성, 대머리, 중년의 얼굴) — wearing: 싸구려 광택이 도는 낡고 통이 넓은 회색 양복과 흙탕물이 튄 구두.; 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 군용 트럭이 타이어 뒤로 흙먼지를 거칠게 뿜어내며 뒤쪽 도로를 향해 질주하는 mid-action 후면 외관.\n\nLOCATION (lock): On the outdoor access road leading away from the refugee settlement's seawall repair site. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: roadway departure direction beyond the truck in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Military truck (Accelerating away with 관리소장 inside) — Its rear and a narrow portion of its side face the camera; used as Receding subject whose decreasing size communicates withdrawal; Departure road (Occupied by the departing truck) — Extends from the lower frame toward the upper-center distance; used as Defines the departure vector and leaves the camera visibly outside it; Tire-thrown earth (Thrown backward behind the accelerating tires); used as Local motion evidence immediately behind the vehicle, not generalized haze.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the established daylight treatment, keeping the truck’s retreat readable without adding a new atmospheric lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military truck is departing, leaving the forklift, construction equipment and repair materials at the halted worksite. The seawall's severe cracks and water leakage remain unresolved.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 관리소장 (50대 남성, 대머리, 중년의 얼굴) — wearing: 싸구려 광택이 도는 낡고 통이 넓은 회색 양복과 흙탕물이 튄 구두.; 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "트럭이 화면 상단 배경을 향해 주행하며 타이어 뒤로 흙먼지를 뿜어냄. 관리소장은 트럭 내부에서 앞을 향하고 있음.",
    "built_space": "좌측에 금이 간 콘크리트 방조제와 바닥의 웅덩이가 배치됨. 발전기와 지게차가 붉은색 기계 부품과 형태를 알 수 없게 융합되어 있음.",
    "entities": "덮개가 없는 군용 트럭, 트럭 내부에 보이는 관리소장. 프롬프트에 없는 다수의 군중이 배경 게이트에 서 있음.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 없는 군중이 배경에 추가됨 (invented people)",
     "[gemini-pro] 지게차와 발전기가 기이하게 융합되어 물리적으로 불가능한 기계가 됨 (invented objects)",
     "[gpt-high] 출입문 앞에 숏에서 요구하지 않은 군중을 추가했다.",
     "[gpt-high] 남겨진 작업장의 지게차를 두 대로 늘려 별도 장비를 추가했다."
    ],
    "physics": "트럭 바퀴가 지면에 닿아 주행을 지탱하며, 타이어 회전에 의해 흙이 튀어 오름."
   },
   {
    "label": "B",
    "direction": "트럭이 화면 안쪽으로 멀어지며 흙먼지를 일으킴. 미연이 트럭을 향해 손가락으로 가리킴.",
    "built_space": "좌측 방조제와 건설 장비들의 배치는 레퍼런스와 유사하나, 우측 전경이 거대한 피사체로 막혀 있음.",
    "entities": "덮개가 있는 트럭, 중경에 위치한 미연. 방조제 쪽에 지시되지 않은 두 남성. 우측 전경에 미연과 동일한 복장의 거대한 다리.",
    "hard_violations": [
     "[gemini-pro] 미연과 똑같은 하의를 입은 다리가 전경에 중복으로 배치됨 (duplicated bodies)",
     "[gemini-pro] 프롬프트에 없는 두 명의 인물이 배경에 추가됨 (invented people)",
     "[gpt-high] 숏이 요구하지 않은 여성 한 명, 배경 인물 두 명, 전경의 별도 인물 하반신을 추가했다."
    ],
    "physics": "튀어 오른 흙먼지가 공중에 떠 있으나 트럭 뒷바퀴의 지면 접촉 상태가 명확하지 않음. 전경의 다리는 바닥을 딛고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "요구된 후면 외관 및 멀어지는 트럭의 와이드 샷 구도를 잘 구현했으나, 트럭 덮개가 사라지고 지시되지 않은 군중이 배경에 등장한 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "트럭의 외형은 레퍼런스와 비슷하지만, 전경에 신체가 중복된 거대한 다리가 나타나 구도를 완전히 망쳤으며 프롬프트에 없는 인물들이 추가되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "트럭이 화면 상단 배경을 향해 주행하며 타이어 뒤로 흙먼지를 뿜어냄. 관리소장은 트럭 내부에서 앞을 향하고 있음.",
        "built_space": "좌측에 금이 간 콘크리트 방조제와 바닥의 웅덩이가 배치됨. 발전기와 지게차가 붉은색 기계 부품과 형태를 알 수 없게 융합되어 있음.",
        "entities": "덮개가 없는 군용 트럭, 트럭 내부에 보이는 관리소장. 프롬프트에 없는 다수의 군중이 배경 게이트에 서 있음.",
        "hard_violations": [
         "프롬프트에 없는 군중이 배경에 추가됨 (invented people)",
         "지게차와 발전기가 기이하게 융합되어 물리적으로 불가능한 기계가 됨 (invented objects)"
        ],
        "physics": "트럭 바퀴가 지면에 닿아 주행을 지탱하며, 타이어 회전에 의해 흙이 튀어 오름."
       },
       {
        "label": "B",
        "direction": "트럭이 화면 안쪽으로 멀어지며 흙먼지를 일으킴. 미연이 트럭을 향해 손가락으로 가리킴.",
        "built_space": "좌측 방조제와 건설 장비들의 배치는 레퍼런스와 유사하나, 우측 전경이 거대한 피사체로 막혀 있음.",
        "entities": "덮개가 있는 트럭, 중경에 위치한 미연. 방조제 쪽에 지시되지 않은 두 남성. 우측 전경에 미연과 동일한 복장의 거대한 다리.",
        "hard_violations": [
         "미연과 똑같은 하의를 입은 다리가 전경에 중복으로 배치됨 (duplicated bodies)",
         "프롬프트에 없는 두 명의 인물이 배경에 추가됨 (invented people)"
        ],
        "physics": "튀어 오른 흙먼지가 공중에 떠 있으나 트럭 뒷바퀴의 지면 접촉 상태가 명확하지 않음. 전경의 다리는 바닥을 딛고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "요구된 후면 외관 및 멀어지는 트럭의 와이드 샷 구도를 잘 구현했으나, 트럭 덮개가 사라지고 지시되지 않은 군중이 배경에 등장한 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "트럭의 외형은 레퍼런스와 비슷하지만, 전경에 신체가 중복된 거대한 다리가 나타나 구도를 완전히 망쳤으며 프롬프트에 없는 인물들이 추가되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "트럭이 화면 상단 배경을 향해 주행하며 타이어 뒤로 흙먼지를 뿜어냄. 관리소장은 트럭 내부에서 앞을 향하고 있음.",
        "built_space": "좌측에 금이 간 콘크리트 방조제와 바닥의 웅덩이가 배치됨. 발전기와 지게차가 붉은색 기계 부품과 형태를 알 수 없게 융합되어 있음.",
        "entities": "덮개가 없는 군용 트럭, 트럭 내부에 보이는 관리소장. 프롬프트에 없는 다수의 군중이 배경 게이트에 서 있음.",
        "hard_violations": [
         "프롬프트에 없는 군중이 배경에 추가됨 (invented people)",
         "지게차와 발전기가 기이하게 융합되어 물리적으로 불가능한 기계가 됨 (invented objects)"
        ],
        "physics": "트럭 바퀴가 지면에 닿아 주행을 지탱하며, 타이어 회전에 의해 흙이 튀어 오름."
       },
       {
        "label": "B",
        "direction": "트럭이 화면 안쪽으로 멀어지며 흙먼지를 일으킴. 미연이 트럭을 향해 손가락으로 가리킴.",
        "built_space": "좌측 방조제와 건설 장비들의 배치는 레퍼런스와 유사하나, 우측 전경이 거대한 피사체로 막혀 있음.",
        "entities": "덮개가 있는 트럭, 중경에 위치한 미연. 방조제 쪽에 지시되지 않은 두 남성. 우측 전경에 미연과 동일한 복장의 거대한 다리.",
        "hard_violations": [
         "미연과 똑같은 하의를 입은 다리가 전경에 중복으로 배치됨 (duplicated bodies)",
         "프롬프트에 없는 두 명의 인물이 배경에 추가됨 (invented people)"
        ],
        "physics": "튀어 오른 흙먼지가 공중에 떠 있으나 트럭 뒷바퀴의 지면 접촉 상태가 명확하지 않음. 전경의 다리는 바닥을 딛고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "불필요한 인물들과 전경의 거대한 하반신이 추가되어, 멀어지는 군용 트럭 중심의 후면 와이드 숏을 훼손한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "후면과 좁은 측면, 바퀴 뒤로 튀는 흙은 더 명확하지만, 군중과 지게차를 추가하고 기존 트럭의 적재함 천막을 없애 연속성을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "트럭은 후면을 카메라에 보이며 화면 상단 중앙보다 오른쪽의 오르막 도로로 이동한다. 흙과 돌은 차량 뒤쪽인 카메라 방향으로 튄다. 왼쪽 여성은 트럭 쪽을 바라보며 손가락을 오른쪽 위로 뻗지만, 그 연장선은 트럭보다 왼쪽의 먼 방향을 향한다. 뒤의 두 사람은 대체로 차량 쪽을 보고, 전경 인물은 얼굴이 보이지 않는다.",
        "built_space": "왼쪽에 균열과 누수 자국이 있는 콘크리트 방조제 하나, 지게차 한 대, 주황색 프레임의 장비 한 대, 팔레트와 적재 자재가 보인다. 젖은 노면과 물웅덩이는 참고 장소의 재질을 유지한다. 도로는 하단에서 우상단으로 이어지며, 극도로 낮고 기울어진 카메라와 오른쪽 전경의 하반신이 도로 및 퇴장 구도를 크게 가린다.",
        "entities": "천막을 씌운 낡은 군용 트럭 한 대는 참고 차량의 대체적인 유형과 맞는다. 운전실 내부는 보이지 않아 관리소장의 탑승이나 외모는 확인할 수 없다. 왼쪽에는 참고 속 미연과 비슷한 중년 동아시아계 여성과 갈색 작업복이 보이지만, 이 숏에는 요구되지 않은 인물이다. 자재 뒤의 남성으로 보이는 두 사람과 오른쪽 전경의 작업복 하반신도 추가되었다. 판독 가능한 글자는 없다.",
        "hard_violations": [
         "숏이 요구하지 않은 여성 한 명, 배경 인물 두 명, 전경의 별도 인물 하반신을 추가했다."
        ],
        "physics": "트럭은 기울어진 차체 아래의 바퀴로 노면을 지지하는 것으로 보이며, 확실한 무지지 부유로 단정할 근거는 없다. 흙과 돌은 타이어 부근에서 뒤로 분출되어 가속 동작으로 설명된다. 여성과 전경 인물의 부츠는 지면에 닿고, 배경 인물들은 자재 뒤에 서 있다. 장비와 자재는 노면이나 팔레트에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "트럭은 후면과 좁은 오른쪽 측면을 보이며 우상단 출입구 방향으로 멀어진다. 운전실에 보이는 사람도 진행 방향을 향한다. 양쪽 뒤 타이어 부근의 흙은 카메라 쪽으로 튀어 요구된 가속 방향과 맞는다. 다만 출발 도로의 소실 방향은 지정된 상단 중앙보다 상당히 오른쪽이다. 먼 군중의 개별 시선은 판별하기 어렵다.",
        "built_space": "왼쪽 방조제의 큰 균열, 누수 흔적, 주황색 장비 한 대, 팔레트와 보수 자재, 전경 물웅덩이가 참고 장소와 이어진다. 지게차는 왼쪽 전경의 붉은 차체와 마스트 한 대 외에 트럭 왼쪽 뒤에도 별도 마스트와 운전석이 보여 총 두 대다. 오른쪽에는 파란 드럼통 세 개가 보이고, 먼 도로에는 출입문과 작은 건물이 있다. 카메라는 도로 가장자리에서 트럭 뒤를 보는 비교적 명료한 와이드 구도지만 차량이 여전히 화면을 크게 차지한다.",
        "entities": "군용 트럭 한 대와 운전실 안의 회색 옷을 입은 남성 뒷모습이 보인다. 뒷모습만으로 관리소장의 정확한 나이, 얼굴, 대머리 여부는 확정하기 어렵다. 참고 차량의 천막 적재함이 사라져 개방형 적재함으로 바뀌었다. 먼 출입문 앞에는 약 열두 명의 추가 인물이 있으며, 작은 크기 때문에 각각의 나이와 민족적 외모는 확인할 수 없다. 미연은 보이지 않으며, 이는 해당 숏에서 문제가 아니다. 판독 가능한 글자는 없다.",
        "hard_violations": [
         "출입문 앞에 숏에서 요구하지 않은 군중을 추가했다.",
         "남겨진 작업장의 지게차를 두 대로 늘려 별도 장비를 추가했다."
        ],
        "physics": "트럭의 바퀴는 먼지가 이는 노면과 접하는 것으로 읽히며, 차체를 지지한다. 두 갈래의 흙과 작은 돌은 뒤 타이어 바로 뒤에서 튀어 올라 가속에 의한 분출로 설명된다. 운전실 인물은 좌석에 앉아 있고, 배경 군중은 도로 위에 서 있다. 지게차, 장비, 드럼통과 자재도 지면이나 팔레트에 놓여 있어 명백한 무지지 부유는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "불필요한 인물들과 전경의 거대한 하반신이 추가되어, 멀어지는 군용 트럭 중심의 후면 와이드 숏을 훼손한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "후면과 좁은 측면, 바퀴 뒤로 튀는 흙은 더 명확하지만, 군중과 지게차를 추가하고 기존 트럭의 적재함 천막을 없애 연속성을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "트럭은 후면을 카메라에 보이며 화면 상단 중앙보다 오른쪽의 오르막 도로로 이동한다. 흙과 돌은 차량 뒤쪽인 카메라 방향으로 튄다. 왼쪽 여성은 트럭 쪽을 바라보며 손가락을 오른쪽 위로 뻗지만, 그 연장선은 트럭보다 왼쪽의 먼 방향을 향한다. 뒤의 두 사람은 대체로 차량 쪽을 보고, 전경 인물은 얼굴이 보이지 않는다.",
        "built_space": "왼쪽에 균열과 누수 자국이 있는 콘크리트 방조제 하나, 지게차 한 대, 주황색 프레임의 장비 한 대, 팔레트와 적재 자재가 보인다. 젖은 노면과 물웅덩이는 참고 장소의 재질을 유지한다. 도로는 하단에서 우상단으로 이어지며, 극도로 낮고 기울어진 카메라와 오른쪽 전경의 하반신이 도로 및 퇴장 구도를 크게 가린다.",
        "entities": "천막을 씌운 낡은 군용 트럭 한 대는 참고 차량의 대체적인 유형과 맞는다. 운전실 내부는 보이지 않아 관리소장의 탑승이나 외모는 확인할 수 없다. 왼쪽에는 참고 속 미연과 비슷한 중년 동아시아계 여성과 갈색 작업복이 보이지만, 이 숏에는 요구되지 않은 인물이다. 자재 뒤의 남성으로 보이는 두 사람과 오른쪽 전경의 작업복 하반신도 추가되었다. 판독 가능한 글자는 없다.",
        "hard_violations": [
         "숏이 요구하지 않은 여성 한 명, 배경 인물 두 명, 전경의 별도 인물 하반신을 추가했다."
        ],
        "physics": "트럭은 기울어진 차체 아래의 바퀴로 노면을 지지하는 것으로 보이며, 확실한 무지지 부유로 단정할 근거는 없다. 흙과 돌은 타이어 부근에서 뒤로 분출되어 가속 동작으로 설명된다. 여성과 전경 인물의 부츠는 지면에 닿고, 배경 인물들은 자재 뒤에 서 있다. 장비와 자재는 노면이나 팔레트에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "트럭은 후면과 좁은 오른쪽 측면을 보이며 우상단 출입구 방향으로 멀어진다. 운전실에 보이는 사람도 진행 방향을 향한다. 양쪽 뒤 타이어 부근의 흙은 카메라 쪽으로 튀어 요구된 가속 방향과 맞는다. 다만 출발 도로의 소실 방향은 지정된 상단 중앙보다 상당히 오른쪽이다. 먼 군중의 개별 시선은 판별하기 어렵다.",
        "built_space": "왼쪽 방조제의 큰 균열, 누수 흔적, 주황색 장비 한 대, 팔레트와 보수 자재, 전경 물웅덩이가 참고 장소와 이어진다. 지게차는 왼쪽 전경의 붉은 차체와 마스트 한 대 외에 트럭 왼쪽 뒤에도 별도 마스트와 운전석이 보여 총 두 대다. 오른쪽에는 파란 드럼통 세 개가 보이고, 먼 도로에는 출입문과 작은 건물이 있다. 카메라는 도로 가장자리에서 트럭 뒤를 보는 비교적 명료한 와이드 구도지만 차량이 여전히 화면을 크게 차지한다.",
        "entities": "군용 트럭 한 대와 운전실 안의 회색 옷을 입은 남성 뒷모습이 보인다. 뒷모습만으로 관리소장의 정확한 나이, 얼굴, 대머리 여부는 확정하기 어렵다. 참고 차량의 천막 적재함이 사라져 개방형 적재함으로 바뀌었다. 먼 출입문 앞에는 약 열두 명의 추가 인물이 있으며, 작은 크기 때문에 각각의 나이와 민족적 외모는 확인할 수 없다. 미연은 보이지 않으며, 이는 해당 숏에서 문제가 아니다. 판독 가능한 글자는 없다.",
        "hard_violations": [
         "출입문 앞에 숏에서 요구하지 않은 군중을 추가했다.",
         "남겨진 작업장의 지게차를 두 대로 늘려 별도 장비를 추가했다."
        ],
        "physics": "트럭의 바퀴는 먼지가 이는 노면과 접하는 것으로 읽히며, 차체를 지지한다. 두 갈래의 흙과 작은 돌은 뒤 타이어 바로 뒤에서 튀어 올라 가속에 의한 분출로 설명된다. 운전실 인물은 좌석에 앉아 있고, 배경 군중은 도로 위에 서 있다. 지게차, 장비, 드럼통과 자재도 지면이나 팔레트에 놓여 있어 명백한 무지지 부유는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.167
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.917
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 없는 군중이 배경에 추가됨 (invented people)",
     "[gemini-pro] 지게차와 발전기가 기이하게 융합되어 물리적으로 불가능한 기계가 됨 (invented objects)",
     "[gpt-high] 출입문 앞에 숏에서 요구하지 않은 군중을 추가했다.",
     "[gpt-high] 남겨진 작업장의 지게차를 두 대로 늘려 별도 장비를 추가했다."
    ],
    "B": [
     "[gemini-pro] 미연과 똑같은 하의를 입은 다리가 전경에 중복으로 배치됨 (duplicated bodies)",
     "[gemini-pro] 프롬프트에 없는 두 명의 인물이 배경에 추가됨 (invented people)",
     "[gpt-high] 숏이 요구하지 않은 여성 한 명, 배경 인물 두 명, 전경의 별도 인물 하반신을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 917
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "요구된 후면 외관 및 멀어지는 트럭의 와이드 샷 구도를 잘 구현했으나, 트럭 덮개가 사라지고 지시되지 않은 군중이 배경에 등장한 점이 감점 요인입니다.  ★위반: [gemini-pro] 프롬프트에 없는 군중이 배경에 추가됨 (invented people) / [gemini-pro] 지게차와 발전기가 기이하게 융합되어 물리적으로 불가능한 기계가 됨 (invented objects) / [gpt-high] 출입문 앞에 숏에서 요구하지 않은 군중을 추가했다. / [gpt-high] 남겨진 작업장의 지게차를 두 대로 늘려 별도 장비를 추가했다."
   },
   {
    "label": "B",
    "score": 917,
    "verdict_ko": "트럭의 외형은 레퍼런스와 비슷하지만, 전경에 신체가 중복된 거대한 다리가 나타나 구도를 완전히 망쳤으며 프롬프트에 없는 인물들이 추가되었습니다.  ★위반: [gemini-pro] 미연과 똑같은 하의를 입은 다리가 전경에 중복으로 배치됨 (duplicated bodies) / [gemini-pro] 프롬프트에 없는 두 명의 인물이 배경에 추가됨 (invented people) / [gpt-high] 숏이 요구하지 않은 여성 한 명, 배경 인물 두 명, 전경의 별도 인물 하반신을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S7sh20_sel.png",
    "asset_id": "713d7291-1da9-4402-a7ed-e91b1e2e3a4a",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd7a-bc82-7ea9-8ea6-b315bb9c8333",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S7sh20"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S7sh26::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:18:19.790815+00:00",
  "fingerprint": "cb3295261b0d6052528e5cc3b8d69a94d00e6cd8a0ad80148a3df16c02391b6b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S7sh26_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S7sh26_sel.png",
  "source_sha256": "b4d8dce02a3a4970143fbf31fc4cc33e22c43cfa9e0dc82ac80a8898b1e60ea6",
  "file": "S7sh26_cine.png",
  "staged_sha256": "cfea8dd1497443ca1bfcb14c1c80bcf4e90493b1ee45c2683ee0522216649744",
  "latency_ms": 11212
 },
 "S8sh1::signage": {
  "fp": "e22e582e5796c9d9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::33f5d6f202a7cbc6": {
  "subjects": [],
  "subject_text": "난민 재판소 내부\n허름한 건물 안의 소박한 재판 공간. 낡은 탁자와 의자가 놓여 있고 탁자 위에는 서류가 높게 쌓여 있다.",
  "identity": "canonical",
  "scope_id": "L16",
  "scope_role": "location_interior",
  "scope_sha": "107b011f5a8b9ea2"
 },
 "S8sh1::bgfirst_bg": {
  "input_fingerprint": "8ce482b141807b24",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두침침한 재판소 안, 아시아계 경찰 두 명에게 양팔이 결박된 채 문 안쪽을 향해 상체가 기울어져 바닥에 무릎을 대고 있는 이현우의 모습.\n\nLOCATION (lock): Just inside the entrance to a shabby refugee tribunal hearing room, in dim interior light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Courtroom entrance (Being traversed by the officers and 이현우) — The inside edge of the entrance appears obliquely at the left margin; used as Marks the entry boundary without adding restraints or blocking the escort; Courtroom floor (Supporting 이현우's knees as he is dragged inward); used as Makes his failure to support his own weight visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the described dim courtroom illumination with sufficient separation to read the arm holds and buckling posture, without specifying an unsupported fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두침침한 재판소 안, 아시아계 경찰 두 명에게 양팔이 결박된 채 문 안쪽을 향해 상체가 기울어져 바닥에 무릎을 대고 있는 이현우의 모습.\n\nLOCATION (lock): Just inside the entrance to a shabby refugee tribunal hearing room, in dim interior light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Courtroom entrance (Being traversed by the officers and 이현우) — The inside edge of the entrance appears obliquely at the left margin; used as Marks the entry boundary without adding restraints or blocking the escort; Courtroom floor (Supporting 이현우's knees as he is dragged inward); used as Makes his failure to support his own weight visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the described dim courtroom illumination with sufficient separation to read the arm holds and buckling posture, without specifying an unsupported fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S8sh1__bgfirst_bg.png",
  "asset_id": "f1cd8ccb-77d6-4fc1-8652-4871c621273a",
  "input_asset_ids": [
   "b648cc69-3d30-4c34-98e9-81184f6f43eb",
   "1e301680-821e-47d5-b2a7-847a335a9307"
  ]
 },
 "S8sh1": {
  "input_fingerprint": "6400e8b49390eea6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두침침한 재판소 안, 아시아계 경찰 두 명에게 양팔이 결박된 채 문 안쪽을 향해 상체가 기울어져 바닥에 무릎을 대고 있는 이현우의 모습.\n\nLOCATION (lock): Just inside the entrance to a shabby refugee tribunal hearing room, in dim interior light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Courtroom entrance (Being traversed by the officers and 이현우) — The inside edge of the entrance appears obliquely at the left margin; used as Marks the entry boundary without adding restraints or blocking the escort; Courtroom floor (Supporting 이현우's knees as he is dragged inward); used as Makes his failure to support his own weight visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the described dim courtroom illumination with sufficient separation to read the arm holds and buckling posture, without specifying an unsupported fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shabby courtroom contains a worn table, a chair and accumulated paperwork. 이현우: Hyunwoo enters with his arms held and his legs unable to support him reliably. His dog-bitten leg remains injured, and he is already feverish and dazed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두침침한 재판소 안, 아시아계 경찰 두 명에게 양팔이 결박된 채 문 안쪽을 향해 상체가 기울어져 바닥에 무릎을 대고 있는 이현우의 모습.\n\nLOCATION (lock): Just inside the entrance to a shabby refugee tribunal hearing room, in dim interior light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Courtroom entrance (Being traversed by the officers and 이현우) — The inside edge of the entrance appears obliquely at the left margin; used as Marks the entry boundary without adding restraints or blocking the escort; Courtroom floor (Supporting 이현우's knees as he is dragged inward); used as Makes his failure to support his own weight visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the described dim courtroom illumination with sufficient separation to read the arm holds and buckling posture, without specifying an unsupported fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shabby courtroom contains a worn table, a chair and accumulated paperwork. 이현우: Hyunwoo enters with his arms held and his legs unable to support him reliably. His dog-bitten leg remains injured, and he is already feverish and dazed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두침침한 재판소 안, 아시아계 경찰 두 명에게 양팔이 결박된 채 문 안쪽을 향해 상체가 기울어져 바닥에 무릎을 대고 있는 이현우의 모습.\n\nLOCATION (lock): Just inside the entrance to a shabby refugee tribunal hearing room, in dim interior light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Courtroom entrance (Being traversed by the officers and 이현우) — The inside edge of the entrance appears obliquely at the left margin; used as Marks the entry boundary without adding restraints or blocking the escort; Courtroom floor (Supporting 이현우's knees as he is dragged inward); used as Makes his failure to support his own weight visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the described dim courtroom illumination with sufficient separation to read the arm holds and buckling posture, without specifying an unsupported fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shabby courtroom contains a worn table, a chair and accumulated paperwork. 이현우: Hyunwoo enters with his arms held and his legs unable to support him reliably. His dog-bitten leg remains injured, and he is already feverish and dazed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S8sh1__bgfirst_bg.png",
     "asset_id": "f1cd8ccb-77d6-4fc1-8652-4871c621273a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S8sh1.png",
     "asset_id": "b648cc69-3d30-4c34-98e9-81184f6f43eb",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L16B01.png",
     "asset_id": "1e301680-821e-47d5-b2a7-847a335a9307",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 2,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "현우는 방 안쪽으로 상체가 기울어져 있고, 경찰들은 그를 이끌며 시선과 몸을 안쪽으로 향하고 있다.",
    "built_space": "참조 사진의 원목 문(전경 좌측), 중앙 책상, 국기, 우측 캐비닛과 의자 배치가 정확히 일치하며 인물들의 동선과 자연스럽게 어우러진다.",
    "entities": "검은 후드와 찢어진 청바지를 입은 이현우와 제복을 입은 두 명의 아시아계 경찰이 정확히 등장한다.",
    "hard_violations": [],
    "physics": "현우의 두 무릎이 바닥에 닿아 하중을 받으며, 두 경찰의 손이 현우의 팔을 단단히 붙잡아 지탱하고 있다."
   },
   {
    "label": "B",
    "direction": "현우는 무릎을 꿇은 채 뒤로 젖혀져 있고, 경찰들은 양옆에서 그를 바라보며 지탱하고 있다.",
    "built_space": "법정 내부는 유사하나 전경 좌측 문에 참조에 없는 유리창이 생겼으며 벽면 구조도 변형되었다.",
    "entities": "이현우와 두 명의 아시아계 경찰이 등장하나, 경찰들이 지시되지 않은 마스크를 착용하고 있다.",
    "hard_violations": [
     "[gemini-pro] 고정된 장소 참조 이미지의 물리적 구조(좌측 문에 유리창 추가)를 임의로 왜곡 및 변형함",
     "[gpt-high] 뒤쪽 법원 표장 중앙에 판독 가능한 한글이 노출되어, 어디에도 읽히는 글자가 없어야 한다는 조건을 위반한다."
    ],
    "physics": "현우는 무릎으로 바닥을 딛고 경찰들이 양팔을 붙잡아 하중을 분산시키고 있어 지지 상태는 정상적이다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 앵글과 공간, 인물의 결박 및 무릎 꿇은 자세를 완벽히 구현했으며 참조 이미지의 구조를 충실히 따랐습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "장소 참조 이미지에 없는 문 유리창을 추가하는 등 공간 구조를 왜곡했고, 양팔이 결박된 상태도 제대로 표현되지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 방 안쪽으로 상체가 기울어져 있고, 경찰들은 그를 이끌며 시선과 몸을 안쪽으로 향하고 있다.",
        "built_space": "참조 사진의 원목 문(전경 좌측), 중앙 책상, 국기, 우측 캐비닛과 의자 배치가 정확히 일치하며 인물들의 동선과 자연스럽게 어우러진다.",
        "entities": "검은 후드와 찢어진 청바지를 입은 이현우와 제복을 입은 두 명의 아시아계 경찰이 정확히 등장한다.",
        "hard_violations": [],
        "physics": "현우의 두 무릎이 바닥에 닿아 하중을 받으며, 두 경찰의 손이 현우의 팔을 단단히 붙잡아 지탱하고 있다."
       },
       {
        "label": "B",
        "direction": "현우는 무릎을 꿇은 채 뒤로 젖혀져 있고, 경찰들은 양옆에서 그를 바라보며 지탱하고 있다.",
        "built_space": "법정 내부는 유사하나 전경 좌측 문에 참조에 없는 유리창이 생겼으며 벽면 구조도 변형되었다.",
        "entities": "이현우와 두 명의 아시아계 경찰이 등장하나, 경찰들이 지시되지 않은 마스크를 착용하고 있다.",
        "hard_violations": [
         "고정된 장소 참조 이미지의 물리적 구조(좌측 문에 유리창 추가)를 임의로 왜곡 및 변형함"
        ],
        "physics": "현우는 무릎으로 바닥을 딛고 경찰들이 양팔을 붙잡아 하중을 분산시키고 있어 지지 상태는 정상적이다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 앵글과 공간, 인물의 결박 및 무릎 꿇은 자세를 완벽히 구현했으며 참조 이미지의 구조를 충실히 따랐습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "장소 참조 이미지에 없는 문 유리창을 추가하는 등 공간 구조를 왜곡했고, 양팔이 결박된 상태도 제대로 표현되지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "현우는 방 안쪽으로 상체가 기울어져 있고, 경찰들은 그를 이끌며 시선과 몸을 안쪽으로 향하고 있다.",
        "built_space": "참조 사진의 원목 문(전경 좌측), 중앙 책상, 국기, 우측 캐비닛과 의자 배치가 정확히 일치하며 인물들의 동선과 자연스럽게 어우러진다.",
        "entities": "검은 후드와 찢어진 청바지를 입은 이현우와 제복을 입은 두 명의 아시아계 경찰이 정확히 등장한다.",
        "hard_violations": [],
        "physics": "현우의 두 무릎이 바닥에 닿아 하중을 받으며, 두 경찰의 손이 현우의 팔을 단단히 붙잡아 지탱하고 있다."
       },
       {
        "label": "B",
        "direction": "현우는 무릎을 꿇은 채 뒤로 젖혀져 있고, 경찰들은 양옆에서 그를 바라보며 지탱하고 있다.",
        "built_space": "법정 내부는 유사하나 전경 좌측 문에 참조에 없는 유리창이 생겼으며 벽면 구조도 변형되었다.",
        "entities": "이현우와 두 명의 아시아계 경찰이 등장하나, 경찰들이 지시되지 않은 마스크를 착용하고 있다.",
        "hard_violations": [
         "고정된 장소 참조 이미지의 물리적 구조(좌측 문에 유리창 추가)를 임의로 왜곡 및 변형함"
        ],
        "physics": "현우는 무릎으로 바닥을 딛고 경찰들이 양팔을 붙잡아 하중을 분산시키고 있어 지지 상태는 정상적이다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "양팔을 붙든 호송과 무릎의 바닥 지지는 보이지만 상체가 실내 쪽으로 기울지 않고, 법원 표장의 읽히는 글자가 무문자 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "참조 재판소의 구조와 양무릎이 꺾인 호송 순간을 더 충실히 구현하지만, 몸이 기우는 방향은 실내가 아니라 카메라와 출입구 쪽이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 고개를 숙이고 화면 왼쪽 아래를 바라보며, 두 경찰은 현우의 팔과 몸을 내려다본다. 경찰들이 현우의 양쪽 위팔을 잡고 있지만 현우의 몸통은 비교적 세워져 있고 카메라 쪽으로 열려 있다. 왼쪽 전경 출입문에서 뒤쪽 법정 내부로 끌려 들어가는 방향은 드러나지 않는다. 겨누는 무기나 물체는 없다.",
        "built_space": "왼쪽에 열린 문 한 짝과 손잡이가 크게 보이고 오른쪽 가장자리에도 문틀이 있어 출입구 경계는 읽힌다. 뒤에는 긴 목제 재판대 하나, 서류 더미, 태극기 하나, 벽면 표장 하나, 창 하나와 금속 수납장 하나가 보이며 왼쪽에는 작은 목제 가구 일부가 있다. 참조의 낡은 벽과 바닥 재질은 유사하지만, 참조에서 재판대 오른쪽에 있던 뒷문과 대기 의자 배치는 이 화면에서 확인되지 않는다. 세 사람의 전신과 바닥을 포함하지만 문짝이 화면 왼쪽을 넓게 차지하여 요구된 가장자리 경계보다 두드러진다.",
        "entities": "인물은 정확히 세 명으로, 젊은 동아시아계 남성 현우와 성인 동아시아계 남성 경찰 두 명이다. 현우의 헝클어진 검은 머리, 마른 체격, 낡은 짙은 후드와 찢어진 청바지, 운동화는 참조에 가깝다. 눈을 반쯤 감은 표정은 쇠약하고 몽롱하게 읽히지만, 노출된 무릎만으로 개에게 물린 상처를 확인하기는 어렵다. 경찰들은 회색 긴팔 제복, 방호조끼와 마스크를 착용한다. 뒤쪽 법원 표장 중앙에는 읽을 수 있는 한글이 남아 있다.",
        "hard_violations": [
         "뒤쪽 법원 표장 중앙에 판독 가능한 한글이 노출되어, 어디에도 읽히는 글자가 없어야 한다는 조건을 위반한다."
        ],
        "physics": "현우는 화면 왼쪽 무릎과 접힌 다리를 바닥에 대고 있으며 반대쪽 다리는 옆으로 벌어져 있다. 양쪽 경찰의 손이 위팔을 붙들어 상체 하중을 나눠 받는다. 경찰들은 신발을 바닥에 딛고 있어 지지 관계는 성립한다. 지지 없이 떠 있는 신체나 물체는 없지만, 몸통이 상당히 세워져 있어 안쪽으로 끌려가며 앞으로 무너지는 동작은 약하다."
       },
       {
        "label": "B",
        "direction": "현우는 머리와 가슴을 카메라 앞쪽 바닥으로 숙인다. 왼쪽 경찰은 붙잡은 팔을 내려다보고, 오른쪽 경찰은 뒤쪽 오른편의 닫힌 문 방향을 돌아본다. 현우와 왼쪽 경찰의 진행 자세는 왼쪽 전경 출입구와 카메라 쪽을 향하므로, 법정 안쪽으로 끌려 들어가는 요구와 반대로 읽힌다. 겨누는 물체는 없다.",
        "built_space": "열린 출입문의 안쪽 면과 손잡이가 왼쪽 가장자리에 비스듬히 잡히고, 세 사람은 그 바로 안쪽 바닥에 배치되어 있다. 뒤에는 긴 목제 재판대 하나와 검은 재판장 의자 하나, 서류 더미, 태극기 하나, 벽면 표장 하나, 왼쪽 창 하나, 오른쪽 뒷문 하나가 보인다. 왼쪽 앞의 작은 목제 책상과 의자, 오른쪽 금속 수납장들과 최소 두 개의 검은 대기 의자도 참조의 배치를 잘 따른다. 인물 전신과 넓은 바닥을 담는 와이드 구도이며, 가구 크기도 인물과 자연스럽게 맞는다.",
        "entities": "현우와 아시아계 남성 경찰 두 명만 등장한다. 현우는 검은 머리의 젊고 마른 동아시아계 남성이며, 얼굴은 고개를 숙여 대부분 가려져 정확한 얼굴 일치는 확인하기 어렵다. 낡은 짙은 후드, 청바지와 찢어진 무릎은 참조에 부합한다. 숙인 머리와 처진 몸은 쇠약함을 전달하지만 발열이나 개에게 물린 상처는 확정할 수 없다. 경찰들은 회색 반팔 제복을 입고 있다. 서류와 벽면 게시물은 존재하지만 문장은 판독되지 않고, 법원 표장 중앙도 글자 없이 처리되어 있다.",
        "hard_violations": [],
        "physics": "현우의 양무릎이 바닥에 닿고, 앞으로 쏠린 상체는 양쪽 위팔을 잡은 경찰들의 손으로 지지된다. 왼쪽 경찰은 앞발을 바닥에 디딘 넓은 보폭이며, 오른쪽 경찰은 앞발 뒤꿈치가 바닥에 닿고 발끝이 들린 보행 순간으로 읽힌다. 접힌 현우의 다리와 경찰의 팔에 하중이 전달되는 관계는 가능하며, 지지 없이 떠 있는 인물은 보이지 않는다. 다만 이 보행과 견인 방향은 요구된 실내 진입 방향과 맞지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "양팔을 붙든 호송과 무릎의 바닥 지지는 보이지만 상체가 실내 쪽으로 기울지 않고, 법원 표장의 읽히는 글자가 무문자 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "참조 재판소의 구조와 양무릎이 꺾인 호송 순간을 더 충실히 구현하지만, 몸이 기우는 방향은 실내가 아니라 카메라와 출입구 쪽이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 고개를 숙이고 화면 왼쪽 아래를 바라보며, 두 경찰은 현우의 팔과 몸을 내려다본다. 경찰들이 현우의 양쪽 위팔을 잡고 있지만 현우의 몸통은 비교적 세워져 있고 카메라 쪽으로 열려 있다. 왼쪽 전경 출입문에서 뒤쪽 법정 내부로 끌려 들어가는 방향은 드러나지 않는다. 겨누는 무기나 물체는 없다.",
        "built_space": "왼쪽에 열린 문 한 짝과 손잡이가 크게 보이고 오른쪽 가장자리에도 문틀이 있어 출입구 경계는 읽힌다. 뒤에는 긴 목제 재판대 하나, 서류 더미, 태극기 하나, 벽면 표장 하나, 창 하나와 금속 수납장 하나가 보이며 왼쪽에는 작은 목제 가구 일부가 있다. 참조의 낡은 벽과 바닥 재질은 유사하지만, 참조에서 재판대 오른쪽에 있던 뒷문과 대기 의자 배치는 이 화면에서 확인되지 않는다. 세 사람의 전신과 바닥을 포함하지만 문짝이 화면 왼쪽을 넓게 차지하여 요구된 가장자리 경계보다 두드러진다.",
        "entities": "인물은 정확히 세 명으로, 젊은 동아시아계 남성 현우와 성인 동아시아계 남성 경찰 두 명이다. 현우의 헝클어진 검은 머리, 마른 체격, 낡은 짙은 후드와 찢어진 청바지, 운동화는 참조에 가깝다. 눈을 반쯤 감은 표정은 쇠약하고 몽롱하게 읽히지만, 노출된 무릎만으로 개에게 물린 상처를 확인하기는 어렵다. 경찰들은 회색 긴팔 제복, 방호조끼와 마스크를 착용한다. 뒤쪽 법원 표장 중앙에는 읽을 수 있는 한글이 남아 있다.",
        "hard_violations": [
         "뒤쪽 법원 표장 중앙에 판독 가능한 한글이 노출되어, 어디에도 읽히는 글자가 없어야 한다는 조건을 위반한다."
        ],
        "physics": "현우는 화면 왼쪽 무릎과 접힌 다리를 바닥에 대고 있으며 반대쪽 다리는 옆으로 벌어져 있다. 양쪽 경찰의 손이 위팔을 붙들어 상체 하중을 나눠 받는다. 경찰들은 신발을 바닥에 딛고 있어 지지 관계는 성립한다. 지지 없이 떠 있는 신체나 물체는 없지만, 몸통이 상당히 세워져 있어 안쪽으로 끌려가며 앞으로 무너지는 동작은 약하다."
       },
       {
        "label": "A",
        "direction": "현우는 머리와 가슴을 카메라 앞쪽 바닥으로 숙인다. 왼쪽 경찰은 붙잡은 팔을 내려다보고, 오른쪽 경찰은 뒤쪽 오른편의 닫힌 문 방향을 돌아본다. 현우와 왼쪽 경찰의 진행 자세는 왼쪽 전경 출입구와 카메라 쪽을 향하므로, 법정 안쪽으로 끌려 들어가는 요구와 반대로 읽힌다. 겨누는 물체는 없다.",
        "built_space": "열린 출입문의 안쪽 면과 손잡이가 왼쪽 가장자리에 비스듬히 잡히고, 세 사람은 그 바로 안쪽 바닥에 배치되어 있다. 뒤에는 긴 목제 재판대 하나와 검은 재판장 의자 하나, 서류 더미, 태극기 하나, 벽면 표장 하나, 왼쪽 창 하나, 오른쪽 뒷문 하나가 보인다. 왼쪽 앞의 작은 목제 책상과 의자, 오른쪽 금속 수납장들과 최소 두 개의 검은 대기 의자도 참조의 배치를 잘 따른다. 인물 전신과 넓은 바닥을 담는 와이드 구도이며, 가구 크기도 인물과 자연스럽게 맞는다.",
        "entities": "현우와 아시아계 남성 경찰 두 명만 등장한다. 현우는 검은 머리의 젊고 마른 동아시아계 남성이며, 얼굴은 고개를 숙여 대부분 가려져 정확한 얼굴 일치는 확인하기 어렵다. 낡은 짙은 후드, 청바지와 찢어진 무릎은 참조에 부합한다. 숙인 머리와 처진 몸은 쇠약함을 전달하지만 발열이나 개에게 물린 상처는 확정할 수 없다. 경찰들은 회색 반팔 제복을 입고 있다. 서류와 벽면 게시물은 존재하지만 문장은 판독되지 않고, 법원 표장 중앙도 글자 없이 처리되어 있다.",
        "hard_violations": [],
        "physics": "현우의 양무릎이 바닥에 닿고, 앞으로 쏠린 상체는 양쪽 위팔을 잡은 경찰들의 손으로 지지된다. 왼쪽 경찰은 앞발을 바닥에 디딘 넓은 보폭이며, 오른쪽 경찰은 앞발 뒤꿈치가 바닥에 닿고 발끝이 들린 보행 순간으로 읽힌다. 접힌 현우의 다리와 경찰의 팔에 하중이 전달되는 관계는 가능하며, 지지 없이 떠 있는 인물은 보이지 않는다. 다만 이 보행과 견인 방향은 요구된 실내 진입 방향과 맞지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.0
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.75
   },
   "violations": {
    "B": [
     "[gemini-pro] 고정된 장소 참조 이미지의 물리적 구조(좌측 문에 유리창 추가)를 임의로 왜곡 및 변형함",
     "[gpt-high] 뒤쪽 법원 표장 중앙에 판독 가능한 한글이 노출되어, 어디에도 읽히는 글자가 없어야 한다는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 앵글과 공간, 인물의 결박 및 무릎 꿇은 자세를 완벽히 구현했으며 참조 이미지의 구조를 충실히 따랐습니다."
   },
   {
    "label": "B",
    "score": 750,
    "verdict_ko": "장소 참조 이미지에 없는 문 유리창을 추가하는 등 공간 구조를 왜곡했고, 양팔이 결박된 상태도 제대로 표현되지 않았습니다.  ★위반: [gemini-pro] 고정된 장소 참조 이미지의 물리적 구조(좌측 문에 유리창 추가)를 임의로 왜곡 및 변형함 / [gpt-high] 뒤쪽 법원 표장 중앙에 판독 가능한 한글이 노출되어, 어디에도 읽히는 글자가 없어야 한다는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L16B01.png",
    "asset_id": "1e301680-821e-47d5-b2a7-847a335a9307",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd7b-7543-76eb-80e1-beb864a8bd5e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S8sh1__bgfirst_bg.png",
   "bg_asset_id": "f1cd8ccb-77d6-4fc1-8652-4871c621273a",
   "bg_record_key": "S8sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S8sh14::signage": {
  "fp": "936305379e17adce",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S8sh14": {
  "input_fingerprint": "9231011b74515be8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 재판관의 얼굴 위로 바짝 엎드린 채 목에 핏대를 세우며 포효하는 광기 어린 표정.\n\nLOCATION (lock): On the floor beside the judge's shabby table and fallen chair inside the dim refugee tribunal hearing room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Shabby courtroom table (Remaining beside the floor confrontation) — A partial end and underside are visible at the far-left edge; used as Anchors the low viewpoint without enlarging a table component in the foreground; Courtroom floor (Supporting the fallen judge); used as Provides a narrow contextual strip beneath the foreground shoulder.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the courtroom’s subdued ambient exposure and shape readable facial contrast without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The worn courtroom table and its accumulated documents remain in place; the judge's chair is now unoccupied. 이현우: Hyunwoo remains feverish with an injured dog-bitten leg, despite his sudden burst of strength. He is bent low with his hands locked in a collar-gripping position. 재판관: The judge is on the floor after falling from his chair, frightened and struggling.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 재판관 (성인 얼굴, 얼굴형 미상, 머리 모양 미상) — wearing: 오랜 사용으로 색이 바래고 구김이 간 검은색 법관용 가운. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 재판관의 얼굴 위로 바짝 엎드린 채 목에 핏대를 세우며 포효하는 광기 어린 표정.\n\nLOCATION (lock): On the floor beside the judge's shabby table and fallen chair inside the dim refugee tribunal hearing room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Shabby courtroom table (Remaining beside the floor confrontation) — A partial end and underside are visible at the far-left edge; used as Anchors the low viewpoint without enlarging a table component in the foreground; Courtroom floor (Supporting the fallen judge); used as Provides a narrow contextual strip beneath the foreground shoulder.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the courtroom’s subdued ambient exposure and shape readable facial contrast without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The worn courtroom table and its accumulated documents remain in place; the judge's chair is now unoccupied. 이현우: Hyunwoo remains feverish with an injured dog-bitten leg, despite his sudden burst of strength. He is bent low with his hands locked in a collar-gripping position. 재판관: The judge is on the floor after falling from his chair, frightened and struggling.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 재판관 (성인 얼굴, 얼굴형 미상, 머리 모양 미상) — wearing: 오랜 사용으로 색이 바래고 구김이 간 검은색 법관용 가운. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 재판관의 얼굴 위로 바짝 엎드린 채 목에 핏대를 세우며 포효하는 광기 어린 표정.\n\nLOCATION (lock): On the floor beside the judge's shabby table and fallen chair inside the dim refugee tribunal hearing room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Shabby courtroom table (Remaining beside the floor confrontation) — A partial end and underside are visible at the far-left edge; used as Anchors the low viewpoint without enlarging a table component in the foreground; Courtroom floor (Supporting the fallen judge); used as Provides a narrow contextual strip beneath the foreground shoulder.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the courtroom’s subdued ambient exposure and shape readable facial contrast without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The worn courtroom table and its accumulated documents remain in place; the judge's chair is now unoccupied. 이현우: Hyunwoo remains feverish with an injured dog-bitten leg, despite his sudden burst of strength. He is bent low with his hands locked in a collar-gripping position. 재판관: The judge is on the floor after falling from his chair, frightened and struggling.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 재판관 (성인 얼굴, 얼굴형 미상, 머리 모양 미상) — wearing: 오랜 사용으로 색이 바래고 구김이 간 검은색 법관용 가운. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선과 입이 재판관의 얼굴을 직접 향하고 있으며, 재판관은 두려운 표정으로 이현우를 올려다보고 있습니다.",
    "built_space": "프레임 좌측 가장자리에 낡은 테이블의 측면과 하단이 배치되어 있으며, 인물들은 법정 바닥에 위치해 구도와 잘 어울립니다.",
    "entities": "이현우는 레퍼런스와 일치하는 후드티와 청바지, 헝클어진 머리를 하고 있으며, 목에 핏대를 세우고 포효하고 있습니다. 재판관은 낡은 검은색 법복을 입고 있습니다.",
    "hard_violations": [],
    "physics": "재판관은 바닥에 완전히 누워 지탱되고 있으며, 이현우는 바닥에 닿은 하체와 재판관의 멱살을 쥔 두 손으로 몸의 무게를 자연스럽게 지탱하고 있습니다."
   },
   {
    "label": "B",
    "direction": "이현우가 재판관의 얼굴 쪽으로 고개를 숙여 시선을 맞추고 있으며, 재판관 역시 위를 향해 시선을 고정하고 있습니다.",
    "built_space": "좌측에 테이블의 다리와 측면이 보이고, 두 인물은 지시된 대로 바닥에 자리 잡고 있습니다.",
    "entities": "이현우의 복장과 외양은 레퍼런스와 일치하나, 입을 벌린 정도가 작아 완전한 포효의 느낌이 부족합니다. 재판관은 지시된 법복을 입고 있습니다.",
    "hard_violations": [],
    "physics": "재판관의 몸은 바닥에 안정적으로 닿아 있으며, 이현우는 바닥과 재판관의 옷깃을 쥔 한쪽 손을 통해 물리적으로 올바른 자세를 유지하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "이현우가 입을 크게 벌리고 목에 핏대를 세우며 포효하는 광기 어린 표정을 텍스트의 묘사에 맞게 가장 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 인물의 외형 조건은 잘 충족했으나, 이현우의 표정이 '포효'라기보다는 단순히 이를 악물고 화를 내는 모습에 가까워 묘사가 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 입이 재판관의 얼굴을 직접 향하고 있으며, 재판관은 두려운 표정으로 이현우를 올려다보고 있습니다.",
        "built_space": "프레임 좌측 가장자리에 낡은 테이블의 측면과 하단이 배치되어 있으며, 인물들은 법정 바닥에 위치해 구도와 잘 어울립니다.",
        "entities": "이현우는 레퍼런스와 일치하는 후드티와 청바지, 헝클어진 머리를 하고 있으며, 목에 핏대를 세우고 포효하고 있습니다. 재판관은 낡은 검은색 법복을 입고 있습니다.",
        "hard_violations": [],
        "physics": "재판관은 바닥에 완전히 누워 지탱되고 있으며, 이현우는 바닥에 닿은 하체와 재판관의 멱살을 쥔 두 손으로 몸의 무게를 자연스럽게 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우가 재판관의 얼굴 쪽으로 고개를 숙여 시선을 맞추고 있으며, 재판관 역시 위를 향해 시선을 고정하고 있습니다.",
        "built_space": "좌측에 테이블의 다리와 측면이 보이고, 두 인물은 지시된 대로 바닥에 자리 잡고 있습니다.",
        "entities": "이현우의 복장과 외양은 레퍼런스와 일치하나, 입을 벌린 정도가 작아 완전한 포효의 느낌이 부족합니다. 재판관은 지시된 법복을 입고 있습니다.",
        "hard_violations": [],
        "physics": "재판관의 몸은 바닥에 안정적으로 닿아 있으며, 이현우는 바닥과 재판관의 옷깃을 쥔 한쪽 손을 통해 물리적으로 올바른 자세를 유지하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "이현우가 입을 크게 벌리고 목에 핏대를 세우며 포효하는 광기 어린 표정을 텍스트의 묘사에 맞게 가장 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 인물의 외형 조건은 잘 충족했으나, 이현우의 표정이 '포효'라기보다는 단순히 이를 악물고 화를 내는 모습에 가까워 묘사가 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 입이 재판관의 얼굴을 직접 향하고 있으며, 재판관은 두려운 표정으로 이현우를 올려다보고 있습니다.",
        "built_space": "프레임 좌측 가장자리에 낡은 테이블의 측면과 하단이 배치되어 있으며, 인물들은 법정 바닥에 위치해 구도와 잘 어울립니다.",
        "entities": "이현우는 레퍼런스와 일치하는 후드티와 청바지, 헝클어진 머리를 하고 있으며, 목에 핏대를 세우고 포효하고 있습니다. 재판관은 낡은 검은색 법복을 입고 있습니다.",
        "hard_violations": [],
        "physics": "재판관은 바닥에 완전히 누워 지탱되고 있으며, 이현우는 바닥에 닿은 하체와 재판관의 멱살을 쥔 두 손으로 몸의 무게를 자연스럽게 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우가 재판관의 얼굴 쪽으로 고개를 숙여 시선을 맞추고 있으며, 재판관 역시 위를 향해 시선을 고정하고 있습니다.",
        "built_space": "좌측에 테이블의 다리와 측면이 보이고, 두 인물은 지시된 대로 바닥에 자리 잡고 있습니다.",
        "entities": "이현우의 복장과 외양은 레퍼런스와 일치하나, 입을 벌린 정도가 작아 완전한 포효의 느낌이 부족합니다. 재판관은 지시된 법복을 입고 있습니다.",
        "hard_violations": [],
        "physics": "재판관의 몸은 바닥에 안정적으로 닿아 있으며, 이현우는 바닥과 재판관의 옷깃을 쥔 한쪽 손을 통해 물리적으로 올바른 자세를 유지하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "재판관 얼굴 바로 위로 낮게 엎드린 밀착 구도는 더 충실하지만, 포효가 이를 악문 표정에 가깝고 탁자가 왼쪽 전경을 지나치게 차지한다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "벌어진 입과 목의 핏대는 포효를 명확히 구현하지만, 상체가 더 세워지고 허벅지와 바닥까지 드러나 얼굴 위에 바짝 엎드린 클로즈업의 집중도가 떨어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우의 얼굴과 눈은 바로 아래 재판관 얼굴을 향하고, 재판관도 현우를 올려다본다. 양손은 재판관의 옷깃을 향해 모여 있다. 공격과 시선의 대상이 서로 분명하다.",
        "built_space": "왼쪽에 탁자 한 개의 상판 끝, 하부 틀, 앞다리와 뒤쪽 지지부가 보인다. 탁자가 화면 왼쪽 약 3분의 1을 차지해 맨 왼쪽 가장자리에 일부만 보여야 한다는 지시보다 크다. 재판관은 탁자 옆 바닥에 누워 있고 현우는 오른쪽에서 그 위로 숙인다. 뒤쪽 목재 가구는 흐려 넘어진 의자인지 확인하기 어렵다. 바닥도 전경 어깨 아래의 좁은 띠보다 넓게 보인다. 반사는 없다.",
        "entities": "두 남성만 보인다. 현우는 젊은 동아시아계 남성의 얼굴, 헝클어진 검은 머리, 오염되고 해진 짙은 후드 상의로 참조의 주요 특징을 따른다. 다만 얼굴에는 참조보다 성숙한 인상이 있다. 재판관은 성인 남성이며 구김 많은 검은 가운을 입고 놀란 표정을 짓는다. 현우의 목 힘줄은 드러나지만 입은 크게 벌린 포효보다 이를 드러낸 으르렁거림에 가깝다. 다리 부상과 탁자 위 서류는 이 구도에서 확인할 수 없으며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "재판관의 등과 어깨는 바닥에 닿고, 들린 머리와 목은 옷깃을 잡아당기는 동작과 연결된다. 현우의 두 손은 실제로 가운을 움켜쥐며 천에 당김과 주름이 생긴다. 현우의 하체 지지점은 화면 밖이지만 상체는 재판관 위로 낮게 기울어진 연속된 자세여서 공중에 떠 있다고 볼 근거는 없다. 탁자는 바닥에 닿은 다리로 지지된다."
       },
       {
        "label": "B",
        "direction": "현우는 눈을 거의 감은 채 얼굴과 벌어진 입을 아래쪽 재판관 얼굴로 향한다. 재판관은 현우를 올려다보며 팔을 그의 몸 쪽으로 뻗는다. 두 손의 옷깃 잡기와 포효의 대상은 일치한다.",
        "built_space": "왼쪽 탁자 한 개에 상판 끝, 하부 가로대와 여러 다리가 보인다. 탁자 하부가 왼쪽 전경을 크게 차지하며, 바닥도 넓게 노출되어 지정된 작은 배경 단서와 다르다. 재판관은 탁자 오른쪽 바닥에 있고 현우는 그 위에 걸터앉듯 몸을 숙인다. 뒤쪽 목재 면과 낮은 노출은 이전 법정 분위기와 유사하지만 넘어진 의자는 확인되지 않는다. 반사는 없다.",
        "entities": "현우와 재판관 두 명만 보인다. 현우의 검은 헝클어진 머리, 젊은 동아시아계 얼굴, 해진 후드와 청바지는 참조의 주요 특징에 부합한다. 재판관은 흐트러진 머리의 성인 남성이며 낡고 구겨진 검은 가운을 입는다. 크게 열린 입과 두드러진 목 힘줄은 포효를 잘 나타내고, 재판관은 두려워하며 저항하는 모습이다. 바닥에는 흩어진 종이와 갈색 종이류가 보이지만 이전 위치에 남아 있어야 할 탁자 위 서류의 상태는 확인되지 않는다. 종이에 작은 인쇄 흔적은 있으나 읽을 수 있는 문구는 식별되지 않는다.",
        "hard_violations": [],
        "physics": "재판관의 등은 바닥에 지지되고 상체는 붙잡힌 옷깃을 따라 약간 들린다. 현우의 양손은 가운을 잡으며, 재판관의 팔은 현우의 몸을 밀거나 붙드는 저항 자세로 이어진다. 현우의 굽힌 허벅지는 화면 아래로 이어지고 최종 무릎 접점은 잘려 있어 보이지 않지만, 떠 있는 신체로 보이지는 않는다. 탁자 다리와 흩어진 종이는 바닥에 닿아 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "재판관 얼굴 바로 위로 낮게 엎드린 밀착 구도는 더 충실하지만, 포효가 이를 악문 표정에 가깝고 탁자가 왼쪽 전경을 지나치게 차지한다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "벌어진 입과 목의 핏대는 포효를 명확히 구현하지만, 상체가 더 세워지고 허벅지와 바닥까지 드러나 얼굴 위에 바짝 엎드린 클로즈업의 집중도가 떨어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우의 얼굴과 눈은 바로 아래 재판관 얼굴을 향하고, 재판관도 현우를 올려다본다. 양손은 재판관의 옷깃을 향해 모여 있다. 공격과 시선의 대상이 서로 분명하다.",
        "built_space": "왼쪽에 탁자 한 개의 상판 끝, 하부 틀, 앞다리와 뒤쪽 지지부가 보인다. 탁자가 화면 왼쪽 약 3분의 1을 차지해 맨 왼쪽 가장자리에 일부만 보여야 한다는 지시보다 크다. 재판관은 탁자 옆 바닥에 누워 있고 현우는 오른쪽에서 그 위로 숙인다. 뒤쪽 목재 가구는 흐려 넘어진 의자인지 확인하기 어렵다. 바닥도 전경 어깨 아래의 좁은 띠보다 넓게 보인다. 반사는 없다.",
        "entities": "두 남성만 보인다. 현우는 젊은 동아시아계 남성의 얼굴, 헝클어진 검은 머리, 오염되고 해진 짙은 후드 상의로 참조의 주요 특징을 따른다. 다만 얼굴에는 참조보다 성숙한 인상이 있다. 재판관은 성인 남성이며 구김 많은 검은 가운을 입고 놀란 표정을 짓는다. 현우의 목 힘줄은 드러나지만 입은 크게 벌린 포효보다 이를 드러낸 으르렁거림에 가깝다. 다리 부상과 탁자 위 서류는 이 구도에서 확인할 수 없으며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "재판관의 등과 어깨는 바닥에 닿고, 들린 머리와 목은 옷깃을 잡아당기는 동작과 연결된다. 현우의 두 손은 실제로 가운을 움켜쥐며 천에 당김과 주름이 생긴다. 현우의 하체 지지점은 화면 밖이지만 상체는 재판관 위로 낮게 기울어진 연속된 자세여서 공중에 떠 있다고 볼 근거는 없다. 탁자는 바닥에 닿은 다리로 지지된다."
       },
       {
        "label": "A",
        "direction": "현우는 눈을 거의 감은 채 얼굴과 벌어진 입을 아래쪽 재판관 얼굴로 향한다. 재판관은 현우를 올려다보며 팔을 그의 몸 쪽으로 뻗는다. 두 손의 옷깃 잡기와 포효의 대상은 일치한다.",
        "built_space": "왼쪽 탁자 한 개에 상판 끝, 하부 가로대와 여러 다리가 보인다. 탁자 하부가 왼쪽 전경을 크게 차지하며, 바닥도 넓게 노출되어 지정된 작은 배경 단서와 다르다. 재판관은 탁자 오른쪽 바닥에 있고 현우는 그 위에 걸터앉듯 몸을 숙인다. 뒤쪽 목재 면과 낮은 노출은 이전 법정 분위기와 유사하지만 넘어진 의자는 확인되지 않는다. 반사는 없다.",
        "entities": "현우와 재판관 두 명만 보인다. 현우의 검은 헝클어진 머리, 젊은 동아시아계 얼굴, 해진 후드와 청바지는 참조의 주요 특징에 부합한다. 재판관은 흐트러진 머리의 성인 남성이며 낡고 구겨진 검은 가운을 입는다. 크게 열린 입과 두드러진 목 힘줄은 포효를 잘 나타내고, 재판관은 두려워하며 저항하는 모습이다. 바닥에는 흩어진 종이와 갈색 종이류가 보이지만 이전 위치에 남아 있어야 할 탁자 위 서류의 상태는 확인되지 않는다. 종이에 작은 인쇄 흔적은 있으나 읽을 수 있는 문구는 식별되지 않는다.",
        "hard_violations": [],
        "physics": "재판관의 등은 바닥에 지지되고 상체는 붙잡힌 옷깃을 따라 약간 들린다. 현우의 양손은 가운을 잡으며, 재판관의 팔은 현우의 몸을 밀거나 붙드는 저항 자세로 이어진다. 현우의 굽힌 허벅지는 화면 아래로 이어지고 최종 무릎 접점은 잘려 있어 보이지 않지만, 떠 있는 신체로 보이지는 않는다. 탁자 다리와 흩어진 종이는 바닥에 닿아 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1857,
   "B": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "이현우가 입을 크게 벌리고 목에 핏대를 세우며 포효하는 광기 어린 표정을 텍스트의 묘사에 맞게 가장 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "구도와 인물의 외형 조건은 잘 충족했으나, 이현우의 표정이 '포효'라기보다는 단순히 이를 악물고 화를 내는 모습에 가까워 묘사가 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S8sh1_sel.png",
    "asset_id": "d2163e37-da25-40cb-8020-db2c103d9c27",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd7c-2fa1-76d6-aeb9-511d886c4164",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S8sh1"
  }
 },
 "S8sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:18:42.987499+00:00",
  "fingerprint": "3c8a76b18734dd3e5577f808025cd21b409e89b69441814dd3cd859707e66017",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S8sh14_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S8sh14_sel.png",
  "source_sha256": "a96c7737570da7347577a0fbaf0c37d354aa12f466b26f3663c9c3b7d4a4b712",
  "file": "S8sh14_cine.png",
  "staged_sha256": "0139f31bd3cff9720c814d8b08799f2ea1f218e791ca4bbdfda938fc795f845b",
  "latency_ms": 11233
 },
 "S8sh19::signage": {
  "fp": "465c14c61239e3f0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S8sh19": {
  "input_fingerprint": "62579e5aa4e06b0d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 두려움에 질린 미연이 어린 이현우와 앰버의 머리를 자신의 양팔로 빈틈없이 감싸 안고 엎드린 모습.\n\nLOCATION (lock): On a city street at night during a violent roundup, shown in a monochrome memory with gunfire flashes nearby. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: American city street ground (The family is kneeling on it during the attack); used as A restrained border around the family establishes the separate remembered location without inventing street furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the explicitly black-and-white nighttime dream treatment with stark tonal separation and a staccato visual impression, without adding an unmentioned local light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The American street flashback is monochrome, set at night, with visible gunfire flashes and fragmented imagery.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 두려움에 질린 미연이 어린 이현우와 앰버의 머리를 자신의 양팔로 빈틈없이 감싸 안고 엎드린 모습.\n\nLOCATION (lock): On a city street at night during a violent roundup, shown in a monochrome memory with gunfire flashes nearby. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: American city street ground (The family is kneeling on it during the attack); used as A restrained border around the family establishes the separate remembered location without inventing street furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the explicitly black-and-white nighttime dream treatment with stark tonal separation and a staccato visual impression, without adding an unmentioned local light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The American street flashback is monochrome, set at night, with visible gunfire flashes and fragmented imagery.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 두려움에 질린 미연이 어린 이현우와 앰버의 머리를 자신의 양팔로 빈틈없이 감싸 안고 엎드린 모습.\n\nLOCATION (lock): On a city street at night during a violent roundup, shown in a monochrome memory with gunfire flashes nearby. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: American city street ground (The family is kneeling on it during the attack); used as A restrained border around the family establishes the separate remembered location without inventing street furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the explicitly black-and-white nighttime dream treatment with stark tonal separation and a staccato visual impression, without adding an unmentioned local light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The American street flashback is monochrome, set at night, with visible gunfire flashes and fragmented imagery.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "여성은 두 아이를 내려다보며 시선을 아래로 향하고 있으며, 주변 바닥에서 총격으로 인한 섬광이 발생하고 있음.",
    "built_space": "우측의 건물, 배경의 철조망, 우측 인도 위의 소화전 등 참조 이미지의 거리 구조와 정확히 일치하며 인물들은 도로 중앙에 웅크리고 있음.",
    "entities": "중년 한국인 여성(미연), 헝클어진 검은 머리의 소년(이현우), 금발의 소녀(앰버) 등 지시된 인물의 외형이 모두 정확하게 나타남.",
    "hard_violations": [],
    "physics": "세 인물 모두 바닥에 무릎을 꿇고 엎드려 지면의 지지를 받고 있으며, 여성의 팔과 손은 아이들의 머리를 단단히 감싸안고 있음."
   },
   {
    "label": "B",
    "direction": "여성은 불안한 표정으로 화면 좌측을 응시하고 있으며, 좌우 상단 허공에서 총구의 화염이 뿜어져 나옴.",
    "built_space": "우측의 건물 구조는 참조와 유사하나, 참조 이미지에는 없는 소화전이 도로 좌측에 임의로 추가되어 있음.",
    "entities": "세 명의 주요 인물 외에 좌측 상단 모서리에 지시문에 없는 인물의 손이 총을 쥐고 등장함.",
    "hard_violations": [
     "[gemini-pro] invented people (지시문에 없는 인물의 손과 총이 좌측 상단에 등장함)",
     "[gemini-pro] built space violation (참조 이미지에 존재하지 않는 소화전이 화면 좌측에 추가됨)",
     "[gpt-high] 왼쪽 위 화면 가장자리에 가족 세 사람과 별개인 인물의 손이 나타나, 지정 인물 외에는 신체 일부도 추가하지 말라는 조건을 위반합니다."
    ],
    "physics": "가족들은 바닥에 무릎을 꿇고 안정적으로 위치해 있으나, 좌측 상단의 손은 화면 밖에서 공중으로 뻗어 나와 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "세 명의 지정된 인물만을 정확히 묘사하였으며, 바닥에 엎드려 아이들을 감싸 안는 동작과 참조 이미지의 공간적 특징을 충실히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시문에 명시되지 않은 인물의 손과 총이 화면에 등장하였으며, 참조 이미지에 없는 소화전을 배경에 추가하는 치명적인 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "여성은 두 아이를 내려다보며 시선을 아래로 향하고 있으며, 주변 바닥에서 총격으로 인한 섬광이 발생하고 있음.",
        "built_space": "우측의 건물, 배경의 철조망, 우측 인도 위의 소화전 등 참조 이미지의 거리 구조와 정확히 일치하며 인물들은 도로 중앙에 웅크리고 있음.",
        "entities": "중년 한국인 여성(미연), 헝클어진 검은 머리의 소년(이현우), 금발의 소녀(앰버) 등 지시된 인물의 외형이 모두 정확하게 나타남.",
        "hard_violations": [],
        "physics": "세 인물 모두 바닥에 무릎을 꿇고 엎드려 지면의 지지를 받고 있으며, 여성의 팔과 손은 아이들의 머리를 단단히 감싸안고 있음."
       },
       {
        "label": "B",
        "direction": "여성은 불안한 표정으로 화면 좌측을 응시하고 있으며, 좌우 상단 허공에서 총구의 화염이 뿜어져 나옴.",
        "built_space": "우측의 건물 구조는 참조와 유사하나, 참조 이미지에는 없는 소화전이 도로 좌측에 임의로 추가되어 있음.",
        "entities": "세 명의 주요 인물 외에 좌측 상단 모서리에 지시문에 없는 인물의 손이 총을 쥐고 등장함.",
        "hard_violations": [
         "invented people (지시문에 없는 인물의 손과 총이 좌측 상단에 등장함)",
         "built space violation (참조 이미지에 존재하지 않는 소화전이 화면 좌측에 추가됨)"
        ],
        "physics": "가족들은 바닥에 무릎을 꿇고 안정적으로 위치해 있으나, 좌측 상단의 손은 화면 밖에서 공중으로 뻗어 나와 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "세 명의 지정된 인물만을 정확히 묘사하였으며, 바닥에 엎드려 아이들을 감싸 안는 동작과 참조 이미지의 공간적 특징을 충실히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시문에 명시되지 않은 인물의 손과 총이 화면에 등장하였으며, 참조 이미지에 없는 소화전을 배경에 추가하는 치명적인 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "여성은 두 아이를 내려다보며 시선을 아래로 향하고 있으며, 주변 바닥에서 총격으로 인한 섬광이 발생하고 있음.",
        "built_space": "우측의 건물, 배경의 철조망, 우측 인도 위의 소화전 등 참조 이미지의 거리 구조와 정확히 일치하며 인물들은 도로 중앙에 웅크리고 있음.",
        "entities": "중년 한국인 여성(미연), 헝클어진 검은 머리의 소년(이현우), 금발의 소녀(앰버) 등 지시된 인물의 외형이 모두 정확하게 나타남.",
        "hard_violations": [],
        "physics": "세 인물 모두 바닥에 무릎을 꿇고 엎드려 지면의 지지를 받고 있으며, 여성의 팔과 손은 아이들의 머리를 단단히 감싸안고 있음."
       },
       {
        "label": "B",
        "direction": "여성은 불안한 표정으로 화면 좌측을 응시하고 있으며, 좌우 상단 허공에서 총구의 화염이 뿜어져 나옴.",
        "built_space": "우측의 건물 구조는 참조와 유사하나, 참조 이미지에는 없는 소화전이 도로 좌측에 임의로 추가되어 있음.",
        "entities": "세 명의 주요 인물 외에 좌측 상단 모서리에 지시문에 없는 인물의 손이 총을 쥐고 등장함.",
        "hard_violations": [
         "invented people (지시문에 없는 인물의 손과 총이 좌측 상단에 등장함)",
         "built space violation (참조 이미지에 존재하지 않는 소화전이 화면 좌측에 추가됨)"
        ],
        "physics": "가족들은 바닥에 무릎을 꿇고 안정적으로 위치해 있으나, 좌측 상단의 손은 화면 밖에서 공중으로 뻗어 나와 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "왼쪽 위에 허용되지 않은 인물의 손이 추가되었으며, 전신에 가까운 구도와 상체를 세운 미연의 자세가 미디엄 숏의 엎드린 포옹에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "미연이 두 아이의 머리를 양팔로 감싸며 깊이 숙인 행동은 더 충실하지만, 발까지 보이는 넓은 구도와 큰 배경 비중은 지정된 미디엄 숏에 미달합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 두 아이 사이의 아래쪽을 바라보고, 왼쪽 현우와 오른쪽 앰버는 얼굴을 안쪽 아래로 숙입니다. 미연의 양손은 각각 두 아이의 정수리와 이마에 닿지만, 미연 자신의 머리와 상체는 비교적 높이 들려 있습니다. 화면 양쪽 위에 섬광이 있으나 총구와 탄도는 명확하지 않아 가족을 향한 발사인지 확인할 수 없습니다.",
        "built_space": "가족은 보도가 아닌 젖은 차도에 무릎을 꿇고 있습니다. 왼쪽 뒤에 맨홀 하나와 소화전 하나가 보이고, 오른쪽에는 연속된 연석과 산업용 건물 벽면, 뒤쪽에는 울타리와 폐자재 더미가 있습니다. 참조의 재질과 산업 거리 분위기는 유사하지만, 참조에서 오른쪽 보도에 있던 소화전이 여기서는 왼쪽에 보입니다. 가족의 발까지 포함하고 건물 배경도 넓게 보여 지면을 절제된 테두리로 쓰는 미디엄 숏은 아닙니다.",
        "entities": "중년의 한국계 여성으로 보이는 미연, 검은 헝클어진 머리의 한국계 청소년 남성으로 보이는 현우, 밝은 금발의 여자아이 앰버가 있습니다. 앰버의 혼혈 정체성과 큰 눈은 숙인 얼굴만으로 확정하기 어렵습니다. 왼쪽 위 가장자리에 가족 세 사람에게 속할 수 없는 별도의 손이 보입니다. 흑백 야간 표현과 섬광은 있으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "왼쪽 위 화면 가장자리에 가족 세 사람과 별개인 인물의 손이 나타나, 지정 인물 외에는 신체 일부도 추가하지 말라는 조건을 위반합니다."
        ],
        "physics": "현우와 앰버의 무릎과 접힌 다리는 차도에 닿아 체중을 지지합니다. 미연도 두 아이 뒤에서 다리를 굽혀 지면에 지지하고 있으며, 양손으로 각 아이의 머리를 실제로 감쌉니다. 포옹 자체는 가능한 자세이고 공중에 떠 있는 몸이나 물체는 없습니다. 다만 상체를 세운 웅크림이어서 엎드려 덮어주는 동작은 약합니다."
       },
       {
        "label": "B",
        "direction": "미연은 얼굴을 두 아이의 머리 사이로 깊이 내리고 아래를 바라봅니다. 현우와 앰버도 얼굴을 미연의 몸 쪽과 지면 쪽으로 숙이며, 미연의 양팔과 손이 각각의 머리를 안쪽으로 감쌉니다. 가족 뒤 좌우 지면의 섬광과 불꽃은 탄착처럼 보이지만, 총구나 발사 방향은 보이지 않습니다.",
        "built_space": "세 사람은 젖은 차도 중앙에 붙어 무릎을 꿇고 있습니다. 오른쪽 보도에는 소화전 하나, 그 앞 연석 가까이에는 배수구 하나가 보입니다. 뒤에는 불이 켜진 가로등 하나와 여러 전주, 울타리, 폐자재 더미가 있고 오른쪽 산업용 건물의 큰 창과 문도 보입니다. 참조의 주요 거리 요소와 배치가 비교적 잘 이어지지만, 양쪽 건물과 도로를 넓게 포함하여 배경이 절제된 테두리 이상을 차지합니다.",
        "entities": "보이는 인물은 세 명뿐입니다. 미연은 중년의 한국계 여성으로, 현우는 검은 머리의 젊은 한국계 남성으로, 앰버는 금발의 여자아이로 읽힙니다. 숙인 얼굴과 가리는 손 때문에 정확한 연령, 앰버의 혼혈 얼굴과 큰 눈은 충분히 확인되지 않습니다. 흑백 야간 기억의 강한 명암은 구현되었고, 지면에 여러 섬광이 있지만 명확한 총구 화염은 아닙니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 아이의 무릎과 정강이, 신발이 차도에 닿아 몸을 지지하고, 미연은 뒤에서 무릎을 낮춘 채 상체를 앞으로 접습니다. 양팔이 아이들의 머리를 둘러싸고 양손이 머리에 밀착되어 보호하는 행동이 물리적으로 성립합니다. 지면의 불꽃은 접촉 지점에서 튀는 모습이며, 지지 없이 떠 있는 인체나 소품은 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "왼쪽 위에 허용되지 않은 인물의 손이 추가되었으며, 전신에 가까운 구도와 상체를 세운 미연의 자세가 미디엄 숏의 엎드린 포옹에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "미연이 두 아이의 머리를 양팔로 감싸며 깊이 숙인 행동은 더 충실하지만, 발까지 보이는 넓은 구도와 큰 배경 비중은 지정된 미디엄 숏에 미달합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "미연은 두 아이 사이의 아래쪽을 바라보고, 왼쪽 현우와 오른쪽 앰버는 얼굴을 안쪽 아래로 숙입니다. 미연의 양손은 각각 두 아이의 정수리와 이마에 닿지만, 미연 자신의 머리와 상체는 비교적 높이 들려 있습니다. 화면 양쪽 위에 섬광이 있으나 총구와 탄도는 명확하지 않아 가족을 향한 발사인지 확인할 수 없습니다.",
        "built_space": "가족은 보도가 아닌 젖은 차도에 무릎을 꿇고 있습니다. 왼쪽 뒤에 맨홀 하나와 소화전 하나가 보이고, 오른쪽에는 연속된 연석과 산업용 건물 벽면, 뒤쪽에는 울타리와 폐자재 더미가 있습니다. 참조의 재질과 산업 거리 분위기는 유사하지만, 참조에서 오른쪽 보도에 있던 소화전이 여기서는 왼쪽에 보입니다. 가족의 발까지 포함하고 건물 배경도 넓게 보여 지면을 절제된 테두리로 쓰는 미디엄 숏은 아닙니다.",
        "entities": "중년의 한국계 여성으로 보이는 미연, 검은 헝클어진 머리의 한국계 청소년 남성으로 보이는 현우, 밝은 금발의 여자아이 앰버가 있습니다. 앰버의 혼혈 정체성과 큰 눈은 숙인 얼굴만으로 확정하기 어렵습니다. 왼쪽 위 가장자리에 가족 세 사람에게 속할 수 없는 별도의 손이 보입니다. 흑백 야간 표현과 섬광은 있으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "왼쪽 위 화면 가장자리에 가족 세 사람과 별개인 인물의 손이 나타나, 지정 인물 외에는 신체 일부도 추가하지 말라는 조건을 위반합니다."
        ],
        "physics": "현우와 앰버의 무릎과 접힌 다리는 차도에 닿아 체중을 지지합니다. 미연도 두 아이 뒤에서 다리를 굽혀 지면에 지지하고 있으며, 양손으로 각 아이의 머리를 실제로 감쌉니다. 포옹 자체는 가능한 자세이고 공중에 떠 있는 몸이나 물체는 없습니다. 다만 상체를 세운 웅크림이어서 엎드려 덮어주는 동작은 약합니다."
       },
       {
        "label": "A",
        "direction": "미연은 얼굴을 두 아이의 머리 사이로 깊이 내리고 아래를 바라봅니다. 현우와 앰버도 얼굴을 미연의 몸 쪽과 지면 쪽으로 숙이며, 미연의 양팔과 손이 각각의 머리를 안쪽으로 감쌉니다. 가족 뒤 좌우 지면의 섬광과 불꽃은 탄착처럼 보이지만, 총구나 발사 방향은 보이지 않습니다.",
        "built_space": "세 사람은 젖은 차도 중앙에 붙어 무릎을 꿇고 있습니다. 오른쪽 보도에는 소화전 하나, 그 앞 연석 가까이에는 배수구 하나가 보입니다. 뒤에는 불이 켜진 가로등 하나와 여러 전주, 울타리, 폐자재 더미가 있고 오른쪽 산업용 건물의 큰 창과 문도 보입니다. 참조의 주요 거리 요소와 배치가 비교적 잘 이어지지만, 양쪽 건물과 도로를 넓게 포함하여 배경이 절제된 테두리 이상을 차지합니다.",
        "entities": "보이는 인물은 세 명뿐입니다. 미연은 중년의 한국계 여성으로, 현우는 검은 머리의 젊은 한국계 남성으로, 앰버는 금발의 여자아이로 읽힙니다. 숙인 얼굴과 가리는 손 때문에 정확한 연령, 앰버의 혼혈 얼굴과 큰 눈은 충분히 확인되지 않습니다. 흑백 야간 기억의 강한 명암은 구현되었고, 지면에 여러 섬광이 있지만 명확한 총구 화염은 아닙니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 아이의 무릎과 정강이, 신발이 차도에 닿아 몸을 지지하고, 미연은 뒤에서 무릎을 낮춘 채 상체를 앞으로 접습니다. 양팔이 아이들의 머리를 둘러싸고 양손이 머리에 밀착되어 보호하는 행동이 물리적으로 성립합니다. 지면의 불꽃은 접촉 지점에서 튀는 모습이며, 지지 없이 떠 있는 인체나 소품은 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.762
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.512
   },
   "violations": {
    "B": [
     "[gemini-pro] invented people (지시문에 없는 인물의 손과 총이 좌측 상단에 등장함)",
     "[gemini-pro] built space violation (참조 이미지에 존재하지 않는 소화전이 화면 좌측에 추가됨)",
     "[gpt-high] 왼쪽 위 화면 가장자리에 가족 세 사람과 별개인 인물의 손이 나타나, 지정 인물 외에는 신체 일부도 추가하지 말라는 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 512
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "세 명의 지정된 인물만을 정확히 묘사하였으며, 바닥에 엎드려 아이들을 감싸 안는 동작과 참조 이미지의 공간적 특징을 충실히 구현함."
   },
   {
    "label": "B",
    "score": 512,
    "verdict_ko": "지시문에 명시되지 않은 인물의 손과 총이 화면에 등장하였으며, 참조 이미지에 없는 소화전을 배경에 추가하는 치명적인 오류가 발생함.  ★위반: [gemini-pro] invented people (지시문에 없는 인물의 손과 총이 좌측 상단에 등장함) / [gemini-pro] built space violation (참조 이미지에 존재하지 않는 소화전이 화면 좌측에 추가됨) / [gpt-high] 왼쪽 위 화면 가장자리에 가족 세 사람과 별개인 인물의 손이 나타나, 지정 인물 외에는 신체 일부도 추가하지 말라는 조건을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L16B03.png",
    "asset_id": "ca69d40c-ce04-4745-a49e-8348a3c48f13",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd7c-e8e8-71cb-8159-10d11b31916c",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S8sh19::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:18:54.572571+00:00",
  "fingerprint": "a9013ce8c7558bc79729e4bb7e096492a1157a69429ab502456490a3d044acc4",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S8sh19_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S8sh19_sel.png",
  "source_sha256": "79703964c7ca5d4c7fe948020fa4a1726f75adefa013b015e278414b5aad76da",
  "file": "S8sh19_cine.png",
  "staged_sha256": "221a04c3b89d10a7ff73ab4c31d8bc6cfe97893f53f99a60ebee90cda638a2e4",
  "latency_ms": 12144
 },
 "S9sh1::signage": {
  "fp": "b48ee8d2a233b629",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::703df31eb71aecb1": {
  "subjects": [],
  "subject_text": "난민 재판소 밖 쓰레기 더미\n허름한 재판소 건물 밖에 폐기물이 어지럽게 쌓인 공간. 쓰레기 더미 사이로 드러난 바닥이 어둠에 잠겨 있다.",
  "identity": "canonical",
  "scope_id": "L18",
  "scope_role": "location_exterior",
  "scope_sha": "7e6d6207d7733b87"
 },
 "S9sh1::bgfirst_bg": {
  "input_fingerprint": "d0f33da7a1910901",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 쓰레기 더미 위, 기절해 있던 이현우가 허리를 튕기듯 상체를 반쯤 들어올린 mid-action 자세로 입을 크게 벌린 찰나.\n\nLOCATION (lock): On a desolate rubbish heap immediately outside the refugee tribunal at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Trash piles outside the tribunal (Surrounding 이현우 where he regains consciousness); used as Frame his raised torso with uneven negative spaces and establish his discarded position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Return from the monochrome dream to restrained nighttime visibility, keeping his awakening expression readable against the dark refuse without specifying a light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 쓰레기 더미 위, 기절해 있던 이현우가 허리를 튕기듯 상체를 반쯤 들어올린 mid-action 자세로 입을 크게 벌린 찰나.\n\nLOCATION (lock): On a desolate rubbish heap immediately outside the refugee tribunal at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Trash piles outside the tribunal (Surrounding 이현우 where he regains consciousness); used as Frame his raised torso with uneven negative spaces and establish his discarded position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Return from the monochrome dream to restrained nighttime visibility, keeping his awakening expression readable against the dark refuse without specifying a light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S9sh1__bgfirst_bg.png",
  "asset_id": "ac5f1bce-ac1d-4444-ab7c-809ed6f14bb1",
  "input_asset_ids": [
   "ece5766a-e7dd-4ef0-a96e-7c9aad2c4594",
   "1267462a-839e-4acd-aed6-8c88a6215016"
  ]
 },
 "S9sh1": {
  "input_fingerprint": "92ff1087f8d351f3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 쓰레기 더미 위, 기절해 있던 이현우가 허리를 튕기듯 상체를 반쯤 들어올린 mid-action 자세로 입을 크게 벌린 찰나.\n\nLOCATION (lock): On a desolate rubbish heap immediately outside the refugee tribunal at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Trash piles outside the tribunal (Surrounding 이현우 where he regains consciousness); used as Frame his raised torso with uneven negative spaces and establish his discarded position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Return from the monochrome dream to restrained nighttime visibility, keeping his awakening expression readable against the dark refuse without specifying a light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It is night outside the refugee court, surrounded by desolate rubbish heaps. 이현우: Hyunwoo is waking after the head blow, with his dog-bitten leg still injured. The judge's wallet is already in his pocket, since he retrieves it without acquiring it after waking.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 쓰레기 더미 위, 기절해 있던 이현우가 허리를 튕기듯 상체를 반쯤 들어올린 mid-action 자세로 입을 크게 벌린 찰나.\n\nLOCATION (lock): On a desolate rubbish heap immediately outside the refugee tribunal at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Trash piles outside the tribunal (Surrounding 이현우 where he regains consciousness); used as Frame his raised torso with uneven negative spaces and establish his discarded position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Return from the monochrome dream to restrained nighttime visibility, keeping his awakening expression readable against the dark refuse without specifying a light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It is night outside the refugee court, surrounded by desolate rubbish heaps. 이현우: Hyunwoo is waking after the head blow, with his dog-bitten leg still injured. The judge's wallet is already in his pocket, since he retrieves it without acquiring it after waking.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 쓰레기 더미 위, 기절해 있던 이현우가 허리를 튕기듯 상체를 반쯤 들어올린 mid-action 자세로 입을 크게 벌린 찰나.\n\nLOCATION (lock): On a desolate rubbish heap immediately outside the refugee tribunal at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Trash piles outside the tribunal (Surrounding 이현우 where he regains consciousness); used as Frame his raised torso with uneven negative spaces and establish his discarded position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Return from the monochrome dream to restrained nighttime visibility, keeping his awakening expression readable against the dark refuse without specifying a light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It is night outside the refugee court, surrounded by desolate rubbish heaps. 이현우: Hyunwoo is waking after the head blow, with his dog-bitten leg still injured. The judge's wallet is already in his pocket, since he retrieves it without acquiring it after waking.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S9sh1__bgfirst_bg.png",
     "asset_id": "ac5f1bce-ac1d-4444-ab7c-809ed6f14bb1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S9sh1.png",
     "asset_id": "ece5766a-e7dd-4ef0-a96e-7c9aad2c4594",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L18B01.png",
     "asset_id": "1267462a-839e-4acd-aed6-8c88a6215016",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 우측을 향하고 있으며, 입은 약간 벌려진 상태입니다.",
    "built_space": "레퍼런스와 일치하는 야간의 쓰레기장이며, 흙바닥과 좌우의 쓰레기 더미, 배경의 펜스와 산이 보입니다.",
    "entities": "이현우의 외모, 체형, 의상(낡은 후드티와 찢어진 청바지)이 레퍼런스와 일치합니다.",
    "hard_violations": [],
    "physics": "땅바닥에 양손과 다리를 짚고 상체를 지탱하며 비스듬히 앉아 있는 자세입니다."
   },
   {
    "label": "B",
    "direction": "시선은 정면 약간 위를 향하고 있으며, 숨을 들이마시듯 입을 크게 벌리고 있습니다.",
    "built_space": "레퍼런스와 일치하는 쓰레기장 배경이며, 인물 주변으로 쓰레기 더미가 자연스럽게 배치되어 있습니다.",
    "entities": "이현우의 얼굴과 의상(후드티, 무릎이 찢어진 청바지)이 레퍼런스와 정확히 일치합니다.",
    "hard_violations": [],
    "physics": "쓰레기 더미 위에서 상체를 꼿꼿이 세워 올린 상태로, 한 손은 바닥의 비닐봉지 위를 짚어 체중을 지탱하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "미디엄 샷의 프레이밍과 허리를 튕기듯 상체를 일으키며 입을 크게 벌리는 mid-action의 순간을 매우 사실적으로 포착했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "미디엄 샷을 요구한 프롬프트와 달리 전신이 다 보이는 풀 샷에 가깝게 프레이밍되었으며, 갑작스럽게 깨어나는 역동성이 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 우측을 향하고 있으며, 입은 약간 벌려진 상태입니다.",
        "built_space": "레퍼런스와 일치하는 야간의 쓰레기장이며, 흙바닥과 좌우의 쓰레기 더미, 배경의 펜스와 산이 보입니다.",
        "entities": "이현우의 외모, 체형, 의상(낡은 후드티와 찢어진 청바지)이 레퍼런스와 일치합니다.",
        "hard_violations": [],
        "physics": "땅바닥에 양손과 다리를 짚고 상체를 지탱하며 비스듬히 앉아 있는 자세입니다."
       },
       {
        "label": "B",
        "direction": "시선은 정면 약간 위를 향하고 있으며, 숨을 들이마시듯 입을 크게 벌리고 있습니다.",
        "built_space": "레퍼런스와 일치하는 쓰레기장 배경이며, 인물 주변으로 쓰레기 더미가 자연스럽게 배치되어 있습니다.",
        "entities": "이현우의 얼굴과 의상(후드티, 무릎이 찢어진 청바지)이 레퍼런스와 정확히 일치합니다.",
        "hard_violations": [],
        "physics": "쓰레기 더미 위에서 상체를 꼿꼿이 세워 올린 상태로, 한 손은 바닥의 비닐봉지 위를 짚어 체중을 지탱하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "미디엄 샷의 프레이밍과 허리를 튕기듯 상체를 일으키며 입을 크게 벌리는 mid-action의 순간을 매우 사실적으로 포착했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "미디엄 샷을 요구한 프롬프트와 달리 전신이 다 보이는 풀 샷에 가깝게 프레이밍되었으며, 갑작스럽게 깨어나는 역동성이 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 우측을 향하고 있으며, 입은 약간 벌려진 상태입니다.",
        "built_space": "레퍼런스와 일치하는 야간의 쓰레기장이며, 흙바닥과 좌우의 쓰레기 더미, 배경의 펜스와 산이 보입니다.",
        "entities": "이현우의 외모, 체형, 의상(낡은 후드티와 찢어진 청바지)이 레퍼런스와 일치합니다.",
        "hard_violations": [],
        "physics": "땅바닥에 양손과 다리를 짚고 상체를 지탱하며 비스듬히 앉아 있는 자세입니다."
       },
       {
        "label": "B",
        "direction": "시선은 정면 약간 위를 향하고 있으며, 숨을 들이마시듯 입을 크게 벌리고 있습니다.",
        "built_space": "레퍼런스와 일치하는 쓰레기장 배경이며, 인물 주변으로 쓰레기 더미가 자연스럽게 배치되어 있습니다.",
        "entities": "이현우의 얼굴과 의상(후드티, 무릎이 찢어진 청바지)이 레퍼런스와 정확히 일치합니다.",
        "hard_violations": [],
        "physics": "쓰레기 더미 위에서 상체를 꼿꼿이 세워 올린 상태로, 한 손은 바닥의 비닐봉지 위를 짚어 체중을 지탱하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "쓰레기 더미에 몸을 둔 채 입을 크게 벌린 상반신 중심 구도가 더 충실하지만, 허리를 튕겨 반쯤 일어나는 순간보다는 이미 상당히 일어나 앉은 모습에 가깝다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "신체 지지는 자연스럽지만, 발까지 포함한 넓은 구도와 옆으로 돌아 양손을 짚은 자세가 요구된 미디엄 숏의 급작스러운 상체 들기보다 주변을 살피는 동작으로 읽힌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 화면 오른쪽 위의 프레임 밖을 향하며, 그곳에 특정 대상은 보이지 않는다. 지시문은 응시 대상을 지정하지 않았다. 입은 크게 열려 있고 몸통은 골반 위로 거의 곧게 올라와 있어, 반쯤 솟는 중간 단계보다는 일어나 앉은 끝자세에 가깝다.",
        "built_space": "왼쪽의 높은 폐기물 더미, 뒤쪽 금속 울타리, 전신주 두 개와 전선, 오른쪽 콘크리트 건물의 가로 창 한 곳 및 외벽 배관이 보인다. 참조 장소의 주요 배치와 재료가 유지된다. 인물은 전경 쓰레기 더미에 앉아 있고 오른쪽에는 폐기된 화면 장치 두 대가 놓여 있다. 이는 고정 시설의 중복이 아니라 폐기물로 보인다. 몸통 주변을 불규칙한 쓰레기 윤곽이 둘러싼다.",
        "entities": "인물은 한 명뿐이며 한국계 청년 남성으로 보이는 얼굴, 헝클어진 검은 머리, 마른 체격이 참조와 대체로 맞는다. 낡은 짙은 회색 후드와 찢어진 청바지도 일치한다. 얼굴은 참조보다 다소 성숙하게 보인다. 무릎은 드러나지만 개에게 물린 상처는 확인되지 않으며, 주머니 속 지갑 역시 가려져 있어 판단할 수 없다. 밤의 어두운 폐기물 장소이며 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "골반과 허벅지가 폐기물 위에 놓이고 화면 왼쪽 손이 검은 봉투 더미를 짚어 몸을 지탱한다. 다리는 아래쪽 프레임 밖으로 이어지며, 몸이 공중에 떠 있는 모습은 아니다. 허리를 펴며 일어나는 동작으로 가능한 자세지만, 거의 수직인 몸통과 짚은 팔 때문에 순간적인 반동보다 지지된 앉은 자세가 강조된다. 주변 봉투와 전자 폐기물도 더미에 받쳐져 있다."
       },
       {
        "label": "B",
        "direction": "고개와 눈은 화면 오른쪽 위의 프레임 밖을 향한다. 보이는 응시 대상은 없으며 몸통도 그쪽으로 비틀려 있다. 입은 열려 있지만 A보다 개방 폭이 작다. 정면으로 허리를 튕겨 깨어나기보다는 옆으로 몸을 일으켜 무언가를 살피는 방향성이 강하다.",
        "built_space": "양옆 쓰레기 더미 사이로 넓은 진흙 통로가 열리고, 뒤쪽 금속 울타리와 전신주 두 개, 오른쪽 콘크리트 건물의 가로 창 한 곳이 보인다. 오른쪽 뒤에는 작은 차양과 점등된 조명, 울타리 앞에는 금속 통 하나가 있다. 참조의 장소 재료와 큰 배치는 유사하지만 통로의 비중이 커졌다. 인물은 높은 쓰레기 더미 위라기보다 통로 가장자리의 낮은 잔해 위에 있으며, 신발까지 들어오는 구도는 요구된 상반신 중심 미디엄 숏보다 넓다.",
        "entities": "추가 인물 없이 젊은 동아시아계 남성 한 명이 보이며, 검은 헝클어진 머리와 얼굴은 참조의 어린 인상에 가깝다. 낡은 회색 후드, 찢어진 청바지, 보이는 낡은 운동화도 참조와 부합한다. 개에게 물린 상처는 식별되지 않고 지갑은 보이지 않아 소지 여부를 판단할 수 없다. 폐금속과 비닐, 젖은 흙은 실제 재료처럼 보이며 야간 설정과 읽을 수 있는 글자 금지도 지킨다.",
        "hard_violations": [],
        "physics": "양손이 앞쪽과 옆쪽 잔해를 짚고 골반 및 접힌 다리가 낮은 쓰레기층에 닿아 있다. 화면 오른쪽으로 뻗은 다리와 신발도 지면 가까이에 놓여 있어 지지가 자연스럽다. 팔로 밀어 옆으로 일어나는 동작은 가능하지만, 요구된 허리 반동에 의한 상체 들기와는 동작의 중심이 다르다. 주변 폐기물은 땅이나 다른 잔해 위에 놓여 있으며 지지 없는 부유는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "쓰레기 더미에 몸을 둔 채 입을 크게 벌린 상반신 중심 구도가 더 충실하지만, 허리를 튕겨 반쯤 일어나는 순간보다는 이미 상당히 일어나 앉은 모습에 가깝다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "신체 지지는 자연스럽지만, 발까지 포함한 넓은 구도와 옆으로 돌아 양손을 짚은 자세가 요구된 미디엄 숏의 급작스러운 상체 들기보다 주변을 살피는 동작으로 읽힌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 화면 오른쪽 위의 프레임 밖을 향하며, 그곳에 특정 대상은 보이지 않는다. 지시문은 응시 대상을 지정하지 않았다. 입은 크게 열려 있고 몸통은 골반 위로 거의 곧게 올라와 있어, 반쯤 솟는 중간 단계보다는 일어나 앉은 끝자세에 가깝다.",
        "built_space": "왼쪽의 높은 폐기물 더미, 뒤쪽 금속 울타리, 전신주 두 개와 전선, 오른쪽 콘크리트 건물의 가로 창 한 곳 및 외벽 배관이 보인다. 참조 장소의 주요 배치와 재료가 유지된다. 인물은 전경 쓰레기 더미에 앉아 있고 오른쪽에는 폐기된 화면 장치 두 대가 놓여 있다. 이는 고정 시설의 중복이 아니라 폐기물로 보인다. 몸통 주변을 불규칙한 쓰레기 윤곽이 둘러싼다.",
        "entities": "인물은 한 명뿐이며 한국계 청년 남성으로 보이는 얼굴, 헝클어진 검은 머리, 마른 체격이 참조와 대체로 맞는다. 낡은 짙은 회색 후드와 찢어진 청바지도 일치한다. 얼굴은 참조보다 다소 성숙하게 보인다. 무릎은 드러나지만 개에게 물린 상처는 확인되지 않으며, 주머니 속 지갑 역시 가려져 있어 판단할 수 없다. 밤의 어두운 폐기물 장소이며 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "골반과 허벅지가 폐기물 위에 놓이고 화면 왼쪽 손이 검은 봉투 더미를 짚어 몸을 지탱한다. 다리는 아래쪽 프레임 밖으로 이어지며, 몸이 공중에 떠 있는 모습은 아니다. 허리를 펴며 일어나는 동작으로 가능한 자세지만, 거의 수직인 몸통과 짚은 팔 때문에 순간적인 반동보다 지지된 앉은 자세가 강조된다. 주변 봉투와 전자 폐기물도 더미에 받쳐져 있다."
       },
       {
        "label": "A",
        "direction": "고개와 눈은 화면 오른쪽 위의 프레임 밖을 향한다. 보이는 응시 대상은 없으며 몸통도 그쪽으로 비틀려 있다. 입은 열려 있지만 A보다 개방 폭이 작다. 정면으로 허리를 튕겨 깨어나기보다는 옆으로 몸을 일으켜 무언가를 살피는 방향성이 강하다.",
        "built_space": "양옆 쓰레기 더미 사이로 넓은 진흙 통로가 열리고, 뒤쪽 금속 울타리와 전신주 두 개, 오른쪽 콘크리트 건물의 가로 창 한 곳이 보인다. 오른쪽 뒤에는 작은 차양과 점등된 조명, 울타리 앞에는 금속 통 하나가 있다. 참조의 장소 재료와 큰 배치는 유사하지만 통로의 비중이 커졌다. 인물은 높은 쓰레기 더미 위라기보다 통로 가장자리의 낮은 잔해 위에 있으며, 신발까지 들어오는 구도는 요구된 상반신 중심 미디엄 숏보다 넓다.",
        "entities": "추가 인물 없이 젊은 동아시아계 남성 한 명이 보이며, 검은 헝클어진 머리와 얼굴은 참조의 어린 인상에 가깝다. 낡은 회색 후드, 찢어진 청바지, 보이는 낡은 운동화도 참조와 부합한다. 개에게 물린 상처는 식별되지 않고 지갑은 보이지 않아 소지 여부를 판단할 수 없다. 폐금속과 비닐, 젖은 흙은 실제 재료처럼 보이며 야간 설정과 읽을 수 있는 글자 금지도 지킨다.",
        "hard_violations": [],
        "physics": "양손이 앞쪽과 옆쪽 잔해를 짚고 골반 및 접힌 다리가 낮은 쓰레기층에 닿아 있다. 화면 오른쪽으로 뻗은 다리와 신발도 지면 가까이에 놓여 있어 지지가 자연스럽다. 팔로 밀어 옆으로 일어나는 동작은 가능하지만, 요구된 허리 반동에 의한 상체 들기와는 동작의 중심이 다르다. 주변 폐기물은 땅이나 다른 잔해 위에 놓여 있으며 지지 없는 부유는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.321,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.321,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1321
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "미디엄 샷의 프레이밍과 허리를 튕기듯 상체를 일으키며 입을 크게 벌리는 mid-action의 순간을 매우 사실적으로 포착했습니다."
   },
   {
    "label": "A",
    "score": 1321,
    "verdict_ko": "미디엄 샷을 요구한 프롬프트와 달리 전신이 다 보이는 풀 샷에 가깝게 프레이밍되었으며, 갑작스럽게 깨어나는 역동성이 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L18B01.png",
    "asset_id": "1267462a-839e-4acd-aed6-8c88a6215016",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd7d-b0a3-7800-9d0e-46568d74224d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S9sh1__bgfirst_bg.png",
   "bg_asset_id": "ac5f1bce-ac1d-4444-ab7c-809ed6f14bb1",
   "bg_record_key": "S9sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S9sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:19:07.131079+00:00",
  "fingerprint": "f4976b5027cddc3595b19f6168c5508ed8419439d7281a6390a8d5aaf657661e",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S9sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S9sh1_sel.png",
  "source_sha256": "f52450faa260f7c4541f2907ccd454cee7252a73c27071fe4c4e5a81b94276ac",
  "file": "S9sh1_cine.png",
  "staged_sha256": "acd37832f8ae3feb78c5b60eb4cd00a386366ebe05ab23bbbdd0591e3337a1da",
  "latency_ms": 14227
 },
 "S9sh6::signage": {
  "fp": "dc9326f5ddb81689",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "신분증명서"
   },
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "재판관"
   }
  ],
  "dropped": []
 },
 "S9sh6": {
  "input_fingerprint": "791751361fcad1c0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 반으로 펼쳐진 지갑 내부의 재판관 사진이 부착된 신분증명서를 이현우가 매섭게 내려다보는 시선.\n\nLOCATION (lock): Beside the rubbish heaps outside the refugee tribunal at night, where the stolen wallet is examined. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open wallet (Held unfolded in 이현우's hands) — Its interior faces upward toward his gaze and the elevated shoulder camera; used as Frames the identification at a realistic hand-held scale; Judge's identification card (Visible inside the open wallet and unobstructed by fingers) — The photograph-bearing face is visible to the camera, showing 재판관's recorded portrait rather than his physical presence; used as Reveals whose wallet 이현우 has taken without inventing readable personal details; Surrounding trash piles (Remaining around 이현우 outside the tribunal); used as A soft peripheral background preserves continuity with his awakening.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established nighttime exposure while preserving local tonal legibility on the identification and hands, without suggesting that the wallet emits light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The opened wallet contains the judge's identification card. The desolate rubbish heaps remain outside the court in the nighttime darkness.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 반으로 펼쳐진 지갑 내부의 재판관 사진이 부착된 신분증명서를 이현우가 매섭게 내려다보는 시선.\n\nLOCATION (lock): Beside the rubbish heaps outside the refugee tribunal at night, where the stolen wallet is examined. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open wallet (Held unfolded in 이현우's hands) — Its interior faces upward toward his gaze and the elevated shoulder camera; used as Frames the identification at a realistic hand-held scale; Judge's identification card (Visible inside the open wallet and unobstructed by fingers) — The photograph-bearing face is visible to the camera, showing 재판관's recorded portrait rather than his physical presence; used as Reveals whose wallet 이현우 has taken without inventing readable personal details; Surrounding trash piles (Remaining around 이현우 outside the tribunal); used as A soft peripheral background preserves continuity with his awakening.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established nighttime exposure while preserving local tonal legibility on the identification and hands, without suggesting that the wallet emits light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The opened wallet contains the judge's identification card. The desolate rubbish heaps remain outside the court in the nighttime darkness.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 반으로 펼쳐진 지갑 내부의 재판관 사진이 부착된 신분증명서를 이현우가 매섭게 내려다보는 시선.\n\nLOCATION (lock): Beside the rubbish heaps outside the refugee tribunal at night, where the stolen wallet is examined. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open wallet (Held unfolded in 이현우's hands) — Its interior faces upward toward his gaze and the elevated shoulder camera; used as Frames the identification at a realistic hand-held scale; Judge's identification card (Visible inside the open wallet and unobstructed by fingers) — The photograph-bearing face is visible to the camera, showing 재판관's recorded portrait rather than his physical presence; used as Reveals whose wallet 이현우 has taken without inventing readable personal details; Surrounding trash piles (Remaining around 이현우 outside the tribunal); used as A soft peripheral background preserves continuity with his awakening.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established nighttime exposure while preserving local tonal legibility on the identification and hands, without suggesting that the wallet emits light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The opened wallet contains the judge's identification card. The desolate rubbish heaps remain outside the court in the nighttime darkness.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선이 손에 든 지갑을 향함.",
    "built_space": "야간의 쓰레기 더미와 부서진 가전제품들이 기준 이미지와 일치함.",
    "entities": "이현우, 펼쳐진 지갑, 재판관 사진과 한글 텍스트가 포함된 신분증명서.",
    "hard_violations": [
     "[gemini-pro] 프롬프트 텍스트('재판관') 화면 유출",
     "[gpt-high] 신분증의 '재판관' 등 읽을 수 있는 문자가 노출되어 무문자 조건을 위반합니다."
    ],
    "physics": "양손으로 지갑을 단단히 잡고 있음."
   },
   {
    "label": "B",
    "direction": "이현우의 시선이 지갑 내부의 신분증을 향함.",
    "built_space": "야간의 쓰레기 더미 배경이 기준 이미지의 장소와 일치함.",
    "entities": "이현우(측면), 펼쳐진 지갑, 인물 사진과 텍스트가 포함된 신분증명서.",
    "hard_violations": [
     "[gpt-high] 신분증에 판독 가능한 한글이 있어 이미지 어디에도 읽히는 글자를 두지 말라는 금지 조건을 위반합니다."
    ],
    "physics": "양손의 손가락으로 지갑의 가장자리를 받치고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 어깨 위 카메라(elevated shoulder camera) 시점과 지갑의 방향을 정확히 구현했으나, 금지된 텍스트가 생성된 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 카메라 앵글을 무시한 정면 구도이며, 프롬프트의 텍스트('재판관')가 화면에 그대로 유출되어 지시를 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 손에 든 지갑을 향함.",
        "built_space": "야간의 쓰레기 더미와 부서진 가전제품들이 기준 이미지와 일치함.",
        "entities": "이현우, 펼쳐진 지갑, 재판관 사진과 한글 텍스트가 포함된 신분증명서.",
        "hard_violations": [
         "프롬프트 텍스트('재판관') 화면 유출"
        ],
        "physics": "양손으로 지갑을 단단히 잡고 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 지갑 내부의 신분증을 향함.",
        "built_space": "야간의 쓰레기 더미 배경이 기준 이미지의 장소와 일치함.",
        "entities": "이현우(측면), 펼쳐진 지갑, 인물 사진과 텍스트가 포함된 신분증명서.",
        "hard_violations": [],
        "physics": "양손의 손가락으로 지갑의 가장자리를 받치고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 어깨 위 카메라(elevated shoulder camera) 시점과 지갑의 방향을 정확히 구현했으나, 금지된 텍스트가 생성된 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 카메라 앵글을 무시한 정면 구도이며, 프롬프트의 텍스트('재판관')가 화면에 그대로 유출되어 지시를 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 손에 든 지갑을 향함.",
        "built_space": "야간의 쓰레기 더미와 부서진 가전제품들이 기준 이미지와 일치함.",
        "entities": "이현우, 펼쳐진 지갑, 재판관 사진과 한글 텍스트가 포함된 신분증명서.",
        "hard_violations": [
         "프롬프트 텍스트('재판관') 화면 유출"
        ],
        "physics": "양손으로 지갑을 단단히 잡고 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 지갑 내부의 신분증을 향함.",
        "built_space": "야간의 쓰레기 더미 배경이 기준 이미지의 장소와 일치함.",
        "entities": "이현우(측면), 펼쳐진 지갑, 인물 사진과 텍스트가 포함된 신분증명서.",
        "hard_violations": [],
        "physics": "양손의 손가락으로 지갑의 가장자리를 받치고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지갑을 향한 하향 시선과 손에 든 신분증의 근접 구도는 더 충실하지만, 읽히는 글자가 금지 조건을 위반하고 얼굴까지 포함해 손·전완 중심의 프레이밍을 벗어납니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "신분증 앞면이 이현우가 아니라 카메라를 향해 독서 방향이 틀렸으며, 읽히는 글자와 정면 상반신 위주의 넓은 구도도 요청에 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 위 이현우의 눈은 왼쪽 아래 지갑의 신분증 쪽을 내려다봅니다. 펼친 지갑 안쪽은 위쪽으로 기울어 얼굴과 비스듬한 상부 카메라에 함께 보이는 방향이어서 읽는 관계가 성립합니다. 다만 매서운 표정보다는 집중한 옆얼굴에 가깝습니다.",
        "built_space": "쓰레기봉투, 부서진 흰 금속 틀과 폐가전이 지갑 뒤를 채웁니다. 중앙에 회색 화면 하나, 오른쪽에 회색 폐가전과 하단의 어두운 화면이 보여 참고 장면과 폐기물 배치는 다릅니다. 건물의 고정 설비는 이 크롭에서 확인하기 어렵습니다. 밤의 폐기물 더미라는 장소감은 유지하지만 배경이 요청한 부드러운 주변부보다 선명하고 넓습니다.",
        "entities": "낡은 갈색 접이식 지갑 하나와 오른쪽 투명 칸의 사진 신분증 하나가 있습니다. 사진은 정장 차림 중년 동아시아계 남성의 기록 초상이며 실제 재판관 몸이 추가된 것은 아닙니다. 이현우의 검은 머리, 젊은 한국계 남성 외모, 더러운 손과 해진 짙은 회색 소매는 참고와 대체로 맞습니다. 얼굴과 어깨까지 들어와 손목·전완까지만 보이라는 제한은 지키지 못합니다. 신분증 상단의 법원 관련 한글과 일부 항목은 읽히며, 오른손 엄지가 카드 일부를 가립니다.",
        "hard_violations": [
         "신분증에 판독 가능한 한글이 있어 이미지 어디에도 읽히는 글자를 두지 말라는 금지 조건을 위반합니다."
        ],
        "physics": "왼손이 지갑 아래쪽과 접힘 부위를 받치고 오른손이 오른쪽 가장자리를 잡습니다. 신분증은 투명 수납칸에 고정되어 있으며 떠 있는 물체는 없습니다. 손목과 팔의 연결도 가능한 자세이고, 지갑 표면은 외부 조명을 반사할 뿐 스스로 빛나지 않습니다."
       },
       {
        "label": "B",
        "direction": "이현우는 눈썹을 찌푸리고 지갑 쪽으로 내려다보지만, 신분증 사진이 있는 안쪽 면은 정면 카메라를 향합니다. 이현우는 그 뒤에 있어 사진 면을 정상적으로 읽을 수 없습니다. 위로 향한 지갑을 어깨 위에서 내려다보는 관계가 아니라 관객에게 신분증을 펼쳐 보이는 관계입니다.",
        "built_space": "양옆 쓰레기 더미, 뒤 울타리와 야산, 오른쪽 건물 외벽이 보이며 오른쪽에는 회색 CRT 한 대와 그 앞의 갈색 TV 한 대가 있습니다. 이 주요 배치는 이전 장면과 가깝습니다. 다만 카메라는 인물 정면의 낮은 축이며, 얼굴·어깨·상반신과 장소를 크게 보여 지갑 중심 근접 프레이밍을 벗어납니다.",
        "entities": "지갑 하나 안에 정장 차림 중년 동아시아계 남성의 증명사진이 든 카드 하나가 있습니다. 재판관은 사진으로만 존재합니다. 이현우의 헝클어진 검은 머리, 젊은 한국계 남성 외모, 오염된 회색 후드와 손은 참고에 대체로 부합합니다. 그러나 얼굴과 몸통까지 포함되어 손·전완 제한과 맞지 않습니다. 카드의 '재판관' 글자가 명확히 읽히고 손가락이 카드 하단을 가립니다.",
        "hard_violations": [
         "신분증의 '재판관' 등 읽을 수 있는 문자가 노출되어 무문자 조건을 위반합니다."
        ],
        "physics": "양손이 지갑의 양쪽 아래를 잡고 있어 지갑은 충분히 지지됩니다. 카드는 수납칸 안에 있으며 손이나 물체가 공중에 뜨는 문제는 없습니다. 이 자세로 지갑을 관객에게 보여주는 것은 가능하지만, 자신이 사진 면을 읽는 행동으로는 방향이 맞지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지갑을 향한 하향 시선과 손에 든 신분증의 근접 구도는 더 충실하지만, 읽히는 글자가 금지 조건을 위반하고 얼굴까지 포함해 손·전완 중심의 프레이밍을 벗어납니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "신분증 앞면이 이현우가 아니라 카메라를 향해 독서 방향이 틀렸으며, 읽히는 글자와 정면 상반신 위주의 넓은 구도도 요청에 어긋납니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 위 이현우의 눈은 왼쪽 아래 지갑의 신분증 쪽을 내려다봅니다. 펼친 지갑 안쪽은 위쪽으로 기울어 얼굴과 비스듬한 상부 카메라에 함께 보이는 방향이어서 읽는 관계가 성립합니다. 다만 매서운 표정보다는 집중한 옆얼굴에 가깝습니다.",
        "built_space": "쓰레기봉투, 부서진 흰 금속 틀과 폐가전이 지갑 뒤를 채웁니다. 중앙에 회색 화면 하나, 오른쪽에 회색 폐가전과 하단의 어두운 화면이 보여 참고 장면과 폐기물 배치는 다릅니다. 건물의 고정 설비는 이 크롭에서 확인하기 어렵습니다. 밤의 폐기물 더미라는 장소감은 유지하지만 배경이 요청한 부드러운 주변부보다 선명하고 넓습니다.",
        "entities": "낡은 갈색 접이식 지갑 하나와 오른쪽 투명 칸의 사진 신분증 하나가 있습니다. 사진은 정장 차림 중년 동아시아계 남성의 기록 초상이며 실제 재판관 몸이 추가된 것은 아닙니다. 이현우의 검은 머리, 젊은 한국계 남성 외모, 더러운 손과 해진 짙은 회색 소매는 참고와 대체로 맞습니다. 얼굴과 어깨까지 들어와 손목·전완까지만 보이라는 제한은 지키지 못합니다. 신분증 상단의 법원 관련 한글과 일부 항목은 읽히며, 오른손 엄지가 카드 일부를 가립니다.",
        "hard_violations": [
         "신분증에 판독 가능한 한글이 있어 이미지 어디에도 읽히는 글자를 두지 말라는 금지 조건을 위반합니다."
        ],
        "physics": "왼손이 지갑 아래쪽과 접힘 부위를 받치고 오른손이 오른쪽 가장자리를 잡습니다. 신분증은 투명 수납칸에 고정되어 있으며 떠 있는 물체는 없습니다. 손목과 팔의 연결도 가능한 자세이고, 지갑 표면은 외부 조명을 반사할 뿐 스스로 빛나지 않습니다."
       },
       {
        "label": "A",
        "direction": "이현우는 눈썹을 찌푸리고 지갑 쪽으로 내려다보지만, 신분증 사진이 있는 안쪽 면은 정면 카메라를 향합니다. 이현우는 그 뒤에 있어 사진 면을 정상적으로 읽을 수 없습니다. 위로 향한 지갑을 어깨 위에서 내려다보는 관계가 아니라 관객에게 신분증을 펼쳐 보이는 관계입니다.",
        "built_space": "양옆 쓰레기 더미, 뒤 울타리와 야산, 오른쪽 건물 외벽이 보이며 오른쪽에는 회색 CRT 한 대와 그 앞의 갈색 TV 한 대가 있습니다. 이 주요 배치는 이전 장면과 가깝습니다. 다만 카메라는 인물 정면의 낮은 축이며, 얼굴·어깨·상반신과 장소를 크게 보여 지갑 중심 근접 프레이밍을 벗어납니다.",
        "entities": "지갑 하나 안에 정장 차림 중년 동아시아계 남성의 증명사진이 든 카드 하나가 있습니다. 재판관은 사진으로만 존재합니다. 이현우의 헝클어진 검은 머리, 젊은 한국계 남성 외모, 오염된 회색 후드와 손은 참고에 대체로 부합합니다. 그러나 얼굴과 몸통까지 포함되어 손·전완 제한과 맞지 않습니다. 카드의 '재판관' 글자가 명확히 읽히고 손가락이 카드 하단을 가립니다.",
        "hard_violations": [
         "신분증의 '재판관' 등 읽을 수 있는 문자가 노출되어 무문자 조건을 위반합니다."
        ],
        "physics": "양손이 지갑의 양쪽 아래를 잡고 있어 지갑은 충분히 지지됩니다. 카드는 수납칸 안에 있으며 손이나 물체가 공중에 뜨는 문제는 없습니다. 이 자세로 지갑을 관객에게 보여주는 것은 가능하지만, 자신이 사진 면을 읽는 행동으로는 방향이 맞지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.929,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.679,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트 텍스트('재판관') 화면 유출",
     "[gpt-high] 신분증의 '재판관' 등 읽을 수 있는 문자가 노출되어 무문자 조건을 위반합니다."
    ],
    "B": [
     "[gpt-high] 신분증에 판독 가능한 한글이 있어 이미지 어디에도 읽히는 글자를 두지 말라는 금지 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 679
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "요구된 어깨 위 카메라(elevated shoulder camera) 시점과 지갑의 방향을 정확히 구현했으나, 금지된 텍스트가 생성된 점이 아쉽습니다.  ★위반: [gpt-high] 신분증에 판독 가능한 한글이 있어 이미지 어디에도 읽히는 글자를 두지 말라는 금지 조건을 위반합니다."
   },
   {
    "label": "A",
    "score": 679,
    "verdict_ko": "지정된 카메라 앵글을 무시한 정면 구도이며, 프롬프트의 텍스트('재판관')가 화면에 그대로 유출되어 지시를 크게 위반했습니다.  ★위반: [gemini-pro] 프롬프트 텍스트('재판관') 화면 유출 / [gpt-high] 신분증의 '재판관' 등 읽을 수 있는 문자가 노출되어 무문자 조건을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S9sh1_sel.png",
    "asset_id": "a1354bce-79a3-4ecf-8817-6899a0d54a08",
    "role": "prev_still"
   },
   {
    "label": "HAND OWNER REFERENCE — 이현우: the person whose hand is on the object in this shot. Use this photograph ONLY to get that hand right — its size, build, skin tone and texture, apparent age, grooming, nails, sleeve and anything worn on the wrist or fingers. Do NOT bring their face, body or clothing into the frame; the framing stays on the object and reaches no further than the wrist and forearm.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd7e-9aed-7a01-9d14-9b18eb18e47e",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S9sh1"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S9sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:19:21.692772+00:00",
  "fingerprint": "f8655999cb393bfcc32bc392ee18abec140e4600ff11b9756f16fd048d7fd86d",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S9sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S9sh6_sel.png",
  "source_sha256": "b8733abeaeba9ad625c1ea42e42f180eacc1924a67a86b9e72a0129815a40e56",
  "file": "S9sh6_cine.png",
  "staged_sha256": "53a3fe24cc7710b8af44f82df49e0c5dc047dfc71ed17b4b711ae78cf73182f9",
  "latency_ms": 13882
 },
 "S9sh8::signage": {
  "fp": "af7ef3350b0775cb",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S9sh8": {
  "input_fingerprint": "4fded33eefa8e612",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 돈이 모두 빠져나간 빈 가죽 지갑이 이현우의 뻗은 손끝을 막 벗어나 쓰레기 더미 위 허공에 떠 있는 mid-action 찰나.\n\nLOCATION (lock): Over the rubbish-strewn ground outside the refugee tribunal at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Discarded leather wallet (Airborne immediately after release, with the money removed) — The opening and an oblique outer face are visible beside the fingertips; used as A small gap from the fingers makes the completed release legible; Trash heap (Below the released wallet outside the refugee tribunal); used as Provides the landing context beneath the hand without obscuring the airborne interval.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure preserves the separation between fingers, wallet, and trash without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the same refuse-covered ground and nighttime lighting from the reference. EXCLUDE any momentarily displaced rubbish from the awakening action and any courtroom furniture.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The discarded wallet has been emptied of money, but the judge's identification card remains inside. Rubbish heaps surround the court exterior at night. 이현우: Hyunwoo has kept the extracted cash and released the wallet. He remains alert but limps on the injured leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 돈이 모두 빠져나간 빈 가죽 지갑이 이현우의 뻗은 손끝을 막 벗어나 쓰레기 더미 위 허공에 떠 있는 mid-action 찰나.\n\nLOCATION (lock): Over the rubbish-strewn ground outside the refugee tribunal at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Discarded leather wallet (Airborne immediately after release, with the money removed) — The opening and an oblique outer face are visible beside the fingertips; used as A small gap from the fingers makes the completed release legible; Trash heap (Below the released wallet outside the refugee tribunal); used as Provides the landing context beneath the hand without obscuring the airborne interval.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure preserves the separation between fingers, wallet, and trash without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the same refuse-covered ground and nighttime lighting from the reference. EXCLUDE any momentarily displaced rubbish from the awakening action and any courtroom furniture.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The discarded wallet has been emptied of money, but the judge's identification card remains inside. Rubbish heaps surround the court exterior at night. 이현우: Hyunwoo has kept the extracted cash and released the wallet. He remains alert but limps on the injured leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 돈이 모두 빠져나간 빈 가죽 지갑이 이현우의 뻗은 손끝을 막 벗어나 쓰레기 더미 위 허공에 떠 있는 mid-action 찰나.\n\nLOCATION (lock): Over the rubbish-strewn ground outside the refugee tribunal at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Discarded leather wallet (Airborne immediately after release, with the money removed) — The opening and an oblique outer face are visible beside the fingertips; used as A small gap from the fingers makes the completed release legible; Trash heap (Below the released wallet outside the refugee tribunal); used as Provides the landing context beneath the hand without obscuring the airborne interval.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure preserves the separation between fingers, wallet, and trash without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the same refuse-covered ground and nighttime lighting from the reference. EXCLUDE any momentarily displaced rubbish from the awakening action and any courtroom furniture.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The discarded wallet has been emptied of money, but the judge's identification card remains inside. Rubbish heaps surround the court exterior at night. 이현우: Hyunwoo has kept the extracted cash and released the wallet. He remains alert but limps on the injured leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "손끝은 쓰레기 더미 쪽 아래를 향하고 있으며, 지갑 역시 낙하 방향에 맞추어 위치함.",
    "built_space": "쓰레기 더미가 화면을 채우는 야간의 야외 공간으로, 이전 컷에 설정된 환경과 일치함.",
    "entities": "이현우의 낡은 후드티 소매와 손이 정확히 묘사되었고, 지갑은 제시된 소품 레퍼런스에 맞는 검은색 가죽 지갑으로 열려 있으며 내부에 신분증이 확인됨.",
    "hard_violations": [],
    "physics": "손가락을 쫙 편 직후의 손 자세와 그 아래로 떨어지는 지갑 사이의 간격(gap)이 자연스럽게 표현되어, 방출(launch) 액션과 중력이 물리적으로 완벽히 설명됨."
   },
   {
    "label": "B",
    "direction": "손바닥은 위를 향해 수평으로 뻗어 있고, 인물의 시선은 허공에 떠 있는 지갑을 향함.",
    "built_space": "야간 야외 쓰레기장 배경이 확인됨.",
    "entities": "이현우의 얼굴이 이전 컷을 그대로 복사한 구도로 등장하며, 지시된 소품 레퍼런스의 검은 지갑이 아닌 이전 컷의 갈색 지갑이 닫힌 상태로 쓰임.",
    "hard_violations": [
     "[gemini-pro] 지갑이 손바닥 위 허공에 떠 있으나 떨어뜨리는 방출(launch) 동작이 전혀 없어 물리적으로 불가능한 공중 부양 상태임"
    ],
    "physics": "손바닥을 위로 한 정적인 자세 위에 지갑이 수평으로 떠 있어, 지갑이 떨어지거나 던져졌다는 물리적 운동(launch)을 전혀 뒷받침하지 못함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "요구된 클로즈업 샷 스케일을 완벽히 준수하였고, 지정된 검은색 가죽 지갑 소품을 사용하여 손끝에서 방금 떨어져 나간 물리적 찰나를 매우 사실적으로 구현함."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "클로즈업 지정 프레임을 무시하고 프레임을 넓혔으며, 소품 레퍼런스(검은색 지갑)를 무시한 데다 지갑이 릴리스 동작 없이 허공에 부자연스럽게 떠 있음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "손끝은 쓰레기 더미 쪽 아래를 향하고 있으며, 지갑 역시 낙하 방향에 맞추어 위치함.",
        "built_space": "쓰레기 더미가 화면을 채우는 야간의 야외 공간으로, 이전 컷에 설정된 환경과 일치함.",
        "entities": "이현우의 낡은 후드티 소매와 손이 정확히 묘사되었고, 지갑은 제시된 소품 레퍼런스에 맞는 검은색 가죽 지갑으로 열려 있으며 내부에 신분증이 확인됨.",
        "hard_violations": [],
        "physics": "손가락을 쫙 편 직후의 손 자세와 그 아래로 떨어지는 지갑 사이의 간격(gap)이 자연스럽게 표현되어, 방출(launch) 액션과 중력이 물리적으로 완벽히 설명됨."
       },
       {
        "label": "B",
        "direction": "손바닥은 위를 향해 수평으로 뻗어 있고, 인물의 시선은 허공에 떠 있는 지갑을 향함.",
        "built_space": "야간 야외 쓰레기장 배경이 확인됨.",
        "entities": "이현우의 얼굴이 이전 컷을 그대로 복사한 구도로 등장하며, 지시된 소품 레퍼런스의 검은 지갑이 아닌 이전 컷의 갈색 지갑이 닫힌 상태로 쓰임.",
        "hard_violations": [
         "지갑이 손바닥 위 허공에 떠 있으나 떨어뜨리는 방출(launch) 동작이 전혀 없어 물리적으로 불가능한 공중 부양 상태임"
        ],
        "physics": "손바닥을 위로 한 정적인 자세 위에 지갑이 수평으로 떠 있어, 지갑이 떨어지거나 던져졌다는 물리적 운동(launch)을 전혀 뒷받침하지 못함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "요구된 클로즈업 샷 스케일을 완벽히 준수하였고, 지정된 검은색 가죽 지갑 소품을 사용하여 손끝에서 방금 떨어져 나간 물리적 찰나를 매우 사실적으로 구현함."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "클로즈업 지정 프레임을 무시하고 프레임을 넓혔으며, 소품 레퍼런스(검은색 지갑)를 무시한 데다 지갑이 릴리스 동작 없이 허공에 부자연스럽게 떠 있음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "손끝은 쓰레기 더미 쪽 아래를 향하고 있으며, 지갑 역시 낙하 방향에 맞추어 위치함.",
        "built_space": "쓰레기 더미가 화면을 채우는 야간의 야외 공간으로, 이전 컷에 설정된 환경과 일치함.",
        "entities": "이현우의 낡은 후드티 소매와 손이 정확히 묘사되었고, 지갑은 제시된 소품 레퍼런스에 맞는 검은색 가죽 지갑으로 열려 있으며 내부에 신분증이 확인됨.",
        "hard_violations": [],
        "physics": "손가락을 쫙 편 직후의 손 자세와 그 아래로 떨어지는 지갑 사이의 간격(gap)이 자연스럽게 표현되어, 방출(launch) 액션과 중력이 물리적으로 완벽히 설명됨."
       },
       {
        "label": "B",
        "direction": "손바닥은 위를 향해 수평으로 뻗어 있고, 인물의 시선은 허공에 떠 있는 지갑을 향함.",
        "built_space": "야간 야외 쓰레기장 배경이 확인됨.",
        "entities": "이현우의 얼굴이 이전 컷을 그대로 복사한 구도로 등장하며, 지시된 소품 레퍼런스의 검은 지갑이 아닌 이전 컷의 갈색 지갑이 닫힌 상태로 쓰임.",
        "hard_violations": [
         "지갑이 손바닥 위 허공에 떠 있으나 떨어뜨리는 방출(launch) 동작이 전혀 없어 물리적으로 불가능한 공중 부양 상태임"
        ],
        "physics": "손바닥을 위로 한 정적인 자세 위에 지갑이 수평으로 떠 있어, 지갑이 떨어지거나 던져졌다는 물리적 운동(launch)을 전혀 뒷받침하지 못함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "손에서 놓인 지갑과 야간 쓰레기장은 구현했지만, 이전 컷처럼 얼굴과 상체까지 담아 손끝·지갑 중심의 클로즈업에서 벗어나며 검은 지갑과 남겨진 신분증도 충실하지 않습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "손끝과 지갑 사이의 작은 틈, 빈 지폐 수납부와 남은 신분증, 바로 아래 쓰레기 더미를 클로즈업으로 담아 요구한 방출 직후의 찰나를 가장 정확히 구현합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 인물의 눈과 왼쪽으로 뻗은 손가락이 중앙 왼쪽 지갑을 향합니다. 지갑은 손끝에서 왼쪽으로 떨어져 비스듬히 펼쳐져 있으며, 아래에는 쓰레기 더미가 있습니다. 카메라에는 지갑의 안쪽 넓은 면이 주로 보이고 요구한 비스듬한 바깥면과 수납부 입구는 분명하지 않습니다.",
        "built_space": "검은 봉투, 투명 봉투, 골판지, 금속 프레임과 폐모니터가 쌓인 야외 공간입니다. 중앙에 회색 폐모니터 하나, 오른쪽에 위아래로 폐모니터 두 개가 식별되어 이전 컷의 배치를 강하게 유지합니다. 법정 가구나 추가 인물은 없습니다. 다만 오른쪽 얼굴과 상체, 넓은 쓰레기장까지 보여 손끝과 지갑을 위한 클로즈업보다 넓습니다.",
        "entities": "한 명의 젊은 한국계 남성 얼굴 일부와 검은 헝클어진 머리, 오염된 피부, 해진 짙은 소매가 보여 이현우의 이전 모습과 잘 맞습니다. 지갑은 하나이며 낡은 갈색 가죽으로 이전 컷과는 가깝지만 지정된 검은 가죽 소품과는 다릅니다. 지폐는 보이지 않으나, 펼쳐진 내부에서 판사의 사진 신분증도 확인되지 않습니다. 카드 슬롯의 얇은 카드 가장자리만 보입니다. 읽을 수 있는 문구는 없습니다.",
        "hard_violations": [],
        "physics": "손목은 화면 오른쪽 팔에 연결되고 손가락은 쥔 상태에서 풀린 모양입니다. 지갑은 손과 접촉하지 않지만 가까운 열린 손이 방출 출발점을 제공하고 바로 아래 쓰레기가 낙하 지점을 제공합니다. 따라서 던져 놓은 직후의 비행으로 설명 가능하며, 근거 없이 정지 부유하는 물체로 볼 필요는 없습니다. 쓰레기들은 더미와 지면에 받쳐져 있습니다."
       },
       {
        "label": "B",
        "direction": "왼쪽 위에서 들어온 손의 손가락들이 오른쪽 아래 지갑을 향해 느슨하게 펴져 있습니다. 지갑은 손끝 바로 아래에서 비스듬히 내려가는 배치이며 아래 쓰레기 더미가 낙하 대상입니다. 카메라에는 벌어진 지폐 수납부와 신분증 면이 보이고, 바깥면은 가장자리 위주로 제한적으로 보입니다. 얼굴과 시선은 프레임 밖입니다.",
        "built_space": "손과 지갑이 전경을 차지하고 아래와 뒤에 검은 봉투, 투명 봉투, 골판지, 금속 잔해가 쌓여 있습니다. 오른쪽에는 회색 폐모니터 하나가 보이며 뒤쪽에는 망가진 프레임이 있습니다. 이전 컷의 폐기물 재질과 어두운 야외 분위기를 유지하면서 카메라를 손·지갑 중심으로 옮겼습니다. 법정 가구, 중복된 고정 설비, 반사는 보이지 않습니다.",
        "entities": "인물은 손과 해진 짙은 소매만 보이며 피부색과 오염 상태는 이현우의 참조와 양립합니다. 얼굴이 없어 정확한 나이와 얼굴 정체성은 확인할 수 없습니다. 검은 가죽 접이식 지갑 하나에 빈 지폐 수납부와 사진 신분증이 보입니다. 신분증의 작은 글자는 비스듬하고 흐려 내용을 확실히 읽을 수 없습니다. 소품 참조의 검은 가죽과 수납 구조에 A보다 가깝습니다. 보관한 현금과 다친 다리는 올바른 크롭 밖이므로 평가 대상이 아닙니다.",
        "hard_violations": [],
        "physics": "손은 화면 밖으로 이어지는 손목과 팔에 연결됩니다. 지갑과 손끝 사이에는 분명한 작은 공기 간격이 있고 손가락이 펴져 있어 방금 놓은 동작의 출발점이 확인됩니다. 지갑은 그 아래에서 회전하며 떨어질 수 있는 자세이고 착지할 쓰레기 더미도 바로 밑에 있어 비행의 물리적 맥락이 자연스럽습니다. 신분증은 지갑 포켓에 고정되어 있으며 쓰레기는 서로와 지면에 받쳐져 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "손에서 놓인 지갑과 야간 쓰레기장은 구현했지만, 이전 컷처럼 얼굴과 상체까지 담아 손끝·지갑 중심의 클로즈업에서 벗어나며 검은 지갑과 남겨진 신분증도 충실하지 않습니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "손끝과 지갑 사이의 작은 틈, 빈 지폐 수납부와 남은 신분증, 바로 아래 쓰레기 더미를 클로즈업으로 담아 요구한 방출 직후의 찰나를 가장 정확히 구현합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 인물의 눈과 왼쪽으로 뻗은 손가락이 중앙 왼쪽 지갑을 향합니다. 지갑은 손끝에서 왼쪽으로 떨어져 비스듬히 펼쳐져 있으며, 아래에는 쓰레기 더미가 있습니다. 카메라에는 지갑의 안쪽 넓은 면이 주로 보이고 요구한 비스듬한 바깥면과 수납부 입구는 분명하지 않습니다.",
        "built_space": "검은 봉투, 투명 봉투, 골판지, 금속 프레임과 폐모니터가 쌓인 야외 공간입니다. 중앙에 회색 폐모니터 하나, 오른쪽에 위아래로 폐모니터 두 개가 식별되어 이전 컷의 배치를 강하게 유지합니다. 법정 가구나 추가 인물은 없습니다. 다만 오른쪽 얼굴과 상체, 넓은 쓰레기장까지 보여 손끝과 지갑을 위한 클로즈업보다 넓습니다.",
        "entities": "한 명의 젊은 한국계 남성 얼굴 일부와 검은 헝클어진 머리, 오염된 피부, 해진 짙은 소매가 보여 이현우의 이전 모습과 잘 맞습니다. 지갑은 하나이며 낡은 갈색 가죽으로 이전 컷과는 가깝지만 지정된 검은 가죽 소품과는 다릅니다. 지폐는 보이지 않으나, 펼쳐진 내부에서 판사의 사진 신분증도 확인되지 않습니다. 카드 슬롯의 얇은 카드 가장자리만 보입니다. 읽을 수 있는 문구는 없습니다.",
        "hard_violations": [],
        "physics": "손목은 화면 오른쪽 팔에 연결되고 손가락은 쥔 상태에서 풀린 모양입니다. 지갑은 손과 접촉하지 않지만 가까운 열린 손이 방출 출발점을 제공하고 바로 아래 쓰레기가 낙하 지점을 제공합니다. 따라서 던져 놓은 직후의 비행으로 설명 가능하며, 근거 없이 정지 부유하는 물체로 볼 필요는 없습니다. 쓰레기들은 더미와 지면에 받쳐져 있습니다."
       },
       {
        "label": "A",
        "direction": "왼쪽 위에서 들어온 손의 손가락들이 오른쪽 아래 지갑을 향해 느슨하게 펴져 있습니다. 지갑은 손끝 바로 아래에서 비스듬히 내려가는 배치이며 아래 쓰레기 더미가 낙하 대상입니다. 카메라에는 벌어진 지폐 수납부와 신분증 면이 보이고, 바깥면은 가장자리 위주로 제한적으로 보입니다. 얼굴과 시선은 프레임 밖입니다.",
        "built_space": "손과 지갑이 전경을 차지하고 아래와 뒤에 검은 봉투, 투명 봉투, 골판지, 금속 잔해가 쌓여 있습니다. 오른쪽에는 회색 폐모니터 하나가 보이며 뒤쪽에는 망가진 프레임이 있습니다. 이전 컷의 폐기물 재질과 어두운 야외 분위기를 유지하면서 카메라를 손·지갑 중심으로 옮겼습니다. 법정 가구, 중복된 고정 설비, 반사는 보이지 않습니다.",
        "entities": "인물은 손과 해진 짙은 소매만 보이며 피부색과 오염 상태는 이현우의 참조와 양립합니다. 얼굴이 없어 정확한 나이와 얼굴 정체성은 확인할 수 없습니다. 검은 가죽 접이식 지갑 하나에 빈 지폐 수납부와 사진 신분증이 보입니다. 신분증의 작은 글자는 비스듬하고 흐려 내용을 확실히 읽을 수 없습니다. 소품 참조의 검은 가죽과 수납 구조에 A보다 가깝습니다. 보관한 현금과 다친 다리는 올바른 크롭 밖이므로 평가 대상이 아닙니다.",
        "hard_violations": [],
        "physics": "손은 화면 밖으로 이어지는 손목과 팔에 연결됩니다. 지갑과 손끝 사이에는 분명한 작은 공기 간격이 있고 손가락이 펴져 있어 방금 놓은 동작의 출발점이 확인됩니다. 지갑은 그 아래에서 회전하며 떨어질 수 있는 자세이고 착지할 쓰레기 더미도 바로 밑에 있어 비행의 물리적 맥락이 자연스럽습니다. 신분증은 지갑 포켓에 고정되어 있으며 쓰레기는 서로와 지면에 받쳐져 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.889
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.639
   },
   "violations": {
    "B": [
     "[gemini-pro] 지갑이 손바닥 위 허공에 떠 있으나 떨어뜨리는 방출(launch) 동작이 전혀 없어 물리적으로 불가능한 공중 부양 상태임"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 639
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "요구된 클로즈업 샷 스케일을 완벽히 준수하였고, 지정된 검은색 가죽 지갑 소품을 사용하여 손끝에서 방금 떨어져 나간 물리적 찰나를 매우 사실적으로 구현함."
   },
   {
    "label": "B",
    "score": 639,
    "verdict_ko": "클로즈업 지정 프레임을 무시하고 프레임을 넓혔으며, 소품 레퍼런스(검은색 지갑)를 무시한 데다 지갑이 릴리스 동작 없이 허공에 부자연스럽게 떠 있음.  ★위반: [gemini-pro] 지갑이 손바닥 위 허공에 떠 있으나 떨어뜨리는 방출(launch) 동작이 전혀 없어 물리적으로 불가능한 공중 부양 상태임"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S9sh6_sel.png",
    "asset_id": "1211d86b-28f9-49c2-8cab-4fa474116bcb",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 검은 가죽 지갑: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1061616>",
    "asset_id": "acf650b0-0bc3-4590-9df1-59326259da0b",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd7f-7e8a-72fc-a0f2-8222145d2679",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S9sh6"
  }
 },
 "S9sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:19:35.920507+00:00",
  "fingerprint": "5e71b4ec4bb2b5845270ff833b8f635652086cd7b504c460309f83da38970221",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S9sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S9sh8_sel.png",
  "source_sha256": "1af75cdfdfcea6de3e09935d38a2f8e79d11a9f987639096e6910fd18c084516",
  "file": "S9sh8_cine.png",
  "staged_sha256": "90d31410e21df09c5fe241509149877c9aa807c4cd0c0650e153eb18ba13054a",
  "latency_ms": 12490
 },
 "S10sh1::signage": {
  "fp": "554f1e411ed29716",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::fc9a47aad2f7eaba": {
  "subjects": [],
  "subject_text": "난민촌 골목 식당 내부\n골목에 자리한 허름한 식당 실내. 낡은 식탁과 좌석이 있으며 입구에서 떨어진 안쪽 구석에도 자리가 마련되어 있다.",
  "identity": "canonical",
  "scope_id": "L19",
  "scope_role": "location_interior",
  "scope_sha": "e6160448485567d0"
 },
 "S10sh1::bgfirst_bg": {
  "input_fingerprint": "392cb99ab4614679",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 식당 안 외진 구석 탁자, 40대 아시아계 남성이 한 손으로 턱을 괸 채 무심하게 앉아있는 구도.\n\nLOCATION (lock): At a secluded corner table inside a shabby alley restaurant, under modest nighttime interior lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Secluded restaurant table (Occupied on one side before 이현우 arrives) — The near edge runs diagonally toward the seated man, with the opposite seating area visible; used as Establishes the exchange axis and the empty place that 이현우 will occupy; Restaurant corner (A secluded area within the shabby restaurant) — The corner is viewed obliquely from the aisle; used as Contains the man in the rear of the composition while leaving approach space in the foreground.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime restaurant keeps the face readable without introducing a specific fixture or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 식당 안 외진 구석 탁자, 40대 아시아계 남성이 한 손으로 턱을 괸 채 무심하게 앉아있는 구도.\n\nLOCATION (lock): At a secluded corner table inside a shabby alley restaurant, under modest nighttime interior lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Secluded restaurant table (Occupied on one side before 이현우 arrives) — The near edge runs diagonally toward the seated man, with the opposite seating area visible; used as Establishes the exchange axis and the empty place that 이현우 will occupy; Restaurant corner (A secluded area within the shabby restaurant) — The corner is viewed obliquely from the aisle; used as Contains the man in the rear of the composition while leaving approach space in the foreground.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime restaurant keeps the face readable without introducing a specific fixture or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S10sh1__bgfirst_bg.png",
  "asset_id": "52fe0f46-bddf-4a55-ab89-784544fa436f",
  "input_asset_ids": [
   "8b93e4f8-749e-40c1-b16a-ac2ae46adec8",
   "f387463f-e36c-4575-8907-f3c62dcde529"
  ]
 },
 "S10sh1": {
  "input_fingerprint": "db064fe69645fa84",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 식당 안 외진 구석 탁자, 40대 아시아계 남성이 한 손으로 턱을 괸 채 무심하게 앉아있는 구도.\n\nLOCATION (lock): At a secluded corner table inside a shabby alley restaurant, under modest nighttime interior lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Secluded restaurant table (Occupied on one side before 이현우 arrives) — The near edge runs diagonally toward the seated man, with the opposite seating area visible; used as Establishes the exchange axis and the empty place that 이현우 will occupy; Restaurant corner (A secluded area within the shabby restaurant) — The corner is viewed obliquely from the aisle; used as Contains the man in the rear of the composition while leaving approach space in the foreground.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime restaurant keeps the face readable without introducing a specific fixture or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The meeting takes place at a secluded table inside the shabby alley restaurant at night.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 식당 안 외진 구석 탁자, 40대 아시아계 남성이 한 손으로 턱을 괸 채 무심하게 앉아있는 구도.\n\nLOCATION (lock): At a secluded corner table inside a shabby alley restaurant, under modest nighttime interior lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Secluded restaurant table (Occupied on one side before 이현우 arrives) — The near edge runs diagonally toward the seated man, with the opposite seating area visible; used as Establishes the exchange axis and the empty place that 이현우 will occupy; Restaurant corner (A secluded area within the shabby restaurant) — The corner is viewed obliquely from the aisle; used as Contains the man in the rear of the composition while leaving approach space in the foreground.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime restaurant keeps the face readable without introducing a specific fixture or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The meeting takes place at a secluded table inside the shabby alley restaurant at night.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 식당 안 외진 구석 탁자, 40대 아시아계 남성이 한 손으로 턱을 괸 채 무심하게 앉아있는 구도.\n\nLOCATION (lock): At a secluded corner table inside a shabby alley restaurant, under modest nighttime interior lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Secluded restaurant table (Occupied on one side before 이현우 arrives) — The near edge runs diagonally toward the seated man, with the opposite seating area visible; used as Establishes the exchange axis and the empty place that 이현우 will occupy; Restaurant corner (A secluded area within the shabby restaurant) — The corner is viewed obliquely from the aisle; used as Contains the man in the rear of the composition while leaving approach space in the foreground.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime restaurant keeps the face readable without introducing a specific fixture or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The meeting takes place at a secluded table inside the shabby alley restaurant at night.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S10sh1__bgfirst_bg.png",
     "asset_id": "52fe0f46-bddf-4a55-ab89-784544fa436f",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S10sh1.png",
     "asset_id": "8b93e4f8-749e-40c1-b16a-ac2ae46adec8",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1260199>",
     "asset_id": "ed57a831-f148-48bd-a65c-5b72dfdc874a",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L19B02.png",
     "asset_id": "f387463f-e36c-4575-8907-f3c62dcde529",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1260199>",
     "asset_id": "ed57a831-f148-48bd-a65c-5b72dfdc874a",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "남성의 시선은 아래쪽 허공을 향하며 무심한 분위기를 연출함.",
    "built_space": "레퍼런스와 동일한 식당 내부 구조. 우측 구석 테이블에 인물이 위치하고 전경에 다른 테이블들이 적절한 스케일로 배치됨.",
    "entities": "40대 아시아계 남성 1명. 샷 텍스트의 지시대로 이현우는 등장하지 않음.",
    "hard_violations": [],
    "physics": "남성의 체중이 의자에 실려 있으며, 한쪽 팔을 테이블과 몸에 지지한 채 손으로 턱을 안정적으로 받치고 있음."
   },
   {
    "label": "B",
    "direction": "앉아있는 남성은 정면 허공을 응시하고, 서 있는 남성은 화면 왼쪽을 향해 걸어감.",
    "built_space": "레퍼런스 사진의 식당 구조 및 가구 배치를 대체로 반영함.",
    "entities": "40대 아시아계 남성과 10대 아시아계 남성(이현우) 총 2명이 등장함.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 존재하지 않는 인물(이현우) 추가 등장",
     "[gpt-high] 숏에 지정되지 않은 젊은 남성 이현우를 통로에 추가했다.",
     "[gpt-high] 오른쪽 벽 메뉴판에 읽을 수 있는 음식명과 가격이 노출되어 문자 금지 조건을 위반한다."
    ],
    "physics": "앉은 인물과 걷고 있는 인물 모두 바닥과 의자에 올바르게 지지되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "샷 텍스트에 명시된 40대 남성만을 정확히 프레임에 담았으며, 지시된 식당 구석 자리와 턱을 괸 자세를 충실히 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "샷 텍스트에 묘사되지 않은 인물(이현우)이 화면에 등장하여, 텍스트에 명시된 인물만 렌더링해야 한다는 원칙을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남성의 시선은 아래쪽 허공을 향하며 무심한 분위기를 연출함.",
        "built_space": "레퍼런스와 동일한 식당 내부 구조. 우측 구석 테이블에 인물이 위치하고 전경에 다른 테이블들이 적절한 스케일로 배치됨.",
        "entities": "40대 아시아계 남성 1명. 샷 텍스트의 지시대로 이현우는 등장하지 않음.",
        "hard_violations": [],
        "physics": "남성의 체중이 의자에 실려 있으며, 한쪽 팔을 테이블과 몸에 지지한 채 손으로 턱을 안정적으로 받치고 있음."
       },
       {
        "label": "B",
        "direction": "앉아있는 남성은 정면 허공을 응시하고, 서 있는 남성은 화면 왼쪽을 향해 걸어감.",
        "built_space": "레퍼런스 사진의 식당 구조 및 가구 배치를 대체로 반영함.",
        "entities": "40대 아시아계 남성과 10대 아시아계 남성(이현우) 총 2명이 등장함.",
        "hard_violations": [
         "샷 텍스트에 존재하지 않는 인물(이현우) 추가 등장"
        ],
        "physics": "앉은 인물과 걷고 있는 인물 모두 바닥과 의자에 올바르게 지지되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "샷 텍스트에 명시된 40대 남성만을 정확히 프레임에 담았으며, 지시된 식당 구석 자리와 턱을 괸 자세를 충실히 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "샷 텍스트에 묘사되지 않은 인물(이현우)이 화면에 등장하여, 텍스트에 명시된 인물만 렌더링해야 한다는 원칙을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "남성의 시선은 아래쪽 허공을 향하며 무심한 분위기를 연출함.",
        "built_space": "레퍼런스와 동일한 식당 내부 구조. 우측 구석 테이블에 인물이 위치하고 전경에 다른 테이블들이 적절한 스케일로 배치됨.",
        "entities": "40대 아시아계 남성 1명. 샷 텍스트의 지시대로 이현우는 등장하지 않음.",
        "hard_violations": [],
        "physics": "남성의 체중이 의자에 실려 있으며, 한쪽 팔을 테이블과 몸에 지지한 채 손으로 턱을 안정적으로 받치고 있음."
       },
       {
        "label": "B",
        "direction": "앉아있는 남성은 정면 허공을 응시하고, 서 있는 남성은 화면 왼쪽을 향해 걸어감.",
        "built_space": "레퍼런스 사진의 식당 구조 및 가구 배치를 대체로 반영함.",
        "entities": "40대 아시아계 남성과 10대 아시아계 남성(이현우) 총 2명이 등장함.",
        "hard_violations": [
         "샷 텍스트에 존재하지 않는 인물(이현우) 추가 등장"
        ],
        "physics": "앉은 인물과 걷고 있는 인물 모두 바닥과 의자에 올바르게 지지되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "중년 남성의 턱 괴기는 맞지만, 등장하면 안 되는 이현우를 추가했고 메뉴와 가격이 선명하게 읽혀 필수 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "외진 구석의 중년 남성 한 명과 빈 맞은편 자리, 전경의 접근 통로를 와이드 숏으로 구현했으나 손이 턱보다 뺨을 받치는 점은 아쉽다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앉은 남성은 화면 오른쪽 위의 벽 쪽을 바라보며 한 손으로 턱을 괸다. 왼쪽 청년 역시 오른쪽을 보지만 중년 남성과 눈을 맞추지는 않는다. 청년의 몸과 발은 통로를 따라 카메라 쪽으로 향한다. 지시문에는 청년의 접근 동작이 없다.",
        "built_space": "오른쪽 후면 구석 탁자, 오른쪽 전경 탁자, 왼쪽 중간 탁자와 왼쪽 가장자리의 탁자 일부가 보인다. 구석 탁자의 뒤편에 중년 남성이 앉고 카메라 쪽에는 빈 의자가 있다. 왼쪽 냉장고 한 대, 후면 커튼 출입구 한 곳, 붉은 벽 장치 하나와 천장 전구 하나가 보여 참고 장소의 주요 설비를 유지한다. 다만 전경 탁자가 크게 차지하고 청년이 접근 통로를 점유한다.",
        "entities": "40대 정도로 보이는 아시아계 남성 한 명은 요구에 맞는다. 검은 헝클어진 머리와 짙은 반팔 차림의 젊은 아시아계 남성도 추가되어 인물 참고의 이현우를 닮았지만, 이 숏에는 있어서는 안 된다. 낡은 목재 탁자와 갈색 의자, 휴지함과 양념통은 참고 장소와 부합한다. 오른쪽 메뉴판의 음식명과 가격이 명확하게 읽힌다.",
        "hard_violations": [
         "숏에 지정되지 않은 젊은 남성 이현우를 통로에 추가했다.",
         "오른쪽 벽 메뉴판에 읽을 수 있는 음식명과 가격이 노출되어 문자 금지 조건을 위반한다."
        ],
        "physics": "중년 남성은 탁자 아래의 좌석에 앉아 있고, 턱을 받친 팔의 팔꿈치는 탁자에 놓여 있어 지지가 자연스럽다. 청년은 앞쪽 신발을 바닥에 딛고 걷는 자세로 보인다. 탁자와 의자는 다리로 바닥에 지지되며, 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "남성의 얼굴과 시선은 화면 왼쪽 아래의 탁자와 빈자리 방향으로 떨어져 있어 무심한 기다림으로 읽힌다. 한 손은 턱 아래보다는 뺨과 턱 옆을 받친다. 맞은편에 시선을 주고받는 다른 인물은 없다.",
        "built_space": "오른쪽 전경과 후면 구석에 탁자가 각각 하나, 왼쪽에 두 개로 총 네 개가 보인다. 구석 탁자 양쪽에 의자가 있고 남성은 오른쪽 벽 가까운 의자에 앉아 있으며 반대쪽 의자는 비어 있다. 중앙 통로가 전경부터 후면까지 열려 있다. 왼쪽 냉장고 한 대, 후면 커튼 출입구 한 곳, 붉은 벽 장치 하나, 천장 전구 하나와 낡은 투톤 벽이 참고 장소와 잘 맞는다. 통로에서 비스듬히 본 와이드 구도 안에 남성이 후경에 배치된다.",
        "entities": "짧은 검은 머리의 40대 정도로 보이는 아시아계 남성 한 명만 등장하여 숏의 인원과 연령 조건에 부합한다. 회갈색 작업복 같은 옷을 입었으며, 이 인물의 복장은 별도로 고정되지 않았다. 젊은 이현우는 등장하지 않는다. 목재 탁자, 갈색 금속틀 의자, 휴지함, 양념통과 냉장고가 참고 식당의 재질과 분위기를 유지한다. 뚜렷하게 읽히는 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "남성의 골반은 의자 좌판에 놓이고 등 뒤에는 등받이가 있으며, 다리는 탁자 아래로 자연스럽게 내려간다. 얼굴을 받친 손의 팔꿈치는 탁자에 지지되고 다른 손은 허벅지 위에 놓인다. 발은 전경 가구에 가려 직접 확인할 수 없다. 보이는 가구와 소품은 바닥이나 탁자에 지지되어 있으며 부유하는 요소는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "중년 남성의 턱 괴기는 맞지만, 등장하면 안 되는 이현우를 추가했고 메뉴와 가격이 선명하게 읽혀 필수 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "외진 구석의 중년 남성 한 명과 빈 맞은편 자리, 전경의 접근 통로를 와이드 숏으로 구현했으나 손이 턱보다 뺨을 받치는 점은 아쉽다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앉은 남성은 화면 오른쪽 위의 벽 쪽을 바라보며 한 손으로 턱을 괸다. 왼쪽 청년 역시 오른쪽을 보지만 중년 남성과 눈을 맞추지는 않는다. 청년의 몸과 발은 통로를 따라 카메라 쪽으로 향한다. 지시문에는 청년의 접근 동작이 없다.",
        "built_space": "오른쪽 후면 구석 탁자, 오른쪽 전경 탁자, 왼쪽 중간 탁자와 왼쪽 가장자리의 탁자 일부가 보인다. 구석 탁자의 뒤편에 중년 남성이 앉고 카메라 쪽에는 빈 의자가 있다. 왼쪽 냉장고 한 대, 후면 커튼 출입구 한 곳, 붉은 벽 장치 하나와 천장 전구 하나가 보여 참고 장소의 주요 설비를 유지한다. 다만 전경 탁자가 크게 차지하고 청년이 접근 통로를 점유한다.",
        "entities": "40대 정도로 보이는 아시아계 남성 한 명은 요구에 맞는다. 검은 헝클어진 머리와 짙은 반팔 차림의 젊은 아시아계 남성도 추가되어 인물 참고의 이현우를 닮았지만, 이 숏에는 있어서는 안 된다. 낡은 목재 탁자와 갈색 의자, 휴지함과 양념통은 참고 장소와 부합한다. 오른쪽 메뉴판의 음식명과 가격이 명확하게 읽힌다.",
        "hard_violations": [
         "숏에 지정되지 않은 젊은 남성 이현우를 통로에 추가했다.",
         "오른쪽 벽 메뉴판에 읽을 수 있는 음식명과 가격이 노출되어 문자 금지 조건을 위반한다."
        ],
        "physics": "중년 남성은 탁자 아래의 좌석에 앉아 있고, 턱을 받친 팔의 팔꿈치는 탁자에 놓여 있어 지지가 자연스럽다. 청년은 앞쪽 신발을 바닥에 딛고 걷는 자세로 보인다. 탁자와 의자는 다리로 바닥에 지지되며, 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "남성의 얼굴과 시선은 화면 왼쪽 아래의 탁자와 빈자리 방향으로 떨어져 있어 무심한 기다림으로 읽힌다. 한 손은 턱 아래보다는 뺨과 턱 옆을 받친다. 맞은편에 시선을 주고받는 다른 인물은 없다.",
        "built_space": "오른쪽 전경과 후면 구석에 탁자가 각각 하나, 왼쪽에 두 개로 총 네 개가 보인다. 구석 탁자 양쪽에 의자가 있고 남성은 오른쪽 벽 가까운 의자에 앉아 있으며 반대쪽 의자는 비어 있다. 중앙 통로가 전경부터 후면까지 열려 있다. 왼쪽 냉장고 한 대, 후면 커튼 출입구 한 곳, 붉은 벽 장치 하나, 천장 전구 하나와 낡은 투톤 벽이 참고 장소와 잘 맞는다. 통로에서 비스듬히 본 와이드 구도 안에 남성이 후경에 배치된다.",
        "entities": "짧은 검은 머리의 40대 정도로 보이는 아시아계 남성 한 명만 등장하여 숏의 인원과 연령 조건에 부합한다. 회갈색 작업복 같은 옷을 입었으며, 이 인물의 복장은 별도로 고정되지 않았다. 젊은 이현우는 등장하지 않는다. 목재 탁자, 갈색 금속틀 의자, 휴지함, 양념통과 냉장고가 참고 식당의 재질과 분위기를 유지한다. 뚜렷하게 읽히는 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "남성의 골반은 의자 좌판에 놓이고 등 뒤에는 등받이가 있으며, 다리는 탁자 아래로 자연스럽게 내려간다. 얼굴을 받친 손의 팔꿈치는 탁자에 지지되고 다른 손은 허벅지 위에 놓인다. 발은 전경 가구에 가려 직접 확인할 수 없다. 보이는 가구와 소품은 바닥이나 탁자에 지지되어 있으며 부유하는 요소는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.679
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.429
   },
   "violations": {
    "B": [
     "[gemini-pro] 샷 텍스트에 존재하지 않는 인물(이현우) 추가 등장",
     "[gpt-high] 숏에 지정되지 않은 젊은 남성 이현우를 통로에 추가했다.",
     "[gpt-high] 오른쪽 벽 메뉴판에 읽을 수 있는 음식명과 가격이 노출되어 문자 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "샷 텍스트에 명시된 40대 남성만을 정확히 프레임에 담았으며, 지시된 식당 구석 자리와 턱을 괸 자세를 충실히 구현했습니다."
   },
   {
    "label": "B",
    "score": 429,
    "verdict_ko": "샷 텍스트에 묘사되지 않은 인물(이현우)이 화면에 등장하여, 텍스트에 명시된 인물만 렌더링해야 한다는 원칙을 심각하게 위반했습니다.  ★위반: [gemini-pro] 샷 텍스트에 존재하지 않는 인물(이현우) 추가 등장 / [gpt-high] 숏에 지정되지 않은 젊은 남성 이현우를 통로에 추가했다. / [gpt-high] 오른쪽 벽 메뉴판에 읽을 수 있는 음식명과 가격이 노출되어 문자 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L19B02.png",
    "asset_id": "f387463f-e36c-4575-8907-f3c62dcde529",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1260199>",
    "asset_id": "ed57a831-f148-48bd-a65c-5b72dfdc874a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd80-4bc7-7cfa-84b4-0fa5973dbb12",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S10sh1__bgfirst_bg.png",
   "bg_asset_id": "52fe0f46-bddf-4a55-ab89-784544fa436f",
   "bg_record_key": "S10sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S10sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:19:48.822881+00:00",
  "fingerprint": "1e26034386373f9030fab8ea572132cade82b2e2e29aaae1e5d151953dc65fcd",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S10sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S10sh1_sel.png",
  "source_sha256": "008a68f7396d33a4024285f12e687ab23692549cd9157ecc86171c5ec6fee0fd",
  "file": "S10sh1_cine.png",
  "staged_sha256": "1e7d5a9dec323638e661f86ec0c134502792805515280f0e76d3c74088ad7019",
  "latency_ms": 10708
 },
 "S10sh5::signage": {
  "fp": "49154a72aa26b8ef",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S10sh5": {
  "input_fingerprint": "353fb9938a7693ee",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 40대 아시아계 남성이 이현우의 손에서 봉투를 거칠게 낚아채는 힘이 실린 순간.\n\nLOCATION (lock): At the secluded transaction table inside the shabby alley restaurant, in subdued interior light at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Envelope being pulled toward the recipient's side of the table in the middle-right of the frame, midground, moves toward recipient's hand at screen right.\n- KEY BACKGROUND ELEMENTS: Money envelope (Being pulled out of 이현우's grasp) — An oblique broad face is visible between the two hands; no readable markings are required; used as Links the opposing arm movements and locates the instant of lost control; Restaurant table (Supporting the ongoing money exchange) — Its top and near edge remain visible beneath the hands; used as Maintains spatial continuity and prevents exaggerated foreground scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the restaurant's subdued ambient treatment, using restrained contrast to distinguish the opposing hands and envelope.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the secluded restaurant table, its chairs, and the surrounding interior finishes and nighttime lighting. EXCLUDE any duplicate envelope or loose money; the transaction requires only the current envelope.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The money envelope contains Hyunwoo's savings, not a newly introduced payment. The meeting remains at the secluded restaurant table. 이현우: Hyunwoo remains seated at the table, with his leg injury and earlier arrest injuries untreated. He has just lost his renewed grip on the envelope.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 40대 아시아계 남성 right now, so 40대 아시아계 남성's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 40대 아시아계 남성: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 40대 아시아계 남성이 이현우의 손에서 봉투를 거칠게 낚아채는 힘이 실린 순간.\n\nLOCATION (lock): At the secluded transaction table inside the shabby alley restaurant, in subdued interior light at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Envelope being pulled toward the recipient's side of the table in the middle-right of the frame, midground, moves toward recipient's hand at screen right.\n- KEY BACKGROUND ELEMENTS: Money envelope (Being pulled out of 이현우's grasp) — An oblique broad face is visible between the two hands; no readable markings are required; used as Links the opposing arm movements and locates the instant of lost control; Restaurant table (Supporting the ongoing money exchange) — Its top and near edge remain visible beneath the hands; used as Maintains spatial continuity and prevents exaggerated foreground scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the restaurant's subdued ambient treatment, using restrained contrast to distinguish the opposing hands and envelope.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the secluded restaurant table, its chairs, and the surrounding interior finishes and nighttime lighting. EXCLUDE any duplicate envelope or loose money; the transaction requires only the current envelope.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The money envelope contains Hyunwoo's savings, not a newly introduced payment. The meeting remains at the secluded restaurant table. 이현우: Hyunwoo remains seated at the table, with his leg injury and earlier arrest injuries untreated. He has just lost his renewed grip on the envelope.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 40대 아시아계 남성 right now, so 40대 아시아계 남성's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 40대 아시아계 남성: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 40대 아시아계 남성이 이현우의 손에서 봉투를 거칠게 낚아채는 힘이 실린 순간.\n\nLOCATION (lock): At the secluded transaction table inside the shabby alley restaurant, in subdued interior light at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Envelope being pulled toward the recipient's side of the table in the middle-right of the frame, midground, moves toward recipient's hand at screen right.\n- KEY BACKGROUND ELEMENTS: Money envelope (Being pulled out of 이현우's grasp) — An oblique broad face is visible between the two hands; no readable markings are required; used as Links the opposing arm movements and locates the instant of lost control; Restaurant table (Supporting the ongoing money exchange) — Its top and near edge remain visible beneath the hands; used as Maintains spatial continuity and prevents exaggerated foreground scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the restaurant's subdued ambient treatment, using restrained contrast to distinguish the opposing hands and envelope.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the secluded restaurant table, its chairs, and the surrounding interior finishes and nighttime lighting. EXCLUDE any duplicate envelope or loose money; the transaction requires only the current envelope.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The money envelope contains Hyunwoo's savings, not a newly introduced payment. The meeting remains at the secluded restaurant table. 이현우: Hyunwoo remains seated at the table, with his leg injury and earlier arrest injuries untreated. He has just lost his renewed grip on the envelope.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 40대 아시아계 남성 right now, so 40대 아시아계 남성's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 40대 아시아계 남성: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "두 남성의 시선과 손의 방향이 테이블 중앙에 있는 봉투를 향해 집중되어 있습니다.",
    "built_space": "지정된 식당 내부와 테이블이 보이나, 배경 좌측에 프롬프트에 명시되지 않은 정체불명의 인물 두 명이 앉아 있습니다.",
    "entities": "이현우와 40대 남성의 기본적인 외형은 나타나지만, 배제해야 할 낱장의 지폐(loose money)가 봉투 밖으로 나와 있습니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 없는 인물(배경의 두 사람) 임의 추가",
     "[gemini-pro] 배제 지시된 지폐(loose money) 노출",
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (손이 얽히고 추가된 손가락 구조)",
     "[gpt-high] 장면에서 요구하지 않은 배경 인물 두 명을 추가했다.",
     "[gpt-high] 현재 봉투만 사용하고 별도 돈은 제외하라는 지시와 달리 봉투 밖으로 지폐를 노출했다."
    ],
    "physics": "봉투를 잡고 있는 인물들의 팔과 손 구조가 비정상적으로 엉켜 있으며, 지폐를 쥔 손의 출처가 불분명해 물리적으로 불가능한 형태입니다."
   },
   {
    "label": "B",
    "direction": "왼쪽의 이현우는 인상을 찌푸린 채 봉투를 바라보고, 오른쪽 남성은 이현우를 응시하며 봉투를 오른쪽으로 강하게 당기고 있습니다.",
    "built_space": "레퍼런스와 일치하는 식당 벽면, 문, 냉장고, 나무 테이블이 정확한 위치에 배치되었고 추가된 인물 없이 공간이 비어 있습니다.",
    "entities": "이현우의 얼굴과 복장이 레퍼런스와 정확히 일치하며, 40대 남성의 외형과 서류 봉투의 형태도 지시사항에 부합합니다.",
    "hard_violations": [],
    "physics": "이현우의 한 손은 테이블에, 다른 한 손은 봉투에 자연스럽게 놓여 있으며, 남성의 팔 역시 어깨로부터 이어져 정상적으로 봉투를 당기고 지지됩니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 인물들만 정확히 등장하며, 봉투를 낚아채는 행동과 지정된 장소의 레퍼런스를 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "배경에 지시되지 않은 인물이 추가되고 노출 금지된 지폐가 등장했으며, 봉투를 쥔 손의 해부학적 구조가 붕괴되는 심각한 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 남성의 시선과 손의 방향이 테이블 중앙에 있는 봉투를 향해 집중되어 있습니다.",
        "built_space": "지정된 식당 내부와 테이블이 보이나, 배경 좌측에 프롬프트에 명시되지 않은 정체불명의 인물 두 명이 앉아 있습니다.",
        "entities": "이현우와 40대 남성의 기본적인 외형은 나타나지만, 배제해야 할 낱장의 지폐(loose money)가 봉투 밖으로 나와 있습니다.",
        "hard_violations": [
         "프롬프트에 없는 인물(배경의 두 사람) 임의 추가",
         "배제 지시된 지폐(loose money) 노출",
         "물리적으로 불가능한 해부학적 구조 (손이 얽히고 추가된 손가락 구조)"
        ],
        "physics": "봉투를 잡고 있는 인물들의 팔과 손 구조가 비정상적으로 엉켜 있으며, 지폐를 쥔 손의 출처가 불분명해 물리적으로 불가능한 형태입니다."
       },
       {
        "label": "B",
        "direction": "왼쪽의 이현우는 인상을 찌푸린 채 봉투를 바라보고, 오른쪽 남성은 이현우를 응시하며 봉투를 오른쪽으로 강하게 당기고 있습니다.",
        "built_space": "레퍼런스와 일치하는 식당 벽면, 문, 냉장고, 나무 테이블이 정확한 위치에 배치되었고 추가된 인물 없이 공간이 비어 있습니다.",
        "entities": "이현우의 얼굴과 복장이 레퍼런스와 정확히 일치하며, 40대 남성의 외형과 서류 봉투의 형태도 지시사항에 부합합니다.",
        "hard_violations": [],
        "physics": "이현우의 한 손은 테이블에, 다른 한 손은 봉투에 자연스럽게 놓여 있으며, 남성의 팔 역시 어깨로부터 이어져 정상적으로 봉투를 당기고 지지됩니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 인물들만 정확히 등장하며, 봉투를 낚아채는 행동과 지정된 장소의 레퍼런스를 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "배경에 지시되지 않은 인물이 추가되고 노출 금지된 지폐가 등장했으며, 봉투를 쥔 손의 해부학적 구조가 붕괴되는 심각한 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 남성의 시선과 손의 방향이 테이블 중앙에 있는 봉투를 향해 집중되어 있습니다.",
        "built_space": "지정된 식당 내부와 테이블이 보이나, 배경 좌측에 프롬프트에 명시되지 않은 정체불명의 인물 두 명이 앉아 있습니다.",
        "entities": "이현우와 40대 남성의 기본적인 외형은 나타나지만, 배제해야 할 낱장의 지폐(loose money)가 봉투 밖으로 나와 있습니다.",
        "hard_violations": [
         "프롬프트에 없는 인물(배경의 두 사람) 임의 추가",
         "배제 지시된 지폐(loose money) 노출",
         "물리적으로 불가능한 해부학적 구조 (손이 얽히고 추가된 손가락 구조)"
        ],
        "physics": "봉투를 잡고 있는 인물들의 팔과 손 구조가 비정상적으로 엉켜 있으며, 지폐를 쥔 손의 출처가 불분명해 물리적으로 불가능한 형태입니다."
       },
       {
        "label": "B",
        "direction": "왼쪽의 이현우는 인상을 찌푸린 채 봉투를 바라보고, 오른쪽 남성은 이현우를 응시하며 봉투를 오른쪽으로 강하게 당기고 있습니다.",
        "built_space": "레퍼런스와 일치하는 식당 벽면, 문, 냉장고, 나무 테이블이 정확한 위치에 배치되었고 추가된 인물 없이 공간이 비어 있습니다.",
        "entities": "이현우의 얼굴과 복장이 레퍼런스와 정확히 일치하며, 40대 남성의 외형과 서류 봉투의 형태도 지시사항에 부합합니다.",
        "hard_violations": [],
        "physics": "이현우의 한 손은 테이블에, 다른 한 손은 봉투에 자연스럽게 놓여 있으며, 남성의 팔 역시 어깨로부터 이어져 정상적으로 봉투를 당기고 지지됩니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "오른쪽 남성이 봉투를 힘주어 당기는 관계와 장소·의상은 잘 맞지만, 클로즈업보다 넓고 현우가 아직 봉투를 단단히 잡고 있다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "배경에 추가 인물을 만들고 봉투 밖 지폐를 노출했으며, 거친 낚아채기보다 양손으로 내용물을 확인하는 거래처럼 보인다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 현우의 얼굴과 오른쪽 중년 남성의 시선이 두 손 사이 봉투를 향한다. 오른쪽 남성이 봉투 오른쪽을 움켜쥐고 자기 쪽으로 당기는 방향은 맞는다. 다만 현우의 왼쪽 화면상 손도 봉투 끝을 확실히 잡고 있어, 이미 놓친 순간보다는 서로 당기는 순간이다.",
        "built_space": "거래용 나무 탁자 하나가 전경을 채우고, 오른쪽 남성 뒤 의자와 봉투 뒤쪽의 작은 등받이 일부가 보인다. 왼쪽 냉장고 하나, 뒤쪽 무늬 커튼 하나, 천장 전구 하나, 낡은 투톤 벽과 당김줄이 이전 장소와 연결된다. 탁자 위 수저통과 작은 용기가 추가로 보인다. 두 사람은 탁자 양쪽에 앉아 있으나 상체와 얼굴까지 넓게 담겨 지정된 손 중심 클로즈업보다 넓다. 상판은 보이지만 카메라 쪽 끝단은 화면 밖이다.",
        "entities": "인물은 두 명이며 왼쪽은 검은 헝클어진 머리의 젊은 아시아계 남성으로 현우 참조와 대체로 맞고, 해진 짙은 후드와 얼굴 상처가 보인다. 오른쪽은 40대 정도의 아시아계 남성으로 이전 스틸의 회갈색 셔츠를 유지한다. 황갈색 종이봉투는 하나이고 두 손 사이에 비스듬한 넓은 면이 드러난다. 별도 지폐나 읽을 수 있는 글씨는 없다. 다리 부상은 프레임 밖이므로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "봉투는 양쪽 손의 실제 집기로 지지되고, 오른쪽 남성의 긴장된 손등과 뻗은 전완이 당기는 힘을 표현한다. 현우의 다른 손과 전완은 탁자에 닿아 있다. 몸통은 앉은 자세로 읽히며 공중에 뜬 물체나 불가능한 관절은 보이지 않는다. 다만 현우의 남은 집기는 '방금 손에서 빠져나감'과 일치하지 않는다."
       },
       {
        "label": "B",
        "direction": "오른쪽 남성은 손 사이의 봉투와 노출된 지폐를 내려다본다. 현우의 눈은 잘려 시선은 확인할 수 없다. 오른쪽 남성의 한 손은 봉투를 잡고 다른 손은 입구와 지폐에 닿아 있어, 오른쪽으로 강하게 낚아채기보다는 봉투를 열거나 돈을 확인하는 동작으로 보인다.",
        "built_space": "전경에 나무 탁자 하나, 왼쪽 뒤에 다른 탁자 일부와 냉장고 하나, 천장 전구 하나, 벽의 붉은 장치 하나 및 배선이 보인다. 낡은 벽과 어두운 조명은 장소 참조와 유사하다. 오른쪽 의자 등받이 일부가 보이고 두 주인물은 탁자를 사이에 둔다. 그러나 뒤쪽 탁자 구역에 별도 인물들이 들어와 있다. 얼굴을 일부 자른 구도는 A보다 가깝지만 상판이 넓게 차지하고 카메라 쪽 탁자 끝단은 보이지 않는다.",
        "entities": "전경에는 젊은 아시아계 남성과 중년 아시아계 남성이 있고, 각각 낡은 후드와 회갈색 셔츠를 입는다. 현우는 얼굴과 머리가 많이 잘려 정확한 참조 일치 확인이 제한된다. 배경에는 등을 보인 검은 옷 인물과 그 왼쪽에 얼굴 일부가 보이는 추가 인물이 있다. 두 손 사이에는 어두운 봉투처럼 보이는 물체 하나와 밖으로 돌출된 지폐가 보인다. 지폐의 국가와 액면은 판독할 수 없다. 봉투의 넓은 종이 면은 A보다 덜 명확하다.",
        "hard_violations": [
         "장면에서 요구하지 않은 배경 인물 두 명을 추가했다.",
         "현재 봉투만 사용하고 별도 돈은 제외하라는 지시와 달리 봉투 밖으로 지폐를 노출했다."
        ],
        "physics": "봉투는 현우의 손과 오른쪽 남성의 손에 지지되며, 돌출된 지폐도 봉투 입구와 손가락에 받쳐져 있다. 현우의 반대쪽 손은 탁자 위에 놓여 있다. 팔과 손의 연결은 물리적으로 가능하고 지지 없는 부유물은 없다. 다만 세 손이 봉투를 안정적으로 붙잡아, 현우가 통제력을 잃고 손을 놓은 순간이나 급격한 낚아채기 동작이 드러나지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "오른쪽 남성이 봉투를 힘주어 당기는 관계와 장소·의상은 잘 맞지만, 클로즈업보다 넓고 현우가 아직 봉투를 단단히 잡고 있다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "배경에 추가 인물을 만들고 봉투 밖 지폐를 노출했으며, 거친 낚아채기보다 양손으로 내용물을 확인하는 거래처럼 보인다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 현우의 얼굴과 오른쪽 중년 남성의 시선이 두 손 사이 봉투를 향한다. 오른쪽 남성이 봉투 오른쪽을 움켜쥐고 자기 쪽으로 당기는 방향은 맞는다. 다만 현우의 왼쪽 화면상 손도 봉투 끝을 확실히 잡고 있어, 이미 놓친 순간보다는 서로 당기는 순간이다.",
        "built_space": "거래용 나무 탁자 하나가 전경을 채우고, 오른쪽 남성 뒤 의자와 봉투 뒤쪽의 작은 등받이 일부가 보인다. 왼쪽 냉장고 하나, 뒤쪽 무늬 커튼 하나, 천장 전구 하나, 낡은 투톤 벽과 당김줄이 이전 장소와 연결된다. 탁자 위 수저통과 작은 용기가 추가로 보인다. 두 사람은 탁자 양쪽에 앉아 있으나 상체와 얼굴까지 넓게 담겨 지정된 손 중심 클로즈업보다 넓다. 상판은 보이지만 카메라 쪽 끝단은 화면 밖이다.",
        "entities": "인물은 두 명이며 왼쪽은 검은 헝클어진 머리의 젊은 아시아계 남성으로 현우 참조와 대체로 맞고, 해진 짙은 후드와 얼굴 상처가 보인다. 오른쪽은 40대 정도의 아시아계 남성으로 이전 스틸의 회갈색 셔츠를 유지한다. 황갈색 종이봉투는 하나이고 두 손 사이에 비스듬한 넓은 면이 드러난다. 별도 지폐나 읽을 수 있는 글씨는 없다. 다리 부상은 프레임 밖이므로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "봉투는 양쪽 손의 실제 집기로 지지되고, 오른쪽 남성의 긴장된 손등과 뻗은 전완이 당기는 힘을 표현한다. 현우의 다른 손과 전완은 탁자에 닿아 있다. 몸통은 앉은 자세로 읽히며 공중에 뜬 물체나 불가능한 관절은 보이지 않는다. 다만 현우의 남은 집기는 '방금 손에서 빠져나감'과 일치하지 않는다."
       },
       {
        "label": "A",
        "direction": "오른쪽 남성은 손 사이의 봉투와 노출된 지폐를 내려다본다. 현우의 눈은 잘려 시선은 확인할 수 없다. 오른쪽 남성의 한 손은 봉투를 잡고 다른 손은 입구와 지폐에 닿아 있어, 오른쪽으로 강하게 낚아채기보다는 봉투를 열거나 돈을 확인하는 동작으로 보인다.",
        "built_space": "전경에 나무 탁자 하나, 왼쪽 뒤에 다른 탁자 일부와 냉장고 하나, 천장 전구 하나, 벽의 붉은 장치 하나 및 배선이 보인다. 낡은 벽과 어두운 조명은 장소 참조와 유사하다. 오른쪽 의자 등받이 일부가 보이고 두 주인물은 탁자를 사이에 둔다. 그러나 뒤쪽 탁자 구역에 별도 인물들이 들어와 있다. 얼굴을 일부 자른 구도는 A보다 가깝지만 상판이 넓게 차지하고 카메라 쪽 탁자 끝단은 보이지 않는다.",
        "entities": "전경에는 젊은 아시아계 남성과 중년 아시아계 남성이 있고, 각각 낡은 후드와 회갈색 셔츠를 입는다. 현우는 얼굴과 머리가 많이 잘려 정확한 참조 일치 확인이 제한된다. 배경에는 등을 보인 검은 옷 인물과 그 왼쪽에 얼굴 일부가 보이는 추가 인물이 있다. 두 손 사이에는 어두운 봉투처럼 보이는 물체 하나와 밖으로 돌출된 지폐가 보인다. 지폐의 국가와 액면은 판독할 수 없다. 봉투의 넓은 종이 면은 A보다 덜 명확하다.",
        "hard_violations": [
         "장면에서 요구하지 않은 배경 인물 두 명을 추가했다.",
         "현재 봉투만 사용하고 별도 돈은 제외하라는 지시와 달리 봉투 밖으로 지폐를 노출했다."
        ],
        "physics": "봉투는 현우의 손과 오른쪽 남성의 손에 지지되며, 돌출된 지폐도 봉투 입구와 손가락에 받쳐져 있다. 현우의 반대쪽 손은 탁자 위에 놓여 있다. 팔과 손의 연결은 물리적으로 가능하고 지지 없는 부유물은 없다. 다만 세 손이 봉투를 안정적으로 붙잡아, 현우가 통제력을 잃고 손을 놓은 순간이나 급격한 낚아채기 동작이 드러나지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.714,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.464,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 없는 인물(배경의 두 사람) 임의 추가",
     "[gemini-pro] 배제 지시된 지폐(loose money) 노출",
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (손이 얽히고 추가된 손가락 구조)",
     "[gpt-high] 장면에서 요구하지 않은 배경 인물 두 명을 추가했다.",
     "[gpt-high] 현재 봉투만 사용하고 별도 돈은 제외하라는 지시와 달리 봉투 밖으로 지폐를 노출했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 464
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 인물들만 정확히 등장하며, 봉투를 낚아채는 행동과 지정된 장소의 레퍼런스를 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 464,
    "verdict_ko": "배경에 지시되지 않은 인물이 추가되고 노출 금지된 지폐가 등장했으며, 봉투를 쥔 손의 해부학적 구조가 붕괴되는 심각한 오류가 있습니다.  ★위반: [gemini-pro] 프롬프트에 없는 인물(배경의 두 사람) 임의 추가 / [gemini-pro] 배제 지시된 지폐(loose money) 노출 / [gemini-pro] 물리적으로 불가능한 해부학적 구조 (손이 얽히고 추가된 손가락 구조) / [gpt-high] 장면에서 요구하지 않은 배경 인물 두 명을 추가했다. / [gpt-high] 현재 봉투만 사용하고 별도 돈은 제외하라는 지시와 달리 봉투 밖으로 지폐를 노출했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S10sh1_sel.png",
    "asset_id": "0a956d69-e644-4541-9b26-14eb91a133dd",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd80-fdce-7dd4-a559-a882529b46fd",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S10sh1"
  }
 },
 "S10sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:19:59.872847+00:00",
  "fingerprint": "5648931ef59fb5d5ab42b49dff13a706935b6c0aaafc212b6bcf952f2ef27673",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S10sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S10sh5_sel.png",
  "source_sha256": "13c7bbfe1ecb0ddb4313d9d44eefbbdd30489bda608ea1e4e4b108e66a8c775b",
  "file": "S10sh5_cine.png",
  "staged_sha256": "15dd93ea12e78059f305228f0af94fbf9425572fe02e7693cfdf31120a092061",
  "latency_ms": 13599
 },
 "S10sh10::signage": {
  "fp": "b96ee1471c83eb4f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S10sh10": {
  "input_fingerprint": "c6f65b178ee99842",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 낡은 식당 벽면에 달린 붉은색 경광등이 강하게 켜져 붉은 빛을 뿜어내는 찰나.\n\nLOCATION (lock): At an interior wall of the shabby alley restaurant, illuminated by the red warning beacon mounted there. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Wall-mounted warning light (Fixed to the restaurant wall) — Its outward face and one side are visible from the oblique table-side position; used as The visible fixture identifies the source of the warning without becoming an oversized isolated object; Old restaurant wall (Supporting the mounted warning fixture) — Viewed diagonally rather than square-on; used as Provides installation context and negative space around the fixture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The warning light ignites strongly red, creating a sudden localized brightness against the otherwise subdued restaurant.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The secluded restaurant table remains the meeting location at night; no visible warning lamp or red illumination is established in the scene text.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 낡은 식당 벽면에 달린 붉은색 경광등이 강하게 켜져 붉은 빛을 뿜어내는 찰나.\n\nLOCATION (lock): At an interior wall of the shabby alley restaurant, illuminated by the red warning beacon mounted there. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Wall-mounted warning light (Fixed to the restaurant wall) — Its outward face and one side are visible from the oblique table-side position; used as The visible fixture identifies the source of the warning without becoming an oversized isolated object; Old restaurant wall (Supporting the mounted warning fixture) — Viewed diagonally rather than square-on; used as Provides installation context and negative space around the fixture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The warning light ignites strongly red, creating a sudden localized brightness against the otherwise subdued restaurant.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The secluded restaurant table remains the meeting location at night; no visible warning lamp or red illumination is established in the scene text.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 낡은 식당 벽면에 달린 붉은색 경광등이 강하게 켜져 붉은 빛을 뿜어내는 찰나.\n\nLOCATION (lock): At an interior wall of the shabby alley restaurant, illuminated by the red warning beacon mounted there. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Wall-mounted warning light (Fixed to the restaurant wall) — Its outward face and one side are visible from the oblique table-side position; used as The visible fixture identifies the source of the warning without becoming an oversized isolated object; Old restaurant wall (Supporting the mounted warning fixture) — Viewed diagonally rather than square-on; used as Provides installation context and negative space around the fixture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The warning light ignites strongly red, creating a sudden localized brightness against the otherwise subdued restaurant.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The secluded restaurant table remains the meeting location at night; no visible warning lamp or red illumination is established in the scene text.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 식당 구석의 벽면을 향하고 있으며, 앉아있는 남자는 벽 쪽을 바라보고 있음.",
    "built_space": "레퍼런스와 유사한 투톤 벽면 구조. 벽면 상단에 꺼져 있는 원래의 알람 장치가 있고, 그 아래에 붉은색의 큰 돔형 경광등이 하나 더 생성되어 있음. 나무 테이블과 의자가 배치됨.",
    "entities": "샷 텍스트에 존재하지 않는 검은 머리의 남성. 강한 붉은 빛을 내는 경광등.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 인물 추가",
     "[gemini-pro] 벽면 경광등 중복 및 임의 생성",
     "[gpt-high] 참고의 기존 경광등을 정면 벽에 유지하면서 오른쪽 벽에 두 번째 경광등을 추가했다.",
     "[gpt-high] 사람을 추가하지 말라는 지시와 달리 이번 숏 텍스트에 등장하지 않는 남성을 화면의 주요 피사체로 포함했다."
    ],
    "physics": "남자는 의자에 지탱되어 앉아 있고, 경광등은 벽면에 고정되어 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 사선 방향으로 벽면과 붉게 빛나는 경광등을 향하고 있음.",
    "built_space": "레퍼런스와 완전히 다른 콘크리트 질감의 벽면과 금속 배관이 존재함. 배관에 붉은 경광등이 부착되어 있으며, 하단에 테이블과 의자들이 놓여 있음.",
    "entities": "강렬한 붉은 빛을 내는 금속망 형태의 경광등. 샷 텍스트에 따라 인물은 배제됨.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 없는 금속 배관 및 콘크리트 벽면 생성 (장소 불일치)",
     "[gpt-high] 뒤 벽의 기존 경광등 형태 기구를 남겨둔 채 전경 벽에 별도의 점등 기구를 추가하여, 참고 장소에 하나였던 고정 경광등이 중복되었다."
    ],
    "physics": "경광등은 배관과 벽면에 의해 지탱되며, 테이블과 의자는 바닥에 정상적으로 놓여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "샷 텍스트의 지시대로 인물 없이 경광등이 켜지는 순간을 정확히 묘사했으나, 벽면의 재질(콘크리트)과 배관 구조가 레퍼런스와 전혀 다릅니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지문에 언급되지 않은 인물이 화면에 등장했으며, 기존 알람 외에 커다란 경광등이 추가로 중복 생성되는 치명적인 위반이 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 식당 구석의 벽면을 향하고 있으며, 앉아있는 남자는 벽 쪽을 바라보고 있음.",
        "built_space": "레퍼런스와 유사한 투톤 벽면 구조. 벽면 상단에 꺼져 있는 원래의 알람 장치가 있고, 그 아래에 붉은색의 큰 돔형 경광등이 하나 더 생성되어 있음. 나무 테이블과 의자가 배치됨.",
        "entities": "샷 텍스트에 존재하지 않는 검은 머리의 남성. 강한 붉은 빛을 내는 경광등.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 인물 추가",
         "벽면 경광등 중복 및 임의 생성"
        ],
        "physics": "남자는 의자에 지탱되어 앉아 있고, 경광등은 벽면에 고정되어 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 사선 방향으로 벽면과 붉게 빛나는 경광등을 향하고 있음.",
        "built_space": "레퍼런스와 완전히 다른 콘크리트 질감의 벽면과 금속 배관이 존재함. 배관에 붉은 경광등이 부착되어 있으며, 하단에 테이블과 의자들이 놓여 있음.",
        "entities": "강렬한 붉은 빛을 내는 금속망 형태의 경광등. 샷 텍스트에 따라 인물은 배제됨.",
        "hard_violations": [
         "레퍼런스에 없는 금속 배관 및 콘크리트 벽면 생성 (장소 불일치)"
        ],
        "physics": "경광등은 배관과 벽면에 의해 지탱되며, 테이블과 의자는 바닥에 정상적으로 놓여 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "샷 텍스트의 지시대로 인물 없이 경광등이 켜지는 순간을 정확히 묘사했으나, 벽면의 재질(콘크리트)과 배관 구조가 레퍼런스와 전혀 다릅니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지문에 언급되지 않은 인물이 화면에 등장했으며, 기존 알람 외에 커다란 경광등이 추가로 중복 생성되는 치명적인 위반이 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 식당 구석의 벽면을 향하고 있으며, 앉아있는 남자는 벽 쪽을 바라보고 있음.",
        "built_space": "레퍼런스와 유사한 투톤 벽면 구조. 벽면 상단에 꺼져 있는 원래의 알람 장치가 있고, 그 아래에 붉은색의 큰 돔형 경광등이 하나 더 생성되어 있음. 나무 테이블과 의자가 배치됨.",
        "entities": "샷 텍스트에 존재하지 않는 검은 머리의 남성. 강한 붉은 빛을 내는 경광등.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 인물 추가",
         "벽면 경광등 중복 및 임의 생성"
        ],
        "physics": "남자는 의자에 지탱되어 앉아 있고, 경광등은 벽면에 고정되어 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 사선 방향으로 벽면과 붉게 빛나는 경광등을 향하고 있음.",
        "built_space": "레퍼런스와 완전히 다른 콘크리트 질감의 벽면과 금속 배관이 존재함. 배관에 붉은 경광등이 부착되어 있으며, 하단에 테이블과 의자들이 놓여 있음.",
        "entities": "강렬한 붉은 빛을 내는 금속망 형태의 경광등. 샷 텍스트에 따라 인물은 배제됨.",
        "hard_violations": [
         "레퍼런스에 없는 금속 배관 및 콘크리트 벽면 생성 (장소 불일치)"
        ],
        "physics": "경광등은 배관과 벽면에 의해 지탱되며, 테이블과 의자는 바닥에 정상적으로 놓여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물 없이 비스듬한 벽면과 강하게 켜진 붉은 등을 담아 해당 순간의 구도에는 더 가깝지만, 뒤쪽 기존 등과 별개로 전경에 등을 추가해 장소의 고정 설비를 바꿨다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "기존 식당 구석의 재현은 더 가깝지만, 경광등을 중복 설치하고 숏 텍스트가 보여주지 않는 남성을 크게 포함해 경광등 중심의 지정 구도를 어겼다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "켜진 등의 발광면은 카메라 쪽을 비스듬히 향하며 오른쪽 측면도 보인다. 붉은 빛은 주변 벽과 아래 탁자·의자에 퍼진다. 사람이나 이동하는 물체는 없다.",
        "built_space": "오른쪽 벽에 큰 점등 기구 하나가 있고, 왼쪽 뒤 벽에는 참고 이미지의 기존 등처럼 보이는 사각 받침의 어두운 원형 기구 하나가 남아 있어 총 두 개가 보인다. 탁자 상판은 전경·중경·후경에 하나씩 보이고 의자들이 주변에 놓였다. 벽의 투톤 도장과 낡은 목제 가구는 유사하지만, 새 등의 설치 위치와 굵은 배관, 창 형태의 벽면 구성이 참고 장소와 다르다.",
        "entities": "붉게 켜진 물리적 벽등과 낡은 식당 벽은 요구에 맞는다. 다만 참고의 작은 원형 렌즈와 사각 받침 대신 보호 철망을 두른 큰 원형 등으로 바뀌었다. 사람은 없어 이번 숏의 인물 제외 요구에 맞는다. 휴지함과 조미료통이 보이며 읽을 수 있는 문구는 없다.",
        "hard_violations": [
         "뒤 벽의 기존 경광등 형태 기구를 남겨둔 채 전경 벽에 별도의 점등 기구를 추가하여, 참고 장소에 하나였던 고정 경광등이 중복되었다."
        ],
        "physics": "점등 기구는 원형 받침으로 벽에 부착되어 있고 배관도 벽 고정쇠로 지지된다. 탁자와 의자는 다리로 바닥에 지지되며 휴지함과 조미료통은 상판 위에 놓였다. 떠 있는 물체나 불가능한 자세는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "오른쪽 벽의 켜진 등은 왼쪽 실내를 향해 돌출되어 발광면과 옆면이 함께 보이고 붉은 빛이 두 벽에 번진다. 남성은 등을 보지 않고 왼쪽 아래 탁자 방향으로 얼굴을 돌리고 있다.",
        "built_space": "정면 벽 위쪽에는 참고의 사각 받침 붉은 경광등 하나가 그대로 있고, 오른쪽 벽에 밝게 켜진 원통형 경광등 하나가 추가되어 총 두 개다. 접속함, 늘어진 배선 끝, 벽 모서리와 왼쪽 냉장고·통로는 참고와 가깝다. 탁자 하나와 맞은편 의자 하나가 보이며 남성은 탁자 오른쪽에 앉아 화면 하단을 크게 차지한다.",
        "entities": "낡은 투톤 식당 벽과 붉은 점등 자체는 맞지만 점등된 기구는 참고의 기존 등이 아니다. 검은 머리의 젊은 동아시아계 남성과 회갈색 셔츠는 장면 인물의 외형에 대체로 가깝지만, 이번 숏 텍스트는 그를 보여주지 않는다. 탁자 위 종이와 휴지함이 보이며 명확히 읽히는 글자는 없다.",
        "hard_violations": [
         "참고의 기존 경광등을 정면 벽에 유지하면서 오른쪽 벽에 두 번째 경광등을 추가했다.",
         "사람을 추가하지 말라는 지시와 달리 이번 숏 텍스트에 등장하지 않는 남성을 화면의 주요 피사체로 포함했다."
        ],
        "physics": "두 경광등과 접속함은 벽에 고정되어 있다. 남성은 탁자 옆 좌석에 앉은 자세이며 앞팔과 손은 상판에 지지된다. 좌석 접촉부는 몸에 가려져 있지만 공중에 떠 있는 모습은 아니다. 종이와 휴지함도 탁자에 놓여 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물 없이 비스듬한 벽면과 강하게 켜진 붉은 등을 담아 해당 순간의 구도에는 더 가깝지만, 뒤쪽 기존 등과 별개로 전경에 등을 추가해 장소의 고정 설비를 바꿨다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "기존 식당 구석의 재현은 더 가깝지만, 경광등을 중복 설치하고 숏 텍스트가 보여주지 않는 남성을 크게 포함해 경광등 중심의 지정 구도를 어겼다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "켜진 등의 발광면은 카메라 쪽을 비스듬히 향하며 오른쪽 측면도 보인다. 붉은 빛은 주변 벽과 아래 탁자·의자에 퍼진다. 사람이나 이동하는 물체는 없다.",
        "built_space": "오른쪽 벽에 큰 점등 기구 하나가 있고, 왼쪽 뒤 벽에는 참고 이미지의 기존 등처럼 보이는 사각 받침의 어두운 원형 기구 하나가 남아 있어 총 두 개가 보인다. 탁자 상판은 전경·중경·후경에 하나씩 보이고 의자들이 주변에 놓였다. 벽의 투톤 도장과 낡은 목제 가구는 유사하지만, 새 등의 설치 위치와 굵은 배관, 창 형태의 벽면 구성이 참고 장소와 다르다.",
        "entities": "붉게 켜진 물리적 벽등과 낡은 식당 벽은 요구에 맞는다. 다만 참고의 작은 원형 렌즈와 사각 받침 대신 보호 철망을 두른 큰 원형 등으로 바뀌었다. 사람은 없어 이번 숏의 인물 제외 요구에 맞는다. 휴지함과 조미료통이 보이며 읽을 수 있는 문구는 없다.",
        "hard_violations": [
         "뒤 벽의 기존 경광등 형태 기구를 남겨둔 채 전경 벽에 별도의 점등 기구를 추가하여, 참고 장소에 하나였던 고정 경광등이 중복되었다."
        ],
        "physics": "점등 기구는 원형 받침으로 벽에 부착되어 있고 배관도 벽 고정쇠로 지지된다. 탁자와 의자는 다리로 바닥에 지지되며 휴지함과 조미료통은 상판 위에 놓였다. 떠 있는 물체나 불가능한 자세는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "오른쪽 벽의 켜진 등은 왼쪽 실내를 향해 돌출되어 발광면과 옆면이 함께 보이고 붉은 빛이 두 벽에 번진다. 남성은 등을 보지 않고 왼쪽 아래 탁자 방향으로 얼굴을 돌리고 있다.",
        "built_space": "정면 벽 위쪽에는 참고의 사각 받침 붉은 경광등 하나가 그대로 있고, 오른쪽 벽에 밝게 켜진 원통형 경광등 하나가 추가되어 총 두 개다. 접속함, 늘어진 배선 끝, 벽 모서리와 왼쪽 냉장고·통로는 참고와 가깝다. 탁자 하나와 맞은편 의자 하나가 보이며 남성은 탁자 오른쪽에 앉아 화면 하단을 크게 차지한다.",
        "entities": "낡은 투톤 식당 벽과 붉은 점등 자체는 맞지만 점등된 기구는 참고의 기존 등이 아니다. 검은 머리의 젊은 동아시아계 남성과 회갈색 셔츠는 장면 인물의 외형에 대체로 가깝지만, 이번 숏 텍스트는 그를 보여주지 않는다. 탁자 위 종이와 휴지함이 보이며 명확히 읽히는 글자는 없다.",
        "hard_violations": [
         "참고의 기존 경광등을 정면 벽에 유지하면서 오른쪽 벽에 두 번째 경광등을 추가했다.",
         "사람을 추가하지 말라는 지시와 달리 이번 숏 텍스트에 등장하지 않는 남성을 화면의 주요 피사체로 포함했다."
        ],
        "physics": "두 경광등과 접속함은 벽에 고정되어 있다. 남성은 탁자 옆 좌석에 앉은 자세이며 앞팔과 손은 상판에 지지된다. 좌석 접촉부는 몸에 가려져 있지만 공중에 떠 있는 모습은 아니다. 종이와 휴지함도 탁자에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.167,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.917,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 인물 추가",
     "[gemini-pro] 벽면 경광등 중복 및 임의 생성",
     "[gpt-high] 참고의 기존 경광등을 정면 벽에 유지하면서 오른쪽 벽에 두 번째 경광등을 추가했다.",
     "[gpt-high] 사람을 추가하지 말라는 지시와 달리 이번 숏 텍스트에 등장하지 않는 남성을 화면의 주요 피사체로 포함했다."
    ],
    "B": [
     "[gemini-pro] 레퍼런스에 없는 금속 배관 및 콘크리트 벽면 생성 (장소 불일치)",
     "[gpt-high] 뒤 벽의 기존 경광등 형태 기구를 남겨둔 채 전경 벽에 별도의 점등 기구를 추가하여, 참고 장소에 하나였던 고정 경광등이 중복되었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 917
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "샷 텍스트의 지시대로 인물 없이 경광등이 켜지는 순간을 정확히 묘사했으나, 벽면의 재질(콘크리트)과 배관 구조가 레퍼런스와 전혀 다릅니다.  ★위반: [gemini-pro] 레퍼런스에 없는 금속 배관 및 콘크리트 벽면 생성 (장소 불일치) / [gpt-high] 뒤 벽의 기존 경광등 형태 기구를 남겨둔 채 전경 벽에 별도의 점등 기구를 추가하여, 참고 장소에 하나였던 고정 경광등이 중복되었다."
   },
   {
    "label": "A",
    "score": 917,
    "verdict_ko": "지문에 언급되지 않은 인물이 화면에 등장했으며, 기존 알람 외에 커다란 경광등이 추가로 중복 생성되는 치명적인 위반이 있습니다.  ★위반: [gemini-pro] 샷 텍스트에 명시되지 않은 인물 추가 / [gemini-pro] 벽면 경광등 중복 및 임의 생성 / [gpt-high] 참고의 기존 경광등을 정면 벽에 유지하면서 오른쪽 벽에 두 번째 경광등을 추가했다. / [gpt-high] 사람을 추가하지 말라는 지시와 달리 이번 숏 텍스트에 등장하지 않는 남성을 화면의 주요 피사체로 포함했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S10sh1_sel.png",
    "asset_id": "0a956d69-e644-4541-9b26-14eb91a133dd",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd81-dcd7-75c0-8d83-4a94bacf9cf2",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S10sh1"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S10sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:20:13.812881+00:00",
  "fingerprint": "30d62690c316d37787332b94f2afe432692cf173276d34681c1e49b2732b4dab",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S10sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S10sh10_sel.png",
  "source_sha256": "029b516e7fcc148e5670c576bd60e93a75bd136ba46e3e75a8bee789e7e2cdb1",
  "file": "S10sh10_cine.png",
  "staged_sha256": "dd78582ba5081cfb03ca0270194d9c5e6ea785dc4dd03ea83d7827a10500b921",
  "latency_ms": 11932
 },
 "S11sh3::signage": {
  "fp": "a3007600abecf039",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S11sh3": {
  "input_fingerprint": "2dc1eeaeaedfdd7d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 상점의 불빛들이 모두 꺼져 칠흑 같은 어둠에 잠긴 황량한 난민 수용소 거리 전경.\n\nLOCATION (lock): Along the refugee settlement's deserted shopping street after the shop lights and streetlights have been switched off. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee camp street (Empty within the frame after curfew shutdown) — Runs diagonally into the distance beneath the elevated camera; used as Provides a broad recession into darkness rather than a symmetrical central corridor; Shopfronts (Closed for curfew) — Their street-facing sides appear obliquely along the road; used as Repeated frontage outlines make the deserted street readable; Container homes (Lining the camp beyond the street) — Street-facing elevations recede into the background; used as Extend the residential scale of the shutdown beyond the shops.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Shop, home, and street lighting are extinguished, leaving near-black nighttime exposure with only minimal spatial legibility and no added visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shops are shut, container-home lights are out, and the streetlights have been switched off, leaving the camp streets dark. Curfew loudspeakers remain installed throughout the camp.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 상점의 불빛들이 모두 꺼져 칠흑 같은 어둠에 잠긴 황량한 난민 수용소 거리 전경.\n\nLOCATION (lock): Along the refugee settlement's deserted shopping street after the shop lights and streetlights have been switched off. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee camp street (Empty within the frame after curfew shutdown) — Runs diagonally into the distance beneath the elevated camera; used as Provides a broad recession into darkness rather than a symmetrical central corridor; Shopfronts (Closed for curfew) — Their street-facing sides appear obliquely along the road; used as Repeated frontage outlines make the deserted street readable; Container homes (Lining the camp beyond the street) — Street-facing elevations recede into the background; used as Extend the residential scale of the shutdown beyond the shops.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Shop, home, and street lighting are extinguished, leaving near-black nighttime exposure with only minimal spatial legibility and no added visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shops are shut, container-home lights are out, and the streetlights have been switched off, leaving the camp streets dark. Curfew loudspeakers remain installed throughout the camp.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 상점의 불빛들이 모두 꺼져 칠흑 같은 어둠에 잠긴 황량한 난민 수용소 거리 전경.\n\nLOCATION (lock): Along the refugee settlement's deserted shopping street after the shop lights and streetlights have been switched off. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee camp street (Empty within the frame after curfew shutdown) — Runs diagonally into the distance beneath the elevated camera; used as Provides a broad recession into darkness rather than a symmetrical central corridor; Shopfronts (Closed for curfew) — Their street-facing sides appear obliquely along the road; used as Repeated frontage outlines make the deserted street readable; Container homes (Lining the camp beyond the street) — Street-facing elevations recede into the background; used as Extend the residential scale of the shutdown beyond the shops.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Shop, home, and street lighting are extinguished, leaving near-black nighttime exposure with only minimal spatial legibility and no added visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shops are shut, container-home lights are out, and the streetlights have been switched off, leaving the camp streets dark. Curfew loudspeakers remain installed throughout the camp.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "해당 없음 (지시된 조준 방향 없음).",
    "built_space": "참조 이미지의 구조와 정확히 일치함. 전면에 흰색 아치 구조물이 있고, 좌측에 확성기가 달린 기둥과 상점들, 우측에 지붕이 있는 경비실과 차단기, 안쪽에 '7-31' 컨테이너가 올바른 위치에 배치됨.",
    "entities": "비어 있는 거리, 불이 꺼진 상점과 컨테이너, 좌측 확성기가 프롬프트대로 묘사됨.",
    "hard_violations": [],
    "physics": "모든 구조물이 지면에 안정적으로 고정되어 있음."
   },
   {
    "label": "B",
    "direction": "해당 없음.",
    "built_space": "공간 구조가 심각하게 왜곡됨. 전면에 있어야 할 흰색 아치가 아주 먼 배경으로 이동했고, 우측 경비실이 누락되었으며, 안쪽에 있어야 할 '7-31' 컨테이너가 우측 전면으로 당겨짐.",
    "entities": "상점과 컨테이너는 존재하며 거리는 비어 있으나, 배치와 묘사가 참조와 전혀 다름.",
    "hard_violations": [
     "[gemini-pro] 고정된 장소의 영구적인 구조(아치, 경비실, 컨테이너 위치)를 마음대로 재배치하여 물리적 공간을 심각하게 훼손함.",
     "[gpt-high] 오른쪽 전경 컨테이너에 ‘7-31’이라는 읽을 수 있는 문자가 노출되어, 읽을 수 있는 글자를 전면 금지한 조건을 위반한다."
    ],
    "physics": "모든 구조물이 지면에 위치함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스 이미지의 고정된 공간 구조를 정확히 유지하면서 프롬프트가 요구한 칠흑 같은 어둠의 야간 상태를 성공적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "참조 이미지의 핵심 구조물인 전면 아치를 배경으로 보내고 경비실을 누락시키는 등 장소 잠금 지시를 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "해당 없음 (지시된 조준 방향 없음).",
        "built_space": "참조 이미지의 구조와 정확히 일치함. 전면에 흰색 아치 구조물이 있고, 좌측에 확성기가 달린 기둥과 상점들, 우측에 지붕이 있는 경비실과 차단기, 안쪽에 '7-31' 컨테이너가 올바른 위치에 배치됨.",
        "entities": "비어 있는 거리, 불이 꺼진 상점과 컨테이너, 좌측 확성기가 프롬프트대로 묘사됨.",
        "hard_violations": [],
        "physics": "모든 구조물이 지면에 안정적으로 고정되어 있음."
       },
       {
        "label": "B",
        "direction": "해당 없음.",
        "built_space": "공간 구조가 심각하게 왜곡됨. 전면에 있어야 할 흰색 아치가 아주 먼 배경으로 이동했고, 우측 경비실이 누락되었으며, 안쪽에 있어야 할 '7-31' 컨테이너가 우측 전면으로 당겨짐.",
        "entities": "상점과 컨테이너는 존재하며 거리는 비어 있으나, 배치와 묘사가 참조와 전혀 다름.",
        "hard_violations": [
         "고정된 장소의 영구적인 구조(아치, 경비실, 컨테이너 위치)를 마음대로 재배치하여 물리적 공간을 심각하게 훼손함."
        ],
        "physics": "모든 구조물이 지면에 위치함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스 이미지의 고정된 공간 구조를 정확히 유지하면서 프롬프트가 요구한 칠흑 같은 어둠의 야간 상태를 성공적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "참조 이미지의 핵심 구조물인 전면 아치를 배경으로 보내고 경비실을 누락시키는 등 장소 잠금 지시를 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "해당 없음 (지시된 조준 방향 없음).",
        "built_space": "참조 이미지의 구조와 정확히 일치함. 전면에 흰색 아치 구조물이 있고, 좌측에 확성기가 달린 기둥과 상점들, 우측에 지붕이 있는 경비실과 차단기, 안쪽에 '7-31' 컨테이너가 올바른 위치에 배치됨.",
        "entities": "비어 있는 거리, 불이 꺼진 상점과 컨테이너, 좌측 확성기가 프롬프트대로 묘사됨.",
        "hard_violations": [],
        "physics": "모든 구조물이 지면에 안정적으로 고정되어 있음."
       },
       {
        "label": "B",
        "direction": "해당 없음.",
        "built_space": "공간 구조가 심각하게 왜곡됨. 전면에 있어야 할 흰색 아치가 아주 먼 배경으로 이동했고, 우측 경비실이 누락되었으며, 안쪽에 있어야 할 '7-31' 컨테이너가 우측 전면으로 당겨짐.",
        "entities": "상점과 컨테이너는 존재하며 거리는 비어 있으나, 배치와 묘사가 참조와 전혀 다름.",
        "hard_violations": [
         "고정된 장소의 영구적인 구조(아치, 경비실, 컨테이너 위치)를 마음대로 재배치하여 물리적 공간을 심각하게 훼손함."
        ],
        "physics": "모든 구조물이 지면에 위치함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "소등된 빈 거리의 어둠은 맞지만, 중앙 대칭에 가까운 원근 구도이고 출입구 배치가 참조와 어긋나며 읽을 수 있는 ‘7-31’이 노출된다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "출입 아치·경비실·차단기·확성기의 배치와 소등된 빈 거리를 더 충실히 보존하지만, 요구한 대각선 거리 구도보다 참조의 정면 출입구 구도에 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "도로가 화면 아래 중앙에서 상단 중앙의 아치로 곧게 수렴한다. 상점은 왼쪽에서 도로를 향하고 주거 컨테이너는 오른쪽에서 도로를 향한다. 요구한 비대칭 대각선 후퇴보다는 중앙 통로 구도다. 사람이나 이동체, 조준 대상은 없다.",
        "built_space": "왼쪽에 연속 상점, 오른쪽에 컨테이너 주택들이 있다. 아치 하나는 거리 먼 끝에 작게 놓였고 그 오른쪽에 작은 경비실 지붕이 보인다. 참조에서는 가까운 출입구와 경비실 뒤로 상점가가 이어지므로, 같은 좌우 배열을 유지하면서 출입구만 먼 끝으로 옮긴 듯한 공간 불일치가 있다. 왼쪽 가로등 세 개와 오른쪽의 뚜렷한 가로등 한 개, 추가 원거리 기둥들이 보이며 모두 꺼져 있다. 차단기와 출입문 세부는 어두워 확인하기 어렵다.",
        "entities": "사람은 없어 빈 거리 조건에 맞으며 이현우도 추가하지 않았다. 닫힌 상점, 골판 금속 컨테이너, 산 배경, 소등된 가로등이 보인다. 먼 아치 왼쪽 기둥에는 작은 확성기 형태가 보이지만 세부 식별은 어렵다. 오른쪽 가까운 컨테이너의 ‘7-31’은 읽을 수 있어 문자 금지 조건을 위반한다. 상점 간판에도 문자 형태가 남아 있다.",
        "hard_violations": [
         "오른쪽 전경 컨테이너에 ‘7-31’이라는 읽을 수 있는 문자가 노출되어, 읽을 수 있는 글자를 전면 금지한 조건을 위반한다."
        ],
        "physics": "건물과 가로등은 지면에 서 있고 실외기들은 지붕이나 바닥에 놓여 있다. 아치는 양쪽 기둥으로 지지된다. 떠 있는 인체나 지지 없는 물체, 움직임에 따른 물리적 모순은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "도로는 넓은 아래쪽에서 화면 중앙 먼 곳으로 수렴하며 약간 비스듬하지만 여전히 중앙 통로 인상이 강하다. 양쪽 상점과 주택의 도로 쪽 입면이 사선으로 보인다. 차단봉은 오른쪽 장치에서 도로를 가로질러 왼쪽으로 뻗고, 왼쪽 출입 기둥의 확성기 두 개는 서로 다른 바깥 방향을 향한다. 사람이나 이동체는 없다.",
        "built_space": "가까운 출입 아치 하나와 양쪽 지주, 오른쪽 경비실 하나, 그 앞 차단기 하나, 왼쪽 출입문 한 짝과 철조망 담장이 보인다. 왼쪽 기둥에는 확성기 두 개가 달려 있다. 그 뒤로 왼쪽 상점열과 오른쪽 컨테이너 주거열이 이어져 참조의 주요 배치와 맞는다. 왼쪽에 구분되는 가로등 세 개, 오른쪽에 한 개와 먼 기둥들이 있으며 모두 소등되어 있다. 지붕이 내려다보이는 높은 시점은 맞지만 정면 출입구를 크게 담아 참조 구도에서 충분히 벗어나지 못했다.",
        "entities": "인물 없이 닫힌 상점, 불 꺼진 컨테이너 주택, 가로등, 통금 확성기, 금속 아치, 경비실, 차단기와 철조망이 보인다. 낡은 금속과 균열 난 포장면도 참조 장소의 재질에 부합한다. 명확하게 읽히는 글자는 확인되지 않는다. 발광하는 조명은 없지만 하늘과 노면의 푸른 밝기가 남아 칠흑 같은 최소 가시성보다는 다소 밝다.",
        "hard_violations": [],
        "physics": "아치는 양쪽 지주가 받치고, 확성기는 기둥에 고정되어 있다. 수평 차단봉은 오른쪽 구동 장치에 연결되어 지지된다. 경비실과 컨테이너는 지면에 놓였으며 지붕 위 장비와 바닥의 판재도 각각 표면에 얹혀 있다. 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "소등된 빈 거리의 어둠은 맞지만, 중앙 대칭에 가까운 원근 구도이고 출입구 배치가 참조와 어긋나며 읽을 수 있는 ‘7-31’이 노출된다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "출입 아치·경비실·차단기·확성기의 배치와 소등된 빈 거리를 더 충실히 보존하지만, 요구한 대각선 거리 구도보다 참조의 정면 출입구 구도에 가깝다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "도로가 화면 아래 중앙에서 상단 중앙의 아치로 곧게 수렴한다. 상점은 왼쪽에서 도로를 향하고 주거 컨테이너는 오른쪽에서 도로를 향한다. 요구한 비대칭 대각선 후퇴보다는 중앙 통로 구도다. 사람이나 이동체, 조준 대상은 없다.",
        "built_space": "왼쪽에 연속 상점, 오른쪽에 컨테이너 주택들이 있다. 아치 하나는 거리 먼 끝에 작게 놓였고 그 오른쪽에 작은 경비실 지붕이 보인다. 참조에서는 가까운 출입구와 경비실 뒤로 상점가가 이어지므로, 같은 좌우 배열을 유지하면서 출입구만 먼 끝으로 옮긴 듯한 공간 불일치가 있다. 왼쪽 가로등 세 개와 오른쪽의 뚜렷한 가로등 한 개, 추가 원거리 기둥들이 보이며 모두 꺼져 있다. 차단기와 출입문 세부는 어두워 확인하기 어렵다.",
        "entities": "사람은 없어 빈 거리 조건에 맞으며 이현우도 추가하지 않았다. 닫힌 상점, 골판 금속 컨테이너, 산 배경, 소등된 가로등이 보인다. 먼 아치 왼쪽 기둥에는 작은 확성기 형태가 보이지만 세부 식별은 어렵다. 오른쪽 가까운 컨테이너의 ‘7-31’은 읽을 수 있어 문자 금지 조건을 위반한다. 상점 간판에도 문자 형태가 남아 있다.",
        "hard_violations": [
         "오른쪽 전경 컨테이너에 ‘7-31’이라는 읽을 수 있는 문자가 노출되어, 읽을 수 있는 글자를 전면 금지한 조건을 위반한다."
        ],
        "physics": "건물과 가로등은 지면에 서 있고 실외기들은 지붕이나 바닥에 놓여 있다. 아치는 양쪽 기둥으로 지지된다. 떠 있는 인체나 지지 없는 물체, 움직임에 따른 물리적 모순은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "도로는 넓은 아래쪽에서 화면 중앙 먼 곳으로 수렴하며 약간 비스듬하지만 여전히 중앙 통로 인상이 강하다. 양쪽 상점과 주택의 도로 쪽 입면이 사선으로 보인다. 차단봉은 오른쪽 장치에서 도로를 가로질러 왼쪽으로 뻗고, 왼쪽 출입 기둥의 확성기 두 개는 서로 다른 바깥 방향을 향한다. 사람이나 이동체는 없다.",
        "built_space": "가까운 출입 아치 하나와 양쪽 지주, 오른쪽 경비실 하나, 그 앞 차단기 하나, 왼쪽 출입문 한 짝과 철조망 담장이 보인다. 왼쪽 기둥에는 확성기 두 개가 달려 있다. 그 뒤로 왼쪽 상점열과 오른쪽 컨테이너 주거열이 이어져 참조의 주요 배치와 맞는다. 왼쪽에 구분되는 가로등 세 개, 오른쪽에 한 개와 먼 기둥들이 있으며 모두 소등되어 있다. 지붕이 내려다보이는 높은 시점은 맞지만 정면 출입구를 크게 담아 참조 구도에서 충분히 벗어나지 못했다.",
        "entities": "인물 없이 닫힌 상점, 불 꺼진 컨테이너 주택, 가로등, 통금 확성기, 금속 아치, 경비실, 차단기와 철조망이 보인다. 낡은 금속과 균열 난 포장면도 참조 장소의 재질에 부합한다. 명확하게 읽히는 글자는 확인되지 않는다. 발광하는 조명은 없지만 하늘과 노면의 푸른 밝기가 남아 칠흑 같은 최소 가시성보다는 다소 밝다.",
        "hard_violations": [],
        "physics": "아치는 양쪽 지주가 받치고, 확성기는 기둥에 고정되어 있다. 수평 차단봉은 오른쪽 구동 장치에 연결되어 지지된다. 경비실과 컨테이너는 지면에 놓였으며 지붕 위 장비와 바닥의 판재도 각각 표면에 얹혀 있다. 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.762
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.512
   },
   "violations": {
    "B": [
     "[gemini-pro] 고정된 장소의 영구적인 구조(아치, 경비실, 컨테이너 위치)를 마음대로 재배치하여 물리적 공간을 심각하게 훼손함.",
     "[gpt-high] 오른쪽 전경 컨테이너에 ‘7-31’이라는 읽을 수 있는 문자가 노출되어, 읽을 수 있는 글자를 전면 금지한 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 512
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스 이미지의 고정된 공간 구조를 정확히 유지하면서 프롬프트가 요구한 칠흑 같은 어둠의 야간 상태를 성공적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 512,
    "verdict_ko": "참조 이미지의 핵심 구조물인 전면 아치를 배경으로 보내고 경비실을 누락시키는 등 장소 잠금 지시를 크게 위반했습니다.  ★위반: [gemini-pro] 고정된 장소의 영구적인 구조(아치, 경비실, 컨테이너 위치)를 마음대로 재배치하여 물리적 공간을 심각하게 훼손함. / [gpt-high] 오른쪽 전경 컨테이너에 ‘7-31’이라는 읽을 수 있는 문자가 노출되어, 읽을 수 있는 글자를 전면 금지한 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_refugee_camp_gate_sel.png",
    "asset_id": "44c554ff-13e3-458c-83bf-864316dac1cf",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd82-a0e6-73b6-adcd-8ab36b9ca80f",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S11sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:20:26.066183+00:00",
  "fingerprint": "61047d6e2ba11a50e03d3e80d4b00445c83942f07464de78fde0e909304cd30e",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S11sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S11sh3_sel.png",
  "source_sha256": "72885a97078c58a22a2fd308ccc8fe964c8e205f3fdc3c9e209b3d056122352b",
  "file": "S11sh3_cine.png",
  "staged_sha256": "4a7382ebac6aba0d5084250f38e3a381da444ea281924494405bcdc707e2db6c",
  "latency_ms": 12503
 },
 "S11sh5::signage": {
  "fp": "c3271e5b677e801c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::a3b476f04f25294e": {
  "subjects": [],
  "subject_text": "인천 난민 수용소 출입문과 입구 도로\n컨테이너 주거지로 이어지는 입구 도로와 개폐식 출입문. 입구에는 ‘난민 거주지역’ 표지판과 방송용 스피커가 설치되어 있다.",
  "identity": "canonical",
  "scope_id": "L20",
  "scope_role": "location_exterior",
  "scope_sha": "a4df5072f1d2dc9a"
 },
 "S11sh5::bgfirst_bg": {
  "input_fingerprint": "97d0cc4800d36c11",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 닫히는 출입문 틈 사이로 한쪽 어깨를 밀어넣고 앞발을 내딛은 mid-action 자세.\n\nLOCATION (lock): In the narrowing gap of the refugee settlement's closing perimeter gate, opening onto the entrance road at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Closing entrance gap behind the incoming shoulder in the middle-right of the frame, midground; Interior landing area for the leading foot in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camp entrance gate (Closing, with a remaining gap through which 이현우 enters) — The inside face and the oblique edge of the narrowing opening are visible; used as Frames his compressed shoulder and establishes the threshold crossed by his leading foot; Ground immediately inside the entrance (Receiving 이현우's leading step); used as Keeps the footfall and interior side of the threshold legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the camp's post-blackout darkness, preserving just enough tonal separation to read his passage without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 닫히는 출입문 틈 사이로 한쪽 어깨를 밀어넣고 앞발을 내딛은 mid-action 자세.\n\nLOCATION (lock): In the narrowing gap of the refugee settlement's closing perimeter gate, opening onto the entrance road at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Closing entrance gap behind the incoming shoulder in the middle-right of the frame, midground; Interior landing area for the leading foot in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camp entrance gate (Closing, with a remaining gap through which 이현우 enters) — The inside face and the oblique edge of the narrowing opening are visible; used as Frames his compressed shoulder and establishes the threshold crossed by his leading foot; Ground immediately inside the entrance (Receiving 이현우's leading step); used as Keeps the footfall and interior side of the threshold legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the camp's post-blackout darkness, preserving just enough tonal separation to read his passage without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S11sh5__bgfirst_bg.png",
  "asset_id": "83ca97fa-dddd-4199-b154-09652aaf1d5c",
  "input_asset_ids": [
   "a844d3b7-5ec7-4214-a8d0-e1e7526f6df3",
   "44c554ff-13e3-458c-83bf-864316dac1cf"
  ]
 },
 "S11sh5": {
  "input_fingerprint": "05aaa2e927284ba2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 닫히는 출입문 틈 사이로 한쪽 어깨를 밀어넣고 앞발을 내딛은 mid-action 자세.\n\nLOCATION (lock): In the narrowing gap of the refugee settlement's closing perimeter gate, opening onto the entrance road at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Closing entrance gap behind the incoming shoulder in the middle-right of the frame, midground; Interior landing area for the leading foot in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camp entrance gate (Closing, with a remaining gap through which 이현우 enters) — The inside face and the oblique edge of the narrowing opening are visible; used as Frames his compressed shoulder and establishes the threshold crossed by his leading foot; Ground immediately inside the entrance (Receiving 이현우's leading step); used as Keeps the footfall and interior side of the threshold legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the camp's post-blackout darkness, preserving just enough tonal separation to read his passage without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp entrance is in the process of closing, while the shops, container-home lights and streetlights remain shut or switched off. 이현우: Hyunwoo is squeezing into the camp before the entrance closes, retaining his injured leg and visible arrest injuries.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 닫히는 출입문 틈 사이로 한쪽 어깨를 밀어넣고 앞발을 내딛은 mid-action 자세.\n\nLOCATION (lock): In the narrowing gap of the refugee settlement's closing perimeter gate, opening onto the entrance road at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Closing entrance gap behind the incoming shoulder in the middle-right of the frame, midground; Interior landing area for the leading foot in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camp entrance gate (Closing, with a remaining gap through which 이현우 enters) — The inside face and the oblique edge of the narrowing opening are visible; used as Frames his compressed shoulder and establishes the threshold crossed by his leading foot; Ground immediately inside the entrance (Receiving 이현우's leading step); used as Keeps the footfall and interior side of the threshold legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the camp's post-blackout darkness, preserving just enough tonal separation to read his passage without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp entrance is in the process of closing, while the shops, container-home lights and streetlights remain shut or switched off. 이현우: Hyunwoo is squeezing into the camp before the entrance closes, retaining his injured leg and visible arrest injuries.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 닫히는 출입문 틈 사이로 한쪽 어깨를 밀어넣고 앞발을 내딛은 mid-action 자세.\n\nLOCATION (lock): In the narrowing gap of the refugee settlement's closing perimeter gate, opening onto the entrance road at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Closing entrance gap behind the incoming shoulder in the middle-right of the frame, midground; Interior landing area for the leading foot in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camp entrance gate (Closing, with a remaining gap through which 이현우 enters) — The inside face and the oblique edge of the narrowing opening are visible; used as Frames his compressed shoulder and establishes the threshold crossed by his leading foot; Ground immediately inside the entrance (Receiving 이현우's leading step); used as Keeps the footfall and interior side of the threshold legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the camp's post-blackout darkness, preserving just enough tonal separation to read his passage without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp entrance is in the process of closing, while the shops, container-home lights and streetlights remain shut or switched off. 이현우: Hyunwoo is squeezing into the camp before the entrance closes, retaining his injured leg and visible arrest injuries.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S11sh5__bgfirst_bg.png",
     "asset_id": "83ca97fa-dddd-4199-b154-09652aaf1d5c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S11sh5.png",
     "asset_id": "a844d3b7-5ec7-4214-a8d0-e1e7526f6df3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_refugee_camp_gate_sel.png",
     "asset_id": "44c554ff-13e3-458c-83bf-864316dac1cf",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 앞쪽 왼쪽을 향함. 왼손은 허공을 향해 내밀고 오른손은 문에 닿아 있음.",
    "built_space": "레퍼런스의 미닫이문과 아치 구조물, 경비 초소가 완전히 사라지고 거대한 양문형 철문으로 대체됨. 뒤쪽으로 컨테이너와 가로등이 늘어선 도로가 있음.",
    "entities": "이현우(얼굴, 착장 일치, 손목에 붕대). 레퍼런스와 전혀 다른 형태의 출입문과 배경.",
    "hard_violations": [
     "[gemini-pro] 가로등과 조명이 모두 꺼져 있어야 한다는 지시를 어기고 가로등이 밝게 켜져 있음.",
     "[gemini-pro] 위치 레퍼런스의 고정된 구조(초소, 문 형태, 아치)를 무시하고 임의의 공간으로 변경함.",
     "[gpt-high] 캠프 내부 도로를 등지고 바깥쪽으로 문턱을 넘어, 요구된 진입 방향과 내부 착지 위치를 반대로 배치했다.",
     "[gpt-high] 참고 복장과 소지 상태에 없는 배낭을 추가했다."
    ],
    "physics": "앞으로 걸어 나오는 자세이며, 뒷발은 땅을 딛고 있고 오른손은 문을 짚어 몸을 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "시선은 정면을 향하고 있으며, 왼손은 닫히고 있는 철문의 가장자리를 꽉 잡고 있음.",
    "built_space": "레퍼런스와 일치하는 3개의 사각형 타공망이 있는 철문, 우측의 경비 초소, 철책 및 아치형 구조물의 하단 기둥이 정확히 배치됨. 조명이 꺼진 어두운 환경을 유지함.",
    "entities": "이현우(얼굴, 착장 일치, 무릎이 찢어진 바지). 레퍼런스와 일치하는 형태의 문과 초소.",
    "hard_violations": [
     "[gpt-high] 캠프 내부 도로를 등지고 외부 전경으로 발을 내딛어, 요구된 캠프 진입과 내부 착지 위치를 반대로 배치했다."
    ],
    "physics": "문 틈 사이로 어깨를 밀어넣는 mid-action 자세이며, 왼손은 문 가장자리를 지탱하고 내딛은 발은 지면에 안정적으로 닿아 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "레퍼런스의 장소 구조(철문, 우측 초소)를 충실히 구현했으며, 가로등이 꺼진 정전 상태의 조명 설정과 프레이밍 지시를 완벽하게 따랐습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "가로등이 꺼져 있어야 한다는 명시적 지시를 어기고 불이 켜져 있으며, 주어진 레퍼런스의 장소 구조를 완전히 무시하고 다른 형태의 문을 생성했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "시선은 정면을 향하고 있으며, 왼손은 닫히고 있는 철문의 가장자리를 꽉 잡고 있음.",
        "built_space": "레퍼런스와 일치하는 3개의 사각형 타공망이 있는 철문, 우측의 경비 초소, 철책 및 아치형 구조물의 하단 기둥이 정확히 배치됨. 조명이 꺼진 어두운 환경을 유지함.",
        "entities": "이현우(얼굴, 착장 일치, 무릎이 찢어진 바지). 레퍼런스와 일치하는 형태의 문과 초소.",
        "hard_violations": [],
        "physics": "문 틈 사이로 어깨를 밀어넣는 mid-action 자세이며, 왼손은 문 가장자리를 지탱하고 내딛은 발은 지면에 안정적으로 닿아 있음."
       },
       {
        "label": "A",
        "direction": "시선은 앞쪽 왼쪽을 향함. 왼손은 허공을 향해 내밀고 오른손은 문에 닿아 있음.",
        "built_space": "레퍼런스의 미닫이문과 아치 구조물, 경비 초소가 완전히 사라지고 거대한 양문형 철문으로 대체됨. 뒤쪽으로 컨테이너와 가로등이 늘어선 도로가 있음.",
        "entities": "이현우(얼굴, 착장 일치, 손목에 붕대). 레퍼런스와 전혀 다른 형태의 출입문과 배경.",
        "hard_violations": [
         "가로등과 조명이 모두 꺼져 있어야 한다는 지시를 어기고 가로등이 밝게 켜져 있음.",
         "위치 레퍼런스의 고정된 구조(초소, 문 형태, 아치)를 무시하고 임의의 공간으로 변경함."
        ],
        "physics": "앞으로 걸어 나오는 자세이며, 뒷발은 땅을 딛고 있고 오른손은 문을 짚어 몸을 지탱하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "레퍼런스의 장소 구조(철문, 우측 초소)를 충실히 구현했으며, 가로등이 꺼진 정전 상태의 조명 설정과 프레이밍 지시를 완벽하게 따랐습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "가로등이 꺼져 있어야 한다는 명시적 지시를 어기고 불이 켜져 있으며, 주어진 레퍼런스의 장소 구조를 완전히 무시하고 다른 형태의 문을 생성했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "시선은 정면을 향하고 있으며, 왼손은 닫히고 있는 철문의 가장자리를 꽉 잡고 있음.",
        "built_space": "레퍼런스와 일치하는 3개의 사각형 타공망이 있는 철문, 우측의 경비 초소, 철책 및 아치형 구조물의 하단 기둥이 정확히 배치됨. 조명이 꺼진 어두운 환경을 유지함.",
        "entities": "이현우(얼굴, 착장 일치, 무릎이 찢어진 바지). 레퍼런스와 일치하는 형태의 문과 초소.",
        "hard_violations": [],
        "physics": "문 틈 사이로 어깨를 밀어넣는 mid-action 자세이며, 왼손은 문 가장자리를 지탱하고 내딛은 발은 지면에 안정적으로 닿아 있음."
       },
       {
        "label": "A",
        "direction": "시선은 앞쪽 왼쪽을 향함. 왼손은 허공을 향해 내밀고 오른손은 문에 닿아 있음.",
        "built_space": "레퍼런스의 미닫이문과 아치 구조물, 경비 초소가 완전히 사라지고 거대한 양문형 철문으로 대체됨. 뒤쪽으로 컨테이너와 가로등이 늘어선 도로가 있음.",
        "entities": "이현우(얼굴, 착장 일치, 손목에 붕대). 레퍼런스와 전혀 다른 형태의 출입문과 배경.",
        "hard_violations": [
         "가로등과 조명이 모두 꺼져 있어야 한다는 지시를 어기고 가로등이 밝게 켜져 있음.",
         "위치 레퍼런스의 고정된 구조(초소, 문 형태, 아치)를 무시하고 임의의 공간으로 변경함."
        ],
        "physics": "앞으로 걸어 나오는 자세이며, 뒷발은 땅을 딛고 있고 오른손은 문을 짚어 몸을 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "와이드 구도와 정전된 밤은 더 충실하지만, 캠프를 등지고 밖으로 나오는 배치이며 어깨를 압박하는 좁은 문틈이 구현되지 않았다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "내딛는 동작은 선명하지만 캠프 밖으로 향하고, 문틈이 넓으며 켜진 가로등·실내등과 추가된 배낭이 명시 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 화면 왼쪽을 향하고, 앞발은 문턱에서 카메라 쪽으로 나와 있다. 인물 뒤로 캠프 내부의 상점과 도로가 이어지므로 보이는 이동 방향은 캠프 안으로 진입하는 방향이 아니라 바깥으로 나오는 방향이다.",
        "built_space": "출입구 기둥 두 개, 상부 아치의 일부, 오른쪽 경비실 한 동, 왼쪽 철망과 상부 철조망이 보인다. 오른쪽 기둥에 연결된 큰 금속 문짝 하나에는 망창 네 개가 보이며, 참고의 왼쪽 문짝에 있는 세 망창과 배치가 다르다. 문짝 끝과 왼쪽 기둥 사이에는 사람보다 훨씬 넓은 통로가 남아 있어 좁아진 틈에 어깨가 끼는 구조가 아니다. 앞발은 하단 중앙, 문짝 모서리는 중앙 오른쪽에 놓여 화면 배치는 비교적 가깝다. 내부 건물과 가로등은 어둡다.",
        "entities": "보이는 사람은 한 명이며, 젊은 동아시아계 남성의 얼굴과 헝클어진 검은 머리, 마른 체격이 참고와 대체로 부합한다. 정확히 18세인지는 시각적으로 확정할 수 없다. 낡은 어두운 후드, 무릎이 찢어진 청바지, 닳은 운동화도 부합한다. 얼굴에 상처성 흔적은 보이지만 다리 부상을 보호하는 동작은 뚜렷하지 않다. 손은 자신의 소매에서 이어져 문 모서리를 잡는다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [
         "캠프 내부 도로를 등지고 외부 전경으로 발을 내딛어, 요구된 캠프 진입과 내부 착지 위치를 반대로 배치했다."
        ],
        "physics": "앞쪽 운동화가 지면에 닿아 몸을 지지하고 손은 문짝의 세로 모서리를 붙잡고 있다. 문짝도 오른쪽 기둥의 경첩으로 지지되어 부유하지 않는다. 몸 일부가 문 뒤에 가려지는 것은 자연스럽지만, 어깨 옆의 반대 경계가 멀어 실제로 좁은 틈을 비집는 압박은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "시선은 화면 왼쪽으로 향하며 상체와 앞발은 카메라 쪽으로 전진한다. 캠프 상점과 컨테이너 도로가 등 뒤에 있으므로 캠프로 들어가기보다 캠프에서 나오는 동작으로 읽힌다.",
        "built_space": "좌우에 금속 문짝 두 개가 있고 각 문짝에서 망창 두 개씩이 적어도 부분적으로 보인다. 문턱의 배수로 한 줄 뒤로 내부 도로와 양쪽 건물이 이어진다. 상부 아치와 전체 기둥은 프레임 밖이어서 확인할 수 없다. 문짝 사이 통로는 인물 여러 명이 통과할 만큼 넓어 닫히기 직전의 좁은 틈이 아니다. 인물은 중앙 오른쪽에 있지만 화면 높이를 크게 차지하고, 앞발은 하단 중앙보다 오른쪽이다. 여러 가로등과 건물 창의 실내등이 실제로 켜져 있어 정전 상태에 어긋난다.",
        "entities": "인물은 한 명이며 젊은 동아시아계 남성, 검은 헝클어진 머리, 낡은 후드와 청바지, 운동화가 참고와 대체로 맞는다. 얼굴은 참고보다 다소 성숙해 보인다. 양 손목의 붕대와 얼굴 상처는 체포 부상을 표현하지만 다리 부상에 따른 절뚝임은 명확하지 않다. 참고와 지시에 없는 배낭 및 어깨끈이 추가되어 있다. 문 가까운 손은 자신의 팔에 연결되어 모서리에 걸쳐 있고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "캠프 내부 도로를 등지고 바깥쪽으로 문턱을 넘어, 요구된 진입 방향과 내부 착지 위치를 반대로 배치했다.",
         "참고 복장과 소지 상태에 없는 배낭을 추가했다."
        ],
        "physics": "앞발이 지면에 거의 닿으며 체중을 받는 보행 순간이고 뒤쪽 다리는 굽혀 따라 나온다. 손은 문 모서리에 걸쳐 있어 몸을 보조할 수 있으며, 명백한 공중 부유는 없다. 배낭은 어깨끈으로 지지되고 문짝에는 경첩 부품이 보인다. 전진 동작 자체는 가능하지만 어깨를 양쪽 문 사이에 압축해 밀어 넣는 자세는 아니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "와이드 구도와 정전된 밤은 더 충실하지만, 캠프를 등지고 밖으로 나오는 배치이며 어깨를 압박하는 좁은 문틈이 구현되지 않았다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "내딛는 동작은 선명하지만 캠프 밖으로 향하고, 문틈이 넓으며 켜진 가로등·실내등과 추가된 배낭이 명시 조건을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 화면 왼쪽을 향하고, 앞발은 문턱에서 카메라 쪽으로 나와 있다. 인물 뒤로 캠프 내부의 상점과 도로가 이어지므로 보이는 이동 방향은 캠프 안으로 진입하는 방향이 아니라 바깥으로 나오는 방향이다.",
        "built_space": "출입구 기둥 두 개, 상부 아치의 일부, 오른쪽 경비실 한 동, 왼쪽 철망과 상부 철조망이 보인다. 오른쪽 기둥에 연결된 큰 금속 문짝 하나에는 망창 네 개가 보이며, 참고의 왼쪽 문짝에 있는 세 망창과 배치가 다르다. 문짝 끝과 왼쪽 기둥 사이에는 사람보다 훨씬 넓은 통로가 남아 있어 좁아진 틈에 어깨가 끼는 구조가 아니다. 앞발은 하단 중앙, 문짝 모서리는 중앙 오른쪽에 놓여 화면 배치는 비교적 가깝다. 내부 건물과 가로등은 어둡다.",
        "entities": "보이는 사람은 한 명이며, 젊은 동아시아계 남성의 얼굴과 헝클어진 검은 머리, 마른 체격이 참고와 대체로 부합한다. 정확히 18세인지는 시각적으로 확정할 수 없다. 낡은 어두운 후드, 무릎이 찢어진 청바지, 닳은 운동화도 부합한다. 얼굴에 상처성 흔적은 보이지만 다리 부상을 보호하는 동작은 뚜렷하지 않다. 손은 자신의 소매에서 이어져 문 모서리를 잡는다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [
         "캠프 내부 도로를 등지고 외부 전경으로 발을 내딛어, 요구된 캠프 진입과 내부 착지 위치를 반대로 배치했다."
        ],
        "physics": "앞쪽 운동화가 지면에 닿아 몸을 지지하고 손은 문짝의 세로 모서리를 붙잡고 있다. 문짝도 오른쪽 기둥의 경첩으로 지지되어 부유하지 않는다. 몸 일부가 문 뒤에 가려지는 것은 자연스럽지만, 어깨 옆의 반대 경계가 멀어 실제로 좁은 틈을 비집는 압박은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "시선은 화면 왼쪽으로 향하며 상체와 앞발은 카메라 쪽으로 전진한다. 캠프 상점과 컨테이너 도로가 등 뒤에 있으므로 캠프로 들어가기보다 캠프에서 나오는 동작으로 읽힌다.",
        "built_space": "좌우에 금속 문짝 두 개가 있고 각 문짝에서 망창 두 개씩이 적어도 부분적으로 보인다. 문턱의 배수로 한 줄 뒤로 내부 도로와 양쪽 건물이 이어진다. 상부 아치와 전체 기둥은 프레임 밖이어서 확인할 수 없다. 문짝 사이 통로는 인물 여러 명이 통과할 만큼 넓어 닫히기 직전의 좁은 틈이 아니다. 인물은 중앙 오른쪽에 있지만 화면 높이를 크게 차지하고, 앞발은 하단 중앙보다 오른쪽이다. 여러 가로등과 건물 창의 실내등이 실제로 켜져 있어 정전 상태에 어긋난다.",
        "entities": "인물은 한 명이며 젊은 동아시아계 남성, 검은 헝클어진 머리, 낡은 후드와 청바지, 운동화가 참고와 대체로 맞는다. 얼굴은 참고보다 다소 성숙해 보인다. 양 손목의 붕대와 얼굴 상처는 체포 부상을 표현하지만 다리 부상에 따른 절뚝임은 명확하지 않다. 참고와 지시에 없는 배낭 및 어깨끈이 추가되어 있다. 문 가까운 손은 자신의 팔에 연결되어 모서리에 걸쳐 있고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "캠프 내부 도로를 등지고 바깥쪽으로 문턱을 넘어, 요구된 진입 방향과 내부 착지 위치를 반대로 배치했다.",
         "참고 복장과 소지 상태에 없는 배낭을 추가했다."
        ],
        "physics": "앞발이 지면에 거의 닿으며 체중을 받는 보행 순간이고 뒤쪽 다리는 굽혀 따라 나온다. 손은 문 모서리에 걸쳐 있어 몸을 보조할 수 있으며, 명백한 공중 부유는 없다. 배낭은 어깨끈으로 지지되고 문짝에는 경첩 부품이 보인다. 전진 동작 자체는 가능하지만 어깨를 양쪽 문 사이에 압축해 밀어 넣는 자세는 아니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.75,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 가로등과 조명이 모두 꺼져 있어야 한다는 지시를 어기고 가로등이 밝게 켜져 있음.",
     "[gemini-pro] 위치 레퍼런스의 고정된 구조(초소, 문 형태, 아치)를 무시하고 임의의 공간으로 변경함.",
     "[gpt-high] 캠프 내부 도로를 등지고 바깥쪽으로 문턱을 넘어, 요구된 진입 방향과 내부 착지 위치를 반대로 배치했다.",
     "[gpt-high] 참고 복장과 소지 상태에 없는 배낭을 추가했다."
    ],
    "B": [
     "[gpt-high] 캠프 내부 도로를 등지고 외부 전경으로 발을 내딛어, 요구된 캠프 진입과 내부 착지 위치를 반대로 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "레퍼런스의 장소 구조(철문, 우측 초소)를 충실히 구현했으며, 가로등이 꺼진 정전 상태의 조명 설정과 프레이밍 지시를 완벽하게 따랐습니다.  ★위반: [gpt-high] 캠프 내부 도로를 등지고 외부 전경으로 발을 내딛어, 요구된 캠프 진입과 내부 착지 위치를 반대로 배치했다."
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "가로등이 꺼져 있어야 한다는 명시적 지시를 어기고 불이 켜져 있으며, 주어진 레퍼런스의 장소 구조를 완전히 무시하고 다른 형태의 문을 생성했습니다.  ★위반: [gemini-pro] 가로등과 조명이 모두 꺼져 있어야 한다는 지시를 어기고 가로등이 밝게 켜져 있음. / [gemini-pro] 위치 레퍼런스의 고정된 구조(초소, 문 형태, 아치)를 무시하고 임의의 공간으로 변경함. / [gpt-high] 캠프 내부 도로를 등지고 바깥쪽으로 문턱을 넘어, 요구된 진입 방향과 내부 착지 위치를 반대로 배치했다. / [gpt-high] 참고 복장과 소지 상태에 없는 배낭을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_refugee_camp_gate_sel.png",
    "asset_id": "44c554ff-13e3-458c-83bf-864316dac1cf",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd83-6e57-7b20-b2c0-07d8b2970be5",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S11sh5__bgfirst_bg.png",
   "bg_asset_id": "83ca97fa-dddd-4199-b154-09652aaf1d5c",
   "bg_record_key": "S11sh5::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S11sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:20:38.986187+00:00",
  "fingerprint": "55d8f9a7af1c262b988b3f4ab0189623ac7e983a20c76049617d33a28efeeafc",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S11sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S11sh5_sel.png",
  "source_sha256": "d499f996d1aaa9b120f1e92d14263c76aab467eec2924562d27c4c1216d6191a",
  "file": "S11sh5_cine.png",
  "staged_sha256": "3f373262a1ff7d24db24b3980558696b45cc80d4368bfbd44bc9035a9515fe3b",
  "latency_ms": 12713
 },
 "S11sh8::signage": {
  "fp": "563937f348b88d01",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::family_container_front": {
  "input_fingerprint": "7639189389aeed01",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "family_container_front",
    "tags": [
     "S11sh8",
     "S24sh3"
    ]
   },
   "context_sig": "fc1477210059daa7"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the exterior door of a family container home in a dark residential alley inside the refugee settlement.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우의 컨테이너 7-31 앞: 7-31이라는 식별 번호가 거칠게 적혀 있는 낡은 철제 컨테이너 외벽과 문고리. (특징: 페인트가 벗겨진 철제 외벽; 7-31 번호 표시와 낡은 문고리) / 인천 난민 수용소 출입문과 입구 도로: 통금 시간을 알리는 거대한 스피커와 철조망, 닫히기 직전의 육중한 철제 출입문이 있는 통제 구역. (특징: 높은 철조망과 육중한 철제 출입문; 곳곳에 설치된 통제용 확성기 스피커)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 컨테이너 번호 7-31, 문고리를 잡고 잠시 머뭇거리다 들어간다.\n- 24. 현우의 컨테이너 – D\n- 잽싸게 집안으로 들어가는데 페드로가 헐레벌떡 따라 들어온다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the exterior door of a family container home in a dark residential alley inside the refugee settlement.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우의 컨테이너 7-31 앞: 7-31이라는 식별 번호가 거칠게 적혀 있는 낡은 철제 컨테이너 외벽과 문고리. (특징: 페인트가 벗겨진 철제 외벽; 7-31 번호 표시와 낡은 문고리) / 인천 난민 수용소 출입문과 입구 도로: 통금 시간을 알리는 거대한 스피커와 철조망, 닫히기 직전의 육중한 철제 출입문이 있는 통제 구역. (특징: 높은 철조망과 육중한 철제 출입문; 곳곳에 설치된 통제용 확성기 스피커)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 컨테이너 번호 7-31, 문고리를 잡고 잠시 머뭇거리다 들어간다.\n- 24. 현우의 컨테이너 – D\n- 잽싸게 집안으로 들어가는데 페드로가 헐레벌떡 따라 들어온다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_family_container_front_d0287f.png",
  "asset_id": "8e091bc8-65c8-4678-9f88-e299bf1a00e4",
  "input_asset_ids": [
   "311ba8df-0946-44de-9fda-2009a2d3abb7"
  ],
  "origin_tag": "S11sh8",
  "place_text": "At the exterior door of a family container home in a dark residential alley inside the refugee settlement.",
  "origin_inputs": {
   "place_text": "At the exterior door of a family container home in a dark residential alley inside the refugee settlement.",
   "time_of_day_en": "night",
   "conti_asset_id": "311ba8df-0946-44de-9fda-2009a2d3abb7"
  }
 },
 "S11sh8::bgfirst_bg": {
  "input_fingerprint": "dd3b4569bb4adc9b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 컨테이너 문의 낡은 쇠문고리를 쥔 채 고개를 푹 숙인 자세.\n\nLOCATION (lock): At the exterior door of a family container home in a dark residential alley inside the refugee settlement.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container entrance door (Not yet opened during 이현우's hesitation) — A limited oblique section of the exterior face is visible beside him; used as Terminates his movement and places the hand at a tangible threshold; Old iron door handle (Held in 이현우's hand) — Its projecting profile is visible beneath his bowed face; used as Provides a small lower-frame counterpoint to the paused face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the exterior in the established post-curfew darkness, with restrained tonal separation on his face and hand and no invented doorway light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 컨테이너 문의 낡은 쇠문고리를 쥔 채 고개를 푹 숙인 자세.\n\nLOCATION (lock): At the exterior door of a family container home in a dark residential alley inside the refugee settlement.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container entrance door (Not yet opened during 이현우's hesitation) — A limited oblique section of the exterior face is visible beside him; used as Terminates his movement and places the hand at a tangible threshold; Old iron door handle (Held in 이현우's hand) — Its projecting profile is visible beneath his bowed face; used as Provides a small lower-frame counterpoint to the paused face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the exterior in the established post-curfew darkness, with restrained tonal separation on his face and hand and no invented doorway light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S11sh8__bgfirst_bg.png",
  "asset_id": "e8e203ba-e45f-45df-86f0-07af1f703e69",
  "input_asset_ids": [
   "311ba8df-0946-44de-9fda-2009a2d3abb7",
   "8e091bc8-65c8-4678-9f88-e299bf1a00e4"
  ]
 },
 "S11sh8": {
  "input_fingerprint": "98e9fcaf54403091",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 컨테이너 문의 낡은 쇠문고리를 쥔 채 고개를 푹 숙인 자세.\n\nLOCATION (lock): At the exterior door of a family container home in a dark residential alley inside the refugee settlement. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container entrance door (Not yet opened during 이현우's hesitation) — A limited oblique section of the exterior face is visible beside him; used as Terminates his movement and places the hand at a tangible threshold; Old iron door handle (Held in 이현우's hand) — Its projecting profile is visible beneath his bowed face; used as Provides a small lower-frame counterpoint to the paused face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the exterior in the established post-curfew darkness, with restrained tonal separation on his face and hand and no invented doorway light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The home container is marked “7-31,” and its door remains closed at the moment of hesitation. The surrounding camp remains under the nighttime blackout. 이현우: Hyunwoo stands at his container entrance with one hand on the door handle, hesitating. His leg and arrest injuries persist.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 컨테이너 문의 낡은 쇠문고리를 쥔 채 고개를 푹 숙인 자세.\n\nLOCATION (lock): At the exterior door of a family container home in a dark residential alley inside the refugee settlement. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container entrance door (Not yet opened during 이현우's hesitation) — A limited oblique section of the exterior face is visible beside him; used as Terminates his movement and places the hand at a tangible threshold; Old iron door handle (Held in 이현우's hand) — Its projecting profile is visible beneath his bowed face; used as Provides a small lower-frame counterpoint to the paused face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the exterior in the established post-curfew darkness, with restrained tonal separation on his face and hand and no invented doorway light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The home container is marked “7-31,” and its door remains closed at the moment of hesitation. The surrounding camp remains under the nighttime blackout. 이현우: Hyunwoo stands at his container entrance with one hand on the door handle, hesitating. His leg and arrest injuries persist.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 컨테이너 문의 낡은 쇠문고리를 쥔 채 고개를 푹 숙인 자세.\n\nLOCATION (lock): At the exterior door of a family container home in a dark residential alley inside the refugee settlement. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container entrance door (Not yet opened during 이현우's hesitation) — A limited oblique section of the exterior face is visible beside him; used as Terminates his movement and places the hand at a tangible threshold; Old iron door handle (Held in 이현우's hand) — Its projecting profile is visible beneath his bowed face; used as Provides a small lower-frame counterpoint to the paused face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the exterior in the established post-curfew darkness, with restrained tonal separation on his face and hand and no invented doorway light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The home container is marked “7-31,” and its door remains closed at the moment of hesitation. The surrounding camp remains under the nighttime blackout. 이현우: Hyunwoo stands at his container entrance with one hand on the door handle, hesitating. His leg and arrest injuries persist.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S11sh8__bgfirst_bg.png",
     "asset_id": "e8e203ba-e45f-45df-86f0-07af1f703e69",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S11sh8.png",
     "asset_id": "311ba8df-0946-44de-9fda-2009a2d3abb7",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_family_container_front_d0287f.png",
     "asset_id": "8e091bc8-65c8-4678-9f88-e299bf1a00e4",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "고개는 아래로 깊게 숙여져 있으며, 오른손은 문고리를 향해 뻗어 쥐고 있음.",
    "built_space": "야간의 컨테이너 골목 외벽. 지시된 낡은 철제 문고리가 적절한 위치에 하나 존재함.",
    "entities": "인물은 헝클어진 검은 머리와 낡은 후드티 등 이현우의 외형과 일치하며, 낡은 문고리도 잘 묘사됨.",
    "hard_violations": [],
    "physics": "인물은 지면에 서서 몸을 지탱하고 있으며, 오른손은 문고리를 단단히 쥐어 하중을 나누고 있음."
   },
   {
    "label": "B",
    "direction": "고개는 아래를 향하나, 오른팔이 전방으로 길게 뻗어 문고리를 향함.",
    "built_space": "야간 컨테이너 배경과 외벽 조명, 문고리가 보임. 위치와 구조는 레퍼런스와 유사함.",
    "entities": "인물의 얼굴과 머리는 일치하나 지정된 후드티의 디테일(후드 부분)이 다소 부족함.",
    "hard_violations": [],
    "physics": "지면에 서서 몸을 지탱하며, 오른손으로 문고리를 잡고 안정적인 자세를 유지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "클로즈업 샷 크기를 준수하고, 숙인 얼굴 바로 아래에 문고리가 위치하는 요구 구도를 정확히 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "프레이밍이 클로즈업보다 넓으며, 팔을 길게 뻗어 문고리가 얼굴 아래가 아닌 화면 우측으로 멀어져 구도 지시를 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개는 아래로 깊게 숙여져 있으며, 오른손은 문고리를 향해 뻗어 쥐고 있음.",
        "built_space": "야간의 컨테이너 골목 외벽. 지시된 낡은 철제 문고리가 적절한 위치에 하나 존재함.",
        "entities": "인물은 헝클어진 검은 머리와 낡은 후드티 등 이현우의 외형과 일치하며, 낡은 문고리도 잘 묘사됨.",
        "hard_violations": [],
        "physics": "인물은 지면에 서서 몸을 지탱하고 있으며, 오른손은 문고리를 단단히 쥐어 하중을 나누고 있음."
       },
       {
        "label": "B",
        "direction": "고개는 아래를 향하나, 오른팔이 전방으로 길게 뻗어 문고리를 향함.",
        "built_space": "야간 컨테이너 배경과 외벽 조명, 문고리가 보임. 위치와 구조는 레퍼런스와 유사함.",
        "entities": "인물의 얼굴과 머리는 일치하나 지정된 후드티의 디테일(후드 부분)이 다소 부족함.",
        "hard_violations": [],
        "physics": "지면에 서서 몸을 지탱하며, 오른손으로 문고리를 잡고 안정적인 자세를 유지함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "클로즈업 샷 크기를 준수하고, 숙인 얼굴 바로 아래에 문고리가 위치하는 요구 구도를 정확히 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "프레이밍이 클로즈업보다 넓으며, 팔을 길게 뻗어 문고리가 얼굴 아래가 아닌 화면 우측으로 멀어져 구도 지시를 어겼습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "고개는 아래로 깊게 숙여져 있으며, 오른손은 문고리를 향해 뻗어 쥐고 있음.",
        "built_space": "야간의 컨테이너 골목 외벽. 지시된 낡은 철제 문고리가 적절한 위치에 하나 존재함.",
        "entities": "인물은 헝클어진 검은 머리와 낡은 후드티 등 이현우의 외형과 일치하며, 낡은 문고리도 잘 묘사됨.",
        "hard_violations": [],
        "physics": "인물은 지면에 서서 몸을 지탱하고 있으며, 오른손은 문고리를 단단히 쥐어 하중을 나누고 있음."
       },
       {
        "label": "B",
        "direction": "고개는 아래를 향하나, 오른팔이 전방으로 길게 뻗어 문고리를 향함.",
        "built_space": "야간 컨테이너 배경과 외벽 조명, 문고리가 보임. 위치와 구조는 레퍼런스와 유사함.",
        "entities": "인물의 얼굴과 머리는 일치하나 지정된 후드티의 디테일(후드 부분)이 다소 부족함.",
        "hard_violations": [],
        "physics": "지면에 서서 몸을 지탱하며, 오른손으로 문고리를 잡고 안정적인 자세를 유지함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "고개를 숙이고 닫힌 문의 손잡이를 쥔 동작과 장소는 맞지만, 허벅지까지 보이는 넓은 구도가 지정된 클로즈업을 벗어나고 켜진 조명이 정전 설정에 어긋난다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "숙인 얼굴과 손잡이를 쥔 손을 더 밀착해 담고 낡은 후드 의상도 잘 맞지만, 골목 노출이 여전히 넓고 켜진 조명들이 야간 정전 설정을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개와 눈이 아래로 향하며, 시선은 손잡이보다 낮은 문 앞 바닥 쪽으로 내려간다. 한쪽 팔은 오른쪽 문으로 뻗어 손잡이를 쥐고 있다. 카메라를 바라보지 않으며 고개를 푹 숙인 망설임은 전달된다.",
        "built_space": "오른쪽에 닫힌 철문 하나, 돌출된 손잡이 하나와 세로 잠금판, 문 가장자리 경첩이 보인다. 인물은 문 바깥에 서 있어 손잡이 접근이 자연스럽다. 녹슨 도장 철판, 인접 창문과 배관, 젖은 골목 및 끝의 철문은 장소 참조와 잘 이어진다. 다만 문 외벽의 제한된 일부 대신 골목 깊이와 인물의 허벅지까지 넓게 담았다. 여러 가로등과 출입구 조명이 켜져 있고 문 위에도 따뜻한 빛이 떨어져 정전 상태가 아니다.",
        "entities": "헝클어진 검은 머리의 젊은 동아시아계 남성 한 명이며, 보이는 범위에서 이현우의 외형과 대체로 부합한다. 숙인 얼굴 때문에 정확한 얼굴 일치는 제한적으로만 확인된다. 어두운 낡은 상의와 청바지, 얼굴과 손의 상처 흔적이 보인다. 참조의 후드 형태는 뚜렷하지 않다. 낡은 금속 손잡이는 실제 문 부속으로 보이며 읽을 수 있는 글씨나 번호는 없다. 화면 하단의 다른 손 곁에는 갈색 고리 모양 일부가 있으나 물체의 종류는 확정하기 어렵다.",
        "hard_violations": [],
        "physics": "손가락이 손잡이를 감싸고 손목과 팔뚝이 자연스럽게 이어진다. 손잡이는 문의 잠금판에 고정되어 있다. 머리는 앞으로 굽힌 목과 상체에 지지되고 몸통은 화면 아래로 이어진다. 발은 구도 밖이지만 공중에 떠 있거나 지지 없이 매달린 징후는 없다. 문이 닫힌 상태에서 손잡이를 잡고 멈춘 자세는 물리적으로 가능하다."
       },
       {
        "label": "B",
        "direction": "얼굴이 깊게 숙여져 눈은 손과 문 아래쪽을 향하고 있다. 오른쪽으로 굽힌 팔 끝의 손이 문손잡이를 감싸며, 문을 열기 직전에 멈춘 방향 관계가 분명하다. 시선은 카메라로 향하지 않는다.",
        "built_space": "오른쪽의 닫힌 철문 하나에 손잡이 하나와 세로 잠금판, 가장자리 경첩이 보이며 인물은 그 바깥에 붙어 서 있다. 손잡이와 인물의 크기 관계는 자연스럽고 고정 부속의 중복은 없다. 녹슨 철판과 문틀, 배관, 컨테이너 골목과 끝의 철문이 참조 장소와 부합한다. 얼굴·어깨·손이 크게 잡혀 A보다 클로즈업에 가깝지만 왼쪽 골목도 상당히 넓게 노출된다. 여러 조명이 켜져 있고 문 상단에 따뜻한 빛이 있어 정전 조건은 충족하지 못한다.",
        "entities": "젊은 동아시아계 남성 한 명만 등장하며 검고 헝클어진 머리, 마른 체격과 낡고 해진 회색 후드가 인물 참조에 잘 맞는다. 옆얼굴은 청소년 후반으로 보이고 얼굴에 상처 흔적이 있다. 보이는 손과 소매도 같은 인물의 것으로 자연스럽게 이어진다. 낡은 철제 손잡이는 있지만 손에 많이 가려져 돌출 윤곽은 A보다 덜 드러난다. 다리 부상은 구도 밖이므로 판단할 수 없다. 읽을 수 있는 번호나 문구는 없다.",
        "hard_violations": [],
        "physics": "굽힌 팔과 손목이 이어지고 손가락들이 문에 고정된 손잡이를 실제로 붙잡고 있다. 머리를 숙이고 어깨를 앞으로 둔 자세는 서서 망설이는 동작으로 가능하다. 하체는 화면 밖으로 이어지며 부유를 나타내는 요소는 없다. 문틈에 열린 공간이 보이지 않아 손잡이만 잡은 채 문을 닫아 둔 상태와 양립한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "고개를 숙이고 닫힌 문의 손잡이를 쥔 동작과 장소는 맞지만, 허벅지까지 보이는 넓은 구도가 지정된 클로즈업을 벗어나고 켜진 조명이 정전 설정에 어긋난다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "숙인 얼굴과 손잡이를 쥔 손을 더 밀착해 담고 낡은 후드 의상도 잘 맞지만, 골목 노출이 여전히 넓고 켜진 조명들이 야간 정전 설정을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "고개와 눈이 아래로 향하며, 시선은 손잡이보다 낮은 문 앞 바닥 쪽으로 내려간다. 한쪽 팔은 오른쪽 문으로 뻗어 손잡이를 쥐고 있다. 카메라를 바라보지 않으며 고개를 푹 숙인 망설임은 전달된다.",
        "built_space": "오른쪽에 닫힌 철문 하나, 돌출된 손잡이 하나와 세로 잠금판, 문 가장자리 경첩이 보인다. 인물은 문 바깥에 서 있어 손잡이 접근이 자연스럽다. 녹슨 도장 철판, 인접 창문과 배관, 젖은 골목 및 끝의 철문은 장소 참조와 잘 이어진다. 다만 문 외벽의 제한된 일부 대신 골목 깊이와 인물의 허벅지까지 넓게 담았다. 여러 가로등과 출입구 조명이 켜져 있고 문 위에도 따뜻한 빛이 떨어져 정전 상태가 아니다.",
        "entities": "헝클어진 검은 머리의 젊은 동아시아계 남성 한 명이며, 보이는 범위에서 이현우의 외형과 대체로 부합한다. 숙인 얼굴 때문에 정확한 얼굴 일치는 제한적으로만 확인된다. 어두운 낡은 상의와 청바지, 얼굴과 손의 상처 흔적이 보인다. 참조의 후드 형태는 뚜렷하지 않다. 낡은 금속 손잡이는 실제 문 부속으로 보이며 읽을 수 있는 글씨나 번호는 없다. 화면 하단의 다른 손 곁에는 갈색 고리 모양 일부가 있으나 물체의 종류는 확정하기 어렵다.",
        "hard_violations": [],
        "physics": "손가락이 손잡이를 감싸고 손목과 팔뚝이 자연스럽게 이어진다. 손잡이는 문의 잠금판에 고정되어 있다. 머리는 앞으로 굽힌 목과 상체에 지지되고 몸통은 화면 아래로 이어진다. 발은 구도 밖이지만 공중에 떠 있거나 지지 없이 매달린 징후는 없다. 문이 닫힌 상태에서 손잡이를 잡고 멈춘 자세는 물리적으로 가능하다."
       },
       {
        "label": "A",
        "direction": "얼굴이 깊게 숙여져 눈은 손과 문 아래쪽을 향하고 있다. 오른쪽으로 굽힌 팔 끝의 손이 문손잡이를 감싸며, 문을 열기 직전에 멈춘 방향 관계가 분명하다. 시선은 카메라로 향하지 않는다.",
        "built_space": "오른쪽의 닫힌 철문 하나에 손잡이 하나와 세로 잠금판, 가장자리 경첩이 보이며 인물은 그 바깥에 붙어 서 있다. 손잡이와 인물의 크기 관계는 자연스럽고 고정 부속의 중복은 없다. 녹슨 철판과 문틀, 배관, 컨테이너 골목과 끝의 철문이 참조 장소와 부합한다. 얼굴·어깨·손이 크게 잡혀 A보다 클로즈업에 가깝지만 왼쪽 골목도 상당히 넓게 노출된다. 여러 조명이 켜져 있고 문 상단에 따뜻한 빛이 있어 정전 조건은 충족하지 못한다.",
        "entities": "젊은 동아시아계 남성 한 명만 등장하며 검고 헝클어진 머리, 마른 체격과 낡고 해진 회색 후드가 인물 참조에 잘 맞는다. 옆얼굴은 청소년 후반으로 보이고 얼굴에 상처 흔적이 있다. 보이는 손과 소매도 같은 인물의 것으로 자연스럽게 이어진다. 낡은 철제 손잡이는 있지만 손에 많이 가려져 돌출 윤곽은 A보다 덜 드러난다. 다리 부상은 구도 밖이므로 판단할 수 없다. 읽을 수 있는 번호나 문구는 없다.",
        "hard_violations": [],
        "physics": "굽힌 팔과 손목이 이어지고 손가락들이 문에 고정된 손잡이를 실제로 붙잡고 있다. 머리를 숙이고 어깨를 앞으로 둔 자세는 서서 망설이는 동작으로 가능하다. 하체는 화면 밖으로 이어지며 부유를 나타내는 요소는 없다. 문틈에 열린 공간이 보이지 않아 손잡이만 잡은 채 문을 닫아 둔 상태와 양립한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.429
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.429
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "클로즈업 샷 크기를 준수하고, 숙인 얼굴 바로 아래에 문고리가 위치하는 요구 구도를 정확히 구현했습니다."
   },
   {
    "label": "B",
    "score": 1429,
    "verdict_ko": "프레이밍이 클로즈업보다 넓으며, 팔을 길게 뻗어 문고리가 얼굴 아래가 아닌 화면 우측으로 멀어져 구도 지시를 어겼습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_family_container_front_d0287f.png",
    "asset_id": "8e091bc8-65c8-4678-9f88-e299bf1a00e4",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd84-4064-796e-b3e0-85337af162ac",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S11sh8__bgfirst_bg.png",
   "bg_asset_id": "e8e203ba-e45f-45df-86f0-07af1f703e69",
   "bg_record_key": "S11sh8::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "family_container_front",
   "groupbg_asset_id": "8e091bc8-65c8-4678-9f88-e299bf1a00e4"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S11sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:20:52.178288+00:00",
  "fingerprint": "2cf37cd4dc8c5ae24e520f8a15fe1cbbc16561de88d35ec7cc03ea0ed3fc96b0",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S11sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S11sh8_sel.png",
  "source_sha256": "a720b41dd3d8af607b8c735ecaabe111170ed33cf5c07c0f8dea27e6852ae293",
  "file": "S11sh8_cine.png",
  "staged_sha256": "c81e66689aceaaff2fdf542853425375b158136a97faee08d19e77c9e5cb63bb",
  "latency_ms": 15690
 },
 "S12sh5::signage": {
  "fp": "c5c63e28987e8461",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::607617a5512240af": {
  "subjects": [],
  "subject_text": "이현우의 컨테이너 내부 식탁과 공용 생활공간\n금속 컨테이너 안의 소박한 공용 생활공간. 식탁 위에 촛대와 식기가 놓여 있고 생활용품과 별실로 이어지는 방문이 보인다.",
  "identity": "canonical",
  "scope_id": "L22",
  "scope_role": "location_interior",
  "scope_sha": "d7d9f0c1e4e684ba"
 },
 "S12sh5::bgfirst_bg": {
  "input_fingerprint": "0ad59765a79482fc",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 바지가 찢겨 피가 묻은 이현우의 다리 쪽에 시선을 둔 채 입을 크게 벌린 놀란 얼굴.\n\nLOCATION (lock): Beside the dining table in a refugee family's container home, lit by the candle set beside the evening meal.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Dining table edge (Beside 이현우 during his meal) — The outer edge recedes obliquely beside 앰버 and above the visible leg; used as Connects the child and seated adult without enlarging the foreground leg; 이현우's torn trousers (Torn and blood-marked around the injured leg) — The damaged area faces obliquely toward the camera at lower right; used as Supplies the visible cause of 앰버's reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established table candle provides intimate warm illumination, with restrained contrast preserving the blood detail and 앰버's shocked expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 바지가 찢겨 피가 묻은 이현우의 다리 쪽에 시선을 둔 채 입을 크게 벌린 놀란 얼굴.\n\nLOCATION (lock): Beside the dining table in a refugee family's container home, lit by the candle set beside the evening meal.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Dining table edge (Beside 이현우 during his meal) — The outer edge recedes obliquely beside 앰버 and above the visible leg; used as Connects the child and seated adult without enlarging the foreground leg; 이현우's torn trousers (Torn and blood-marked around the injured leg) — The damaged area faces obliquely toward the camera at lower right; used as Supplies the visible cause of 앰버's reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established table candle provides intimate warm illumination, with restrained contrast preserving the blood detail and 앰버's shocked expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S12sh5__bgfirst_bg.png",
  "asset_id": "ae902a22-17cb-462b-99b5-3a080f1db9b2",
  "input_asset_ids": [
   "f03378b3-eaad-4a06-a08a-e250aa20d05a",
   "26032ccf-781f-486e-8ac7-e58a4659868d"
  ]
 },
 "S12sh5": {
  "input_fingerprint": "80c2e3a3dfe3206f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 앰버가 바지가 찢겨 피가 묻은 이현우의 다리 쪽에 시선을 둔 채 입을 크게 벌린 놀란 얼굴.\n\nLOCATION (lock): Beside the dining table in a refugee family's container home, lit by the candle set beside the evening meal. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Dining table edge (Beside 이현우 during his meal) — The outer edge recedes obliquely beside 앰버 and above the visible leg; used as Connects the child and seated adult without enlarging the foreground leg; 이현우's torn trousers (Torn and blood-marked around the injured leg) — The damaged area faces obliquely toward the camera at lower right; used as Supplies the visible cause of 앰버's reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established table candle provides intimate warm illumination, with restrained contrast preserving the blood detail and 앰버's shocked expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A burning candle lights the dining table, where Hyunwoo's meal has been laid out. The old clothes being mended remain at the table. 이현우: Hyunwoo has removed his outer top and is eating at the table. His face is injured, and his dog-bitten leg is visibly bleeding. 앰버: Amber remains near the dining table, visibly startled and still coughing intermittently.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 앰버가 바지가 찢겨 피가 묻은 이현우의 다리 쪽에 시선을 둔 채 입을 크게 벌린 놀란 얼굴.\n\nLOCATION (lock): Beside the dining table in a refugee family's container home, lit by the candle set beside the evening meal. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Dining table edge (Beside 이현우 during his meal) — The outer edge recedes obliquely beside 앰버 and above the visible leg; used as Connects the child and seated adult without enlarging the foreground leg; 이현우's torn trousers (Torn and blood-marked around the injured leg) — The damaged area faces obliquely toward the camera at lower right; used as Supplies the visible cause of 앰버's reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established table candle provides intimate warm illumination, with restrained contrast preserving the blood detail and 앰버's shocked expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A burning candle lights the dining table, where Hyunwoo's meal has been laid out. The old clothes being mended remain at the table. 이현우: Hyunwoo has removed his outer top and is eating at the table. His face is injured, and his dog-bitten leg is visibly bleeding. 앰버: Amber remains near the dining table, visibly startled and still coughing intermittently.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 앰버가 바지가 찢겨 피가 묻은 이현우의 다리 쪽에 시선을 둔 채 입을 크게 벌린 놀란 얼굴.\n\nLOCATION (lock): Beside the dining table in a refugee family's container home, lit by the candle set beside the evening meal. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Dining table edge (Beside 이현우 during his meal) — The outer edge recedes obliquely beside 앰버 and above the visible leg; used as Connects the child and seated adult without enlarging the foreground leg; 이현우's torn trousers (Torn and blood-marked around the injured leg) — The damaged area faces obliquely toward the camera at lower right; used as Supplies the visible cause of 앰버's reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established table candle provides intimate warm illumination, with restrained contrast preserving the blood detail and 앰버's shocked expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A burning candle lights the dining table, where Hyunwoo's meal has been laid out. The old clothes being mended remain at the table. 이현우: Hyunwoo has removed his outer top and is eating at the table. His face is injured, and his dog-bitten leg is visibly bleeding. 앰버: Amber remains near the dining table, visibly startled and still coughing intermittently.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S12sh5__bgfirst_bg.png",
     "asset_id": "ae902a22-17cb-462b-99b5-3a080f1db9b2",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S12sh5.png",
     "asset_id": "f03378b3-eaad-4a06-a08a-e250aa20d05a",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L22B02.png",
     "asset_id": "26032ccf-781f-486e-8ac7-e58a4659868d",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "앰버는 이현우의 다리 쪽을 바라보나, 다리에 아무런 상처가 없음.",
    "built_space": "식탁, 촛불, 배경의 선반과 문 등 레퍼런스 공간이 넓게 포착됨.",
    "entities": "앰버(금발, 놀란 표정), 이현우(반팔 티셔츠 착용). 이현우의 바지가 찢어지거나 피가 묻지 않음.",
    "hard_violations": [],
    "physics": "앰버는 바닥에 무릎을 꿇고 있고, 이현우는 의자에 앉아 젓가락을 들고 있음."
   },
   {
    "label": "B",
    "direction": "앰버의 시선이 이현우의 찢어지고 피가 묻은 무릎을 향하고 있음.",
    "built_space": "식탁, 촛불, 배경의 액자 등 레퍼런스의 공간적 특징이 클로즈업 스케일에 맞게 배치됨.",
    "entities": "앰버(금발, 크게 벌린 입), 이현우(찢어지고 피 묻은 청바지, 단 상의를 벗지 않고 후드티 착용).",
    "hard_violations": [],
    "physics": "앰버는 상체를 기울여 지지하고 있으며, 이현우는 앉은 자세로 다리를 편안하게 두고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 클로즈업 프레이밍을 완벽히 따르며 찢어지고 피가 묻은 다리와 이에 놀란 앰버의 표정을 정확히 담아냈으나, 이현우가 겉옷을 벗지 않은 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프레이밍이 너무 넓고, 장면의 핵심인 이현우의 찢어지고 피가 묻은 다리 묘사가 완전히 누락되어 앰버의 놀란 반응이 무색해졌습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 시선이 이현우의 찢어지고 피가 묻은 무릎을 향하고 있음.",
        "built_space": "식탁, 촛불, 배경의 액자 등 레퍼런스의 공간적 특징이 클로즈업 스케일에 맞게 배치됨.",
        "entities": "앰버(금발, 크게 벌린 입), 이현우(찢어지고 피 묻은 청바지, 단 상의를 벗지 않고 후드티 착용).",
        "hard_violations": [],
        "physics": "앰버는 상체를 기울여 지지하고 있으며, 이현우는 앉은 자세로 다리를 편안하게 두고 있음."
       },
       {
        "label": "A",
        "direction": "앰버는 이현우의 다리 쪽을 바라보나, 다리에 아무런 상처가 없음.",
        "built_space": "식탁, 촛불, 배경의 선반과 문 등 레퍼런스 공간이 넓게 포착됨.",
        "entities": "앰버(금발, 놀란 표정), 이현우(반팔 티셔츠 착용). 이현우의 바지가 찢어지거나 피가 묻지 않음.",
        "hard_violations": [],
        "physics": "앰버는 바닥에 무릎을 꿇고 있고, 이현우는 의자에 앉아 젓가락을 들고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 클로즈업 프레이밍을 완벽히 따르며 찢어지고 피가 묻은 다리와 이에 놀란 앰버의 표정을 정확히 담아냈으나, 이현우가 겉옷을 벗지 않은 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프레이밍이 너무 넓고, 장면의 핵심인 이현우의 찢어지고 피가 묻은 다리 묘사가 완전히 누락되어 앰버의 놀란 반응이 무색해졌습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 시선이 이현우의 찢어지고 피가 묻은 무릎을 향하고 있음.",
        "built_space": "식탁, 촛불, 배경의 액자 등 레퍼런스의 공간적 특징이 클로즈업 스케일에 맞게 배치됨.",
        "entities": "앰버(금발, 크게 벌린 입), 이현우(찢어지고 피 묻은 청바지, 단 상의를 벗지 않고 후드티 착용).",
        "hard_violations": [],
        "physics": "앰버는 상체를 기울여 지지하고 있으며, 이현우는 앉은 자세로 다리를 편안하게 두고 있음."
       },
       {
        "label": "A",
        "direction": "앰버는 이현우의 다리 쪽을 바라보나, 다리에 아무런 상처가 없음.",
        "built_space": "식탁, 촛불, 배경의 선반과 문 등 레퍼런스 공간이 넓게 포착됨.",
        "entities": "앰버(금발, 놀란 표정), 이현우(반팔 티셔츠 착용). 이현우의 바지가 찢어지거나 피가 묻지 않음.",
        "hard_violations": [],
        "physics": "앰버는 바닥에 무릎을 꿇고 있고, 이현우는 의자에 앉아 젓가락을 들고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "앰버의 크게 벌린 입을 근접 촬영하고 오른쪽 아래에 피 묻은 찢어진 바지를 보여 핵심 순간을 구현하지만, 다리가 전경에서 과장되고 이현우가 겉옷을 그대로 입고 있다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "식사 중인 이현우와 다리를 내려다보는 앰버의 관계는 자연스럽지만, 클로즈업보다 넓은 구도이며 화면에 나온 바지에 찢김과 출혈이 없어 놀람의 원인이 빠졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 카메라가 아니라 오른쪽 아래 이현우의 무릎이 있는 쪽을 바라본다. 다리 쪽을 향하는 방향은 맞지만 눈동자가 상처 자체에 정확히 고정됐는지는 다소 모호하다. 이현우의 눈은 프레임 밖이라 시선은 확인할 수 없다.",
        "built_space": "나무 식탁 하나가 두 사람 사이 뒤쪽에 있고, 상판 위에 불붙은 양초 하나, 음식 그릇 두 개, 금속 주전자와 옷가지가 보인다. 왼쪽 책장과 화분, 뒤쪽 벽 액자 일부는 장소 참고와 부합한다. 앰버는 식탁 왼쪽 가까이, 이현우는 오른쪽에 앉아 있다. 식탁 가장자리는 다리 위에 놓이지만 비교적 수평이고, 전경 다리가 하단을 크게 차지하여 요구된 절제된 다리 크기와는 차이가 있다.",
        "entities": "앰버는 금발에 큰 눈을 가진 열 살 안팎 여자아이로 보이며, 참고의 얼굴 인상과 녹색 겉옷을 대체로 따른다. 입을 크게 벌린 놀란 표정이 선명하다. 이현우는 턱과 손, 몸통 및 다리만 보여 얼굴 정체성이나 머리 모양은 판단하기 어렵다. 찢어진 청바지에 노출된 무릎 상처와 피부 및 천에 묻은 피가 실제 질감으로 보인다. 다만 벗어야 할 어두운 후드 겉옷을 입고 있으며 식사 동작은 보이지 않는다. 식사, 촛불, 수선 중인 옷감은 있고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우의 허벅지는 오른쪽 몸통에서 무릎으로 이어지고 종아리는 왼쪽 아래로 내려가 앉은 자세로 성립한다. 좌석과 발은 화면 밖이지만 몸이 떠 있다고 볼 근거는 없다. 손은 무릎 뒤쪽에 자연스럽게 내려와 있다. 앰버의 하체 지지는 잘렸으며 상체 자세는 가능하다. 양초와 그릇, 주전자, 옷감은 식탁에 놓여 있고 피는 피부와 청바지 표면을 따른다."
       },
       {
        "label": "B",
        "direction": "앰버는 고개와 눈을 오른쪽 아래로 향해 이현우의 화면에 보이는 무릎과 다리 쪽을 내려다본다. 그러나 그 시선의 대상에는 상처나 피가 보이지 않는다. 이현우도 아래쪽을 보며, 손에 든 젓가락은 식탁의 음식 그릇 쪽으로 향한다.",
        "built_space": "식탁 하나, 상판의 양초 하나, 뒤쪽과 식탁 아래에 걸쳐 나무 걸상 두 개가 보인다. 왼쪽 출입구와 전자레인지, 벽 사진과 중앙 액자, 오른쪽 커튼과 꽃병 등은 장소 참고를 잘 따른다. 앰버는 식탁 왼쪽에서 몸을 낮추고 이현우는 오른쪽에 앉아 있으며, 식탁 가장자리가 다리 위를 비스듬히 지난다. 다만 방과 두 사람의 신체를 넓게 보여 앰버 얼굴 중심의 클로즈업이 아니다.",
        "entities": "앰버는 금발의 어린 여자아이이고 큰 눈과 얼굴 인상은 참고에 가깝지만 녹색 겉옷과 회색 원피스 대신 반소매 티셔츠와 바지를 입었다. 입은 벌어져 있으나 요구된 크게 벌린 놀람은 A보다 약하다. 이현우는 헝클어진 검은 머리의 젊은 한국계 남성이라는 설정과 참고 인상에 대체로 맞으며 겉옷을 벗고 젓가락을 들고 있다. 얼굴 부상은 선명하지 않다. 화면에 나온 회갈색 바지는 찢어지지 않았고 출혈도 보이지 않아 핵심 소품 상태가 빠졌다. 촛불, 식사와 옷가지는 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앰버는 몸을 낮춘 채 한 손을 입 가까이에 두고 다른 팔을 아래로 내려 기침 직후처럼 가능한 자세를 취한다. 하체와 발의 접지는 프레임 밖이며 부유하는 모습은 아니다. 이현우의 허벅지와 굽힌 무릎은 오른쪽 좌석에 앉은 몸에서 자연스럽게 이어지고, 젓가락은 손가락으로 잡고 있다. 식탁의 음식과 촛불, 옷감은 모두 상판에 지지되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "앰버의 크게 벌린 입을 근접 촬영하고 오른쪽 아래에 피 묻은 찢어진 바지를 보여 핵심 순간을 구현하지만, 다리가 전경에서 과장되고 이현우가 겉옷을 그대로 입고 있다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "식사 중인 이현우와 다리를 내려다보는 앰버의 관계는 자연스럽지만, 클로즈업보다 넓은 구도이며 화면에 나온 바지에 찢김과 출혈이 없어 놀람의 원인이 빠졌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 카메라가 아니라 오른쪽 아래 이현우의 무릎이 있는 쪽을 바라본다. 다리 쪽을 향하는 방향은 맞지만 눈동자가 상처 자체에 정확히 고정됐는지는 다소 모호하다. 이현우의 눈은 프레임 밖이라 시선은 확인할 수 없다.",
        "built_space": "나무 식탁 하나가 두 사람 사이 뒤쪽에 있고, 상판 위에 불붙은 양초 하나, 음식 그릇 두 개, 금속 주전자와 옷가지가 보인다. 왼쪽 책장과 화분, 뒤쪽 벽 액자 일부는 장소 참고와 부합한다. 앰버는 식탁 왼쪽 가까이, 이현우는 오른쪽에 앉아 있다. 식탁 가장자리는 다리 위에 놓이지만 비교적 수평이고, 전경 다리가 하단을 크게 차지하여 요구된 절제된 다리 크기와는 차이가 있다.",
        "entities": "앰버는 금발에 큰 눈을 가진 열 살 안팎 여자아이로 보이며, 참고의 얼굴 인상과 녹색 겉옷을 대체로 따른다. 입을 크게 벌린 놀란 표정이 선명하다. 이현우는 턱과 손, 몸통 및 다리만 보여 얼굴 정체성이나 머리 모양은 판단하기 어렵다. 찢어진 청바지에 노출된 무릎 상처와 피부 및 천에 묻은 피가 실제 질감으로 보인다. 다만 벗어야 할 어두운 후드 겉옷을 입고 있으며 식사 동작은 보이지 않는다. 식사, 촛불, 수선 중인 옷감은 있고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우의 허벅지는 오른쪽 몸통에서 무릎으로 이어지고 종아리는 왼쪽 아래로 내려가 앉은 자세로 성립한다. 좌석과 발은 화면 밖이지만 몸이 떠 있다고 볼 근거는 없다. 손은 무릎 뒤쪽에 자연스럽게 내려와 있다. 앰버의 하체 지지는 잘렸으며 상체 자세는 가능하다. 양초와 그릇, 주전자, 옷감은 식탁에 놓여 있고 피는 피부와 청바지 표면을 따른다."
       },
       {
        "label": "A",
        "direction": "앰버는 고개와 눈을 오른쪽 아래로 향해 이현우의 화면에 보이는 무릎과 다리 쪽을 내려다본다. 그러나 그 시선의 대상에는 상처나 피가 보이지 않는다. 이현우도 아래쪽을 보며, 손에 든 젓가락은 식탁의 음식 그릇 쪽으로 향한다.",
        "built_space": "식탁 하나, 상판의 양초 하나, 뒤쪽과 식탁 아래에 걸쳐 나무 걸상 두 개가 보인다. 왼쪽 출입구와 전자레인지, 벽 사진과 중앙 액자, 오른쪽 커튼과 꽃병 등은 장소 참고를 잘 따른다. 앰버는 식탁 왼쪽에서 몸을 낮추고 이현우는 오른쪽에 앉아 있으며, 식탁 가장자리가 다리 위를 비스듬히 지난다. 다만 방과 두 사람의 신체를 넓게 보여 앰버 얼굴 중심의 클로즈업이 아니다.",
        "entities": "앰버는 금발의 어린 여자아이이고 큰 눈과 얼굴 인상은 참고에 가깝지만 녹색 겉옷과 회색 원피스 대신 반소매 티셔츠와 바지를 입었다. 입은 벌어져 있으나 요구된 크게 벌린 놀람은 A보다 약하다. 이현우는 헝클어진 검은 머리의 젊은 한국계 남성이라는 설정과 참고 인상에 대체로 맞으며 겉옷을 벗고 젓가락을 들고 있다. 얼굴 부상은 선명하지 않다. 화면에 나온 회갈색 바지는 찢어지지 않았고 출혈도 보이지 않아 핵심 소품 상태가 빠졌다. 촛불, 식사와 옷가지는 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앰버는 몸을 낮춘 채 한 손을 입 가까이에 두고 다른 팔을 아래로 내려 기침 직후처럼 가능한 자세를 취한다. 하체와 발의 접지는 프레임 밖이며 부유하는 모습은 아니다. 이현우의 허벅지와 굽힌 무릎은 오른쪽 좌석에 앉은 몸에서 자연스럽게 이어지고, 젓가락은 손가락으로 잡고 있다. 식탁의 음식과 촛불, 옷감은 모두 상판에 지지되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.0,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 클로즈업 프레이밍을 완벽히 따르며 찢어지고 피가 묻은 다리와 이에 놀란 앰버의 표정을 정확히 담아냈으나, 이현우가 겉옷을 벗지 않은 점이 아쉽습니다."
   },
   {
    "label": "A",
    "score": 1000,
    "verdict_ko": "프레이밍이 너무 넓고, 장면의 핵심인 이현우의 찢어지고 피가 묻은 다리 묘사가 완전히 누락되어 앰버의 놀란 반응이 무색해졌습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L22B02.png",
    "asset_id": "26032ccf-781f-486e-8ac7-e58a4659868d",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd85-42f6-786a-a473-4c78601051a6",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S12sh5__bgfirst_bg.png",
   "bg_asset_id": "ae902a22-17cb-462b-99b5-3a080f1db9b2",
   "bg_record_key": "S12sh5::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S12sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:21:08.275522+00:00",
  "fingerprint": "dc48da605920bf5edb808f40aa5e27d8ab0f07822dca2760d52275b7db0ff766",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S12sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S12sh5_sel.png",
  "source_sha256": "cb9a70bb4d642deb23918ca215fc48e8e25b546aa00a96e8aed690b097d1f144",
  "file": "S12sh5_cine.png",
  "staged_sha256": "f9de22b25edad238a490ef6805eea7d17c8040a03f072a3c31d501511538342c",
  "latency_ms": 12929
 },
 "S12sh20::signage": {
  "fp": "45b487c484c308f8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S12sh20": {
  "input_fingerprint": "e6c9d5ecd4385b24",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연이 당황한 눈으로, 문 쪽을 향해 발을 내디디며 뒷모습을 보인 이현우의 굽은 등짝을 바라보는 구도.\n\nLOCATION (lock): In the shared living area between the candlelit dining table and the exit of the family container home. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Dining table corner separating the seated mother from the departure route in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Dining table corner (Still beside the seated 미연 as 이현우 leaves) — An oblique corner projects between 미연's seated position and 이현우's departure route; used as Anchors their shared room and keeps the separation in depth understandable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established candlelit interior, allowing restrained warmth and facial contrast to carry 미연's alarm without adding entrance lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The table candle remains lit, with the meal and mending materials still present. The medicine container has been opened, but no treatment has been shown. 이현우: Hyunwoo has put his outer top back on and risen to leave. His facial injuries and bleeding dog-bitten leg remain untreated on screen. 미연: Miyeon remains inside near the table and the opened medicine container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연이 당황한 눈으로, 문 쪽을 향해 발을 내디디며 뒷모습을 보인 이현우의 굽은 등짝을 바라보는 구도.\n\nLOCATION (lock): In the shared living area between the candlelit dining table and the exit of the family container home. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Dining table corner separating the seated mother from the departure route in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Dining table corner (Still beside the seated 미연 as 이현우 leaves) — An oblique corner projects between 미연's seated position and 이현우's departure route; used as Anchors their shared room and keeps the separation in depth understandable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established candlelit interior, allowing restrained warmth and facial contrast to carry 미연's alarm without adding entrance lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The table candle remains lit, with the meal and mending materials still present. The medicine container has been opened, but no treatment has been shown. 이현우: Hyunwoo has put his outer top back on and risen to leave. His facial injuries and bleeding dog-bitten leg remain untreated on screen. 미연: Miyeon remains inside near the table and the opened medicine container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연이 당황한 눈으로, 문 쪽을 향해 발을 내디디며 뒷모습을 보인 이현우의 굽은 등짝을 바라보는 구도.\n\nLOCATION (lock): In the shared living area between the candlelit dining table and the exit of the family container home. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Dining table corner separating the seated mother from the departure route in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Dining table corner (Still beside the seated 미연 as 이현우 leaves) — An oblique corner projects between 미연's seated position and 이현우's departure route; used as Anchors their shared room and keeps the separation in depth understandable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established candlelit interior, allowing restrained warmth and facial contrast to carry 미연's alarm without adding entrance lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The table candle remains lit, with the meal and mending materials still present. The medicine container has been opened, but no treatment has been shown. 이현우: Hyunwoo has put his outer top back on and risen to leave. His facial injuries and bleeding dog-bitten leg remain untreated on screen. 미연: Miyeon remains inside near the table and the opened medicine container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "미연의 시선은 화면 우측을 향해 걸어가는 이현우의 등을 정확히 향하고 있으며, 이현우는 열린 문 쪽을 향해 뒷모습을 보이고 있습니다.",
    "built_space": "좌측 하단에 식탁 모서리가 배치되어 미연과 이현우의 동선을 분리하는 구도를 잘 형성하고 있으며, 배경의 컨테이너 구조도 자연스럽습니다.",
    "entities": "미연과 이현우의 의상 및 외양 레퍼런스가 잘 일치하며, 촛불, 음식, 약통 등의 명시된 소품들이 테이블 위에 존재합니다. 이현우의 다리에 혈흔이 묘사되어 있습니다.",
    "hard_violations": [],
    "physics": "미연은 의자에 자연스럽게 앉아 테이블에 지지하고 있으며, 이현우는 바닥을 딛고 안정적으로 걸어가고 있습니다."
   },
   {
    "label": "B",
    "direction": "이현우는 문을 향해 걸어가고 있으나, 미연은 그를 보지 않고 카메라 렌즈를 정면으로 응시하고 있습니다.",
    "built_space": "중앙에 식탁이 위치하며 미연이 그 뒤에 앉아 있고 뒷배경으로 문이 열려 있으나, 요구된 동선 분리 구도가 명확하지 않습니다.",
    "entities": "인물들의 외양 레퍼런스는 일치하나, 미연이 약통을 손에 들고 있어 '아무런 치료가 나타나지 않음'이라는 상태와 충돌할 여지가 있습니다.",
    "hard_violations": [],
    "physics": "미연과 이현우 모두 바닥과 의자 등에 물리적으로 올바르게 지지되어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미연이 이현우의 뒷모습을 바라보는 샷의 핵심 구도와 시선 방향을 정확히 연출하여 프롬프트의 요구사항을 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미연이 이현우를 바라보지 않고 정면의 카메라를 응시하여 프롬프트의 가장 중요한 시선 지시를 완전히 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연의 시선은 화면 우측을 향해 걸어가는 이현우의 등을 정확히 향하고 있으며, 이현우는 열린 문 쪽을 향해 뒷모습을 보이고 있습니다.",
        "built_space": "좌측 하단에 식탁 모서리가 배치되어 미연과 이현우의 동선을 분리하는 구도를 잘 형성하고 있으며, 배경의 컨테이너 구조도 자연스럽습니다.",
        "entities": "미연과 이현우의 의상 및 외양 레퍼런스가 잘 일치하며, 촛불, 음식, 약통 등의 명시된 소품들이 테이블 위에 존재합니다. 이현우의 다리에 혈흔이 묘사되어 있습니다.",
        "hard_violations": [],
        "physics": "미연은 의자에 자연스럽게 앉아 테이블에 지지하고 있으며, 이현우는 바닥을 딛고 안정적으로 걸어가고 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우는 문을 향해 걸어가고 있으나, 미연은 그를 보지 않고 카메라 렌즈를 정면으로 응시하고 있습니다.",
        "built_space": "중앙에 식탁이 위치하며 미연이 그 뒤에 앉아 있고 뒷배경으로 문이 열려 있으나, 요구된 동선 분리 구도가 명확하지 않습니다.",
        "entities": "인물들의 외양 레퍼런스는 일치하나, 미연이 약통을 손에 들고 있어 '아무런 치료가 나타나지 않음'이라는 상태와 충돌할 여지가 있습니다.",
        "hard_violations": [],
        "physics": "미연과 이현우 모두 바닥과 의자 등에 물리적으로 올바르게 지지되어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미연이 이현우의 뒷모습을 바라보는 샷의 핵심 구도와 시선 방향을 정확히 연출하여 프롬프트의 요구사항을 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미연이 이현우를 바라보지 않고 정면의 카메라를 응시하여 프롬프트의 가장 중요한 시선 지시를 완전히 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "미연의 시선은 화면 우측을 향해 걸어가는 이현우의 등을 정확히 향하고 있으며, 이현우는 열린 문 쪽을 향해 뒷모습을 보이고 있습니다.",
        "built_space": "좌측 하단에 식탁 모서리가 배치되어 미연과 이현우의 동선을 분리하는 구도를 잘 형성하고 있으며, 배경의 컨테이너 구조도 자연스럽습니다.",
        "entities": "미연과 이현우의 의상 및 외양 레퍼런스가 잘 일치하며, 촛불, 음식, 약통 등의 명시된 소품들이 테이블 위에 존재합니다. 이현우의 다리에 혈흔이 묘사되어 있습니다.",
        "hard_violations": [],
        "physics": "미연은 의자에 자연스럽게 앉아 테이블에 지지하고 있으며, 이현우는 바닥을 딛고 안정적으로 걸어가고 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우는 문을 향해 걸어가고 있으나, 미연은 그를 보지 않고 카메라 렌즈를 정면으로 응시하고 있습니다.",
        "built_space": "중앙에 식탁이 위치하며 미연이 그 뒤에 앉아 있고 뒷배경으로 문이 열려 있으나, 요구된 동선 분리 구도가 명확하지 않습니다.",
        "entities": "인물들의 외양 레퍼런스는 일치하나, 미연이 약통을 손에 들고 있어 '아무런 치료가 나타나지 않음'이라는 상태와 충돌할 여지가 있습니다.",
        "hard_violations": [],
        "physics": "미연과 이현우 모두 바닥과 의자 등에 물리적으로 올바르게 지지되어 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "굽은 등을 보이며 문으로 내딛는 동작은 잘 구현했지만, 전신까지 넓힌 구도와 현우가 아닌 카메라 쪽을 보는 미연의 시선이 핵심 지시에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "요구한 미디엄 숏에 상대적으로 가까운 크롭과 식탁 모서리의 분리는 우세하지만, 미연의 시선이 현우의 등에 닿지 않고 현우의 굽은 등과 출발 동작도 약합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 눈을 크게 뜨고 화면 정면의 카메라 부근을 봅니다. 오른쪽 뒤에 있는 현우의 등을 보는 방향이 아닙니다. 현우는 카메라에 등을 보이고 중앙 뒤쪽의 열린 출입구를 향해 몸과 발을 옮깁니다.",
        "built_space": "왼쪽에 나무 식탁 하나, 미연의 좌석 하나와 전경의 빈 의자 하나가 보입니다. 뒤쪽 출입구 하나에는 양쪽으로 열린 문짝 두 개가 있고, 오른쪽에는 냄비를 놓은 조리대와 수납장이 있습니다. 식탁의 비스듬한 모서리가 미연과 출구로 이어지는 통로를 나눕니다. 따뜻한 벽과 낡은 나무 식탁은 이전 장면과 비교적 이어지지만, 현우의 신발까지 들어오는 넓은 구도는 지정된 미디엄 숏이 아닙니다. 출입구 주변에는 차가운 빛이 보입니다.",
        "entities": "인물은 중년 한국인 여성으로 보이는 미연과 검은 머리의 젊은 남성 현우, 두 명뿐입니다. 미연의 단발머리와 갈색 작업 재킷은 참조와 유사하지만 모자는 없습니다. 현우의 얼굴은 뒷모습 때문에 확인할 수 없으며, 헝클어진 검은 머리와 낡은 짙은 후드 상의, 청바지는 참조에 부합합니다. 식탁에는 켜진 초 하나, 음식 접시들, 열린 흰 약통과 포장재가 있습니다. 바느질 재료는 분명하지 않습니다. 얼굴과 다리 상처는 이 방향에서 확인되지 않으므로 치료 여부도 판단할 수 없습니다. 명확하게 읽히는 문구는 식별되지 않습니다.",
        "hard_violations": [],
        "physics": "미연의 하체는 식탁 아래 의자에 앉아 있고 두 손은 식탁과 약품 부근에 놓여 있습니다. 현우는 오른발로 바닥을 지지하면서 왼발 뒤꿈치를 들어 보폭을 옮기는 모습이라 문으로 걸어가는 동작이 성립합니다. 어깨와 등이 앞으로 굽어 있습니다. 초, 약통과 접시는 식탁 위에 놓여 있으며 떠 있는 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "미연의 얼굴은 약간 오른쪽으로 돌아가 있지만 눈은 화면 정면 또는 카메라 가까운 쪽을 향합니다. 오른쪽 뒤의 현우 등짝을 응시한다고 보기 어렵습니다. 현우는 열린 출입구 바깥을 향해 등을 보이고 서 있습니다.",
        "built_space": "왼쪽 아래에 식탁 하나가 크게 놓이고 비스듬한 오른쪽 모서리가 미연의 좌석과 현우의 출구 동선을 나눕니다. 미연 뒤 좌석의 등받이와 왼쪽 빈 의자, 왼쪽 전경의 일부 가려진 의자가 보입니다. 오른쪽 출입구 하나에 열린 금속 문짝 두 개가 있습니다. 미연 뒤에는 작은 작업대가 있습니다. 현우를 정강이 부근에서 자른 구도는 A보다 미디엄 숏에 가깝지만 여전히 넓습니다. 노출된 단열재와 금속 벽은 이전 장면의 마감된 따뜻한 벽과 시각적 연속성이 약하고, 출입구의 차가운 빛도 두드러집니다.",
        "entities": "미연과 현우 두 명만 보입니다. 미연은 중년 한국인 여성의 외형, 검은 단발, 갈색 재킷과 회색 상의를 유지하지만 참조의 모자는 없습니다. 현우는 검은 헝클어진 머리, 낡은 짙은 후드 상의와 청바지를 입었으며 얼굴은 의도된 뒷모습으로 가려져 있습니다. 식탁에는 켜진 초 하나, 음식 접시들, 열린 흰 약통과 검은 작은 용기가 있고, 미연은 수선 중인 듯한 천을 잡고 있습니다. 현우의 왼쪽 무릎 부근에는 붉은 상처가 보이지만 이전 장면보다 상처 위치와 찢어진 청바지 형태가 다릅니다. 명확하게 읽히는 문구는 식별되지 않습니다.",
        "hard_violations": [],
        "physics": "미연은 의자에 앉아 양손으로 천을 잡고 있어 지지가 자연스럽습니다. 현우의 발은 프레임 밖이지만 다리는 아래로 이어지고 몸은 수직으로 서 있어 공중에 떠 있는 징후는 없습니다. 다만 양팔이 내려가고 어깨가 비교적 수평이며 보폭 변화가 약해, 등을 굽혀 발을 내딛는 순간보다 출입구 앞에 멈춘 자세에 가깝습니다. 식탁 위 물체들은 상판에 지지됩니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "굽은 등을 보이며 문으로 내딛는 동작은 잘 구현했지만, 전신까지 넓힌 구도와 현우가 아닌 카메라 쪽을 보는 미연의 시선이 핵심 지시에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "요구한 미디엄 숏에 상대적으로 가까운 크롭과 식탁 모서리의 분리는 우세하지만, 미연의 시선이 현우의 등에 닿지 않고 현우의 굽은 등과 출발 동작도 약합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "미연은 눈을 크게 뜨고 화면 정면의 카메라 부근을 봅니다. 오른쪽 뒤에 있는 현우의 등을 보는 방향이 아닙니다. 현우는 카메라에 등을 보이고 중앙 뒤쪽의 열린 출입구를 향해 몸과 발을 옮깁니다.",
        "built_space": "왼쪽에 나무 식탁 하나, 미연의 좌석 하나와 전경의 빈 의자 하나가 보입니다. 뒤쪽 출입구 하나에는 양쪽으로 열린 문짝 두 개가 있고, 오른쪽에는 냄비를 놓은 조리대와 수납장이 있습니다. 식탁의 비스듬한 모서리가 미연과 출구로 이어지는 통로를 나눕니다. 따뜻한 벽과 낡은 나무 식탁은 이전 장면과 비교적 이어지지만, 현우의 신발까지 들어오는 넓은 구도는 지정된 미디엄 숏이 아닙니다. 출입구 주변에는 차가운 빛이 보입니다.",
        "entities": "인물은 중년 한국인 여성으로 보이는 미연과 검은 머리의 젊은 남성 현우, 두 명뿐입니다. 미연의 단발머리와 갈색 작업 재킷은 참조와 유사하지만 모자는 없습니다. 현우의 얼굴은 뒷모습 때문에 확인할 수 없으며, 헝클어진 검은 머리와 낡은 짙은 후드 상의, 청바지는 참조에 부합합니다. 식탁에는 켜진 초 하나, 음식 접시들, 열린 흰 약통과 포장재가 있습니다. 바느질 재료는 분명하지 않습니다. 얼굴과 다리 상처는 이 방향에서 확인되지 않으므로 치료 여부도 판단할 수 없습니다. 명확하게 읽히는 문구는 식별되지 않습니다.",
        "hard_violations": [],
        "physics": "미연의 하체는 식탁 아래 의자에 앉아 있고 두 손은 식탁과 약품 부근에 놓여 있습니다. 현우는 오른발로 바닥을 지지하면서 왼발 뒤꿈치를 들어 보폭을 옮기는 모습이라 문으로 걸어가는 동작이 성립합니다. 어깨와 등이 앞으로 굽어 있습니다. 초, 약통과 접시는 식탁 위에 놓여 있으며 떠 있는 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "미연의 얼굴은 약간 오른쪽으로 돌아가 있지만 눈은 화면 정면 또는 카메라 가까운 쪽을 향합니다. 오른쪽 뒤의 현우 등짝을 응시한다고 보기 어렵습니다. 현우는 열린 출입구 바깥을 향해 등을 보이고 서 있습니다.",
        "built_space": "왼쪽 아래에 식탁 하나가 크게 놓이고 비스듬한 오른쪽 모서리가 미연의 좌석과 현우의 출구 동선을 나눕니다. 미연 뒤 좌석의 등받이와 왼쪽 빈 의자, 왼쪽 전경의 일부 가려진 의자가 보입니다. 오른쪽 출입구 하나에 열린 금속 문짝 두 개가 있습니다. 미연 뒤에는 작은 작업대가 있습니다. 현우를 정강이 부근에서 자른 구도는 A보다 미디엄 숏에 가깝지만 여전히 넓습니다. 노출된 단열재와 금속 벽은 이전 장면의 마감된 따뜻한 벽과 시각적 연속성이 약하고, 출입구의 차가운 빛도 두드러집니다.",
        "entities": "미연과 현우 두 명만 보입니다. 미연은 중년 한국인 여성의 외형, 검은 단발, 갈색 재킷과 회색 상의를 유지하지만 참조의 모자는 없습니다. 현우는 검은 헝클어진 머리, 낡은 짙은 후드 상의와 청바지를 입었으며 얼굴은 의도된 뒷모습으로 가려져 있습니다. 식탁에는 켜진 초 하나, 음식 접시들, 열린 흰 약통과 검은 작은 용기가 있고, 미연은 수선 중인 듯한 천을 잡고 있습니다. 현우의 왼쪽 무릎 부근에는 붉은 상처가 보이지만 이전 장면보다 상처 위치와 찢어진 청바지 형태가 다릅니다. 명확하게 읽히는 문구는 식별되지 않습니다.",
        "hard_violations": [],
        "physics": "미연은 의자에 앉아 양손으로 천을 잡고 있어 지지가 자연스럽습니다. 현우의 발은 프레임 밖이지만 다리는 아래로 이어지고 몸은 수직으로 서 있어 공중에 떠 있는 징후는 없습니다. 다만 양팔이 내려가고 어깨가 비교적 수평이며 보폭 변화가 약해, 등을 굽혀 발을 내딛는 순간보다 출입구 앞에 멈춘 자세에 가깝습니다. 식탁 위 물체들은 상판에 지지됩니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.229
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.229
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1229
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "미연이 이현우의 뒷모습을 바라보는 샷의 핵심 구도와 시선 방향을 정확히 연출하여 프롬프트의 요구사항을 충실히 반영했습니다."
   },
   {
    "label": "B",
    "score": 1229,
    "verdict_ko": "미연이 이현우를 바라보지 않고 정면의 카메라를 응시하여 프롬프트의 가장 중요한 시선 지시를 완전히 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S12sh5_sel.png",
    "asset_id": "fb40c30a-760b-43d4-a8a2-7e63de357faf",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:878964>",
    "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd86-1872-7628-ba09-8fa78701e6f2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S12sh5"
  }
 },
 "S12sh20::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:21:21.538187+00:00",
  "fingerprint": "935725ecd6de1b63b3e2d049717f10e65cb944744ac495b29a2c43709d164973",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S12sh20_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S12sh20_sel.png",
  "source_sha256": "77d82efd01947e8af728318164cf446139217373d8ca9ceedecc7af237b42a1b",
  "file": "S12sh20_cine.png",
  "staged_sha256": "a0fcf6847adb82e7a49096aca6d69890dca6e8161148faeb2c5ad87061306dee",
  "latency_ms": 11856
 },
 "S12sh23::signage": {
  "fp": "24f2eb241a675ba1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S12sh23::bgfirst_bg": {
  "input_fingerprint": "a078f12420d543eb",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 방 안쪽에서 헝겊 커튼 틈새로 빼꼼히 내민 앰버의 눈동자.\n\nLOCATION (lock): Just inside the girl's sleeping room in the family container home, at the curtained doorway adjoining the candlelit common area.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cloth curtain at 앰버's room opening (Parted enough for 앰버 to peek through) — The common-room-facing folds flank an actual opening through which her face is seen, rather than viewed through the cloth; used as Frames and partially conceals her face while preserving the direction of her outward attention; 앰버's room opening (Separating her concealed position from the common room) — Viewed obliquely from the common-room side; used as Establishes that the camera remains outside her room and discovers her listening position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the threshold dim with restrained facial contrast, preserving the established interior exposure without inventing a separate light inside 앰버's room.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 방 안쪽에서 헝겊 커튼 틈새로 빼꼼히 내민 앰버의 눈동자.\n\nLOCATION (lock): Just inside the girl's sleeping room in the family container home, at the curtained doorway adjoining the candlelit common area.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cloth curtain at 앰버's room opening (Parted enough for 앰버 to peek through) — The common-room-facing folds flank an actual opening through which her face is seen, rather than viewed through the cloth; used as Frames and partially conceals her face while preserving the direction of her outward attention; 앰버's room opening (Separating her concealed position from the common room) — Viewed obliquely from the common-room side; used as Establishes that the camera remains outside her room and discovers her listening position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the threshold dim with restrained facial contrast, preserving the established interior exposure without inventing a separate light inside 앰버's room.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S12sh23__bgfirst_bg.png",
  "asset_id": "0cedf53e-180f-4dfd-94ef-8318fced84db",
  "input_asset_ids": [
   "42ff5719-3ccf-48de-9c46-47ebee06bc65",
   "26032ccf-781f-486e-8ac7-e58a4659868d"
  ]
 },
 "S12sh23": {
  "input_fingerprint": "c7dd83b6ede25db9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 방 안쪽에서 헝겊 커튼 틈새로 빼꼼히 내민 앰버의 눈동자.\n\nLOCATION (lock): Just inside the girl's sleeping room in the family container home, at the curtained doorway adjoining the candlelit common area. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cloth curtain at 앰버's room opening (Parted enough for 앰버 to peek through) — The common-room-facing folds flank an actual opening through which her face is seen, rather than viewed through the cloth; used as Frames and partially conceals her face while preserving the direction of her outward attention; 앰버's room opening (Separating her concealed position from the common room) — Viewed obliquely from the common-room side; used as Establishes that the camera remains outside her room and discovers her listening position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the threshold dim with restrained facial contrast, preserving the established interior exposure without inventing a separate light inside 앰버's room.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The container's entrance door is now slammed shut, while the dining-table candle remains lit beside the meal, mending materials and opened medicine container. Amber's listening position is behind a room door; no fabric curtain is established. 앰버: Amber remains behind her room door, awake and listening.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 방 안쪽에서 헝겊 커튼 틈새로 빼꼼히 내민 앰버의 눈동자.\n\nLOCATION (lock): Just inside the girl's sleeping room in the family container home, at the curtained doorway adjoining the candlelit common area. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cloth curtain at 앰버's room opening (Parted enough for 앰버 to peek through) — The common-room-facing folds flank an actual opening through which her face is seen, rather than viewed through the cloth; used as Frames and partially conceals her face while preserving the direction of her outward attention; 앰버's room opening (Separating her concealed position from the common room) — Viewed obliquely from the common-room side; used as Establishes that the camera remains outside her room and discovers her listening position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the threshold dim with restrained facial contrast, preserving the established interior exposure without inventing a separate light inside 앰버's room.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The container's entrance door is now slammed shut, while the dining-table candle remains lit beside the meal, mending materials and opened medicine container. Amber's listening position is behind a room door; no fabric curtain is established. 앰버: Amber remains behind her room door, awake and listening.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 방 안쪽에서 헝겊 커튼 틈새로 빼꼼히 내민 앰버의 눈동자.\n\nLOCATION (lock): Just inside the girl's sleeping room in the family container home, at the curtained doorway adjoining the candlelit common area. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cloth curtain at 앰버's room opening (Parted enough for 앰버 to peek through) — The common-room-facing folds flank an actual opening through which her face is seen, rather than viewed through the cloth; used as Frames and partially conceals her face while preserving the direction of her outward attention; 앰버's room opening (Separating her concealed position from the common room) — Viewed obliquely from the common-room side; used as Establishes that the camera remains outside her room and discovers her listening position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the threshold dim with restrained facial contrast, preserving the established interior exposure without inventing a separate light inside 앰버's room.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The container's entrance door is now slammed shut, while the dining-table candle remains lit beside the meal, mending materials and opened medicine container. Amber's listening position is behind a room door; no fabric curtain is established. 앰버: Amber remains behind her room door, awake and listening.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S12sh23__bgfirst_bg.png",
     "asset_id": "0cedf53e-180f-4dfd-94ef-8318fced84db",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S12sh23.png",
     "asset_id": "42ff5719-3ccf-48de-9c46-47ebee06bc65",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L22B02.png",
     "asset_id": "26032ccf-781f-486e-8ac7-e58a4659868d",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "B",
    "direction": "보이는 눈은 렌즈 정면보다 화면 오른쪽의 프레임 밖을 향한다. 커튼 뒤에서 공동 공간 쪽을 살피는 시선으로 읽히며, 구체적인 상대는 보이지 않는다.",
    "built_space": "화면 양끝에 목재 문틀이 있고, 하나의 출입구를 덮는 천의 양쪽 주름 사이에 좁은 실제 틈이 있다. 앰버는 그 틈 뒤의 어두운 방에 있으며 얼굴 일부가 천에 가려진다. 천의 직조와 가로 줄무늬, 목재 재질은 장소 참조와 잘 맞는다. 공동 공간 쪽에서 보는 구성으로 읽히지만 사선 각도는 약하다. 식탁과 현관은 이 클로즈업 밖이므로 확인할 수 없다.",
    "entities": "금발 계열의 긴 머리와 큰 정상적인 눈을 지닌 여자아이 한 명만 보이며, 외관상 약 10세라는 설정에 부합한다. 보이는 얼굴 부분은 참조와 대체로 유사하지만 혼혈 정체성이나 정확한 동일 인물 여부를 외모만으로 확정할 수는 없다. 어두운 옷 일부만 보여 재킷과 원피스의 일치는 판단하기 어렵고, 참조의 흰 마스크는 없다. 읽을 수 있는 글자는 없다.",
    "hard_violations": [],
    "physics": "커튼은 화면 위쪽으로 이어져 매달린 채 자연스럽게 늘어진다. 틈 가장자리에는 손끝 일부가 보이며 천을 살짝 벌린 동작으로 읽힌다. 얼굴과 머리카락은 화면 아래로 이어지는 몸에 자연스럽게 연결된다. 발은 프레임 밖이고, 공중에 뜬 몸이나 지지 없는 물체는 보이지 않는다."
   },
   {
    "label": "A",
    "direction": "앰버의 눈은 화면 왼쪽의 카메라 가까운 방향을 향한다. 식탁 자체를 응시하기보다는 문틈 밖 가까운 곳을 살피는 모습이며, 시선의 구체적인 상대는 화면에 없다.",
    "built_space": "오른쪽 전경에 커튼 틈과 얼굴이 있고 왼쪽에는 식탁 하나, 켜진 촛불 하나, 의자 세 개, 책장 하나, 현관 쪽 개구부 하나가 보인다. 뒤쪽 액자와 오른쪽 보조 탁자 일부도 참조를 따른다. 그러나 공동 공간 대부분을 얼굴 옆으로 깊게 펼쳐 보여, 참조의 방 문턱을 공동 공간 쪽에서 비스듬히 관찰하는 관계가 명료하지 않다. 얼굴은 크게 잡혔지만 전체 화면은 공간 소개에 상당 부분 할애된다.",
    "entities": "금발 머리와 큰 눈을 가진 여자아이 한 명으로 연령과 성별 표현은 설정에 부합한다. 참조 얼굴과 어느 정도 유사하지만 정확한 정체성은 확정하기 어렵다. 마스크는 없으며 의상은 거의 가려져 비교할 수 없다. 배경에는 식사 그릇, 주전자, 천과 작은 약통 형태가 보이지만 약통이 열렸는지는 식별되지 않는다. 읽을 수 있는 글자는 없다.",
    "hard_violations": [],
    "physics": "손가락이 커튼 가장자리를 실제로 감싸 쥐고 있어 천을 벌리는 동작의 지지가 명확하다. 손과 얼굴의 연결 및 자세는 자연스럽다. 커튼은 위에서 늘어지고 식기와 촛불은 식탁 위에 놓여 있으며 가구는 바닥에 지지된다. 몸의 하부는 프레임 밖이고, 떠 있는 인체나 물체는 없다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "커튼의 실제 틈으로 눈과 반쪽 얼굴만 내민 클로즈업이 핵심 순간에 충실하다. 다만 요구된 사선보다 정면에 가깝고 참조의 마스크는 없다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "커튼을 잡고 엿보는 행동은 맞지만 공동 공간 전경까지 펼쳐 보여 눈동자 중심의 클로즈업과 방 밖에서 발견하는 시점이 약해졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "보이는 눈은 렌즈 정면보다 화면 오른쪽의 프레임 밖을 향한다. 커튼 뒤에서 공동 공간 쪽을 살피는 시선으로 읽히며, 구체적인 상대는 보이지 않는다.",
        "built_space": "화면 양끝에 목재 문틀이 있고, 하나의 출입구를 덮는 천의 양쪽 주름 사이에 좁은 실제 틈이 있다. 앰버는 그 틈 뒤의 어두운 방에 있으며 얼굴 일부가 천에 가려진다. 천의 직조와 가로 줄무늬, 목재 재질은 장소 참조와 잘 맞는다. 공동 공간 쪽에서 보는 구성으로 읽히지만 사선 각도는 약하다. 식탁과 현관은 이 클로즈업 밖이므로 확인할 수 없다.",
        "entities": "금발 계열의 긴 머리와 큰 정상적인 눈을 지닌 여자아이 한 명만 보이며, 외관상 약 10세라는 설정에 부합한다. 보이는 얼굴 부분은 참조와 대체로 유사하지만 혼혈 정체성이나 정확한 동일 인물 여부를 외모만으로 확정할 수는 없다. 어두운 옷 일부만 보여 재킷과 원피스의 일치는 판단하기 어렵고, 참조의 흰 마스크는 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "커튼은 화면 위쪽으로 이어져 매달린 채 자연스럽게 늘어진다. 틈 가장자리에는 손끝 일부가 보이며 천을 살짝 벌린 동작으로 읽힌다. 얼굴과 머리카락은 화면 아래로 이어지는 몸에 자연스럽게 연결된다. 발은 프레임 밖이고, 공중에 뜬 몸이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "앰버의 눈은 화면 왼쪽의 카메라 가까운 방향을 향한다. 식탁 자체를 응시하기보다는 문틈 밖 가까운 곳을 살피는 모습이며, 시선의 구체적인 상대는 화면에 없다.",
        "built_space": "오른쪽 전경에 커튼 틈과 얼굴이 있고 왼쪽에는 식탁 하나, 켜진 촛불 하나, 의자 세 개, 책장 하나, 현관 쪽 개구부 하나가 보인다. 뒤쪽 액자와 오른쪽 보조 탁자 일부도 참조를 따른다. 그러나 공동 공간 대부분을 얼굴 옆으로 깊게 펼쳐 보여, 참조의 방 문턱을 공동 공간 쪽에서 비스듬히 관찰하는 관계가 명료하지 않다. 얼굴은 크게 잡혔지만 전체 화면은 공간 소개에 상당 부분 할애된다.",
        "entities": "금발 머리와 큰 눈을 가진 여자아이 한 명으로 연령과 성별 표현은 설정에 부합한다. 참조 얼굴과 어느 정도 유사하지만 정확한 정체성은 확정하기 어렵다. 마스크는 없으며 의상은 거의 가려져 비교할 수 없다. 배경에는 식사 그릇, 주전자, 천과 작은 약통 형태가 보이지만 약통이 열렸는지는 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "손가락이 커튼 가장자리를 실제로 감싸 쥐고 있어 천을 벌리는 동작의 지지가 명확하다. 손과 얼굴의 연결 및 자세는 자연스럽다. 커튼은 위에서 늘어지고 식기와 촛불은 식탁 위에 놓여 있으며 가구는 바닥에 지지된다. 몸의 하부는 프레임 밖이고, 떠 있는 인체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "커튼의 실제 틈으로 눈과 반쪽 얼굴만 내민 클로즈업이 핵심 순간에 충실하다. 다만 요구된 사선보다 정면에 가깝고 참조의 마스크는 없다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "커튼을 잡고 엿보는 행동은 맞지만 공동 공간 전경까지 펼쳐 보여 눈동자 중심의 클로즈업과 방 밖에서 발견하는 시점이 약해졌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "보이는 눈은 렌즈 정면보다 화면 오른쪽의 프레임 밖을 향한다. 커튼 뒤에서 공동 공간 쪽을 살피는 시선으로 읽히며, 구체적인 상대는 보이지 않는다.",
        "built_space": "화면 양끝에 목재 문틀이 있고, 하나의 출입구를 덮는 천의 양쪽 주름 사이에 좁은 실제 틈이 있다. 앰버는 그 틈 뒤의 어두운 방에 있으며 얼굴 일부가 천에 가려진다. 천의 직조와 가로 줄무늬, 목재 재질은 장소 참조와 잘 맞는다. 공동 공간 쪽에서 보는 구성으로 읽히지만 사선 각도는 약하다. 식탁과 현관은 이 클로즈업 밖이므로 확인할 수 없다.",
        "entities": "금발 계열의 긴 머리와 큰 정상적인 눈을 지닌 여자아이 한 명만 보이며, 외관상 약 10세라는 설정에 부합한다. 보이는 얼굴 부분은 참조와 대체로 유사하지만 혼혈 정체성이나 정확한 동일 인물 여부를 외모만으로 확정할 수는 없다. 어두운 옷 일부만 보여 재킷과 원피스의 일치는 판단하기 어렵고, 참조의 흰 마스크는 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "커튼은 화면 위쪽으로 이어져 매달린 채 자연스럽게 늘어진다. 틈 가장자리에는 손끝 일부가 보이며 천을 살짝 벌린 동작으로 읽힌다. 얼굴과 머리카락은 화면 아래로 이어지는 몸에 자연스럽게 연결된다. 발은 프레임 밖이고, 공중에 뜬 몸이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "앰버의 눈은 화면 왼쪽의 카메라 가까운 방향을 향한다. 식탁 자체를 응시하기보다는 문틈 밖 가까운 곳을 살피는 모습이며, 시선의 구체적인 상대는 화면에 없다.",
        "built_space": "오른쪽 전경에 커튼 틈과 얼굴이 있고 왼쪽에는 식탁 하나, 켜진 촛불 하나, 의자 세 개, 책장 하나, 현관 쪽 개구부 하나가 보인다. 뒤쪽 액자와 오른쪽 보조 탁자 일부도 참조를 따른다. 그러나 공동 공간 대부분을 얼굴 옆으로 깊게 펼쳐 보여, 참조의 방 문턱을 공동 공간 쪽에서 비스듬히 관찰하는 관계가 명료하지 않다. 얼굴은 크게 잡혔지만 전체 화면은 공간 소개에 상당 부분 할애된다.",
        "entities": "금발 머리와 큰 눈을 가진 여자아이 한 명으로 연령과 성별 표현은 설정에 부합한다. 참조 얼굴과 어느 정도 유사하지만 정확한 정체성은 확정하기 어렵다. 마스크는 없으며 의상은 거의 가려져 비교할 수 없다. 배경에는 식사 그릇, 주전자, 천과 작은 약통 형태가 보이지만 약통이 열렸는지는 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "손가락이 커튼 가장자리를 실제로 감싸 쥐고 있어 천을 벌리는 동작의 지지가 명확하다. 손과 얼굴의 연결 및 자세는 자연스럽다. 커튼은 위에서 늘어지고 식기와 촛불은 식탁 위에 놓여 있으며 가구는 바닥에 지지된다. 몸의 하부는 프레임 밖이고, 떠 있는 인체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "totals": {
   "B": 8,
   "A": 5
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "커튼의 실제 틈으로 눈과 반쪽 얼굴만 내민 클로즈업이 핵심 순간에 충실하다. 다만 요구된 사선보다 정면에 가깝고 참조의 마스크는 없다."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "커튼을 잡고 엿보는 행동은 맞지만 공동 공간 전경까지 펼쳐 보여 눈동자 중심의 클로즈업과 방 밖에서 발견하는 시점이 약해졌다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L22B02.png",
    "asset_id": "26032ccf-781f-486e-8ac7-e58a4659868d",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd86-db84-7c59-8b52-bc073d933f27",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S12sh23__bgfirst_bg.png",
   "bg_asset_id": "0cedf53e-180f-4dfd-94ef-8318fced84db",
   "bg_record_key": "S12sh23::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S12sh23::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:21:33.804914+00:00",
  "fingerprint": "01ef251e580322712cc1e68a374fb495ac0f84fe6380b7960060f7600be55eed",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S12sh23_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S12sh23_sel.png",
  "source_sha256": "796488300875fa4f46c8c8f7b24b2f988eded804598714a036b31579fa25ecb0",
  "file": "S12sh23_cine.png",
  "staged_sha256": "4785ca8d7c1d59b67aa5c1d7baeb4b694be87db23ec3533c58dddd08d7049d06",
  "latency_ms": 11080
 },
 "S13sh10::signage": {
  "fp": "10df2bfddee9b430",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::3438213eaad04f9a": {
  "subjects": [],
  "subject_text": "쿠마 일행이 모인 난민촌 노상 모닥불 주변\n인공제방 바로 옆의 노상 공터. 양철통 안의 불이 주변을 주황빛으로 비추며 바비큐 그릴과 맥주 용기가 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L24",
  "scope_role": "location_exterior",
  "scope_sha": "2540796420c682dc"
 },
 "groupbg::roadside_hideout": {
  "input_fingerprint": "0bba906847684caf",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "roadside_hideout",
    "tags": [
     "S13sh10"
    ]
   },
   "context_sig": "a7cad84018a0950a"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At an outdoor gang gathering beside the refugee settlement's seawall, near burning metal barrels and a barbecue grill.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n쿠마 일행이 모인 난민촌 노상 모닥불 주변: 인공제방 바로 옆 어두운 도로가에 양철통을 두고 피운 불 주위로 사람들이 모여 있는 노상 아지트. (특징: 타오르는 드럼통 모닥불; 바비큐 그릴과 닭꼬치; 조잡하게 개조된 몽둥이와 총기 무장)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 인공제방 바로 옆 노상에 쿠마의 아지트가 있다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At an outdoor gang gathering beside the refugee settlement's seawall, near burning metal barrels and a barbecue grill.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n쿠마 일행이 모인 난민촌 노상 모닥불 주변: 인공제방 바로 옆 어두운 도로가에 양철통을 두고 피운 불 주위로 사람들이 모여 있는 노상 아지트. (특징: 타오르는 드럼통 모닥불; 바비큐 그릴과 닭꼬치; 조잡하게 개조된 몽둥이와 총기 무장)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 인공제방 바로 옆 노상에 쿠마의 아지트가 있다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_roadside_hideout_f5126d.png",
  "asset_id": "7257deaa-5076-4cec-b938-3f5afb7c84e0",
  "input_asset_ids": [
   "4ed837b5-18ce-4b3f-a221-547b32869e91"
  ],
  "origin_tag": "S13sh10",
  "place_text": "At an outdoor gang gathering beside the refugee settlement's seawall, near burning metal barrels and a barbecue grill.",
  "origin_inputs": {
   "place_text": "At an outdoor gang gathering beside the refugee settlement's seawall, near burning metal barrels and a barbecue grill.",
   "time_of_day_en": "night",
   "conti_asset_id": "4ed837b5-18ce-4b3f-a221-547b32869e91"
  }
 },
 "S13sh10::bgfirst_bg": {
  "input_fingerprint": "335c030b1d6e3b56",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 빼앗은 소총의 총구를 쿠마의 얼굴을 향해 정확히 겨눈 매서운 구도.\n\nLOCATION (lock): At an outdoor gang gathering beside the refugee settlement's seawall, near burning metal barrels and a barbecue grill.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 behind the rifle in the middle-left of the frame, foreground, looks toward 쿠마 beyond the muzzle; 쿠마 beyond the muzzle in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Seized rifle (Held steadily with its muzzle aimed at 쿠마's face) — Seen obliquely from behind its holder, with the barrel extending toward the right midground; used as Connects the foreground hands to the threatened face without exaggerated muzzle size; Barbecue grill (Present where 쿠마 has been eating chicken skewers) — A partial side and upper edge remain below 쿠마; used as Provides a restrained contextual trace of the interrupted meal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established barrel fire provides uneven illumination against the nighttime darkness, keeping both faces and the line of aim legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 빼앗은 소총의 총구를 쿠마의 얼굴을 향해 정확히 겨눈 매서운 구도.\n\nLOCATION (lock): At an outdoor gang gathering beside the refugee settlement's seawall, near burning metal barrels and a barbecue grill.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 behind the rifle in the middle-left of the frame, foreground, looks toward 쿠마 beyond the muzzle; 쿠마 beyond the muzzle in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Seized rifle (Held steadily with its muzzle aimed at 쿠마's face) — Seen obliquely from behind its holder, with the barrel extending toward the right midground; used as Connects the foreground hands to the threatened face without exaggerated muzzle size; Barbecue grill (Present where 쿠마 has been eating chicken skewers) — A partial side and upper edge remain below 쿠마; used as Provides a restrained contextual trace of the interrupted meal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established barrel fire provides uneven illumination against the nighttime darkness, keeping both faces and the line of aim legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S13sh10__bgfirst_bg.png",
  "asset_id": "4d3cf60a-94be-467f-a657-7dae51d668e6",
  "input_asset_ids": [
   "4ed837b5-18ce-4b3f-a221-547b32869e91",
   "7257deaa-5076-4cec-b938-3f5afb7c84e0"
  ]
 },
 "S13sh10": {
  "input_fingerprint": "5557aec864bfd3c7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 빼앗은 소총의 총구를 쿠마의 얼굴을 향해 정확히 겨눈 매서운 구도.\n\nLOCATION (lock): At an outdoor gang gathering beside the refugee settlement's seawall, near burning metal barrels and a barbecue grill. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 behind the rifle in the middle-left of the frame, foreground, looks toward 쿠마 beyond the muzzle; 쿠마 beyond the muzzle in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Seized rifle (Held steadily with its muzzle aimed at 쿠마's face) — Seen obliquely from behind its holder, with the barrel extending toward the right midground; used as Connects the foreground hands to the threatened face without exaggerated muzzle size; Barbecue grill (Present where 쿠마 has been eating chicken skewers) — A partial side and upper edge remain below 쿠마; used as Provides a restrained contextual trace of the interrupted meal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established barrel fire provides uneven illumination against the nighttime darkness, keeping both faces and the line of aim legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hideout stands beside the cracked, leaking seawall at night; fires burn in metal drums, and chicken skewers remain on the barbecue grill. 이현우: He has put his outer top back on, and his facial injuries and untreated dog-bitten leg remain visible where exposed. He holds the seized gun raised and aimed. 쿠마: He is at the barbecue area, still holding a chicken skewer before discarding it and standing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 빼앗은 소총의 총구를 쿠마의 얼굴을 향해 정확히 겨눈 매서운 구도.\n\nLOCATION (lock): At an outdoor gang gathering beside the refugee settlement's seawall, near burning metal barrels and a barbecue grill. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 behind the rifle in the middle-left of the frame, foreground, looks toward 쿠마 beyond the muzzle; 쿠마 beyond the muzzle in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Seized rifle (Held steadily with its muzzle aimed at 쿠마's face) — Seen obliquely from behind its holder, with the barrel extending toward the right midground; used as Connects the foreground hands to the threatened face without exaggerated muzzle size; Barbecue grill (Present where 쿠마 has been eating chicken skewers) — A partial side and upper edge remain below 쿠마; used as Provides a restrained contextual trace of the interrupted meal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established barrel fire provides uneven illumination against the nighttime darkness, keeping both faces and the line of aim legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hideout stands beside the cracked, leaking seawall at night; fires burn in metal drums, and chicken skewers remain on the barbecue grill. 이현우: He has put his outer top back on, and his facial injuries and untreated dog-bitten leg remain visible where exposed. He holds the seized gun raised and aimed. 쿠마: He is at the barbecue area, still holding a chicken skewer before discarding it and standing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 빼앗은 소총의 총구를 쿠마의 얼굴을 향해 정확히 겨눈 매서운 구도.\n\nLOCATION (lock): At an outdoor gang gathering beside the refugee settlement's seawall, near burning metal barrels and a barbecue grill. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 behind the rifle in the middle-left of the frame, foreground, looks toward 쿠마 beyond the muzzle; 쿠마 beyond the muzzle in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Seized rifle (Held steadily with its muzzle aimed at 쿠마's face) — Seen obliquely from behind its holder, with the barrel extending toward the right midground; used as Connects the foreground hands to the threatened face without exaggerated muzzle size; Barbecue grill (Present where 쿠마 has been eating chicken skewers) — A partial side and upper edge remain below 쿠마; used as Provides a restrained contextual trace of the interrupted meal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established barrel fire provides uneven illumination against the nighttime darkness, keeping both faces and the line of aim legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hideout stands beside the cracked, leaking seawall at night; fires burn in metal drums, and chicken skewers remain on the barbecue grill. 이현우: He has put his outer top back on, and his facial injuries and untreated dog-bitten leg remain visible where exposed. He holds the seized gun raised and aimed. 쿠마: He is at the barbecue area, still holding a chicken skewer before discarding it and standing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S13sh10__bgfirst_bg.png",
     "asset_id": "4d3cf60a-94be-467f-a657-7dae51d668e6",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S13sh10.png",
     "asset_id": "4ed837b5-18ce-4b3f-a221-547b32869e91",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904970>",
     "asset_id": "afafdeae-9f05-477d-8a3f-6cedcc1096d3",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:667686>",
     "asset_id": "fcb46f42-5d47-4486-85bb-3542f9f839f4",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_roadside_hideout_f5126d.png",
     "asset_id": "7257deaa-5076-4cec-b938-3f5afb7c84e0",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904970>",
     "asset_id": "afafdeae-9f05-477d-8a3f-6cedcc1096d3",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:667686>",
     "asset_id": "fcb46f42-5d47-4486-85bb-3542f9f839f4",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 총구가 쿠마의 얼굴을 향함.",
    "built_space": "방파제, 불타는 드럼통, 바비큐 그릴이 참조 사진과 일치하게 배치됨.",
    "entities": "이현우와 쿠마가 인식되나 쿠마의 모자가 없음. 배경에 3명의 미지정 인물이 추가됨.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 없는 추가 인물 등장",
     "[gpt-high] 지정된 두 인물 외에 배경 인물 네 명이 추가되었다.",
     "[gpt-high] 참조 장소의 단일 바비큐 그릴 외에 두 번째 독립된 그릴이 오른쪽 아래에 추가되었다."
    ],
    "physics": "이현우가 양손으로 총을 파지하고 있으며, 쿠마의 오른손이 꼬치를 들고 있음."
   },
   {
    "label": "B",
    "direction": "소총의 총구가 쿠마의 얼굴을 정확히 겨누고 있음.",
    "built_space": "방파제, 드럼통, 바비큐 그릴의 위치와 환경이 참조와 일치함.",
    "entities": "두 인물의 외형 및 의상(모자 포함)이 참조와 일치함. 배경에 2명의 미지정 인물이 추가됨.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 없는 추가 인물 등장",
     "[gpt-high] 이현우와 쿠마 외에 최소 세 명의 배경 인물이 추가되어 두 사람만 허용한 조건을 위반한다.",
     "[gpt-high] 쿠마의 모자에 식별 가능한 숫자 ‘77’이 보여 읽을 수 있는 표기 금지 조건을 위반한다."
    ],
    "physics": "이현우의 양손이 총을 안정적으로 지지하고, 쿠마가 오른손으로 꼬치를 쥠."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정되지 않은 배경 인물이 추가된 치명적 위반이 있으며, 쿠마의 모자가 누락됨."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "의상 및 구도 구현은 더 우수하나, 텍스트에 없는 배경 인물이 추가되어 치명적 위반임."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 총구가 쿠마의 얼굴을 향함.",
        "built_space": "방파제, 불타는 드럼통, 바비큐 그릴이 참조 사진과 일치하게 배치됨.",
        "entities": "이현우와 쿠마가 인식되나 쿠마의 모자가 없음. 배경에 3명의 미지정 인물이 추가됨.",
        "hard_violations": [
         "샷 텍스트에 없는 추가 인물 등장"
        ],
        "physics": "이현우가 양손으로 총을 파지하고 있으며, 쿠마의 오른손이 꼬치를 들고 있음."
       },
       {
        "label": "B",
        "direction": "소총의 총구가 쿠마의 얼굴을 정확히 겨누고 있음.",
        "built_space": "방파제, 드럼통, 바비큐 그릴의 위치와 환경이 참조와 일치함.",
        "entities": "두 인물의 외형 및 의상(모자 포함)이 참조와 일치함. 배경에 2명의 미지정 인물이 추가됨.",
        "hard_violations": [
         "샷 텍스트에 없는 추가 인물 등장"
        ],
        "physics": "이현우의 양손이 총을 안정적으로 지지하고, 쿠마가 오른손으로 꼬치를 쥠."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정되지 않은 배경 인물이 추가된 치명적 위반이 있으며, 쿠마의 모자가 누락됨."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "의상 및 구도 구현은 더 우수하나, 텍스트에 없는 배경 인물이 추가되어 치명적 위반임."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 총구가 쿠마의 얼굴을 향함.",
        "built_space": "방파제, 불타는 드럼통, 바비큐 그릴이 참조 사진과 일치하게 배치됨.",
        "entities": "이현우와 쿠마가 인식되나 쿠마의 모자가 없음. 배경에 3명의 미지정 인물이 추가됨.",
        "hard_violations": [
         "샷 텍스트에 없는 추가 인물 등장"
        ],
        "physics": "이현우가 양손으로 총을 파지하고 있으며, 쿠마의 오른손이 꼬치를 들고 있음."
       },
       {
        "label": "B",
        "direction": "소총의 총구가 쿠마의 얼굴을 정확히 겨누고 있음.",
        "built_space": "방파제, 드럼통, 바비큐 그릴의 위치와 환경이 참조와 일치함.",
        "entities": "두 인물의 외형 및 의상(모자 포함)이 참조와 일치함. 배경에 2명의 미지정 인물이 추가됨.",
        "hard_violations": [
         "샷 텍스트에 없는 추가 인물 등장"
        ],
        "physics": "이현우의 양손이 총을 안정적으로 지지하고, 쿠마가 오른손으로 꼬치를 쥠."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴을 향한 조준선과 전경·중경 배치, 쿠마의 모자 및 꼬치 상태는 더 충실하지만, 금지된 배경 인물들과 읽히는 모자 숫자 때문에 실사용은 어렵습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "총구는 쿠마의 얼굴을 향하지만, 미디엄 숏보다 넓어진 구도에 추가 인물 네 명과 두 번째 그릴이 들어가고 쿠마도 이미 서 있어 지정 순간에서 더 멀어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 왼쪽 전경에서 오른쪽의 쿠마를 바라보며 견착하고 있다. 총열의 연장선은 쿠마의 입과 볼 부근에 닿아 얼굴 조준을 충족한다. 쿠마의 눈도 왼쪽의 이현우와 총 쪽을 향한다. 쿠마가 쥔 닭꼬치는 가슴 앞에서 위로 향한다.",
        "built_space": "왼쪽에는 기둥과 천막을 갖춘 은신처, 뒤에는 난간과 소파블록이 놓인 균열진 방조제, 오른쪽에는 바다가 있다. 불이 보이는 드럼통은 왼쪽 인물 뒤, 중앙 뒤, 쿠마 오른쪽 뒤에 하나씩 있다. 그릴은 오른쪽에 하나이며 쿠마 바로 앞에 놓였다. 장소의 주요 재료와 배치는 참조와 유사하지만, 그릴은 상단과 옆면의 절제된 흔적을 넘어 다리까지 크게 드러난다. 이현우의 뒤쪽 측면에서 보는 전경·중경 관계와 상반신 중심 구도는 B보다 요청에 가깝다.",
        "entities": "이현우는 헝클어진 검은 머리, 젊은 동아시아계 남성 얼굴, 낡은 어두운 후드 상의와 얼굴 상처를 갖췄다. 쿠마는 참조와 유사한 젊은 성인 남성 얼굴에 장식 모자, 스터드 조끼, 무늬 셔츠를 착용하고 닭꼬치를 들고 있다. 외모만으로 구체적 혈통을 확정할 수는 없다. 소총은 스코프, 긴 총열, 볼트와 골격형 개머리판을 갖춰 참조의 종류와 형태에 가깝지만 표면이 더 갈색으로 보인다. 그릴 위에도 꼬치가 남아 있다. 배경에는 지정되지 않은 사람이 최소 세 명 보이며 모자 앞의 숫자 ‘77’도 식별된다. 다리 상처는 프레임에서 확인할 수 없다.",
        "hard_violations": [
         "이현우와 쿠마 외에 최소 세 명의 배경 인물이 추가되어 두 사람만 허용한 조건을 위반한다.",
         "쿠마의 모자에 식별 가능한 숫자 ‘77’이 보여 읽을 수 있는 표기 금지 조건을 위반한다."
        ],
        "physics": "소총은 이현우의 방아쇠 쪽 손과 총열 아래를 받친 손, 어깨에 댄 개머리판으로 지지된다. 손과 소매도 이현우의 신체 및 의상에 연결된다. 쿠마의 꼬치는 손에 잡혀 있고 조리 중인 꼬치들은 석쇠에 놓여 있다. 그릴은 금속 다리로 바닥을 딛는다. 쿠마는 하체를 굽힌 낮은 자세로 보이지만 좌석 접촉은 가려져 있어 앉은 상태인지는 확정하기 어렵다. 두 주인공의 발은 잘려 있으며, 이것만으로 공중에 떠 있다고 판단할 근거는 없다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 쿠마를 바라보며 소총을 견착한다. 거의 수평인 총열의 연장선은 쿠마의 코와 눈 아래 부근을 향하므로 얼굴 조준 자체는 맞는다. 쿠마는 왼쪽의 이현우를 응시하며 닭꼬치를 오른손에 쥐고 아래로 늘어뜨린다.",
        "built_space": "천막 은신처, 균열진 방조제, 난간과 소파블록, 바다가 참조와 유사하게 배치되어 있다. 불타는 드럼통은 이현우 뒤, 중앙, 쿠마 오른쪽 뒤에 세 개 보인다. 오른쪽에는 큰 그릴 외에 별도의 낮은 그릴이 하나 더 있다. 이현우는 허벅지까지, 쿠마는 무릎 아래까지 보여 요청한 미디엄 숏보다 넓으며 바닥과 배경 인물의 비중도 크다. 쿠마는 그릴 왼쪽에 이미 똑바로 서 있다.",
        "entities": "이현우는 검은 헝클어진 머리와 낡은 후드 상의, 뚜렷한 얼굴 상처를 갖춘 젊은 동아시아계 남성으로 보인다. 쿠마는 무늬 셔츠와 스터드 조끼를 입었지만 참조의 모자가 없고 얼굴도 참조보다 성숙하고 각지게 보인다. 구체적 혼혈 정체성은 외모만으로 확인할 수 없다. 소총은 참조와 유사한 스코프 장착 볼트액션형이다. 쿠마가 든 닭꼬치와 그릴 위 꼬치는 존재한다. 왼쪽 서 있는 사람, 중앙에 앉은 사람, 그 뒤 서 있는 사람, 오른쪽 사람까지 추가 인물이 네 명 있다.",
        "hard_violations": [
         "지정된 두 인물 외에 배경 인물 네 명이 추가되었다.",
         "참조 장소의 단일 바비큐 그릴 외에 두 번째 독립된 그릴이 오른쪽 아래에 추가되었다."
        ],
        "physics": "소총은 양손과 어깨로 안정적으로 지지되며 닭꼬치는 쿠마의 손에 잡혀 있다. 배경의 앉은 사람은 상자형 받침 위에 엉덩이를 두고 발을 바닥에 댄다. 그릴과 드럼통도 바닥에 놓여 있어 지지 없는 물체는 보이지 않는다. 쿠마의 서 있는 자세 자체는 물리적으로 가능하지만, 꼬치를 버리고 일어서기 전이라는 요청보다 행동이 진행된 상태다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "얼굴을 향한 조준선과 전경·중경 배치, 쿠마의 모자 및 꼬치 상태는 더 충실하지만, 금지된 배경 인물들과 읽히는 모자 숫자 때문에 실사용은 어렵습니다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "총구는 쿠마의 얼굴을 향하지만, 미디엄 숏보다 넓어진 구도에 추가 인물 네 명과 두 번째 그릴이 들어가고 쿠마도 이미 서 있어 지정 순간에서 더 멀어집니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 왼쪽 전경에서 오른쪽의 쿠마를 바라보며 견착하고 있다. 총열의 연장선은 쿠마의 입과 볼 부근에 닿아 얼굴 조준을 충족한다. 쿠마의 눈도 왼쪽의 이현우와 총 쪽을 향한다. 쿠마가 쥔 닭꼬치는 가슴 앞에서 위로 향한다.",
        "built_space": "왼쪽에는 기둥과 천막을 갖춘 은신처, 뒤에는 난간과 소파블록이 놓인 균열진 방조제, 오른쪽에는 바다가 있다. 불이 보이는 드럼통은 왼쪽 인물 뒤, 중앙 뒤, 쿠마 오른쪽 뒤에 하나씩 있다. 그릴은 오른쪽에 하나이며 쿠마 바로 앞에 놓였다. 장소의 주요 재료와 배치는 참조와 유사하지만, 그릴은 상단과 옆면의 절제된 흔적을 넘어 다리까지 크게 드러난다. 이현우의 뒤쪽 측면에서 보는 전경·중경 관계와 상반신 중심 구도는 B보다 요청에 가깝다.",
        "entities": "이현우는 헝클어진 검은 머리, 젊은 동아시아계 남성 얼굴, 낡은 어두운 후드 상의와 얼굴 상처를 갖췄다. 쿠마는 참조와 유사한 젊은 성인 남성 얼굴에 장식 모자, 스터드 조끼, 무늬 셔츠를 착용하고 닭꼬치를 들고 있다. 외모만으로 구체적 혈통을 확정할 수는 없다. 소총은 스코프, 긴 총열, 볼트와 골격형 개머리판을 갖춰 참조의 종류와 형태에 가깝지만 표면이 더 갈색으로 보인다. 그릴 위에도 꼬치가 남아 있다. 배경에는 지정되지 않은 사람이 최소 세 명 보이며 모자 앞의 숫자 ‘77’도 식별된다. 다리 상처는 프레임에서 확인할 수 없다.",
        "hard_violations": [
         "이현우와 쿠마 외에 최소 세 명의 배경 인물이 추가되어 두 사람만 허용한 조건을 위반한다.",
         "쿠마의 모자에 식별 가능한 숫자 ‘77’이 보여 읽을 수 있는 표기 금지 조건을 위반한다."
        ],
        "physics": "소총은 이현우의 방아쇠 쪽 손과 총열 아래를 받친 손, 어깨에 댄 개머리판으로 지지된다. 손과 소매도 이현우의 신체 및 의상에 연결된다. 쿠마의 꼬치는 손에 잡혀 있고 조리 중인 꼬치들은 석쇠에 놓여 있다. 그릴은 금속 다리로 바닥을 딛는다. 쿠마는 하체를 굽힌 낮은 자세로 보이지만 좌석 접촉은 가려져 있어 앉은 상태인지는 확정하기 어렵다. 두 주인공의 발은 잘려 있으며, 이것만으로 공중에 떠 있다고 판단할 근거는 없다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 쿠마를 바라보며 소총을 견착한다. 거의 수평인 총열의 연장선은 쿠마의 코와 눈 아래 부근을 향하므로 얼굴 조준 자체는 맞는다. 쿠마는 왼쪽의 이현우를 응시하며 닭꼬치를 오른손에 쥐고 아래로 늘어뜨린다.",
        "built_space": "천막 은신처, 균열진 방조제, 난간과 소파블록, 바다가 참조와 유사하게 배치되어 있다. 불타는 드럼통은 이현우 뒤, 중앙, 쿠마 오른쪽 뒤에 세 개 보인다. 오른쪽에는 큰 그릴 외에 별도의 낮은 그릴이 하나 더 있다. 이현우는 허벅지까지, 쿠마는 무릎 아래까지 보여 요청한 미디엄 숏보다 넓으며 바닥과 배경 인물의 비중도 크다. 쿠마는 그릴 왼쪽에 이미 똑바로 서 있다.",
        "entities": "이현우는 검은 헝클어진 머리와 낡은 후드 상의, 뚜렷한 얼굴 상처를 갖춘 젊은 동아시아계 남성으로 보인다. 쿠마는 무늬 셔츠와 스터드 조끼를 입었지만 참조의 모자가 없고 얼굴도 참조보다 성숙하고 각지게 보인다. 구체적 혼혈 정체성은 외모만으로 확인할 수 없다. 소총은 참조와 유사한 스코프 장착 볼트액션형이다. 쿠마가 든 닭꼬치와 그릴 위 꼬치는 존재한다. 왼쪽 서 있는 사람, 중앙에 앉은 사람, 그 뒤 서 있는 사람, 오른쪽 사람까지 추가 인물이 네 명 있다.",
        "hard_violations": [
         "지정된 두 인물 외에 배경 인물 네 명이 추가되었다.",
         "참조 장소의 단일 바비큐 그릴 외에 두 번째 독립된 그릴이 오른쪽 아래에 추가되었다."
        ],
        "physics": "소총은 양손과 어깨로 안정적으로 지지되며 닭꼬치는 쿠마의 손에 잡혀 있다. 배경의 앉은 사람은 상자형 받침 위에 엉덩이를 두고 발을 바닥에 댄다. 그릴과 드럼통도 바닥에 놓여 있어 지지 없는 물체는 보이지 않는다. 쿠마의 서 있는 자세 자체는 물리적으로 가능하지만, 꼬치를 버리고 일어서기 전이라는 요청보다 행동이 진행된 상태다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.083,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.833,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 샷 텍스트에 없는 추가 인물 등장",
     "[gpt-high] 지정된 두 인물 외에 배경 인물 네 명이 추가되었다.",
     "[gpt-high] 참조 장소의 단일 바비큐 그릴 외에 두 번째 독립된 그릴이 오른쪽 아래에 추가되었다."
    ],
    "B": [
     "[gemini-pro] 샷 텍스트에 없는 추가 인물 등장",
     "[gpt-high] 이현우와 쿠마 외에 최소 세 명의 배경 인물이 추가되어 두 사람만 허용한 조건을 위반한다.",
     "[gpt-high] 쿠마의 모자에 식별 가능한 숫자 ‘77’이 보여 읽을 수 있는 표기 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 833,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 833,
    "verdict_ko": "지정되지 않은 배경 인물이 추가된 치명적 위반이 있으며, 쿠마의 모자가 누락됨.  ★위반: [gemini-pro] 샷 텍스트에 없는 추가 인물 등장 / [gpt-high] 지정된 두 인물 외에 배경 인물 네 명이 추가되었다. / [gpt-high] 참조 장소의 단일 바비큐 그릴 외에 두 번째 독립된 그릴이 오른쪽 아래에 추가되었다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "의상 및 구도 구현은 더 우수하나, 텍스트에 없는 배경 인물이 추가되어 치명적 위반임.  ★위반: [gemini-pro] 샷 텍스트에 없는 추가 인물 등장 / [gpt-high] 이현우와 쿠마 외에 최소 세 명의 배경 인물이 추가되어 두 사람만 허용한 조건을 위반한다. / [gpt-high] 쿠마의 모자에 식별 가능한 숫자 ‘77’이 보여 읽을 수 있는 표기 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_roadside_hideout_f5126d.png",
    "asset_id": "7257deaa-5076-4cec-b938-3f5afb7c84e0",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:904970>",
    "asset_id": "afafdeae-9f05-477d-8a3f-6cedcc1096d3",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:667686>",
    "asset_id": "fcb46f42-5d47-4486-85bb-3542f9f839f4",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd87-9346-7950-be87-d57d11c7d711",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S13sh10__bgfirst_bg.png",
   "bg_asset_id": "4d3cf60a-94be-467f-a657-7dae51d668e6",
   "bg_record_key": "S13sh10::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "roadside_hideout",
   "groupbg_asset_id": "7257deaa-5076-4cec-b938-3f5afb7c84e0"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S13sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:21:45.364033+00:00",
  "fingerprint": "67ec574238584cce9b6de8a9940b531fe233943bb0d96d5de2bde65d0c22af6c",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S13sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S13sh10_sel.png",
  "source_sha256": "2a455b412e1983e567a8c4db2d70187abaf2556d31e8bf86610f0aed59c02fdd",
  "file": "S13sh10_cine.png",
  "staged_sha256": "ba27d985b85f7b9d15c8a7b56db813b2e7f2e54604d6804fdd38aeeb50df69d4",
  "latency_ms": 14997
 },
 "S13sh14::signage": {
  "fp": "6288331eb205a8ae",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S13sh14": {
  "input_fingerprint": "7ef839277385c5a3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어둠 속, 길게 금이 간 제방 벽 틈새와 지지대로 검은 바닷물이 줄줄 흘러내리는 표면 클로즈업.\n\nLOCATION (lock): Against the exterior face of the refugee settlement's seawall at night, where seawater leaks through cracks and runs down the supports. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cracked embankment wall section (Split by a long fissure through which seawater seeps) — The damaged face recedes obliquely away from the camera; used as The fissure establishes the left-side focal detail and leads toward the structural junction; Adjacent embankment support (Water runs down its surface) — Its side and connection to the wall are visible on the right; used as Provides a second structural plane that makes the leakage spatially intelligible; Seeping seawater (Flowing from the cracks and down the support); used as Supplies the only active movement within the close structural observation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the nighttime darkness with restrained contrast sufficient to distinguish the cracks and the nearly black seawater descending along the support.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall remains deeply cracked in the darkness, with seawater seeping through its fissures and running over its supports.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어둠 속, 길게 금이 간 제방 벽 틈새와 지지대로 검은 바닷물이 줄줄 흘러내리는 표면 클로즈업.\n\nLOCATION (lock): Against the exterior face of the refugee settlement's seawall at night, where seawater leaks through cracks and runs down the supports. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cracked embankment wall section (Split by a long fissure through which seawater seeps) — The damaged face recedes obliquely away from the camera; used as The fissure establishes the left-side focal detail and leads toward the structural junction; Adjacent embankment support (Water runs down its surface) — Its side and connection to the wall are visible on the right; used as Provides a second structural plane that makes the leakage spatially intelligible; Seeping seawater (Flowing from the cracks and down the support); used as Supplies the only active movement within the close structural observation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the nighttime darkness with restrained contrast sufficient to distinguish the cracks and the nearly black seawater descending along the support.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall remains deeply cracked in the darkness, with seawater seeping through its fissures and running over its supports.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어둠 속, 길게 금이 간 제방 벽 틈새와 지지대로 검은 바닷물이 줄줄 흘러내리는 표면 클로즈업.\n\nLOCATION (lock): Against the exterior face of the refugee settlement's seawall at night, where seawater leaks through cracks and runs down the supports. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cracked embankment wall section (Split by a long fissure through which seawater seeps) — The damaged face recedes obliquely away from the camera; used as The fissure establishes the left-side focal detail and leads toward the structural junction; Adjacent embankment support (Water runs down its surface) — Its side and connection to the wall are visible on the right; used as Provides a second structural plane that makes the leakage spatially intelligible; Seeping seawater (Flowing from the cracks and down the support); used as Supplies the only active movement within the close structural observation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the nighttime darkness with restrained contrast sufficient to distinguish the cracks and the nearly black seawater descending along the support.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall remains deeply cracked in the darkness, with seawater seeping through its fissures and running over its supports.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "물이 제방의 갈라진 틈과 우측 구조물의 표면을 따라 아래로 똑바로 흘러내립니다.",
    "built_space": "경사진 콘크리트 제방 벽이 카메라에서 멀어지도록 배치되어 있으며, 우측에는 벽과 연결된 수직 지지대 구조물이 있습니다.",
    "entities": "갈라진 제방 벽, 검은색에 가까운 바닷물, 우측 지지대가 프롬프트의 설명대로 묘사되었습니다.",
    "hard_violations": [
     "[gpt-high] 참조에서 보행로와 경사면이 이어지는 제방 상단을 연속된 높은 콘크리트 벽으로 바꾸어 고정 구조를 추가했습니다."
    ],
    "physics": "물이 표면을 타고 중력 방향으로 자연스럽게 흘러내리고 있어 물리적으로 타당합니다."
   },
   {
    "label": "B",
    "direction": "물이 틈새를 따라 흐르며, 우측 콘크리트 구조물에서는 허공으로 쏟아지듯 아래로 떨어집니다.",
    "built_space": "경사진 콘크리트 제방이 있으며, 상단 둔치 쪽에 레퍼런스 사진(탁 트인 포장도로)에 존재하지 않는 금속 가드레일 기둥이 세워져 있습니다.",
    "entities": "갈라진 벽과 바닷물은 존재하나, 프롬프트와 레퍼런스에 없는 가드레일이 포함되었습니다.",
    "hard_violations": [
     "[gemini-pro] invented objects (레퍼런스에 없는 상단 금속 가드레일 추가)",
     "[gpt-high] 고정 장소 참조에 없는 금속 방호난간과 이를 받치는 기둥들을 제방 위에 추가했습니다."
    ],
    "physics": "물의 낙하와 흐름을 지탱하는 표면 등 물리적 법칙에는 어긋남이 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 클로즈업 구도를 잘 살렸으며, 갈라진 틈과 우측 지지대를 표면을 따라 흐르는 검은 물의 묘사가 레퍼런스의 환경과 자연스럽게 부합합니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "장소의 영구적인 구조(LOCATION lock)를 유지해야 함에도 레퍼런스에 없는 금속 가드레일을 임의로 추가하여 프롬프트 지시를 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "물이 제방의 갈라진 틈과 우측 구조물의 표면을 따라 아래로 똑바로 흘러내립니다.",
        "built_space": "경사진 콘크리트 제방 벽이 카메라에서 멀어지도록 배치되어 있으며, 우측에는 벽과 연결된 수직 지지대 구조물이 있습니다.",
        "entities": "갈라진 제방 벽, 검은색에 가까운 바닷물, 우측 지지대가 프롬프트의 설명대로 묘사되었습니다.",
        "hard_violations": [],
        "physics": "물이 표면을 타고 중력 방향으로 자연스럽게 흘러내리고 있어 물리적으로 타당합니다."
       },
       {
        "label": "B",
        "direction": "물이 틈새를 따라 흐르며, 우측 콘크리트 구조물에서는 허공으로 쏟아지듯 아래로 떨어집니다.",
        "built_space": "경사진 콘크리트 제방이 있으며, 상단 둔치 쪽에 레퍼런스 사진(탁 트인 포장도로)에 존재하지 않는 금속 가드레일 기둥이 세워져 있습니다.",
        "entities": "갈라진 벽과 바닷물은 존재하나, 프롬프트와 레퍼런스에 없는 가드레일이 포함되었습니다.",
        "hard_violations": [
         "invented objects (레퍼런스에 없는 상단 금속 가드레일 추가)"
        ],
        "physics": "물의 낙하와 흐름을 지탱하는 표면 등 물리적 법칙에는 어긋남이 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 클로즈업 구도를 잘 살렸으며, 갈라진 틈과 우측 지지대를 표면을 따라 흐르는 검은 물의 묘사가 레퍼런스의 환경과 자연스럽게 부합합니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "장소의 영구적인 구조(LOCATION lock)를 유지해야 함에도 레퍼런스에 없는 금속 가드레일을 임의로 추가하여 프롬프트 지시를 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "물이 제방의 갈라진 틈과 우측 구조물의 표면을 따라 아래로 똑바로 흘러내립니다.",
        "built_space": "경사진 콘크리트 제방 벽이 카메라에서 멀어지도록 배치되어 있으며, 우측에는 벽과 연결된 수직 지지대 구조물이 있습니다.",
        "entities": "갈라진 제방 벽, 검은색에 가까운 바닷물, 우측 지지대가 프롬프트의 설명대로 묘사되었습니다.",
        "hard_violations": [],
        "physics": "물이 표면을 타고 중력 방향으로 자연스럽게 흘러내리고 있어 물리적으로 타당합니다."
       },
       {
        "label": "B",
        "direction": "물이 틈새를 따라 흐르며, 우측 콘크리트 구조물에서는 허공으로 쏟아지듯 아래로 떨어집니다.",
        "built_space": "경사진 콘크리트 제방이 있으며, 상단 둔치 쪽에 레퍼런스 사진(탁 트인 포장도로)에 존재하지 않는 금속 가드레일 기둥이 세워져 있습니다.",
        "entities": "갈라진 벽과 바닷물은 존재하나, 프롬프트와 레퍼런스에 없는 가드레일이 포함되었습니다.",
        "hard_violations": [
         "invented objects (레퍼런스에 없는 상단 금속 가드레일 추가)"
        ],
        "physics": "물의 낙하와 흐름을 지탱하는 표면 등 물리적 법칙에는 어긋남이 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "밤의 누수와 경사진 콘크리트는 표현했지만, 참조에 없는 금속 방호난간을 추가했고 균열보다 오른쪽 월류가 강조되어 지정된 표면 클로즈업에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "왼쪽 긴 균열에서 오른쪽 지지대 접합부로 이어지는 근접 구도와 하향 누수가 더 충실하지만, 참조에 없는 연속 콘크리트 상부벽을 만들어 장소 구조 고정을 위반합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "긴 균열이 화면 위쪽 중앙에서 아래쪽 중앙으로 내려오며 오른쪽 구조물 옆에 이릅니다. 오른쪽 물줄기는 구조물 윗모서리에서 아래 경사면으로 떨어지고, 경사면의 물은 왼쪽 아래로 흐릅니다. 균열을 통한 누수보다 윗면에서 넘치는 물이 더 뚜렷합니다. 사람이나 조준 대상은 없습니다.",
        "built_space": "왼쪽과 중앙에 경사진 콘크리트 면 하나, 오른쪽에 돌출된 콘크리트 구조물 하나가 보입니다. 그 위에는 금속 방호난간과 기둥 세 개가 보이고, 아래에는 어두운 개구부가 있습니다. 참조의 제방은 보행로와 이어지는 연속 경사면이며 이런 난간은 없습니다. 수평선 불빛과 넓은 경사면까지 포함하여 표면 관찰의 범위도 비교적 넓습니다.",
        "entities": "거친 콘크리트, 깊고 긴 균열, 오른쪽 지지 구조, 광택 있는 어두운 물이 보입니다. 물의 염분은 이미지로 확인할 수 없지만 바닷가 누수로 읽힙니다. 밤은 맞고, 인물이나 읽을 수 있는 글자는 없습니다. 금속 방호난간은 프롬프트와 장소 참조에 없는 추가 시설입니다.",
        "hard_violations": [
         "고정 장소 참조에 없는 금속 방호난간과 이를 받치는 기둥들을 제방 위에 추가했습니다."
        ],
        "physics": "물은 콘크리트 윗모서리에서 중력 방향으로 낙하해 아래 경사면에 닿고, 이후 경사를 따라 흐릅니다. 난간 기둥은 콘크리트 받침에 고정되어 있습니다. 지지 없이 떠 있는 물체는 없으며, 물의 낙하 자체는 가능하지만 주된 유출 시작점은 균열보다 구조물 상단으로 보입니다."
       },
       {
        "label": "B",
        "direction": "균열이 왼쪽 중간에서 오른쪽 아래 접합부까지 비스듬히 이어집니다. 균열 여러 지점에서 물이 경사면 아래쪽으로 흘러내리고, 오른쪽 구조물의 윗접합부에서도 물이 수직으로 내려갑니다. 왼쪽 균열과 오른쪽 지지대라는 지정 관계가 A보다 명확합니다. 사람이나 조준 대상은 없습니다.",
        "built_space": "왼쪽의 경사진 손상 면 하나, 오른쪽의 수평 콘크리트 부재 하나와 그 아래 수직 지지 면 하나가 크게 보입니다. 두 면의 접합부와 어두운 틈이 드러납니다. 상단에는 연속된 높은 콘크리트 턱벽이 있고 왼쪽 뒤에는 가느다란 기둥 하나가 보입니다. 참조의 보행로 가장자리에는 이러한 연속 상부벽이 없어 고정 구조가 일치하지 않습니다. 균열과 접합부를 크게 잡은 구도 자체는 요청에 가깝습니다.",
        "entities": "긴 파열 균열, 마모된 콘크리트, 오른쪽 지지 부재, 거의 검은 누수가 모두 보입니다. 균열 아래의 젖은 자국과 흐르는 물도 구별됩니다. 야간이며 인물과 판독 가능한 글자는 없습니다. 다만 상부의 연속 콘크리트 벽은 참조 장소의 시설이 아닙니다.",
        "hard_violations": [
         "참조에서 보행로와 경사면이 이어지는 제방 상단을 연속된 높은 콘크리트 벽으로 바꾸어 고정 구조를 추가했습니다."
        ],
        "physics": "균열에서 나온 물이 콘크리트 표면을 따라 아래로 흐르고, 오른쪽 물줄기는 윗모서리에서 아래 지지 면으로 떨어집니다. 수평 부재는 아래 수직 구조에 받쳐져 있고 파손 경사판도 주변 콘크리트와 연결되어 있습니다. 지지 없이 떠 있는 물체나 불가능한 물의 진행 방향은 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "밤의 누수와 경사진 콘크리트는 표현했지만, 참조에 없는 금속 방호난간을 추가했고 균열보다 오른쪽 월류가 강조되어 지정된 표면 클로즈업에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "왼쪽 긴 균열에서 오른쪽 지지대 접합부로 이어지는 근접 구도와 하향 누수가 더 충실하지만, 참조에 없는 연속 콘크리트 상부벽을 만들어 장소 구조 고정을 위반합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "긴 균열이 화면 위쪽 중앙에서 아래쪽 중앙으로 내려오며 오른쪽 구조물 옆에 이릅니다. 오른쪽 물줄기는 구조물 윗모서리에서 아래 경사면으로 떨어지고, 경사면의 물은 왼쪽 아래로 흐릅니다. 균열을 통한 누수보다 윗면에서 넘치는 물이 더 뚜렷합니다. 사람이나 조준 대상은 없습니다.",
        "built_space": "왼쪽과 중앙에 경사진 콘크리트 면 하나, 오른쪽에 돌출된 콘크리트 구조물 하나가 보입니다. 그 위에는 금속 방호난간과 기둥 세 개가 보이고, 아래에는 어두운 개구부가 있습니다. 참조의 제방은 보행로와 이어지는 연속 경사면이며 이런 난간은 없습니다. 수평선 불빛과 넓은 경사면까지 포함하여 표면 관찰의 범위도 비교적 넓습니다.",
        "entities": "거친 콘크리트, 깊고 긴 균열, 오른쪽 지지 구조, 광택 있는 어두운 물이 보입니다. 물의 염분은 이미지로 확인할 수 없지만 바닷가 누수로 읽힙니다. 밤은 맞고, 인물이나 읽을 수 있는 글자는 없습니다. 금속 방호난간은 프롬프트와 장소 참조에 없는 추가 시설입니다.",
        "hard_violations": [
         "고정 장소 참조에 없는 금속 방호난간과 이를 받치는 기둥들을 제방 위에 추가했습니다."
        ],
        "physics": "물은 콘크리트 윗모서리에서 중력 방향으로 낙하해 아래 경사면에 닿고, 이후 경사를 따라 흐릅니다. 난간 기둥은 콘크리트 받침에 고정되어 있습니다. 지지 없이 떠 있는 물체는 없으며, 물의 낙하 자체는 가능하지만 주된 유출 시작점은 균열보다 구조물 상단으로 보입니다."
       },
       {
        "label": "A",
        "direction": "균열이 왼쪽 중간에서 오른쪽 아래 접합부까지 비스듬히 이어집니다. 균열 여러 지점에서 물이 경사면 아래쪽으로 흘러내리고, 오른쪽 구조물의 윗접합부에서도 물이 수직으로 내려갑니다. 왼쪽 균열과 오른쪽 지지대라는 지정 관계가 A보다 명확합니다. 사람이나 조준 대상은 없습니다.",
        "built_space": "왼쪽의 경사진 손상 면 하나, 오른쪽의 수평 콘크리트 부재 하나와 그 아래 수직 지지 면 하나가 크게 보입니다. 두 면의 접합부와 어두운 틈이 드러납니다. 상단에는 연속된 높은 콘크리트 턱벽이 있고 왼쪽 뒤에는 가느다란 기둥 하나가 보입니다. 참조의 보행로 가장자리에는 이러한 연속 상부벽이 없어 고정 구조가 일치하지 않습니다. 균열과 접합부를 크게 잡은 구도 자체는 요청에 가깝습니다.",
        "entities": "긴 파열 균열, 마모된 콘크리트, 오른쪽 지지 부재, 거의 검은 누수가 모두 보입니다. 균열 아래의 젖은 자국과 흐르는 물도 구별됩니다. 야간이며 인물과 판독 가능한 글자는 없습니다. 다만 상부의 연속 콘크리트 벽은 참조 장소의 시설이 아닙니다.",
        "hard_violations": [
         "참조에서 보행로와 경사면이 이어지는 제방 상단을 연속된 높은 콘크리트 벽으로 바꾸어 고정 구조를 추가했습니다."
        ],
        "physics": "균열에서 나온 물이 콘크리트 표면을 따라 아래로 흐르고, 오른쪽 물줄기는 윗모서리에서 아래 지지 면으로 떨어집니다. 수평 부재는 아래 수직 구조에 받쳐져 있고 파손 경사판도 주변 콘크리트와 연결되어 있습니다. 지지 없이 떠 있는 물체나 불가능한 물의 진행 방향은 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.321
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.071
   },
   "violations": {
    "B": [
     "[gemini-pro] invented objects (레퍼런스에 없는 상단 금속 가드레일 추가)",
     "[gpt-high] 고정 장소 참조에 없는 금속 방호난간과 이를 받치는 기둥들을 제방 위에 추가했습니다."
    ],
    "A": [
     "[gpt-high] 참조에서 보행로와 경사면이 이어지는 제방 상단을 연속된 높은 콘크리트 벽으로 바꾸어 고정 구조를 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1071
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "프롬프트가 요구한 클로즈업 구도를 잘 살렸으며, 갈라진 틈과 우측 지지대를 표면을 따라 흐르는 검은 물의 묘사가 레퍼런스의 환경과 자연스럽게 부합합니다.  ★위반: [gpt-high] 참조에서 보행로와 경사면이 이어지는 제방 상단을 연속된 높은 콘크리트 벽으로 바꾸어 고정 구조를 추가했습니다."
   },
   {
    "label": "B",
    "score": 1071,
    "verdict_ko": "장소의 영구적인 구조(LOCATION lock)를 유지해야 함에도 레퍼런스에 없는 금속 가드레일을 임의로 추가하여 프롬프트 지시를 위반했습니다.  ★위반: [gemini-pro] invented objects (레퍼런스에 없는 상단 금속 가드레일 추가) / [gpt-high] 고정 장소 참조에 없는 금속 방호난간과 이를 받치는 기둥들을 제방 위에 추가했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_roadside_barrel_fire_sel.png",
    "asset_id": "02baceb0-aacc-4af3-8b75-a81171fa4c57",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd88-8abf-79a4-94a7-9766491030e7",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S13sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:22:00.681821+00:00",
  "fingerprint": "a76458da2d5863df0f4eb8aca8ef6496e223638a331ef5df1a9c3874e8edf5ad",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S13sh14_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S13sh14_sel.png",
  "source_sha256": "1fe08b0b1bc794da43b79b4383717b1fe46d854e63408d761477ed22b211cc24",
  "file": "S13sh14_cine.png",
  "staged_sha256": "cad14d8a0dccd79781c9657824136f9eaf96dc59fcace1bd9d17091c262b76ed",
  "latency_ms": 14047
 },
 "S13sh19::signage": {
  "fp": "c48da3ef07bb5b89",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S13sh19": {
  "input_fingerprint": "2a84aa630eabe213",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 쿠마가 빳빳한 지폐 뭉치를 이현우의 가슴 앞으로 불쑥 뻗어 멈춘 손끝.\n\nLOCATION (lock): At the foot of the leaking seawall beside the refugee settlement, outdoors at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Banknote bundle (Held out by 쿠마, not yet taken by 이현우) — Seen at an oblique angle that exposes the layered edges without inventing denominations or printed details; used as Forms the small, sharply readable focal point between the two bodies; Embankment behind the exchange (Cracked in the established location) — A limited section of its face lies beyond the hands; used as Soft background context ties the payment to the preceding warning.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination, using restrained tonal separation to distinguish the hands, banknotes, and clothing without introducing another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall remains cracked and wet with seawater, including its leaking supports; the setting is still dark. 이현우: He retains his outer top, facial injuries, dog-bitten leg, and the seized gun. The offered payment has not yet been taken. 쿠마: His palm is wet from wiping the seawall. He holds out money taken from inside his clothing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 쿠마 right now, so 쿠마's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 쿠마: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 쿠마가 빳빳한 지폐 뭉치를 이현우의 가슴 앞으로 불쑥 뻗어 멈춘 손끝.\n\nLOCATION (lock): At the foot of the leaking seawall beside the refugee settlement, outdoors at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Banknote bundle (Held out by 쿠마, not yet taken by 이현우) — Seen at an oblique angle that exposes the layered edges without inventing denominations or printed details; used as Forms the small, sharply readable focal point between the two bodies; Embankment behind the exchange (Cracked in the established location) — A limited section of its face lies beyond the hands; used as Soft background context ties the payment to the preceding warning.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination, using restrained tonal separation to distinguish the hands, banknotes, and clothing without introducing another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall remains cracked and wet with seawater, including its leaking supports; the setting is still dark. 이현우: He retains his outer top, facial injuries, dog-bitten leg, and the seized gun. The offered payment has not yet been taken. 쿠마: His palm is wet from wiping the seawall. He holds out money taken from inside his clothing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 쿠마 right now, so 쿠마's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 쿠마: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 쿠마가 빳빳한 지폐 뭉치를 이현우의 가슴 앞으로 불쑥 뻗어 멈춘 손끝.\n\nLOCATION (lock): At the foot of the leaking seawall beside the refugee settlement, outdoors at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Banknote bundle (Held out by 쿠마, not yet taken by 이현우) — Seen at an oblique angle that exposes the layered edges without inventing denominations or printed details; used as Forms the small, sharply readable focal point between the two bodies; Embankment behind the exchange (Cracked in the established location) — A limited section of its face lies beyond the hands; used as Soft background context ties the payment to the preceding warning.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination, using restrained tonal separation to distinguish the hands, banknotes, and clothing without introducing another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall remains cracked and wet with seawater, including its leaking supports; the setting is still dark. 이현우: He retains his outer top, facial injuries, dog-bitten leg, and the seized gun. The offered payment has not yet been taken. 쿠마: His palm is wet from wiping the seawall. He holds out money taken from inside his clothing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 쿠마 right now, so 쿠마's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 쿠마: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "쿠마가 이현우의 가슴 앞쪽으로 지폐를 내밀고, 이현우는 시선을 내려 그 돈을 바라봄.",
    "built_space": "물이 새어나오는 갈라진 방파제가 인물들 바로 뒤에 위치함.",
    "entities": "이현우는 상처 난 얼굴과 후드티, 총을 들고 있음. 쿠마는 셔츠와 조끼를 입었으나 조끼에 영문 텍스트가 있고, 쥔 지폐에는 현대 1만원권의 선명한 도안이 보임.",
    "hard_violations": [
     "[gemini-pro] 조끼에 읽을 수 있는 텍스트 노출",
     "[gemini-pro] 지폐에 명확한 액면가 및 인쇄 디테일 노출",
     "[gpt-high] 쿠마의 조끼에 판독 가능한 영문 글자열 일부가 노출되어, 이미지 어디에도 읽을 수 있는 글씨가 없어야 한다는 금지 조건을 위반합니다."
    ],
    "physics": "쿠마가 한 손으로 돈을 쥐고 팔을 뻗고 있으며, 이현우는 서서 한 손에 총을 쥐고 있음."
   },
   {
    "label": "B",
    "direction": "쿠마가 이현우의 가슴 쪽으로 지폐 뭉치를 뻗어 내밀고 있음.",
    "built_space": "야외 야간, 물이 흐르는 갈라진 콘크리트 방파제가 두 사람 뒤에 배경으로 자리함.",
    "entities": "쿠마의 외모와 의상이 레퍼런스와 일치하며 지폐 뭉치는 텍스트나 디테일이 흐릿하게 처리됨. 이현우는 낡은 후드티를 입고 턱과 가슴 위주로 프레임에 들어옴.",
    "hard_violations": [],
    "physics": "쿠마의 손이 지폐를 쥔 채 허공에 안정적으로 뻗어 있으며, 이현우 역시 바닥에 서서 체중을 지탱함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "클로즈업 구도 안에서 지폐의 디테일을 규칙에 맞게 억제하고 두 인물의 물리적 상호작용을 사실적으로 담아내어 지시문을 훌륭히 수행함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "장면의 분위기는 준수하나, 읽을 수 없는 텍스트와 가려진 화폐 디테일을 요구한 프롬프트를 어기고 조끼와 지폐에 선명한 문자와 도안이 노출되어 치명적인 감점 요소가 됨."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "쿠마가 이현우의 가슴 쪽으로 지폐 뭉치를 뻗어 내밀고 있음.",
        "built_space": "야외 야간, 물이 흐르는 갈라진 콘크리트 방파제가 두 사람 뒤에 배경으로 자리함.",
        "entities": "쿠마의 외모와 의상이 레퍼런스와 일치하며 지폐 뭉치는 텍스트나 디테일이 흐릿하게 처리됨. 이현우는 낡은 후드티를 입고 턱과 가슴 위주로 프레임에 들어옴.",
        "hard_violations": [],
        "physics": "쿠마의 손이 지폐를 쥔 채 허공에 안정적으로 뻗어 있으며, 이현우 역시 바닥에 서서 체중을 지탱함."
       },
       {
        "label": "A",
        "direction": "쿠마가 이현우의 가슴 앞쪽으로 지폐를 내밀고, 이현우는 시선을 내려 그 돈을 바라봄.",
        "built_space": "물이 새어나오는 갈라진 방파제가 인물들 바로 뒤에 위치함.",
        "entities": "이현우는 상처 난 얼굴과 후드티, 총을 들고 있음. 쿠마는 셔츠와 조끼를 입었으나 조끼에 영문 텍스트가 있고, 쥔 지폐에는 현대 1만원권의 선명한 도안이 보임.",
        "hard_violations": [
         "조끼에 읽을 수 있는 텍스트 노출",
         "지폐에 명확한 액면가 및 인쇄 디테일 노출"
        ],
        "physics": "쿠마가 한 손으로 돈을 쥐고 팔을 뻗고 있으며, 이현우는 서서 한 손에 총을 쥐고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "클로즈업 구도 안에서 지폐의 디테일을 규칙에 맞게 억제하고 두 인물의 물리적 상호작용을 사실적으로 담아내어 지시문을 훌륭히 수행함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "장면의 분위기는 준수하나, 읽을 수 없는 텍스트와 가려진 화폐 디테일을 요구한 프롬프트를 어기고 조끼와 지폐에 선명한 문자와 도안이 노출되어 치명적인 감점 요소가 됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "쿠마가 이현우의 가슴 쪽으로 지폐 뭉치를 뻗어 내밀고 있음.",
        "built_space": "야외 야간, 물이 흐르는 갈라진 콘크리트 방파제가 두 사람 뒤에 배경으로 자리함.",
        "entities": "쿠마의 외모와 의상이 레퍼런스와 일치하며 지폐 뭉치는 텍스트나 디테일이 흐릿하게 처리됨. 이현우는 낡은 후드티를 입고 턱과 가슴 위주로 프레임에 들어옴.",
        "hard_violations": [],
        "physics": "쿠마의 손이 지폐를 쥔 채 허공에 안정적으로 뻗어 있으며, 이현우 역시 바닥에 서서 체중을 지탱함."
       },
       {
        "label": "A",
        "direction": "쿠마가 이현우의 가슴 앞쪽으로 지폐를 내밀고, 이현우는 시선을 내려 그 돈을 바라봄.",
        "built_space": "물이 새어나오는 갈라진 방파제가 인물들 바로 뒤에 위치함.",
        "entities": "이현우는 상처 난 얼굴과 후드티, 총을 들고 있음. 쿠마는 셔츠와 조끼를 입었으나 조끼에 영문 텍스트가 있고, 쥔 지폐에는 현대 1만원권의 선명한 도안이 보임.",
        "hard_violations": [
         "조끼에 읽을 수 있는 텍스트 노출",
         "지폐에 명확한 액면가 및 인쇄 디테일 노출"
        ],
        "physics": "쿠마가 한 손으로 돈을 쥐고 팔을 뻗고 있으며, 이현우는 서서 한 손에 총을 쥐고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이현우의 가슴 앞에서 멈춘 쿠마의 손과 지폐에 더 밀착한 구도이며, 다만 손의 젖음이 약하고 방조제와 지폐 인쇄가 필요 이상으로 선명합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "젖은 손으로 가슴 앞에 돈을 내미는 동작은 맞지만, 조끼의 읽을 수 있는 영문이 금지 조건을 위반하고 얼굴·총까지 포함한 구도가 손끝 중심에서 벗어납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 쿠마의 팔과 지폐 끝이 오른쪽 이현우의 가슴을 향하고, 지폐는 가슴에 닿기 전에 멈춰 있습니다. 쿠마의 눈은 오른쪽 이현우 쪽을 향합니다. 이현우의 눈은 화면 밖이며, 돈을 받는 손은 보이지 않습니다.",
        "built_space": "두 사람 사이 뒤편에 갈라진 경사 콘크리트 벽면 하나, 상단 턱 한 줄, 오른쪽 누수 지지부 일부가 보입니다. 균열과 흘러내리는 물의 배치는 장소 참조와 부합하며 중복된 구조물은 없습니다. 다만 배경 벽면의 노출 면적과 선명도가 커서 제한적이고 부드러운 배경이라는 지시에는 덜 맞습니다.",
        "entities": "보이는 사람은 쿠마와 이현우 두 명뿐입니다. 쿠마의 젊은 성인 남성 얼굴 일부, 무늬 셔츠와 걷힌 소매, 가죽 조끼는 참조와 대체로 맞습니다. 이현우는 턱 아래와 낡은 짙은 후드 상의만 보여 얼굴 정체성과 상처는 판단할 수 없습니다. 다리와 총은 구도 밖이므로 누락으로 판단하지 않습니다. 지폐 한 뭉치의 겹친 가장자리가 비스듬히 보이지만, 윗장의 인쇄 흔적도 드러나 있습니다. 국가와 액면은 확실히 판독되지 않습니다. 쿠마 손의 물기는 뚜렷하지 않습니다.",
        "hard_violations": [],
        "physics": "쿠마의 엄지와 굽힌 손가락이 지폐 아래쪽을 집어 지지하며, 손목과 전완이 왼쪽 소매로 자연스럽게 이어집니다. 지폐 끝이 약간 벌어지는 형태도 손으로 쥔 종이 뭉치로 가능합니다. 이현우에게 아직 전달되지 않은 상태이고, 지지 없이 떠 있는 물체나 불가능한 관절은 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "쿠마의 팔이 왼쪽에서 오른쪽으로 뻗고 지폐 끝은 이현우의 윗가슴 앞을 향합니다. 이현우는 왼쪽 아래의 돈 쪽을 바라보며 받으려는 손을 내밀지 않습니다. 아래쪽 총의 보이는 축은 왼쪽 아래로 기울어 있으며 쿠마의 몸이나 지폐를 겨누지 않습니다. 쿠마의 눈은 잘려 있어 시선은 확인할 수 없습니다.",
        "built_space": "배경에 경사 방조제 벽면 하나와 상단 턱 한 줄, 오른쪽 누수 지지부 일부가 보입니다. 중앙 균열과 젖은 콘크리트는 장소 참조와 잘 이어지고 추가되거나 중복된 고정 구조물은 없습니다. 그러나 벽면이 넓고 비교적 선명하게 드러나 손 너머의 제한적인 연성 배경 역할보다 강하게 부각됩니다.",
        "entities": "쿠마와 이현우 두 명이 보입니다. 쿠마의 무늬 셔츠와 걷힌 소매는 참조에 맞으며 손과 팔에는 물기가 분명합니다. 이현우의 젊은 한국계 남성 외관, 검은 머리, 얼굴 상처와 낡은 후드 상의가 보이고 참조에 대체로 부합합니다. 아래쪽에는 손에 든 총 일부가 보이며 다리 부상은 구도 밖입니다. 지폐는 겹친 가장자리가 보이는 한 뭉치이지만 윗장에 인물 초상과 인쇄 도안이 노출되어 인쇄 세부를 만들지 말라는 지시와 어긋납니다. 쿠마 조끼의 왼쪽 아래에는 읽을 수 있는 영문 글자열 일부가 드러납니다.",
        "hard_violations": [
         "쿠마의 조끼에 판독 가능한 영문 글자열 일부가 노출되어, 이미지 어디에도 읽을 수 있는 글씨가 없어야 한다는 금지 조건을 위반합니다."
        ],
        "physics": "쿠마의 손가락과 엄지가 지폐를 잡고 있으며 젖은 전완은 소매와 자연스럽게 연결됩니다. 뻗은 팔로 가슴 앞에 돈을 멈춰 들고 있는 자세는 가능합니다. 아래쪽 총은 이현우의 손과 접촉해 지지되고 있습니다. 돈은 이현우 손에 넘어가지 않았고, 근거 없이 떠 있는 물체나 신체는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이현우의 가슴 앞에서 멈춘 쿠마의 손과 지폐에 더 밀착한 구도이며, 다만 손의 젖음이 약하고 방조제와 지폐 인쇄가 필요 이상으로 선명합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "젖은 손으로 가슴 앞에 돈을 내미는 동작은 맞지만, 조끼의 읽을 수 있는 영문이 금지 조건을 위반하고 얼굴·총까지 포함한 구도가 손끝 중심에서 벗어납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 쿠마의 팔과 지폐 끝이 오른쪽 이현우의 가슴을 향하고, 지폐는 가슴에 닿기 전에 멈춰 있습니다. 쿠마의 눈은 오른쪽 이현우 쪽을 향합니다. 이현우의 눈은 화면 밖이며, 돈을 받는 손은 보이지 않습니다.",
        "built_space": "두 사람 사이 뒤편에 갈라진 경사 콘크리트 벽면 하나, 상단 턱 한 줄, 오른쪽 누수 지지부 일부가 보입니다. 균열과 흘러내리는 물의 배치는 장소 참조와 부합하며 중복된 구조물은 없습니다. 다만 배경 벽면의 노출 면적과 선명도가 커서 제한적이고 부드러운 배경이라는 지시에는 덜 맞습니다.",
        "entities": "보이는 사람은 쿠마와 이현우 두 명뿐입니다. 쿠마의 젊은 성인 남성 얼굴 일부, 무늬 셔츠와 걷힌 소매, 가죽 조끼는 참조와 대체로 맞습니다. 이현우는 턱 아래와 낡은 짙은 후드 상의만 보여 얼굴 정체성과 상처는 판단할 수 없습니다. 다리와 총은 구도 밖이므로 누락으로 판단하지 않습니다. 지폐 한 뭉치의 겹친 가장자리가 비스듬히 보이지만, 윗장의 인쇄 흔적도 드러나 있습니다. 국가와 액면은 확실히 판독되지 않습니다. 쿠마 손의 물기는 뚜렷하지 않습니다.",
        "hard_violations": [],
        "physics": "쿠마의 엄지와 굽힌 손가락이 지폐 아래쪽을 집어 지지하며, 손목과 전완이 왼쪽 소매로 자연스럽게 이어집니다. 지폐 끝이 약간 벌어지는 형태도 손으로 쥔 종이 뭉치로 가능합니다. 이현우에게 아직 전달되지 않은 상태이고, 지지 없이 떠 있는 물체나 불가능한 관절은 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "쿠마의 팔이 왼쪽에서 오른쪽으로 뻗고 지폐 끝은 이현우의 윗가슴 앞을 향합니다. 이현우는 왼쪽 아래의 돈 쪽을 바라보며 받으려는 손을 내밀지 않습니다. 아래쪽 총의 보이는 축은 왼쪽 아래로 기울어 있으며 쿠마의 몸이나 지폐를 겨누지 않습니다. 쿠마의 눈은 잘려 있어 시선은 확인할 수 없습니다.",
        "built_space": "배경에 경사 방조제 벽면 하나와 상단 턱 한 줄, 오른쪽 누수 지지부 일부가 보입니다. 중앙 균열과 젖은 콘크리트는 장소 참조와 잘 이어지고 추가되거나 중복된 고정 구조물은 없습니다. 그러나 벽면이 넓고 비교적 선명하게 드러나 손 너머의 제한적인 연성 배경 역할보다 강하게 부각됩니다.",
        "entities": "쿠마와 이현우 두 명이 보입니다. 쿠마의 무늬 셔츠와 걷힌 소매는 참조에 맞으며 손과 팔에는 물기가 분명합니다. 이현우의 젊은 한국계 남성 외관, 검은 머리, 얼굴 상처와 낡은 후드 상의가 보이고 참조에 대체로 부합합니다. 아래쪽에는 손에 든 총 일부가 보이며 다리 부상은 구도 밖입니다. 지폐는 겹친 가장자리가 보이는 한 뭉치이지만 윗장에 인물 초상과 인쇄 도안이 노출되어 인쇄 세부를 만들지 말라는 지시와 어긋납니다. 쿠마 조끼의 왼쪽 아래에는 읽을 수 있는 영문 글자열 일부가 드러납니다.",
        "hard_violations": [
         "쿠마의 조끼에 판독 가능한 영문 글자열 일부가 노출되어, 이미지 어디에도 읽을 수 있는 글씨가 없어야 한다는 금지 조건을 위반합니다."
        ],
        "physics": "쿠마의 손가락과 엄지가 지폐를 잡고 있으며 젖은 전완은 소매와 자연스럽게 연결됩니다. 뻗은 팔로 가슴 앞에 돈을 멈춰 들고 있는 자세는 가능합니다. 아래쪽 총은 이현우의 손과 접촉해 지지되고 있습니다. 돈은 이현우 손에 넘어가지 않았고, 근거 없이 떠 있는 물체나 신체는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.75,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 조끼에 읽을 수 있는 텍스트 노출",
     "[gemini-pro] 지폐에 명확한 액면가 및 인쇄 디테일 노출",
     "[gpt-high] 쿠마의 조끼에 판독 가능한 영문 글자열 일부가 노출되어, 이미지 어디에도 읽을 수 있는 글씨가 없어야 한다는 금지 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "클로즈업 구도 안에서 지폐의 디테일을 규칙에 맞게 억제하고 두 인물의 물리적 상호작용을 사실적으로 담아내어 지시문을 훌륭히 수행함."
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "장면의 분위기는 준수하나, 읽을 수 없는 텍스트와 가려진 화폐 디테일을 요구한 프롬프트를 어기고 조끼와 지폐에 선명한 문자와 도안이 노출되어 치명적인 감점 요소가 됨.  ★위반: [gemini-pro] 조끼에 읽을 수 있는 텍스트 노출 / [gemini-pro] 지폐에 명확한 액면가 및 인쇄 디테일 노출 / [gpt-high] 쿠마의 조끼에 판독 가능한 영문 글자열 일부가 노출되어, 이미지 어디에도 읽을 수 있는 글씨가 없어야 한다는 금지 조건을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S13sh14_sel.png",
    "asset_id": "b240a4b4-0a16-41aa-af5d-3a8aba55ef31",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:904970>",
    "asset_id": "afafdeae-9f05-477d-8a3f-6cedcc1096d3",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd89-70ca-79de-b9f2-f5066505e45d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S13sh14"
  }
 },
 "S13sh19::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:22:15.060540+00:00",
  "fingerprint": "53797fbbc1c11422ef0f6c915d6ab7bf04dabed29f379c029937f0e7f479209b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S13sh19_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S13sh19_sel.png",
  "source_sha256": "4e423a7132e5944722dacf3724b99d55e32bdf52587e4bab80bcf03e0eecf41c",
  "file": "S13sh19_cine.png",
  "staged_sha256": "cd9186d1c173689746eb587222bcc4e592cbdb0deed0172e194570b06935d0e5",
  "latency_ms": 12688
 },
 "S14sh2::signage": {
  "fp": "8335de23d22453d7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S14sh2": {
  "input_fingerprint": "5a6358496b9b7f41",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 거대한 고철 로봇 찰리의 뭉툭한 발밑에서 도자기 화병이 산산조각 나며 파편이 사방으로 튀어 오르는 타격의 찰나.\n\nLOCATION (lock): On the floor of the family container home's shared living area, in daytime ambient light beside scattered vase fragments. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ceramic vase fragments (Breaking apart and scattering at 찰리's foot); used as Distributed fragments make the impact legible without a single oversized prop; Container interior floor (Receiving the broken vase fragments); used as Provides negative space around the impact and a shared scale reference for the foot and debris.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient illumination appropriate to the daytime container interior, with clear tonal separation between the ceramic fragments, 찰리's foot, and the floor.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Inside container home 7-31, the vase is breaking and its fragments are beginning to scatter. Charlie retains his old, dirty gorilla-shaped metal body, blue-lit eyes, worn coat and hat, and the already peeling UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 거대한 고철 로봇 찰리의 뭉툭한 발밑에서 도자기 화병이 산산조각 나며 파편이 사방으로 튀어 오르는 타격의 찰나.\n\nLOCATION (lock): On the floor of the family container home's shared living area, in daytime ambient light beside scattered vase fragments. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ceramic vase fragments (Breaking apart and scattering at 찰리's foot); used as Distributed fragments make the impact legible without a single oversized prop; Container interior floor (Receiving the broken vase fragments); used as Provides negative space around the impact and a shared scale reference for the foot and debris.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient illumination appropriate to the daytime container interior, with clear tonal separation between the ceramic fragments, 찰리's foot, and the floor.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Inside container home 7-31, the vase is breaking and its fragments are beginning to scatter. Charlie retains his old, dirty gorilla-shaped metal body, blue-lit eyes, worn coat and hat, and the already peeling UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 거대한 고철 로봇 찰리의 뭉툭한 발밑에서 도자기 화병이 산산조각 나며 파편이 사방으로 튀어 오르는 타격의 찰나.\n\nLOCATION (lock): On the floor of the family container home's shared living area, in daytime ambient light beside scattered vase fragments. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ceramic vase fragments (Breaking apart and scattering at 찰리's foot); used as Distributed fragments make the impact legible without a single oversized prop; Container interior floor (Receiving the broken vase fragments); used as Provides negative space around the impact and a shared scale reference for the foot and debris.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient illumination appropriate to the daytime container interior, with clear tonal separation between the ceramic fragments, 찰리's foot, and the floor.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Inside container home 7-31, the vase is breaking and its fragments are beginning to scatter. Charlie retains his old, dirty gorilla-shaped metal body, blue-lit eyes, worn coat and hat, and the already peeling UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "금속 재질의 사지가 도자기 화병을 내리찍어 파편이 바깥쪽으로 튀어 오르고 있음.",
    "built_space": "레퍼런스 이미지와 일치하는 나무 바닥, 식탁, 의자, 문, 선반 등의 거실 구조가 보임.",
    "entities": "고릴라 형태의 기계 몸, 낡은 코트, UBIK 로고가 있는 가슴판, 산산조각 나는 도자기 화병이 일치함.",
    "hard_violations": [
     "[gpt-high] 가슴의 UBIK 문자가 크고 선명하게 판독되어 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
    ],
    "physics": "금속 사지가 화병을 짓누르는 힘에 의해 파편이 바닥에 흩어지고 공중으로 튀어 오르는 물리 법칙이 타당함."
   },
   {
    "label": "B",
    "direction": "가죽 부츠를 신은 발이 화병을 밟아 부수며 파편이 사방으로 튀고 있음.",
    "built_space": "나무 바닥과 배경의 가구 다리 일부만 제한적으로 보임.",
    "entities": "깨지는 도자기 화병은 일치하나, 뭉툭한 금속 발 대신 가죽 부츠가 등장하며 배경에 UBIK 문구가 적힌 옷을 입은 인물이 보임.",
    "hard_violations": [
     "[gemini-pro] 고철 로봇 찰리의 발을 금속이 아닌 인간의 가죽 부츠로 잘못 묘사함.",
     "[gpt-high] 가슴에 UBIK 문자가 판독 가능하게 노출되어 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
    ],
    "physics": "부츠가 화병에 가하는 압력과 파편이 비산하는 형태는 물리적으로 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "고릴라 형태의 고철 로봇 특징과 레퍼런스 배경을 훌륭히 구현했으나 샷의 크기가 요구된 클로즈업보다 다소 넓게 잡혔습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트가 명시한 금속 기계 발 대신 가죽 부츠를 묘사하여 캐릭터의 핵심 정체성을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "금속 재질의 사지가 도자기 화병을 내리찍어 파편이 바깥쪽으로 튀어 오르고 있음.",
        "built_space": "레퍼런스 이미지와 일치하는 나무 바닥, 식탁, 의자, 문, 선반 등의 거실 구조가 보임.",
        "entities": "고릴라 형태의 기계 몸, 낡은 코트, UBIK 로고가 있는 가슴판, 산산조각 나는 도자기 화병이 일치함.",
        "hard_violations": [],
        "physics": "금속 사지가 화병을 짓누르는 힘에 의해 파편이 바닥에 흩어지고 공중으로 튀어 오르는 물리 법칙이 타당함."
       },
       {
        "label": "B",
        "direction": "가죽 부츠를 신은 발이 화병을 밟아 부수며 파편이 사방으로 튀고 있음.",
        "built_space": "나무 바닥과 배경의 가구 다리 일부만 제한적으로 보임.",
        "entities": "깨지는 도자기 화병은 일치하나, 뭉툭한 금속 발 대신 가죽 부츠가 등장하며 배경에 UBIK 문구가 적힌 옷을 입은 인물이 보임.",
        "hard_violations": [
         "고철 로봇 찰리의 발을 금속이 아닌 인간의 가죽 부츠로 잘못 묘사함."
        ],
        "physics": "부츠가 화병에 가하는 압력과 파편이 비산하는 형태는 물리적으로 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "고릴라 형태의 고철 로봇 특징과 레퍼런스 배경을 훌륭히 구현했으나 샷의 크기가 요구된 클로즈업보다 다소 넓게 잡혔습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트가 명시한 금속 기계 발 대신 가죽 부츠를 묘사하여 캐릭터의 핵심 정체성을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "금속 재질의 사지가 도자기 화병을 내리찍어 파편이 바깥쪽으로 튀어 오르고 있음.",
        "built_space": "레퍼런스 이미지와 일치하는 나무 바닥, 식탁, 의자, 문, 선반 등의 거실 구조가 보임.",
        "entities": "고릴라 형태의 기계 몸, 낡은 코트, UBIK 로고가 있는 가슴판, 산산조각 나는 도자기 화병이 일치함.",
        "hard_violations": [],
        "physics": "금속 사지가 화병을 짓누르는 힘에 의해 파편이 바닥에 흩어지고 공중으로 튀어 오르는 물리 법칙이 타당함."
       },
       {
        "label": "B",
        "direction": "가죽 부츠를 신은 발이 화병을 밟아 부수며 파편이 사방으로 튀고 있음.",
        "built_space": "나무 바닥과 배경의 가구 다리 일부만 제한적으로 보임.",
        "entities": "깨지는 도자기 화병은 일치하나, 뭉툭한 금속 발 대신 가죽 부츠가 등장하며 배경에 UBIK 문구가 적힌 옷을 입은 인물이 보임.",
        "hard_violations": [
         "고철 로봇 찰리의 발을 금속이 아닌 인간의 가죽 부츠로 잘못 묘사함."
        ],
        "physics": "부츠가 화병에 가하는 압력과 파편이 비산하는 형태는 물리적으로 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "발로 화병을 내리누르는 타격과 근접 구도는 더 충실하지만, 읽히는 UBIK 문자 때문에 실격이며 가죽 부츠와 크게 남은 화병 몸통도 요구와 다릅니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "금속 몸체와 실내 재현은 가깝지만 발이 아닌 주먹으로 화병을 깨고 구도도 넓어졌으며, 선명한 UBIK 문자까지 노출되어 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 위에서 내려온 발의 밑창이 화면 중앙의 화병 몸통을 직접 누릅니다. 파편은 접촉점에서 위쪽과 오른쪽, 전경으로 퍼져 발밑에서 깨지는 방향 관계가 맞습니다. 얼굴과 시선은 보이지 않습니다.",
        "built_space": "목재 무늬 바닥, 밝은 벽과 갈색 걸레받이, 오른쪽 뒤의 좁은 목재 탁자 한 개가 보입니다. 탁자 아래에도 파편이 놓여 있어 참고 장소의 오른쪽 벽 주변과 대체로 연결됩니다. 다른 가구는 근접 프레임 밖이며, 중복된 고정 설비나 반사는 보이지 않습니다.",
        "entities": "금속 정강이와 낡은 가죽 끈 부츠, 뒤쪽의 낡은 외투와 가슴 일부가 보입니다. 발로 식별되지만 요청한 뭉툭한 고철 로봇 발보다는 사람의 작업화에 가깝습니다. 도자기 화병과 파편은 분명하나 큰 몸통이 상당 부분 남았고, 참고의 청백색 장식과 달리 다색 꽃무늬입니다. 얼굴·눈·모자는 프레임 밖입니다. 가슴의 UBIK 문자는 읽을 수 있습니다.",
        "hard_violations": [
         "가슴에 UBIK 문자가 판독 가능하게 노출되어 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
        ],
        "physics": "발은 정강이에 연결되어 있고 밑창이 화병에 닿아 하중을 전달합니다. 화병 아래쪽은 바닥에 닿아 있으며, 떠오른 파편은 이 압착 충격으로 설명됩니다. 바닥의 파편은 바닥에 지지됩니다. 몸 전체의 지지 자세는 잘려 있지만 근거 없이 공중에 떠 있는 몸이나 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "외투 소매에서 이어진 금속 팔과 주먹이 중앙의 화병을 아래로 누릅니다. 옆으로 감긴 엄지와 앞쪽 손가락 마디가 보여 타격 부위는 발이 아니라 손입니다. 금속 발은 오른쪽 바닥에 따로 놓여 있습니다. 파편은 주먹 아래에서 사방으로 튀며 얼굴과 시선은 보이지 않습니다.",
        "built_space": "왼쪽 출입구 한 곳과 그 너머 선반, 실내 책장 한 개와 휴지통 한 개, 중앙 식탁 한 개와 주변 의자들, 오른쪽 벽의 좁은 탁자 한 개가 보입니다. 목재 바닥과 밝은 벽, 왼쪽에서 들어오는 주간광은 참고 장소와 가깝습니다. 다만 발과 파편의 클로즈업보다 몸통과 방 전체 배치가 많이 포함된 넓은 구도입니다.",
        "entities": "낡은 외투, 오염된 금속 몸체, 커다란 금속 손과 발은 고릴라형 로봇 설정에 가깝습니다. 청백색 도자기 화병과 흩어진 파편도 보입니다. 그러나 화병을 깨는 것은 요청된 발이 아닌 주먹입니다. 얼굴·눈·모자는 보이지 않으며 가슴의 벗겨진 UBIK 문자가 크게 읽힙니다. 추가 인물은 보이지 않습니다.",
        "hard_violations": [
         "가슴의 UBIK 문자가 크고 선명하게 판독되어 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
        ],
        "physics": "오른쪽 금속 발이 바닥에 붙어 몸을 지지하고, 구부린 팔의 주먹이 바닥에 놓인 화병을 압착합니다. 파편과 분진의 공중 운동에는 명확한 충격원이 있고 떨어진 조각은 바닥에 놓여 있습니다. 물리적으로 설명 가능한 자세이지만 요청된 발 구르기 동작은 아닙니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "발로 화병을 내리누르는 타격과 근접 구도는 더 충실하지만, 읽히는 UBIK 문자 때문에 실격이며 가죽 부츠와 크게 남은 화병 몸통도 요구와 다릅니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "금속 몸체와 실내 재현은 가깝지만 발이 아닌 주먹으로 화병을 깨고 구도도 넓어졌으며, 선명한 UBIK 문자까지 노출되어 실격입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 위에서 내려온 발의 밑창이 화면 중앙의 화병 몸통을 직접 누릅니다. 파편은 접촉점에서 위쪽과 오른쪽, 전경으로 퍼져 발밑에서 깨지는 방향 관계가 맞습니다. 얼굴과 시선은 보이지 않습니다.",
        "built_space": "목재 무늬 바닥, 밝은 벽과 갈색 걸레받이, 오른쪽 뒤의 좁은 목재 탁자 한 개가 보입니다. 탁자 아래에도 파편이 놓여 있어 참고 장소의 오른쪽 벽 주변과 대체로 연결됩니다. 다른 가구는 근접 프레임 밖이며, 중복된 고정 설비나 반사는 보이지 않습니다.",
        "entities": "금속 정강이와 낡은 가죽 끈 부츠, 뒤쪽의 낡은 외투와 가슴 일부가 보입니다. 발로 식별되지만 요청한 뭉툭한 고철 로봇 발보다는 사람의 작업화에 가깝습니다. 도자기 화병과 파편은 분명하나 큰 몸통이 상당 부분 남았고, 참고의 청백색 장식과 달리 다색 꽃무늬입니다. 얼굴·눈·모자는 프레임 밖입니다. 가슴의 UBIK 문자는 읽을 수 있습니다.",
        "hard_violations": [
         "가슴에 UBIK 문자가 판독 가능하게 노출되어 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
        ],
        "physics": "발은 정강이에 연결되어 있고 밑창이 화병에 닿아 하중을 전달합니다. 화병 아래쪽은 바닥에 닿아 있으며, 떠오른 파편은 이 압착 충격으로 설명됩니다. 바닥의 파편은 바닥에 지지됩니다. 몸 전체의 지지 자세는 잘려 있지만 근거 없이 공중에 떠 있는 몸이나 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "외투 소매에서 이어진 금속 팔과 주먹이 중앙의 화병을 아래로 누릅니다. 옆으로 감긴 엄지와 앞쪽 손가락 마디가 보여 타격 부위는 발이 아니라 손입니다. 금속 발은 오른쪽 바닥에 따로 놓여 있습니다. 파편은 주먹 아래에서 사방으로 튀며 얼굴과 시선은 보이지 않습니다.",
        "built_space": "왼쪽 출입구 한 곳과 그 너머 선반, 실내 책장 한 개와 휴지통 한 개, 중앙 식탁 한 개와 주변 의자들, 오른쪽 벽의 좁은 탁자 한 개가 보입니다. 목재 바닥과 밝은 벽, 왼쪽에서 들어오는 주간광은 참고 장소와 가깝습니다. 다만 발과 파편의 클로즈업보다 몸통과 방 전체 배치가 많이 포함된 넓은 구도입니다.",
        "entities": "낡은 외투, 오염된 금속 몸체, 커다란 금속 손과 발은 고릴라형 로봇 설정에 가깝습니다. 청백색 도자기 화병과 흩어진 파편도 보입니다. 그러나 화병을 깨는 것은 요청된 발이 아닌 주먹입니다. 얼굴·눈·모자는 보이지 않으며 가슴의 벗겨진 UBIK 문자가 크게 읽힙니다. 추가 인물은 보이지 않습니다.",
        "hard_violations": [
         "가슴의 UBIK 문자가 크고 선명하게 판독되어 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
        ],
        "physics": "오른쪽 금속 발이 바닥에 붙어 몸을 지지하고, 구부린 팔의 주먹이 바닥에 놓인 화병을 압착합니다. 파편과 분진의 공중 운동에는 명확한 충격원이 있고 떨어진 조각은 바닥에 놓여 있습니다. 물리적으로 설명 가능한 자세이지만 요청된 발 구르기 동작은 아닙니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.179
   },
   "violations": {
    "B": [
     "[gemini-pro] 고철 로봇 찰리의 발을 금속이 아닌 인간의 가죽 부츠로 잘못 묘사함.",
     "[gpt-high] 가슴에 UBIK 문자가 판독 가능하게 노출되어 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
    ],
    "A": [
     "[gpt-high] 가슴의 UBIK 문자가 크고 선명하게 판독되어 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1417,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "고릴라 형태의 고철 로봇 특징과 레퍼런스 배경을 훌륭히 구현했으나 샷의 크기가 요구된 클로즈업보다 다소 넓게 잡혔습니다.  ★위반: [gpt-high] 가슴의 UBIK 문자가 크고 선명하게 판독되어 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "프롬프트가 명시한 금속 기계 발 대신 가죽 부츠를 묘사하여 캐릭터의 핵심 정체성을 심각하게 위반했습니다.  ★위반: [gemini-pro] 고철 로봇 찰리의 발을 금속이 아닌 인간의 가죽 부츠로 잘못 묘사함. / [gpt-high] 가슴에 UBIK 문자가 판독 가능하게 노출되어 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L22B01.png",
    "asset_id": "d4424321-29fd-4712-9827-332afb2fecb5",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd8a-411b-7d61-aa65-c2bed1c175a1",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S14sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:22:28.081237+00:00",
  "fingerprint": "9917bdca288990292b70b1fb625924faaded1b084d89883a9da9370de15f392d",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S14sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S14sh2_sel.png",
  "source_sha256": "ce07acd7a0a1f7bd56d247bdba80d4b734d0cd8c9733883547537a408d7a3ece",
  "file": "S14sh2_cine.png",
  "staged_sha256": "2dfc30572f9dbfa4e5a7f470c81cefdea1611bb5083195efa63515432696510a",
  "latency_ms": 11341
 },
 "S14sh5::signage": {
  "fp": "3f7991b2b81c252d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S14sh5": {
  "input_fingerprint": "7401380835722154",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 눈가 푸른빛이 희미해지는 진행 중의 한 순간, 두 어깨를 아래로 축 늘어뜨린 시무룩한 자태.\n\nLOCATION (lock): In the family container home's shared living area, among the household belongings in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container interior wall (Visible behind 찰리) — A partial interior face is seen obliquely behind his upper body; used as Provides a quiet background that preserves the silhouette of both lowered shoulders.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime interior ambience steady while the blue illumination around 찰리's eyes fades, preserving enough facial contrast to read his dejection.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken vase fragments remain on the floor, with some pushed aside in an attempt to conceal them. Charlie remains dirty and weathered, wearing the old coat and hat, with his peeling UBIK chest logo unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 눈가 푸른빛이 희미해지는 진행 중의 한 순간, 두 어깨를 아래로 축 늘어뜨린 시무룩한 자태.\n\nLOCATION (lock): In the family container home's shared living area, among the household belongings in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container interior wall (Visible behind 찰리) — A partial interior face is seen obliquely behind his upper body; used as Provides a quiet background that preserves the silhouette of both lowered shoulders.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime interior ambience steady while the blue illumination around 찰리's eyes fades, preserving enough facial contrast to read his dejection.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken vase fragments remain on the floor, with some pushed aside in an attempt to conceal them. Charlie remains dirty and weathered, wearing the old coat and hat, with his peeling UBIK chest logo unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 눈가 푸른빛이 희미해지는 진행 중의 한 순간, 두 어깨를 아래로 축 늘어뜨린 시무룩한 자태.\n\nLOCATION (lock): In the family container home's shared living area, among the household belongings in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container interior wall (Visible behind 찰리) — A partial interior face is seen obliquely behind his upper body; used as Provides a quiet background that preserves the silhouette of both lowered shoulders.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime interior ambience steady while the blue illumination around 찰리's eyes fades, preserving enough facial contrast to read his dejection.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken vase fragments remain on the floor, with some pushed aside in an attempt to conceal them. Charlie remains dirty and weathered, wearing the old coat and hat, with his peeling UBIK chest logo unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "로봇이 시선을 아래로 향하고 양 어깨를 축 늘어뜨리고 있음.",
    "built_space": "참조 이미지의 거실 공간(문, 선반, 테이블 등)이 보이나, 프롬프트에서 요구한 컨테이너 내벽이 배경에 존재하지 않음.",
    "entities": "UBIK 로고가 있는 코트 등은 일치하나, 고릴라형 머리가 아닌 평평하고 단순한 형태의 로봇 얼굴을 하고 있음.",
    "hard_violations": [
     "[gpt-high] 가슴의 UBIK 로고가 크게 읽혀, 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
    ],
    "physics": "몸을 앞으로 숙인 채 서 있으며 자세와 균형에 물리적 오류가 없음. 바닥에 도자기 파편이 흩어져 있음."
   },
   {
    "label": "B",
    "direction": "로봇이 바닥 쪽으로 시선을 두고 어깨를 늘어뜨린 채 웅크리고 있음.",
    "built_space": "참조 이미지의 거실 요소가 유지되었으며, 화면 우측 배경에 지시된 컨테이너의 주름진 철판 내벽이 배치됨.",
    "entities": "흰 얼굴판, 점 형태의 눈, 선 형태의 입을 갖춘 고릴라형 기계 머리와 UBIK 로고가 있는 낡은 코트 등 묘사가 텍스트와 정확히 일치함.",
    "hard_violations": [
     "[gpt-high] 가슴의 UBIK 로고가 명확하게 읽혀, 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
    ],
    "physics": "무릎을 꿇고 양손으로 바닥을 짚어 무게를 안정적으로 지탱하고 있음. 주변 바닥에 깨진 도자기 파편이 놓여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레이밍이 약간 넓게 잡혔으나, 프롬프트가 명시한 고릴라형 기계 머리의 형태와 배경의 컨테이너 내벽 묘사를 충실히 구현하여 프롬프트 부합도가 높습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "미디엄 샷의 구도는 준수하나, 핵심 지시사항인 고릴라형 머리 디자인과 배경의 컨테이너 내벽 묘사를 완전히 누락하여 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇이 시선을 아래로 향하고 양 어깨를 축 늘어뜨리고 있음.",
        "built_space": "참조 이미지의 거실 공간(문, 선반, 테이블 등)이 보이나, 프롬프트에서 요구한 컨테이너 내벽이 배경에 존재하지 않음.",
        "entities": "UBIK 로고가 있는 코트 등은 일치하나, 고릴라형 머리가 아닌 평평하고 단순한 형태의 로봇 얼굴을 하고 있음.",
        "hard_violations": [],
        "physics": "몸을 앞으로 숙인 채 서 있으며 자세와 균형에 물리적 오류가 없음. 바닥에 도자기 파편이 흩어져 있음."
       },
       {
        "label": "B",
        "direction": "로봇이 바닥 쪽으로 시선을 두고 어깨를 늘어뜨린 채 웅크리고 있음.",
        "built_space": "참조 이미지의 거실 요소가 유지되었으며, 화면 우측 배경에 지시된 컨테이너의 주름진 철판 내벽이 배치됨.",
        "entities": "흰 얼굴판, 점 형태의 눈, 선 형태의 입을 갖춘 고릴라형 기계 머리와 UBIK 로고가 있는 낡은 코트 등 묘사가 텍스트와 정확히 일치함.",
        "hard_violations": [],
        "physics": "무릎을 꿇고 양손으로 바닥을 짚어 무게를 안정적으로 지탱하고 있음. 주변 바닥에 깨진 도자기 파편이 놓여 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레이밍이 약간 넓게 잡혔으나, 프롬프트가 명시한 고릴라형 기계 머리의 형태와 배경의 컨테이너 내벽 묘사를 충실히 구현하여 프롬프트 부합도가 높습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "미디엄 샷의 구도는 준수하나, 핵심 지시사항인 고릴라형 머리 디자인과 배경의 컨테이너 내벽 묘사를 완전히 누락하여 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "로봇이 시선을 아래로 향하고 양 어깨를 축 늘어뜨리고 있음.",
        "built_space": "참조 이미지의 거실 공간(문, 선반, 테이블 등)이 보이나, 프롬프트에서 요구한 컨테이너 내벽이 배경에 존재하지 않음.",
        "entities": "UBIK 로고가 있는 코트 등은 일치하나, 고릴라형 머리가 아닌 평평하고 단순한 형태의 로봇 얼굴을 하고 있음.",
        "hard_violations": [],
        "physics": "몸을 앞으로 숙인 채 서 있으며 자세와 균형에 물리적 오류가 없음. 바닥에 도자기 파편이 흩어져 있음."
       },
       {
        "label": "B",
        "direction": "로봇이 바닥 쪽으로 시선을 두고 어깨를 늘어뜨린 채 웅크리고 있음.",
        "built_space": "참조 이미지의 거실 요소가 유지되었으며, 화면 우측 배경에 지시된 컨테이너의 주름진 철판 내벽이 배치됨.",
        "entities": "흰 얼굴판, 점 형태의 눈, 선 형태의 입을 갖춘 고릴라형 기계 머리와 UBIK 로고가 있는 낡은 코트 등 묘사가 텍스트와 정확히 일치함.",
        "hard_violations": [],
        "physics": "무릎을 꿇고 양손으로 바닥을 짚어 무게를 안정적으로 지탱하고 있음. 주변 바닥에 깨진 도자기 파편이 놓여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "읽히는 가슴 로고가 금지 조건을 위반하며, 바닥을 짚은 손과 발까지 보여주는 넓은 구도와 구체적인 고릴라 얼굴이 지정된 미디엄 숏·점 눈·선 입에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "읽히는 가슴 로고로 탈락 조건에 해당하지만, 상반신 중심 미디엄 숏과 각진 흰 얼굴판·점 눈·선 입, 고개를 숙인 시무룩한 모습은 A보다 충실합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 머리를 아래로 숙이고 화면 왼쪽 아래의 바닥 파편 쪽을 바라봅니다. 시무룩한 시선 방향은 맞지만, 눈의 푸른 점광과 주변 발광은 상당히 선명하여 희미해지는 순간이라는 표현은 약합니다. 무기나 겨누는 물건은 없습니다.",
        "built_space": "왼쪽에 밝은 창이 보이는 출입구 하나, 그 너머 낮은 선반 하나, 방 안 뒤쪽 책장 하나, 왼쪽 전경 수납장 하나가 보입니다. 뒤에는 나무 탁자 하나와 스툴 하나가 분명히 보이며, 오른쪽에는 낡은 밝은색 벽면이 비스듬히 놓입니다. 참고의 목재 바닥·가구·왼쪽 자연광은 대체로 이어지지만, 방과 바닥을 많이 보여줘 두 어깨 뒤의 조용한 벽면을 강조하는 미디엄 숏보다 넓습니다. 불가능한 반사나 명백한 고정 설비 중복은 보이지 않습니다.",
        "entities": "등장 인물은 기계 몸체의 찰리 하나뿐입니다. 낡고 더러운 갈색 코트와 마모된 금속 몸체, 벗겨진 가슴 도안, 바닥의 도자기 파편은 유지됩니다. 다만 머리는 둥근 두개 외장과 입체적인 고릴라 코·주둥이를 가지며, 지정된 각진 머리와 점 눈·선 입보다 동물 얼굴에 가깝습니다. 머리 전체가 보이지만 모자는 없습니다. 가슴의 UBIK 글자가 읽혀 글자와 로고를 읽을 수 없게 하라는 최종 지시를 어깁니다.",
        "hard_violations": [
         "가슴의 UBIK 로고가 명확하게 읽혀, 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
        ],
        "physics": "화면 왼쪽의 주먹 관절과 중앙의 넓은 발이 바닥에 닿아 앞으로 숙인 몸을 지탱합니다. 반대쪽 손은 하단에서 일부 잘리지만 몸 전체가 떠 있는 모습은 아닙니다. 코트는 몸과 팔을 따라 처지고 파편은 바닥에 놓여 있습니다. 다만 팔로 체중을 받는 자세라 두 어깨를 힘없이 아래로 늘어뜨린 동작보다는 바닥에 몸을 버티는 자세로 읽힙니다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴판이 아래로 기울어 있고 점 형태의 눈은 카메라 앞쪽의 낮은 바닥 방향을 향합니다. 특정 소품을 주시한다기보다 고개를 떨군 모습입니다. 눈 자체는 어둡고 눈 아래와 관자 부근에 작은 푸른 불빛이 남아 있어 A보다 잔광으로 읽히지만, 정지 화면만으로 빛이 감소하는 과정까지 확정할 수는 없습니다.",
        "built_space": "뒤쪽에 나무 탁자 하나, 왼쪽에 분명한 스툴 하나와 책장 하나가 보이고, 책장 아래 옆에는 작은 직육면체 생활용품이 놓여 있습니다. 벽에는 작은 사진들과 종이, 일부 가려진 액자가 있으며 오른쪽에는 밝은 커튼이 보입니다. 참고의 밝은 벽·목재 바닥·생활 가구와 낮의 자연광이 이어집니다. 찰리의 상반신이 화면을 채우고 벽면이 어깨 뒤에 남아 있어 지정된 미디엄 숏에 더 가깝습니다. 불가능한 반사나 명백한 설비 중복은 보이지 않습니다.",
        "entities": "찰리 하나만 등장하며, 각진 기계 머리 외장과 흰 얼굴판, 작은 원형 눈 두 개, 짧은 수평선 입이 명시된 디자인에 부합합니다. 더러운 갈색 코트와 긁히고 벗겨진 가슴판도 참고와 이어집니다. 머리가 노출되어 있지만 모자는 없습니다. 배경 바닥에는 깨진 도자기 조각들이 보이나 일부를 숨기려고 밀어놓았는지는 분명하지 않습니다. 가슴의 UBIK 글자는 크게 읽혀 최종 무문자 조건에 어긋납니다.",
        "hard_violations": [
         "가슴의 UBIK 로고가 크게 읽혀, 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
        ],
        "physics": "상체가 앞으로 기울고 두 팔이 아래로 내려와 있으며 머리는 목 기계부에 연결되어 있습니다. 손과 발의 접촉점은 미디엄 숏 바깥이므로 지지 방식을 직접 확인할 수 없지만, 몸이 공중에 떠 있다는 증거도 없습니다. 코트의 무게와 주름은 자연스럽고 배경 파편은 바닥에 놓여 있습니다. 어깨는 여전히 크고 둥글게 솟은 실루엣이어서 완전히 축 내려앉은 느낌은 제한적이지만, 숙인 머리와 내려온 팔은 시무룩한 자세를 전달합니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "읽히는 가슴 로고가 금지 조건을 위반하며, 바닥을 짚은 손과 발까지 보여주는 넓은 구도와 구체적인 고릴라 얼굴이 지정된 미디엄 숏·점 눈·선 입에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "읽히는 가슴 로고로 탈락 조건에 해당하지만, 상반신 중심 미디엄 숏과 각진 흰 얼굴판·점 눈·선 입, 고개를 숙인 시무룩한 모습은 A보다 충실합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 머리를 아래로 숙이고 화면 왼쪽 아래의 바닥 파편 쪽을 바라봅니다. 시무룩한 시선 방향은 맞지만, 눈의 푸른 점광과 주변 발광은 상당히 선명하여 희미해지는 순간이라는 표현은 약합니다. 무기나 겨누는 물건은 없습니다.",
        "built_space": "왼쪽에 밝은 창이 보이는 출입구 하나, 그 너머 낮은 선반 하나, 방 안 뒤쪽 책장 하나, 왼쪽 전경 수납장 하나가 보입니다. 뒤에는 나무 탁자 하나와 스툴 하나가 분명히 보이며, 오른쪽에는 낡은 밝은색 벽면이 비스듬히 놓입니다. 참고의 목재 바닥·가구·왼쪽 자연광은 대체로 이어지지만, 방과 바닥을 많이 보여줘 두 어깨 뒤의 조용한 벽면을 강조하는 미디엄 숏보다 넓습니다. 불가능한 반사나 명백한 고정 설비 중복은 보이지 않습니다.",
        "entities": "등장 인물은 기계 몸체의 찰리 하나뿐입니다. 낡고 더러운 갈색 코트와 마모된 금속 몸체, 벗겨진 가슴 도안, 바닥의 도자기 파편은 유지됩니다. 다만 머리는 둥근 두개 외장과 입체적인 고릴라 코·주둥이를 가지며, 지정된 각진 머리와 점 눈·선 입보다 동물 얼굴에 가깝습니다. 머리 전체가 보이지만 모자는 없습니다. 가슴의 UBIK 글자가 읽혀 글자와 로고를 읽을 수 없게 하라는 최종 지시를 어깁니다.",
        "hard_violations": [
         "가슴의 UBIK 로고가 명확하게 읽혀, 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
        ],
        "physics": "화면 왼쪽의 주먹 관절과 중앙의 넓은 발이 바닥에 닿아 앞으로 숙인 몸을 지탱합니다. 반대쪽 손은 하단에서 일부 잘리지만 몸 전체가 떠 있는 모습은 아닙니다. 코트는 몸과 팔을 따라 처지고 파편은 바닥에 놓여 있습니다. 다만 팔로 체중을 받는 자세라 두 어깨를 힘없이 아래로 늘어뜨린 동작보다는 바닥에 몸을 버티는 자세로 읽힙니다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴판이 아래로 기울어 있고 점 형태의 눈은 카메라 앞쪽의 낮은 바닥 방향을 향합니다. 특정 소품을 주시한다기보다 고개를 떨군 모습입니다. 눈 자체는 어둡고 눈 아래와 관자 부근에 작은 푸른 불빛이 남아 있어 A보다 잔광으로 읽히지만, 정지 화면만으로 빛이 감소하는 과정까지 확정할 수는 없습니다.",
        "built_space": "뒤쪽에 나무 탁자 하나, 왼쪽에 분명한 스툴 하나와 책장 하나가 보이고, 책장 아래 옆에는 작은 직육면체 생활용품이 놓여 있습니다. 벽에는 작은 사진들과 종이, 일부 가려진 액자가 있으며 오른쪽에는 밝은 커튼이 보입니다. 참고의 밝은 벽·목재 바닥·생활 가구와 낮의 자연광이 이어집니다. 찰리의 상반신이 화면을 채우고 벽면이 어깨 뒤에 남아 있어 지정된 미디엄 숏에 더 가깝습니다. 불가능한 반사나 명백한 설비 중복은 보이지 않습니다.",
        "entities": "찰리 하나만 등장하며, 각진 기계 머리 외장과 흰 얼굴판, 작은 원형 눈 두 개, 짧은 수평선 입이 명시된 디자인에 부합합니다. 더러운 갈색 코트와 긁히고 벗겨진 가슴판도 참고와 이어집니다. 머리가 노출되어 있지만 모자는 없습니다. 배경 바닥에는 깨진 도자기 조각들이 보이나 일부를 숨기려고 밀어놓았는지는 분명하지 않습니다. 가슴의 UBIK 글자는 크게 읽혀 최종 무문자 조건에 어긋납니다.",
        "hard_violations": [
         "가슴의 UBIK 로고가 크게 읽혀, 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
        ],
        "physics": "상체가 앞으로 기울고 두 팔이 아래로 내려와 있으며 머리는 목 기계부에 연결되어 있습니다. 손과 발의 접촉점은 미디엄 숏 바깥이므로 지지 방식을 직접 확인할 수 없지만, 몸이 공중에 떠 있다는 증거도 없습니다. 코트의 무게와 주름은 자연스럽고 배경 파편은 바닥에 놓여 있습니다. 어깨는 여전히 크고 둥글게 솟은 실루엣이어서 완전히 축 내려앉은 느낌은 제한적이지만, 숙인 머리와 내려온 팔은 시무룩한 자세를 전달합니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.321,
    "B": 1.5
   },
   "violations": {
    "B": [
     "[gpt-high] 가슴의 UBIK 로고가 명확하게 읽혀, 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
    ],
    "A": [
     "[gpt-high] 가슴의 UBIK 로고가 크게 읽혀, 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1500,
   "A": 1321
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "프레이밍이 약간 넓게 잡혔으나, 프롬프트가 명시한 고릴라형 기계 머리의 형태와 배경의 컨테이너 내벽 묘사를 충실히 구현하여 프롬프트 부합도가 높습니다.  ★위반: [gpt-high] 가슴의 UBIK 로고가 명확하게 읽혀, 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
   },
   {
    "label": "A",
    "score": 1321,
    "verdict_ko": "미디엄 샷의 구도는 준수하나, 핵심 지시사항인 고릴라형 머리 디자인과 배경의 컨테이너 내벽 묘사를 완전히 누락하여 아쉽습니다.  ★위반: [gpt-high] 가슴의 UBIK 로고가 크게 읽혀, 읽을 수 있는 글자와 로고를 금지한 조건을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S14sh2_sel.png",
    "asset_id": "1ed31a43-9445-43a2-8b0b-d202a7584840",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd8a-fbd4-75b6-9182-c1f86666a195",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S14sh2"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S14sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:22:39.748567+00:00",
  "fingerprint": "0a55c8ee0713ae83897a1f54d075eab9f9c36e8d2733ac7c68726e1b2ceda18d",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S14sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S14sh5_sel.png",
  "source_sha256": "c2d6c91928f53f0bb6f86c78be62c282417afc0c09086ac4c4fec6b5b25c41bd",
  "file": "S14sh5_cine.png",
  "staged_sha256": "ef1c4943c7d05aebd079ee61f69f50b6607bd9110e38100566d87531107ba1e8",
  "latency_ms": 14675
 },
 "S14sh8::signage": {
  "fp": "d984b200cd196819",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S14sh8": {
  "input_fingerprint": "a02fed99fc87d7c7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 험상궂은 표정으로 현관문 쪽을 가리키며 입을 크게 벌려 소리치는 찰나.\n\nLOCATION (lock): In the common living area of the family container home, facing its entrance door in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the middle-right of the frame, midground, points to entrance doorway; entrance doorway in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Entrance doorway (Indicated by 이현우 as the destination; door position unspecified) — The interior-facing doorway is visible obliquely in the left background; used as Anchors the pointing gesture independently of his gaze toward 찰리.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established daytime interior ambience, shaping readable facial contrast without a new source or color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken vase fragments have not been cleaned up; the pieces pushed aside remain displaced. Charlie retains his worn coat and hat, dirty metal exterior, blue-lit eyes, and peeling chest logo. 이현우: He has stood up inside the container. His facial injuries and dog-bitten leg persist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 험상궂은 표정으로 현관문 쪽을 가리키며 입을 크게 벌려 소리치는 찰나.\n\nLOCATION (lock): In the common living area of the family container home, facing its entrance door in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the middle-right of the frame, midground, points to entrance doorway; entrance doorway in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Entrance doorway (Indicated by 이현우 as the destination; door position unspecified) — The interior-facing doorway is visible obliquely in the left background; used as Anchors the pointing gesture independently of his gaze toward 찰리.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established daytime interior ambience, shaping readable facial contrast without a new source or color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken vase fragments have not been cleaned up; the pieces pushed aside remain displaced. Charlie retains his worn coat and hat, dirty metal exterior, blue-lit eyes, and peeling chest logo. 이현우: He has stood up inside the container. His facial injuries and dog-bitten leg persist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 험상궂은 표정으로 현관문 쪽을 가리키며 입을 크게 벌려 소리치는 찰나.\n\nLOCATION (lock): In the common living area of the family container home, facing its entrance door in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the middle-right of the frame, midground, points to entrance doorway; entrance doorway in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Entrance doorway (Indicated by 이현우 as the destination; door position unspecified) — The interior-facing doorway is visible obliquely in the left background; used as Anchors the pointing gesture independently of his gaze toward 찰리.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established daytime interior ambience, shaping readable facial contrast without a new source or color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken vase fragments have not been cleaned up; the pieces pushed aside remain displaced. Charlie retains his worn coat and hat, dirty metal exterior, blue-lit eyes, and peeling chest logo. 이현우: He has stood up inside the container. His facial injuries and dog-bitten leg persist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선과 뻗은 오른손이 화면 왼쪽 배경의 현관문을 정확히 향하고 있음.",
    "built_space": "참조 이미지와 동일한 목재 바닥, 왼쪽의 낮은 선반과 노란색 바구니, 높은 책장, 탁자와 머그컵, 우측의 녹색 페인트가 묻은 철판 벽면이 모두 올바르게 배치됨. 이현우는 화면 중앙 우측에 위치함.",
    "entities": "이현우의 얼굴과 복장(낡은 후드, 찢어진 청바지)이 참조와 일치하며, 왼쪽 무릎에 핏자국이 있는 붕대로 다리 부상을 표현함. 왼쪽 전경에 찰리의 낡은 코트 어깨 부분이 보임.",
    "hard_violations": [
     "[gpt-high] 이현우 외 인물을 보여서는 안 되는데 왼쪽 전경에 다른 인물의 외투 차림 어깨와 몸통 일부가 등장한다."
    ],
    "physics": "이현우는 바닥에 안정적으로 서 있고, 뻗은 팔은 몸통에 자연스럽게 연결됨. 깨진 화분 조각들이 바닥에 정상적으로 흩어져 있음."
   },
   {
    "label": "B",
    "direction": "이현우의 시선과 뻗은 오른손이 화면 왼쪽 배경의 현관문을 향하고 있음.",
    "built_space": "참조 이미지의 공간 구조(노란 바구니가 있는 낮은 선반, 현관문, 탁자, 우측 철판 벽)가 완벽히 재현됨. 이현우는 화면 중앙 우측에 서 있음.",
    "entities": "이현우의 얼굴과 복장은 참조와 일치하나, 다리 부상(개에게 물린 자국)은 단지 찢어진 바지로만 나타남. 화면 왼쪽 전경에 찰리의 코트와 파란색 빛이 나는 금속 머리 일부가 존재함.",
    "hard_violations": [
     "[gpt-high] 이현우만 보여야 하는 숏에 찰리의 금속 머리 일부와 외투 차림 몸통이 왼쪽 전경을 크게 차지한다."
    ],
    "physics": "이현우는 두 발로 바닥을 딛고 자연스럽게 서 있으며, 화분 파편들은 바닥에 안정적으로 놓여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지문에 명시된 '험상궂은 표정'과 '크게 벌려 소리치는' 찰나의 액션을 가장 역동적으로 구현했으며, 개에게 물린 다리의 부상 상태도 정확하게 반영했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "공간 구조와 인물의 지시 동작은 잘 표현되었으나, 소리치는 표정의 강렬함이 다소 부족하고 다리의 상처가 명확히 드러나지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 뻗은 오른손이 화면 왼쪽 배경의 현관문을 정확히 향하고 있음.",
        "built_space": "참조 이미지와 동일한 목재 바닥, 왼쪽의 낮은 선반과 노란색 바구니, 높은 책장, 탁자와 머그컵, 우측의 녹색 페인트가 묻은 철판 벽면이 모두 올바르게 배치됨. 이현우는 화면 중앙 우측에 위치함.",
        "entities": "이현우의 얼굴과 복장(낡은 후드, 찢어진 청바지)이 참조와 일치하며, 왼쪽 무릎에 핏자국이 있는 붕대로 다리 부상을 표현함. 왼쪽 전경에 찰리의 낡은 코트 어깨 부분이 보임.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 안정적으로 서 있고, 뻗은 팔은 몸통에 자연스럽게 연결됨. 깨진 화분 조각들이 바닥에 정상적으로 흩어져 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선과 뻗은 오른손이 화면 왼쪽 배경의 현관문을 향하고 있음.",
        "built_space": "참조 이미지의 공간 구조(노란 바구니가 있는 낮은 선반, 현관문, 탁자, 우측 철판 벽)가 완벽히 재현됨. 이현우는 화면 중앙 우측에 서 있음.",
        "entities": "이현우의 얼굴과 복장은 참조와 일치하나, 다리 부상(개에게 물린 자국)은 단지 찢어진 바지로만 나타남. 화면 왼쪽 전경에 찰리의 코트와 파란색 빛이 나는 금속 머리 일부가 존재함.",
        "hard_violations": [],
        "physics": "이현우는 두 발로 바닥을 딛고 자연스럽게 서 있으며, 화분 파편들은 바닥에 안정적으로 놓여 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지문에 명시된 '험상궂은 표정'과 '크게 벌려 소리치는' 찰나의 액션을 가장 역동적으로 구현했으며, 개에게 물린 다리의 부상 상태도 정확하게 반영했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "공간 구조와 인물의 지시 동작은 잘 표현되었으나, 소리치는 표정의 강렬함이 다소 부족하고 다리의 상처가 명확히 드러나지 않습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 뻗은 오른손이 화면 왼쪽 배경의 현관문을 정확히 향하고 있음.",
        "built_space": "참조 이미지와 동일한 목재 바닥, 왼쪽의 낮은 선반과 노란색 바구니, 높은 책장, 탁자와 머그컵, 우측의 녹색 페인트가 묻은 철판 벽면이 모두 올바르게 배치됨. 이현우는 화면 중앙 우측에 위치함.",
        "entities": "이현우의 얼굴과 복장(낡은 후드, 찢어진 청바지)이 참조와 일치하며, 왼쪽 무릎에 핏자국이 있는 붕대로 다리 부상을 표현함. 왼쪽 전경에 찰리의 낡은 코트 어깨 부분이 보임.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 안정적으로 서 있고, 뻗은 팔은 몸통에 자연스럽게 연결됨. 깨진 화분 조각들이 바닥에 정상적으로 흩어져 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선과 뻗은 오른손이 화면 왼쪽 배경의 현관문을 향하고 있음.",
        "built_space": "참조 이미지의 공간 구조(노란 바구니가 있는 낮은 선반, 현관문, 탁자, 우측 철판 벽)가 완벽히 재현됨. 이현우는 화면 중앙 우측에 서 있음.",
        "entities": "이현우의 얼굴과 복장은 참조와 일치하나, 다리 부상(개에게 물린 자국)은 단지 찢어진 바지로만 나타남. 화면 왼쪽 전경에 찰리의 코트와 파란색 빛이 나는 금속 머리 일부가 존재함.",
        "hard_violations": [],
        "physics": "이현우는 두 발로 바닥을 딛고 자연스럽게 서 있으며, 화분 파편들은 바닥에 안정적으로 놓여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "금지된 찰리의 신체가 등장해 부적합하지만, B보다 인물을 크게 잡아 요구된 미디엄 숏에 가깝고 오른쪽 인물·왼쪽 출입구 배치를 유지한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "금지된 인물의 외투 차림 신체가 전경에 등장하며, 무릎 아래까지 넓힌 구도가 요구된 미디엄 숏에서 A보다 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 눈과 얼굴은 왼쪽 전경의 찰리를 향한다. 뻗은 검지는 왼쪽 출입구 개구부 앞에 겹치지만 왼쪽 전경 인물 방향으로 이어져, 시선과 독립적으로 뒤쪽 문을 지목한다는 깊이 관계는 명확하지 않다.",
        "built_space": "왼쪽 뒤에 목재 문틀의 출입구 하나, 그 너머 밝은 창과 낮은 선반 하나가 보인다. 실내에는 왼쪽 낮은 수납장 하나, 뒤쪽 높은 책장 하나와 휴지통 하나, 오른쪽 식탁 하나와 스툴 두 개가 보인다. 목재 바닥과 밝은 벽, 왼쪽 자연광은 참조와 대체로 이어진다. 이현우는 중앙 오른쪽에 서 있지만, 화면은 허벅지까지 포함해 지정된 미디엄 숏보다 넓다.",
        "entities": "이현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 참조의 얼굴·체격과 대체로 부합한다. 해진 짙은 후드와 청바지, 얼굴의 찰과상, 크게 벌린 입과 찡그린 표정이 보인다. 다리 부상 부위는 프레임 밖이다. 바닥 가장자리에는 흰 도자기 파편이 남아 있다. 왼쪽 전경에 금속 머리 일부와 낡은 외투를 입은 찰리의 몸이 추가로 등장한다.",
        "hard_violations": [
         "이현우만 보여야 하는 숏에 찰리의 금속 머리 일부와 외투 차림 몸통이 왼쪽 전경을 크게 차지한다."
        ],
        "physics": "이현우의 몸통은 하체 위에 이어지고, 상체를 앞으로 기울인 채 어깨와 팔꿈치로 지시하는 팔을 지탱한다. 발은 화면 밖이지만 부유하거나 지지 없이 매달린 징후는 없다. 전경 찰리도 화면 밖으로 이어지는 몸통 일부이며 떠 있는 물체는 아니다. 도자기 파편은 바닥에 놓이고 식기와 병은 수납장 상판에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 화면 왼쪽 전경의 외투 입은 상대를 노려본다. 검지는 왼쪽으로 뻗어 뒤쪽 출입구 상부 높이와 대략 정렬되지만, 상대가 있는 전경 방향도 함께 가리켜 문 자체를 목적지로 지시하는 동작은 모호하다.",
        "built_space": "왼쪽 뒤에 출입구 하나와 그 너머 창·선반이 있고, 왼쪽 실내 수납장 하나, 뒤쪽 책장 하나와 휴지통 하나, 식탁 하나와 보이는 스툴 두 개가 배치되어 있다. 오른쪽 낡은 세로무늬 벽과 목재 바닥은 장소 참조에 가깝다. 이현우는 중앙 오른쪽에 있으나 무릎 아래까지 드러나며 바닥 면적도 커서 미디엄 숏보다 넓다.",
        "entities": "젊은 동아시아계 남성의 검은 헝클어진 머리, 얼굴 윤곽, 해진 짙은 후드와 청바지는 이현우 참조에 대체로 부합한다. 얼굴 상처와 험상궂게 입을 크게 벌린 표정이 보이고, 찢어진 청바지 무릎에는 붕대가 있다. 흰 도자기 파편 여러 개가 바닥에 남아 있다. 왼쪽 전경의 흐릿한 갈색 외투 차림 신체는 이현우의 의상과 다르며 별도 인물이다.",
        "hard_violations": [
         "이현우 외 인물을 보여서는 안 되는데 왼쪽 전경에 다른 인물의 외투 차림 어깨와 몸통 일부가 등장한다."
        ],
        "physics": "벌린 두 다리 위로 몸통이 이어지며 앞으로 기울여 소리치는 자세는 가능한 동작이다. 발은 잘렸지만 하체는 바닥 방향으로 자연스럽게 이어진다. 뻗은 팔과 내려놓은 손의 연결도 가능하며, 지지 없이 뜬 신체나 물체는 보이지 않는다. 도자기 조각은 바닥, 화병과 식기는 수납장, 컵은 식탁이 지탱한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "금지된 찰리의 신체가 등장해 부적합하지만, B보다 인물을 크게 잡아 요구된 미디엄 숏에 가깝고 오른쪽 인물·왼쪽 출입구 배치를 유지한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "금지된 인물의 외투 차림 신체가 전경에 등장하며, 무릎 아래까지 넓힌 구도가 요구된 미디엄 숏에서 A보다 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 눈과 얼굴은 왼쪽 전경의 찰리를 향한다. 뻗은 검지는 왼쪽 출입구 개구부 앞에 겹치지만 왼쪽 전경 인물 방향으로 이어져, 시선과 독립적으로 뒤쪽 문을 지목한다는 깊이 관계는 명확하지 않다.",
        "built_space": "왼쪽 뒤에 목재 문틀의 출입구 하나, 그 너머 밝은 창과 낮은 선반 하나가 보인다. 실내에는 왼쪽 낮은 수납장 하나, 뒤쪽 높은 책장 하나와 휴지통 하나, 오른쪽 식탁 하나와 스툴 두 개가 보인다. 목재 바닥과 밝은 벽, 왼쪽 자연광은 참조와 대체로 이어진다. 이현우는 중앙 오른쪽에 서 있지만, 화면은 허벅지까지 포함해 지정된 미디엄 숏보다 넓다.",
        "entities": "이현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 참조의 얼굴·체격과 대체로 부합한다. 해진 짙은 후드와 청바지, 얼굴의 찰과상, 크게 벌린 입과 찡그린 표정이 보인다. 다리 부상 부위는 프레임 밖이다. 바닥 가장자리에는 흰 도자기 파편이 남아 있다. 왼쪽 전경에 금속 머리 일부와 낡은 외투를 입은 찰리의 몸이 추가로 등장한다.",
        "hard_violations": [
         "이현우만 보여야 하는 숏에 찰리의 금속 머리 일부와 외투 차림 몸통이 왼쪽 전경을 크게 차지한다."
        ],
        "physics": "이현우의 몸통은 하체 위에 이어지고, 상체를 앞으로 기울인 채 어깨와 팔꿈치로 지시하는 팔을 지탱한다. 발은 화면 밖이지만 부유하거나 지지 없이 매달린 징후는 없다. 전경 찰리도 화면 밖으로 이어지는 몸통 일부이며 떠 있는 물체는 아니다. 도자기 파편은 바닥에 놓이고 식기와 병은 수납장 상판에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "이현우는 화면 왼쪽 전경의 외투 입은 상대를 노려본다. 검지는 왼쪽으로 뻗어 뒤쪽 출입구 상부 높이와 대략 정렬되지만, 상대가 있는 전경 방향도 함께 가리켜 문 자체를 목적지로 지시하는 동작은 모호하다.",
        "built_space": "왼쪽 뒤에 출입구 하나와 그 너머 창·선반이 있고, 왼쪽 실내 수납장 하나, 뒤쪽 책장 하나와 휴지통 하나, 식탁 하나와 보이는 스툴 두 개가 배치되어 있다. 오른쪽 낡은 세로무늬 벽과 목재 바닥은 장소 참조에 가깝다. 이현우는 중앙 오른쪽에 있으나 무릎 아래까지 드러나며 바닥 면적도 커서 미디엄 숏보다 넓다.",
        "entities": "젊은 동아시아계 남성의 검은 헝클어진 머리, 얼굴 윤곽, 해진 짙은 후드와 청바지는 이현우 참조에 대체로 부합한다. 얼굴 상처와 험상궂게 입을 크게 벌린 표정이 보이고, 찢어진 청바지 무릎에는 붕대가 있다. 흰 도자기 파편 여러 개가 바닥에 남아 있다. 왼쪽 전경의 흐릿한 갈색 외투 차림 신체는 이현우의 의상과 다르며 별도 인물이다.",
        "hard_violations": [
         "이현우 외 인물을 보여서는 안 되는데 왼쪽 전경에 다른 인물의 외투 차림 어깨와 몸통 일부가 등장한다."
        ],
        "physics": "벌린 두 다리 위로 몸통이 이어지며 앞으로 기울여 소리치는 자세는 가능한 동작이다. 발은 잘렸지만 하체는 바닥 방향으로 자연스럽게 이어진다. 뻗은 팔과 내려놓은 손의 연결도 가능하며, 지지 없이 뜬 신체나 물체는 보이지 않는다. 도자기 조각은 바닥, 화병과 식기는 수납장, 컵은 식탁이 지탱한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.5
   },
   "violations": {
    "B": [
     "[gpt-high] 이현우만 보여야 하는 숏에 찰리의 금속 머리 일부와 외투 차림 몸통이 왼쪽 전경을 크게 차지한다."
    ],
    "A": [
     "[gpt-high] 이현우 외 인물을 보여서는 안 되는데 왼쪽 전경에 다른 인물의 외투 차림 어깨와 몸통 일부가 등장한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1417,
   "B": 1500
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "지문에 명시된 '험상궂은 표정'과 '크게 벌려 소리치는' 찰나의 액션을 가장 역동적으로 구현했으며, 개에게 물린 다리의 부상 상태도 정확하게 반영했습니다.  ★위반: [gpt-high] 이현우 외 인물을 보여서는 안 되는데 왼쪽 전경에 다른 인물의 외투 차림 어깨와 몸통 일부가 등장한다."
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "공간 구조와 인물의 지시 동작은 잘 표현되었으나, 소리치는 표정의 강렬함이 다소 부족하고 다리의 상처가 명확히 드러나지 않습니다.  ★위반: [gpt-high] 이현우만 보여야 하는 숏에 찰리의 금속 머리 일부와 외투 차림 몸통이 왼쪽 전경을 크게 차지한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S14sh5_sel.png",
    "asset_id": "b6be8250-0627-4681-8004-70212dc0e5ab",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd8b-ebf2-7990-b142-7cc779e160ed",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S14sh5"
  }
 },
 "S14sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:22:54.756738+00:00",
  "fingerprint": "410659b7d290ca7e2715382525e8def839bb3c37204c6bfda798852abde9c816",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S14sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S14sh8_sel.png",
  "source_sha256": "6cf51a5ee8b7485ac7b9255ebdff5b8e37276527bf4806bb82d7a3683782abd7",
  "file": "S14sh8_cine.png",
  "staged_sha256": "c9f7c7e0fa9f797e4952b4e1fd38aacb16a0c62ad58fd5e23d12ea3ac6b56537",
  "latency_ms": 11657
 },
 "S15sh7::signage": {
  "fp": "3d809294f938fbed",
  "inscriptions": [
   {
    "text_native": "5억 원 현상금",
    "source": "scene_text_quoted",
    "reason_ko": "모니터 화면에 붉은 글씨로 표시되는 뉴스 기사의 문구가 명시되어 있다.",
    "source_quote": "5억 원 현상금"
   }
  ],
  "cues": [],
  "dropped": []
 },
 "S15sh7": {
  "input_fingerprint": "76201765584184d1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 오래된 모니터 화면 위로 '5억 원 현상금'이라는 붉은 글씨의 뉴스 기사가 밝게 떠오른 클로즈업.\n\nLOCATION (lock): At the computer workstation inside a refugee-settlement junk shop, where a bright monitor displays a news article amid old electrical goods. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: monitor display with readable red reward text in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Old computer monitor (Displaying the news article about 찰리 and the reward) — The content-bearing display faces the camera obliquely, with '5억 원 현상금' readable in red and the bezel visible; used as Contains the mediated news reveal and preserves the distinction between article content and physical shop space; Piled appliances and scrap (Accumulated throughout the shop); used as Softly visible around the monitor as environmental context and scale reference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The bright display and its red reward lettering stand out against restrained daytime shop ambience without turning the surrounding shop red.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop is crowded with appliances and scrap; its computer browser displays articles identifying Charlie as a unique UBIK model missing near Haenam and advertising a 500-million-won reward. Charlie remains dirty and weathered, with the worn coat and hat and peeling UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"5억 원 현상금\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 오래된 모니터 화면 위로 '5억 원 현상금'이라는 붉은 글씨의 뉴스 기사가 밝게 떠오른 클로즈업.\n\nLOCATION (lock): At the computer workstation inside a refugee-settlement junk shop, where a bright monitor displays a news article amid old electrical goods. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: monitor display with readable red reward text in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Old computer monitor (Displaying the news article about 찰리 and the reward) — The content-bearing display faces the camera obliquely, with '5억 원 현상금' readable in red and the bezel visible; used as Contains the mediated news reveal and preserves the distinction between article content and physical shop space; Piled appliances and scrap (Accumulated throughout the shop); used as Softly visible around the monitor as environmental context and scale reference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The bright display and its red reward lettering stand out against restrained daytime shop ambience without turning the surrounding shop red.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop is crowded with appliances and scrap; its computer browser displays articles identifying Charlie as a unique UBIK model missing near Haenam and advertising a 500-million-won reward. Charlie remains dirty and weathered, with the worn coat and hat and peeling UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"5억 원 현상금\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 오래된 모니터 화면 위로 '5억 원 현상금'이라는 붉은 글씨의 뉴스 기사가 밝게 떠오른 클로즈업.\n\nLOCATION (lock): At the computer workstation inside a refugee-settlement junk shop, where a bright monitor displays a news article amid old electrical goods. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: monitor display with readable red reward text in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Old computer monitor (Displaying the news article about 찰리 and the reward) — The content-bearing display faces the camera obliquely, with '5억 원 현상금' readable in red and the bezel visible; used as Contains the mediated news reveal and preserves the distinction between article content and physical shop space; Piled appliances and scrap (Accumulated throughout the shop); used as Softly visible around the monitor as environmental context and scale reference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The bright display and its red reward lettering stand out against restrained daytime shop ambience without turning the surrounding shop red.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop is crowded with appliances and scrap; its computer browser displays articles identifying Charlie as a unique UBIK model missing near Haenam and advertising a 500-million-won reward. Charlie remains dirty and weathered, with the worn coat and hat and peeling UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"5억 원 현상금\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 우측에서 모니터를 비스듬히 향하며, '5억 원 현상금' 문구가 화면 중앙에 위치함.",
    "built_space": "고물상 내부 책상. 모니터, 키보드 일부, 나사 그릇이 있으며 배경에 전자제품이 쌓여 있음.",
    "entities": "붉은색 글씨가 있는 검은색 구형 모니터. 지시되지 않은 영어 및 한국어 텍스트가 화면에 선명하게 나타남.",
    "hard_violations": [
     "[gemini-pro] 허용된 문구 외에 읽을 수 있는 임의의 텍스트를 화면에 생성함",
     "[gpt-high] 유일하게 허용된 '5억 원 현상금' 외에 읽을 수 있는 영문 기사, 한글 메뉴, 방송사명과 기기 로고가 노출되어 명시적인 추가 문자 금지를 위반한다."
    ],
    "physics": "모든 물체가 책상 표면에 안정적으로 놓여 있음."
   },
   {
    "label": "B",
    "direction": "카메라가 모니터를 거의 정면에서 바라봄. '5억 원 현상금' 문구가 중앙에 있음.",
    "built_space": "고물상 내부. 키보드, 마우스, 펜꽂이 등 레퍼런스 사진과 동일한 사물 배치를 보임.",
    "entities": "붉은 글씨가 있는 검은색 모니터. 레퍼런스의 한국어 기사 내용을 그대로 복사해 선명하게 표시함.",
    "hard_violations": [
     "[gemini-pro] 허용된 문구 외에 읽을 수 있는 다른 텍스트를 화면에 표시함",
     "[gemini-pro] 레퍼런스 이미지의 카메라 프레이밍을 복사하지 말라는 지시를 위반함",
     "[gpt-high] 유일하게 허용된 '5억 원 현상금' 외에 방송사명, 기사 제목·본문, 메뉴, 기기 로고 등 다수의 읽을 수 있는 문자가 노출되어 명시적인 추가 문자 금지를 위반한다."
    ],
    "physics": "책상 위 물체들이 표면에 정상적으로 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지정된 문구 외의 텍스트가 선명하게 렌더링된 오류가 있으나, 요구된 '클로즈업' 및 '비스듬한 각도' 프레이밍 지시를 비교적 잘 따랐습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "레퍼런스 이미지의 카메라 프레이밍과 화면 텍스트를 그대로 복사하여 지시사항을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 우측에서 모니터를 비스듬히 향하며, '5억 원 현상금' 문구가 화면 중앙에 위치함.",
        "built_space": "고물상 내부 책상. 모니터, 키보드 일부, 나사 그릇이 있으며 배경에 전자제품이 쌓여 있음.",
        "entities": "붉은색 글씨가 있는 검은색 구형 모니터. 지시되지 않은 영어 및 한국어 텍스트가 화면에 선명하게 나타남.",
        "hard_violations": [
         "허용된 문구 외에 읽을 수 있는 임의의 텍스트를 화면에 생성함"
        ],
        "physics": "모든 물체가 책상 표면에 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 모니터를 거의 정면에서 바라봄. '5억 원 현상금' 문구가 중앙에 있음.",
        "built_space": "고물상 내부. 키보드, 마우스, 펜꽂이 등 레퍼런스 사진과 동일한 사물 배치를 보임.",
        "entities": "붉은 글씨가 있는 검은색 모니터. 레퍼런스의 한국어 기사 내용을 그대로 복사해 선명하게 표시함.",
        "hard_violations": [
         "허용된 문구 외에 읽을 수 있는 다른 텍스트를 화면에 표시함",
         "레퍼런스 이미지의 카메라 프레이밍을 복사하지 말라는 지시를 위반함"
        ],
        "physics": "책상 위 물체들이 표면에 정상적으로 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지정된 문구 외의 텍스트가 선명하게 렌더링된 오류가 있으나, 요구된 '클로즈업' 및 '비스듬한 각도' 프레이밍 지시를 비교적 잘 따랐습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "레퍼런스 이미지의 카메라 프레이밍과 화면 텍스트를 그대로 복사하여 지시사항을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 우측에서 모니터를 비스듬히 향하며, '5억 원 현상금' 문구가 화면 중앙에 위치함.",
        "built_space": "고물상 내부 책상. 모니터, 키보드 일부, 나사 그릇이 있으며 배경에 전자제품이 쌓여 있음.",
        "entities": "붉은색 글씨가 있는 검은색 구형 모니터. 지시되지 않은 영어 및 한국어 텍스트가 화면에 선명하게 나타남.",
        "hard_violations": [
         "허용된 문구 외에 읽을 수 있는 임의의 텍스트를 화면에 생성함"
        ],
        "physics": "모든 물체가 책상 표면에 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 모니터를 거의 정면에서 바라봄. '5억 원 현상금' 문구가 중앙에 있음.",
        "built_space": "고물상 내부. 키보드, 마우스, 펜꽂이 등 레퍼런스 사진과 동일한 사물 배치를 보임.",
        "entities": "붉은 글씨가 있는 검은색 모니터. 레퍼런스의 한국어 기사 내용을 그대로 복사해 선명하게 표시함.",
        "hard_violations": [
         "허용된 문구 외에 읽을 수 있는 다른 텍스트를 화면에 표시함",
         "레퍼런스 이미지의 카메라 프레이밍을 복사하지 말라는 지시를 위반함"
        ],
        "physics": "책상 위 물체들이 표면에 정상적으로 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "허용되지 않은 기사·로고 문구가 선명하게 노출되며, 화면 중심의 클로즈업보다 작업대와 주변을 넓게 보여 주고 현상금 문구도 중앙에서 왼쪽 위로 벗어난다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "비스듬한 모니터 클로즈업과 중앙의 붉은 현상금 문구는 더 정확하지만, 추가로 읽히는 영문 기사·메뉴·로고 때문에 필수 문자 제한을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "모니터의 표시 면이 카메라를 비스듬히 향해 붉은 현상금 문구를 읽을 수 있다. 실제 매장 안 인물이나 시선·무기·이동 동작은 보이지 않는다.",
        "built_space": "주 모니터 한 대가 작업대 위에 있고, 아래에는 키보드 한 개, 오른쪽에는 마우스 한 개가 보인다. 모니터 오른쪽의 공구통 한 개, 흰 고리형 부품 한 개, 금속 부품 그릇 한 개는 참조 작업대의 구성을 따른다. 뒤쪽 선반과 가전 적재물, 낮빛이 들어오는 개구부가 보이지만 참조보다 오른쪽 냉장고와 적재물이 크게 드러난다. 화면뿐 아니라 받침대와 작업대까지 넓게 담았다.",
        "entities": "낡고 먼지 낀 검은 모니터와 가전·고철 매장의 재질은 참조에 부합한다. 붉은 '5억 원 현상금'은 표시되지만, 기사 내용은 찰리의 해남 실종보다 폐전자제품 불법 유출 조직 수배를 다룬다. 방송사명, 기사 제목·본문, 메뉴와 기기 로고 등 지정되지 않은 글자도 명확히 읽힌다. 매장 안의 실제 인물은 없다.",
        "hard_violations": [
         "유일하게 허용된 '5억 원 현상금' 외에 방송사명, 기사 제목·본문, 메뉴, 기기 로고 등 다수의 읽을 수 있는 문자가 노출되어 명시적인 추가 문자 금지를 위반한다."
        ],
        "physics": "모니터는 목과 넓은 받침대로 작업대에 지지된다. 키보드·마우스·그릇은 작업대에 놓이고 공구는 통 안에 꽂혀 있다. 뒤쪽 가전은 선반이나 다른 적재물에 받쳐져 있으며, 지지 없이 떠 있는 물체나 신체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "모니터 표시 면이 카메라를 뚜렷하게 비스듬히 향한다. 붉은 '5억 원 현상금'은 화면 중앙 부근에서 읽히며, 표시 면과 베젤의 원근 방향이 일치한다. 매장 안 인물이나 방향을 검증할 이동 동작은 없다.",
        "built_space": "주 모니터 한 대가 프레임 대부분을 차지하고 상단과 받침대 일부가 잘려 클로즈업에 가깝다. 오른쪽 뒤에는 공구통 한 개, 흰 고리형 부품 한 개, 금속 부품 그릇 한 개가 있고, 양쪽 주변에는 선반과 낡은 가전이 흐리게 보인다. 참조 작업대의 주요 소품 관계를 유지하며 배경을 제한한다. 불가능한 반사는 보이지 않는다.",
        "entities": "낡은 검은 모니터, 밝은 기사 화면, 붉은 현상금 문구와 주변 폐가전이 요청된 대상에 해당한다. 기사에 찰리·유비크 모델·해남을 언급하는 영문이 보이지만, 고유 모델 여부를 부정하는 문장도 있어 내용이 일관되지 않다. 영문 본문과 한글 메뉴, 방송사명 및 기기 로고가 추가로 읽힌다. 실제 매장 인물이나 찰리의 실물은 없다.",
        "hard_violations": [
         "유일하게 허용된 '5억 원 현상금' 외에 읽을 수 있는 영문 기사, 한글 메뉴, 방송사명과 기기 로고가 노출되어 명시적인 추가 문자 금지를 위반한다."
        ],
        "physics": "모니터 아래의 목과 일부 받침대가 작업대 위 지지를 보여 준다. 오른쪽 그릇과 부품은 작업대 또는 쌓인 상자에 놓여 있고, 공구는 통에 지지된다. 배경 가전도 선반과 적재물에 받쳐져 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "허용되지 않은 기사·로고 문구가 선명하게 노출되며, 화면 중심의 클로즈업보다 작업대와 주변을 넓게 보여 주고 현상금 문구도 중앙에서 왼쪽 위로 벗어난다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "비스듬한 모니터 클로즈업과 중앙의 붉은 현상금 문구는 더 정확하지만, 추가로 읽히는 영문 기사·메뉴·로고 때문에 필수 문자 제한을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "모니터의 표시 면이 카메라를 비스듬히 향해 붉은 현상금 문구를 읽을 수 있다. 실제 매장 안 인물이나 시선·무기·이동 동작은 보이지 않는다.",
        "built_space": "주 모니터 한 대가 작업대 위에 있고, 아래에는 키보드 한 개, 오른쪽에는 마우스 한 개가 보인다. 모니터 오른쪽의 공구통 한 개, 흰 고리형 부품 한 개, 금속 부품 그릇 한 개는 참조 작업대의 구성을 따른다. 뒤쪽 선반과 가전 적재물, 낮빛이 들어오는 개구부가 보이지만 참조보다 오른쪽 냉장고와 적재물이 크게 드러난다. 화면뿐 아니라 받침대와 작업대까지 넓게 담았다.",
        "entities": "낡고 먼지 낀 검은 모니터와 가전·고철 매장의 재질은 참조에 부합한다. 붉은 '5억 원 현상금'은 표시되지만, 기사 내용은 찰리의 해남 실종보다 폐전자제품 불법 유출 조직 수배를 다룬다. 방송사명, 기사 제목·본문, 메뉴와 기기 로고 등 지정되지 않은 글자도 명확히 읽힌다. 매장 안의 실제 인물은 없다.",
        "hard_violations": [
         "유일하게 허용된 '5억 원 현상금' 외에 방송사명, 기사 제목·본문, 메뉴, 기기 로고 등 다수의 읽을 수 있는 문자가 노출되어 명시적인 추가 문자 금지를 위반한다."
        ],
        "physics": "모니터는 목과 넓은 받침대로 작업대에 지지된다. 키보드·마우스·그릇은 작업대에 놓이고 공구는 통 안에 꽂혀 있다. 뒤쪽 가전은 선반이나 다른 적재물에 받쳐져 있으며, 지지 없이 떠 있는 물체나 신체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "모니터 표시 면이 카메라를 뚜렷하게 비스듬히 향한다. 붉은 '5억 원 현상금'은 화면 중앙 부근에서 읽히며, 표시 면과 베젤의 원근 방향이 일치한다. 매장 안 인물이나 방향을 검증할 이동 동작은 없다.",
        "built_space": "주 모니터 한 대가 프레임 대부분을 차지하고 상단과 받침대 일부가 잘려 클로즈업에 가깝다. 오른쪽 뒤에는 공구통 한 개, 흰 고리형 부품 한 개, 금속 부품 그릇 한 개가 있고, 양쪽 주변에는 선반과 낡은 가전이 흐리게 보인다. 참조 작업대의 주요 소품 관계를 유지하며 배경을 제한한다. 불가능한 반사는 보이지 않는다.",
        "entities": "낡은 검은 모니터, 밝은 기사 화면, 붉은 현상금 문구와 주변 폐가전이 요청된 대상에 해당한다. 기사에 찰리·유비크 모델·해남을 언급하는 영문이 보이지만, 고유 모델 여부를 부정하는 문장도 있어 내용이 일관되지 않다. 영문 본문과 한글 메뉴, 방송사명 및 기기 로고가 추가로 읽힌다. 실제 매장 인물이나 찰리의 실물은 없다.",
        "hard_violations": [
         "유일하게 허용된 '5억 원 현상금' 외에 읽을 수 있는 영문 기사, 한글 메뉴, 방송사명과 기기 로고가 노출되어 명시적인 추가 문자 금지를 위반한다."
        ],
        "physics": "모니터 아래의 목과 일부 받침대가 작업대 위 지지를 보여 준다. 오른쪽 그릇과 부품은 작업대 또는 쌓인 상자에 놓여 있고, 공구는 통에 지지된다. 배경 가전도 선반과 적재물에 받쳐져 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.167
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.917
   },
   "violations": {
    "A": [
     "[gemini-pro] 허용된 문구 외에 읽을 수 있는 임의의 텍스트를 화면에 생성함",
     "[gpt-high] 유일하게 허용된 '5억 원 현상금' 외에 읽을 수 있는 영문 기사, 한글 메뉴, 방송사명과 기기 로고가 노출되어 명시적인 추가 문자 금지를 위반한다."
    ],
    "B": [
     "[gemini-pro] 허용된 문구 외에 읽을 수 있는 다른 텍스트를 화면에 표시함",
     "[gemini-pro] 레퍼런스 이미지의 카메라 프레이밍을 복사하지 말라는 지시를 위반함",
     "[gpt-high] 유일하게 허용된 '5억 원 현상금' 외에 방송사명, 기사 제목·본문, 메뉴, 기기 로고 등 다수의 읽을 수 있는 문자가 노출되어 명시적인 추가 문자 금지를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 917
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지정된 문구 외의 텍스트가 선명하게 렌더링된 오류가 있으나, 요구된 '클로즈업' 및 '비스듬한 각도' 프레이밍 지시를 비교적 잘 따랐습니다.  ★위반: [gemini-pro] 허용된 문구 외에 읽을 수 있는 임의의 텍스트를 화면에 생성함 / [gpt-high] 유일하게 허용된 '5억 원 현상금' 외에 읽을 수 있는 영문 기사, 한글 메뉴, 방송사명과 기기 로고가 노출되어 명시적인 추가 문자 금지를 위반한다."
   },
   {
    "label": "B",
    "score": 917,
    "verdict_ko": "레퍼런스 이미지의 카메라 프레이밍과 화면 텍스트를 그대로 복사하여 지시사항을 심각하게 위반했습니다.  ★위반: [gemini-pro] 허용된 문구 외에 읽을 수 있는 다른 텍스트를 화면에 표시함 / [gemini-pro] 레퍼런스 이미지의 카메라 프레이밍을 복사하지 말라는 지시를 위반함 / [gpt-high] 유일하게 허용된 '5억 원 현상금' 외에 방송사명, 기사 제목·본문, 메뉴, 기기 로고 등 다수의 읽을 수 있는 문자가 노출되어 명시적인 추가 문자 금지를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L25B02.png",
    "asset_id": "15004adb-566d-4b90-af50-d06a3101d7a8",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd8c-abd6-7e15-9e24-13e795113fc1",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S15sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:23:06.753649+00:00",
  "fingerprint": "c94eea9986e30edfb506788f7dc9448b8f2f132481b85f35c0e709eb0894fc6b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S15sh7_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S15sh7_sel.png",
  "source_sha256": "78e9ef3eaa43fd619211680e590df94f90ba441632b4e5023667f2ebddda57d8",
  "file": "S15sh7_cine.png",
  "staged_sha256": "4cb082d542c28d4427c902b05bc536b5e0c34bd9ef25e826d94f835c4c312097",
  "latency_ms": 13510
 },
 "S15sh11::signage": {
  "fp": "f5d1788127d8a14a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::4281d07542359cc7": {
  "subjects": [],
  "subject_text": "마츠다의 고물상 내부\n온갖 낡은 가전제품과 고철이 빽빽하게 쌓인 상점. 수리용 책상 위에 전기제품이 놓여 있고 컴퓨터와 텔레비전이 보인다.",
  "identity": "canonical",
  "scope_id": "L25",
  "scope_role": "location_interior",
  "scope_sha": "06c301103f675b1b"
 },
 "S15sh11::bgfirst_bg": {
  "input_fingerprint": "77fe105ea9622b0f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 찰리의 거대한 금속 팔을 쥔 채 상점 밖을 향해 몸을 확 돌린 역동적인 구도.\n\nLOCATION (lock): In the junk shop's interior passage toward the street entrance, with daylight at the doorway and discarded appliances piled nearby.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이현우 turning toward the exit in the middle-left of the frame, midground, moves toward exit route at the left edge; 찰리 following the arm pull in the middle-right of the frame, midground, moves toward exit route at the left edge.\n- KEY BACKGROUND ELEMENTS: Shop entrance and exit route (Being used by 이현우 and 찰리 to leave) — Seen from inside the shop along the left side of the composition; used as Makes the direction of the turn and impending departure immediately readable; Piled appliances and scrap (Remaining stacked inside the shop); used as Frames the route without obstructing the hand-to-arm contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime shop illumination continuous, using stable contrast to separate the gripping hand from 찰리's metal arm and the surrounding clutter.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 찰리의 거대한 금속 팔을 쥔 채 상점 밖을 향해 몸을 확 돌린 역동적인 구도.\n\nLOCATION (lock): In the junk shop's interior passage toward the street entrance, with daylight at the doorway and discarded appliances piled nearby.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이현우 turning toward the exit in the middle-left of the frame, midground, moves toward exit route at the left edge; 찰리 following the arm pull in the middle-right of the frame, midground, moves toward exit route at the left edge.\n- KEY BACKGROUND ELEMENTS: Shop entrance and exit route (Being used by 이현우 and 찰리 to leave) — Seen from inside the shop along the left side of the composition; used as Makes the direction of the turn and impending departure immediately readable; Piled appliances and scrap (Remaining stacked inside the shop); used as Frames the route without obstructing the hand-to-arm contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime shop illumination continuous, using stable contrast to separate the gripping hand from 찰리's metal arm and the surrounding clutter.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S15sh11__bgfirst_bg.png",
  "asset_id": "28bbef6d-c5d4-4371-888b-bea3ba16935d",
  "input_asset_ids": [
   "8c4866d2-deb8-481d-b29b-99ab1897d5b9",
   "7433b9c2-c294-46c1-bd2c-18bf7fa04322"
  ]
 },
 "S15sh11": {
  "input_fingerprint": "c9e61cb0c3ecb2bf",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 거대한 금속 팔을 쥔 채 상점 밖을 향해 몸을 확 돌린 역동적인 구도.\n\nLOCATION (lock): In the junk shop's interior passage toward the street entrance, with daylight at the doorway and discarded appliances piled nearby. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이현우 turning toward the exit in the middle-left of the frame, midground, moves toward exit route at the left edge; 찰리 following the arm pull in the middle-right of the frame, midground, moves toward exit route at the left edge.\n- KEY BACKGROUND ELEMENTS: Shop entrance and exit route (Being used by 이현우 and 찰리 to leave) — Seen from inside the shop along the left side of the composition; used as Makes the direction of the turn and impending departure immediately readable; Piled appliances and scrap (Remaining stacked inside the shop); used as Frames the route without obstructing the hand-to-arm contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime shop illumination continuous, using stable contrast to separate the gripping hand from 찰리's metal arm and the surrounding clutter.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The browser still shows the UBIK disappearance and reward articles, with appliances and scrap piled throughout the shop. Charlie's dirty exterior, worn coat and hat, and peeling UBIK logo remain unchanged. 이현우: He is leaving the shop, still bearing his facial injuries and dog-bitten leg.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 거대한 금속 팔을 쥔 채 상점 밖을 향해 몸을 확 돌린 역동적인 구도.\n\nLOCATION (lock): In the junk shop's interior passage toward the street entrance, with daylight at the doorway and discarded appliances piled nearby. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이현우 turning toward the exit in the middle-left of the frame, midground, moves toward exit route at the left edge; 찰리 following the arm pull in the middle-right of the frame, midground, moves toward exit route at the left edge.\n- KEY BACKGROUND ELEMENTS: Shop entrance and exit route (Being used by 이현우 and 찰리 to leave) — Seen from inside the shop along the left side of the composition; used as Makes the direction of the turn and impending departure immediately readable; Piled appliances and scrap (Remaining stacked inside the shop); used as Frames the route without obstructing the hand-to-arm contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime shop illumination continuous, using stable contrast to separate the gripping hand from 찰리's metal arm and the surrounding clutter.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The browser still shows the UBIK disappearance and reward articles, with appliances and scrap piled throughout the shop. Charlie's dirty exterior, worn coat and hat, and peeling UBIK logo remain unchanged. 이현우: He is leaving the shop, still bearing his facial injuries and dog-bitten leg.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 거대한 금속 팔을 쥔 채 상점 밖을 향해 몸을 확 돌린 역동적인 구도.\n\nLOCATION (lock): In the junk shop's interior passage toward the street entrance, with daylight at the doorway and discarded appliances piled nearby. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이현우 turning toward the exit in the middle-left of the frame, midground, moves toward exit route at the left edge; 찰리 following the arm pull in the middle-right of the frame, midground, moves toward exit route at the left edge.\n- KEY BACKGROUND ELEMENTS: Shop entrance and exit route (Being used by 이현우 and 찰리 to leave) — Seen from inside the shop along the left side of the composition; used as Makes the direction of the turn and impending departure immediately readable; Piled appliances and scrap (Remaining stacked inside the shop); used as Frames the route without obstructing the hand-to-arm contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime shop illumination continuous, using stable contrast to separate the gripping hand from 찰리's metal arm and the surrounding clutter.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The browser still shows the UBIK disappearance and reward articles, with appliances and scrap piled throughout the shop. Charlie's dirty exterior, worn coat and hat, and peeling UBIK logo remain unchanged. 이현우: He is leaving the shop, still bearing his facial injuries and dog-bitten leg.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S15sh11__bgfirst_bg.png",
     "asset_id": "28bbef6d-c5d4-4371-888b-bea3ba16935d",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S15sh11.png",
     "asset_id": "8c4866d2-deb8-481d-b29b-99ab1897d5b9",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1597480>",
     "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L25B04.png",
     "asset_id": "7433b9c2-c294-46c1-bd2c-18bf7fa04322",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1597480>",
     "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 배경 좌측의 출구를 바라보며 이동하고, 찰리 역시 왼쪽으로 이끌려감. 이현우의 오른손은 찰리의 오른팔을 정확히 겨냥해 쥐고 있음.",
    "built_space": "고물상 내부 중앙 통로. 배경 좌측에 1개의 유리 출입문이 보이며, 양옆으로 다수의 냉장고와 폐가전이 레퍼런스와 일치하게 늘어서 있음. 두 인물의 위치와 공간 구조가 카메라 구도와 자연스럽게 어울림.",
    "entities": "이현우(18세 한국인, 까치집 머리, 후드티, 찢어진 청바지와 얼굴 상처 일치), 찰리(고릴라형 금속 머리, 흰 얼굴판, 코트, 모자, 가슴의 UBIK 로고 일치) 모두 지시사항에 부합함.",
    "hard_violations": [
     "[gpt-high] 찰리 가슴에 판독 가능한 UBIK 표기가 남아 있어 글자·로고 금지 조건을 위반합니다."
    ],
    "physics": "이현우는 왼발로 바닥을 딛고 오른발을 든 채 체중을 실어 당기는 포즈를 취함. 이현우의 오른손이 찰리의 팔을 물리적으로 단단히 지탱하며, 찰리도 두 발로 바닥을 딛고 걸어감."
   },
   {
    "label": "B",
    "direction": "이현우와 찰리 모두 좌측 출구를 향해 시선을 두고 이동함. 이현우 측의 기형적인 팔이 찰리의 오른손을 쥐고 있음.",
    "built_space": "화면 왼쪽 가장자리에 1개의 유리 출입문이 배치되고 우측에 폐가전이 쌓여 있음. 프레이밍 지시는 따랐으나 출구 위치가 레퍼런스(정면 끝)와 달리 측면으로 변경되어 공간 구조가 왜곡됨.",
    "entities": "이현우(얼굴과 의상 특징 일치), 찰리(고릴라형 머리, 코트, 중절모, 낡은 UBIK 로고 일치) 모두 외형 지시를 충족함.",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy (이현우의 오른팔이 몸통 옆으로 늘어져 있음에도 불구하고, 등 뒤에서 불가능한 구조의 또 다른 팔이 길게 뻗어 나와 찰리를 잡고 있음)",
     "[gemini-pro] leaked markers/diagrams/text (유리문에 '고물상'이라는 식별 가능한 한글 텍스트가 노출됨)",
     "[gpt-high] 찰리 가슴의 UBIK 로고, 유리문의 한글 및 벽 포스터의 판독 가능한 글자가 노출되어 글자·로고 금지 조건을 위반합니다."
    ],
    "physics": "두 인물의 발은 바닥을 지탱하고 있으나, 이현우의 몸에서 뻗어 나온 팔의 관절과 연결 부위가 해부학적으로 전혀 불가능한 형태로 공중에 꼬여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "장소의 실제 구조를 유지하며 캐릭터 특징과 역동적인 당기기 자세를 물리적 오류 없이 훌륭하게 구현했으나, 출구를 화면 좌측 가장자리에 밀착시키라는 구도 지시에는 약간 미치지 못함."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프레이밍 지시는 잘 따랐으나, 이현우의 오른팔이 아래로 내려와 있음에도 등 뒤에서 기형적인 팔이 뻗어 나오는 치명적인 해부학적 오류와 금지된 텍스트가 포함됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 배경 좌측의 출구를 바라보며 이동하고, 찰리 역시 왼쪽으로 이끌려감. 이현우의 오른손은 찰리의 오른팔을 정확히 겨냥해 쥐고 있음.",
        "built_space": "고물상 내부 중앙 통로. 배경 좌측에 1개의 유리 출입문이 보이며, 양옆으로 다수의 냉장고와 폐가전이 레퍼런스와 일치하게 늘어서 있음. 두 인물의 위치와 공간 구조가 카메라 구도와 자연스럽게 어울림.",
        "entities": "이현우(18세 한국인, 까치집 머리, 후드티, 찢어진 청바지와 얼굴 상처 일치), 찰리(고릴라형 금속 머리, 흰 얼굴판, 코트, 모자, 가슴의 UBIK 로고 일치) 모두 지시사항에 부합함.",
        "hard_violations": [],
        "physics": "이현우는 왼발로 바닥을 딛고 오른발을 든 채 체중을 실어 당기는 포즈를 취함. 이현우의 오른손이 찰리의 팔을 물리적으로 단단히 지탱하며, 찰리도 두 발로 바닥을 딛고 걸어감."
       },
       {
        "label": "B",
        "direction": "이현우와 찰리 모두 좌측 출구를 향해 시선을 두고 이동함. 이현우 측의 기형적인 팔이 찰리의 오른손을 쥐고 있음.",
        "built_space": "화면 왼쪽 가장자리에 1개의 유리 출입문이 배치되고 우측에 폐가전이 쌓여 있음. 프레이밍 지시는 따랐으나 출구 위치가 레퍼런스(정면 끝)와 달리 측면으로 변경되어 공간 구조가 왜곡됨.",
        "entities": "이현우(얼굴과 의상 특징 일치), 찰리(고릴라형 머리, 코트, 중절모, 낡은 UBIK 로고 일치) 모두 외형 지시를 충족함.",
        "hard_violations": [
         "physically impossible anatomy (이현우의 오른팔이 몸통 옆으로 늘어져 있음에도 불구하고, 등 뒤에서 불가능한 구조의 또 다른 팔이 길게 뻗어 나와 찰리를 잡고 있음)",
         "leaked markers/diagrams/text (유리문에 '고물상'이라는 식별 가능한 한글 텍스트가 노출됨)"
        ],
        "physics": "두 인물의 발은 바닥을 지탱하고 있으나, 이현우의 몸에서 뻗어 나온 팔의 관절과 연결 부위가 해부학적으로 전혀 불가능한 형태로 공중에 꼬여 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "장소의 실제 구조를 유지하며 캐릭터 특징과 역동적인 당기기 자세를 물리적 오류 없이 훌륭하게 구현했으나, 출구를 화면 좌측 가장자리에 밀착시키라는 구도 지시에는 약간 미치지 못함."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프레이밍 지시는 잘 따랐으나, 이현우의 오른팔이 아래로 내려와 있음에도 등 뒤에서 기형적인 팔이 뻗어 나오는 치명적인 해부학적 오류와 금지된 텍스트가 포함됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 배경 좌측의 출구를 바라보며 이동하고, 찰리 역시 왼쪽으로 이끌려감. 이현우의 오른손은 찰리의 오른팔을 정확히 겨냥해 쥐고 있음.",
        "built_space": "고물상 내부 중앙 통로. 배경 좌측에 1개의 유리 출입문이 보이며, 양옆으로 다수의 냉장고와 폐가전이 레퍼런스와 일치하게 늘어서 있음. 두 인물의 위치와 공간 구조가 카메라 구도와 자연스럽게 어울림.",
        "entities": "이현우(18세 한국인, 까치집 머리, 후드티, 찢어진 청바지와 얼굴 상처 일치), 찰리(고릴라형 금속 머리, 흰 얼굴판, 코트, 모자, 가슴의 UBIK 로고 일치) 모두 지시사항에 부합함.",
        "hard_violations": [],
        "physics": "이현우는 왼발로 바닥을 딛고 오른발을 든 채 체중을 실어 당기는 포즈를 취함. 이현우의 오른손이 찰리의 팔을 물리적으로 단단히 지탱하며, 찰리도 두 발로 바닥을 딛고 걸어감."
       },
       {
        "label": "B",
        "direction": "이현우와 찰리 모두 좌측 출구를 향해 시선을 두고 이동함. 이현우 측의 기형적인 팔이 찰리의 오른손을 쥐고 있음.",
        "built_space": "화면 왼쪽 가장자리에 1개의 유리 출입문이 배치되고 우측에 폐가전이 쌓여 있음. 프레이밍 지시는 따랐으나 출구 위치가 레퍼런스(정면 끝)와 달리 측면으로 변경되어 공간 구조가 왜곡됨.",
        "entities": "이현우(얼굴과 의상 특징 일치), 찰리(고릴라형 머리, 코트, 중절모, 낡은 UBIK 로고 일치) 모두 외형 지시를 충족함.",
        "hard_violations": [
         "physically impossible anatomy (이현우의 오른팔이 몸통 옆으로 늘어져 있음에도 불구하고, 등 뒤에서 불가능한 구조의 또 다른 팔이 길게 뻗어 나와 찰리를 잡고 있음)",
         "leaked markers/diagrams/text (유리문에 '고물상'이라는 식별 가능한 한글 텍스트가 노출됨)"
        ],
        "physics": "두 인물의 발은 바닥을 지탱하고 있으나, 이현우의 몸에서 뻗어 나온 팔의 관절과 연결 부위가 해부학적으로 전혀 불가능한 형태로 공중에 꼬여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "왼쪽 출구를 향한 동작은 읽히지만 이현우가 금속 팔을 붙잡기보다 찰리에게 손을 잡힌 모습이며, 선명한 글자와 로고가 금지 조건을 위반합니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "이현우의 금속 팔 파지와 찰리가 끌려오는 동작, 상점 통로가 더 충실하지만 중경보다 가까운 인물 배치와 판독 가능한 가슴 로고가 남습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 얼굴과 시선, 앞선 발은 화면 왼쪽의 밝은 출입문을 향합니다. 찰리도 왼쪽 이현우 쪽으로 머리와 뻗은 팔을 향합니다. 퇴장 방향은 분명하지만 접촉부에서는 찰리의 큰 손이 이현우의 손을 감싸는 형태여서, 이현우가 찰리의 금속 팔을 쥐고 당긴다는 핵심 관계가 뒤집혀 보입니다.",
        "built_space": "왼쪽에 유리 출입구 하나, 중앙에 별도 공간으로 이어지는 사각 개구부 하나가 보입니다. 출입구 옆에는 의자 하나, 왼쪽 아래에는 전화기 하나, 오른쪽 전경에는 상부 투입식 세탁기 하나가 있으며 여러 냉장고·화면·전자레인지가 통로 양옆에 쌓여 있습니다. 낡은 벽과 콘크리트 바닥은 비슷하지만 중앙 개구부로 공간이 나뉘는 구조는 참고의 긴 연속 통로와 다릅니다. 이현우는 왼쪽, 찰리는 오른쪽에 있으나 특히 찰리가 중경보다 전경에 크게 배치됩니다.",
        "entities": "등장 주체는 이현우와 찰리 둘뿐입니다. 이현우는 젊은 동아시아계 남성 외형, 헝클어진 검은 머리, 얼굴 상처, 낡은 어두운 후드와 청바지를 갖춰 참고와 대체로 맞습니다. 다리의 개 물림 상처는 분명히 확인되지 않습니다. 찰리는 각진 기계 머리, 흰 얼굴판, 점 눈과 선 입, 거대한 금속 팔, 더러운 코트와 모자를 갖췄습니다. 가슴의 UBIK 로고와 출입문 글자, 벽 포스터 글자가 읽힙니다. 기사 내용이 확인되는 브라우저는 보이지 않습니다.",
        "hard_violations": [
         "찰리 가슴의 UBIK 로고, 유리문의 한글 및 벽 포스터의 판독 가능한 글자가 노출되어 글자·로고 금지 조건을 위반합니다."
        ],
        "physics": "이현우는 앞쪽 신발을 바닥에 디디고 뒤쪽 발을 들어 보폭을 옮기며, 찰리도 화면 아래에 보이는 금속 발로 바닥을 지지합니다. 두 인물의 손은 실제로 접촉하고 있어 부유하지 않지만, 금속 팔을 쥐는 이현우의 손가락 동작은 확인되지 않고 오히려 찰리의 손이 그를 잡습니다. 가전은 바닥이나 다른 가전 위에 놓여 있고 공중에 떠 있는 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "이현우는 얼굴과 상체를 왼쪽으로 돌리고 화면 왼쪽 통로 쪽으로 발을 내딛습니다. 실제 출입문은 왼쪽 뒤에 있어 몸의 진행 방향이 문 정면으로 곧장 이어지지는 않지만, 왼쪽 퇴장 동선은 읽힙니다. 찰리는 이현우와 잡힌 팔 쪽으로 고개를 숙이고 왼쪽으로 따라옵니다. 이현우의 손이 찰리의 금속 손목 바로 위를 감싸 당기는 관계가 명확합니다.",
        "built_space": "왼쪽 후경에 상부 채광창을 갖춘 유리 출입구 하나가 있고, 천장에는 형광등 기구 세 개가 보입니다. 양쪽 가전 선반 사이로 콘크리트 통로가 이어지며 오른쪽에는 전면 투입식 세탁기 하나와 선풍기 하나, 여러 냉장고·전자레인지·화면이 있습니다. 문 가까이 작업대와 의자가 있고 전경 작업대 오른쪽에 전화기 하나가 놓입니다. 참고의 긴 통로와 가전 적재 구조를 A보다 잘 유지합니다. 두 인물은 좌우 배치를 지키지만 요청한 중경보다 카메라에 가깝고 크게 보입니다.",
        "entities": "이현우와 찰리 외의 인물은 없습니다. 이현우는 젊은 동아시아계 남성 외형, 헝클어진 검은 머리, 얼굴의 붉은 상처와 낡은 청바지를 갖췄습니다. 무릎의 찢어진 부분과 붉은 손상이 보이지만 개 물림인지는 확정할 수 없습니다. 참고의 후드 위에 열린 겉옷이 추가되어 복장은 완전히 같지 않습니다. 찰리의 흰 얼굴판, 각진 금속 머리, 점 눈과 선 입, 낡은 코트와 모자는 요구에 부합하지만 팔의 거대함은 A보다 약합니다. 가슴의 UBIK 표기가 읽히며, 배경 모니터에서는 기사 내용을 확인할 수 없습니다.",
        "hard_violations": [
         "찰리 가슴에 판독 가능한 UBIK 표기가 남아 있어 글자·로고 금지 조건을 위반합니다."
        ],
        "physics": "이현우는 뒤쪽 다리에 체중을 남긴 채 앞발을 내딛고 상체를 비튼 보행 순간으로 읽힙니다. 찰리는 앞쪽 발로 바닥을 지지하고 뒤쪽 발꿈치를 들어 따라오는 자세입니다. 이현우의 맨손과 손목, 드러난 팔뚝이 금속 전완으로 자연스럽게 이어지고 손가락이 이를 감싸므로 당김의 접촉과 힘 전달이 보입니다. 코트 자락의 들림도 보행에 맞으며, 지지 없이 떠 있는 인물이나 물체는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "왼쪽 출구를 향한 동작은 읽히지만 이현우가 금속 팔을 붙잡기보다 찰리에게 손을 잡힌 모습이며, 선명한 글자와 로고가 금지 조건을 위반합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "이현우의 금속 팔 파지와 찰리가 끌려오는 동작, 상점 통로가 더 충실하지만 중경보다 가까운 인물 배치와 판독 가능한 가슴 로고가 남습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 얼굴과 시선, 앞선 발은 화면 왼쪽의 밝은 출입문을 향합니다. 찰리도 왼쪽 이현우 쪽으로 머리와 뻗은 팔을 향합니다. 퇴장 방향은 분명하지만 접촉부에서는 찰리의 큰 손이 이현우의 손을 감싸는 형태여서, 이현우가 찰리의 금속 팔을 쥐고 당긴다는 핵심 관계가 뒤집혀 보입니다.",
        "built_space": "왼쪽에 유리 출입구 하나, 중앙에 별도 공간으로 이어지는 사각 개구부 하나가 보입니다. 출입구 옆에는 의자 하나, 왼쪽 아래에는 전화기 하나, 오른쪽 전경에는 상부 투입식 세탁기 하나가 있으며 여러 냉장고·화면·전자레인지가 통로 양옆에 쌓여 있습니다. 낡은 벽과 콘크리트 바닥은 비슷하지만 중앙 개구부로 공간이 나뉘는 구조는 참고의 긴 연속 통로와 다릅니다. 이현우는 왼쪽, 찰리는 오른쪽에 있으나 특히 찰리가 중경보다 전경에 크게 배치됩니다.",
        "entities": "등장 주체는 이현우와 찰리 둘뿐입니다. 이현우는 젊은 동아시아계 남성 외형, 헝클어진 검은 머리, 얼굴 상처, 낡은 어두운 후드와 청바지를 갖춰 참고와 대체로 맞습니다. 다리의 개 물림 상처는 분명히 확인되지 않습니다. 찰리는 각진 기계 머리, 흰 얼굴판, 점 눈과 선 입, 거대한 금속 팔, 더러운 코트와 모자를 갖췄습니다. 가슴의 UBIK 로고와 출입문 글자, 벽 포스터 글자가 읽힙니다. 기사 내용이 확인되는 브라우저는 보이지 않습니다.",
        "hard_violations": [
         "찰리 가슴의 UBIK 로고, 유리문의 한글 및 벽 포스터의 판독 가능한 글자가 노출되어 글자·로고 금지 조건을 위반합니다."
        ],
        "physics": "이현우는 앞쪽 신발을 바닥에 디디고 뒤쪽 발을 들어 보폭을 옮기며, 찰리도 화면 아래에 보이는 금속 발로 바닥을 지지합니다. 두 인물의 손은 실제로 접촉하고 있어 부유하지 않지만, 금속 팔을 쥐는 이현우의 손가락 동작은 확인되지 않고 오히려 찰리의 손이 그를 잡습니다. 가전은 바닥이나 다른 가전 위에 놓여 있고 공중에 떠 있는 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "이현우는 얼굴과 상체를 왼쪽으로 돌리고 화면 왼쪽 통로 쪽으로 발을 내딛습니다. 실제 출입문은 왼쪽 뒤에 있어 몸의 진행 방향이 문 정면으로 곧장 이어지지는 않지만, 왼쪽 퇴장 동선은 읽힙니다. 찰리는 이현우와 잡힌 팔 쪽으로 고개를 숙이고 왼쪽으로 따라옵니다. 이현우의 손이 찰리의 금속 손목 바로 위를 감싸 당기는 관계가 명확합니다.",
        "built_space": "왼쪽 후경에 상부 채광창을 갖춘 유리 출입구 하나가 있고, 천장에는 형광등 기구 세 개가 보입니다. 양쪽 가전 선반 사이로 콘크리트 통로가 이어지며 오른쪽에는 전면 투입식 세탁기 하나와 선풍기 하나, 여러 냉장고·전자레인지·화면이 있습니다. 문 가까이 작업대와 의자가 있고 전경 작업대 오른쪽에 전화기 하나가 놓입니다. 참고의 긴 통로와 가전 적재 구조를 A보다 잘 유지합니다. 두 인물은 좌우 배치를 지키지만 요청한 중경보다 카메라에 가깝고 크게 보입니다.",
        "entities": "이현우와 찰리 외의 인물은 없습니다. 이현우는 젊은 동아시아계 남성 외형, 헝클어진 검은 머리, 얼굴의 붉은 상처와 낡은 청바지를 갖췄습니다. 무릎의 찢어진 부분과 붉은 손상이 보이지만 개 물림인지는 확정할 수 없습니다. 참고의 후드 위에 열린 겉옷이 추가되어 복장은 완전히 같지 않습니다. 찰리의 흰 얼굴판, 각진 금속 머리, 점 눈과 선 입, 낡은 코트와 모자는 요구에 부합하지만 팔의 거대함은 A보다 약합니다. 가슴의 UBIK 표기가 읽히며, 배경 모니터에서는 기사 내용을 확인할 수 없습니다.",
        "hard_violations": [
         "찰리 가슴에 판독 가능한 UBIK 표기가 남아 있어 글자·로고 금지 조건을 위반합니다."
        ],
        "physics": "이현우는 뒤쪽 다리에 체중을 남긴 채 앞발을 내딛고 상체를 비튼 보행 순간으로 읽힙니다. 찰리는 앞쪽 발로 바닥을 지지하고 뒤쪽 발꿈치를 들어 따라오는 자세입니다. 이현우의 맨손과 손목, 드러난 팔뚝이 금속 전완으로 자연스럽게 이어지고 손가락이 이를 감싸므로 당김의 접촉과 힘 전달이 보입니다. 코트 자락의 들림도 보행에 맞으며, 지지 없이 떠 있는 인물이나 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.85
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.6
   },
   "violations": {
    "B": [
     "[gemini-pro] physically impossible anatomy (이현우의 오른팔이 몸통 옆으로 늘어져 있음에도 불구하고, 등 뒤에서 불가능한 구조의 또 다른 팔이 길게 뻗어 나와 찰리를 잡고 있음)",
     "[gemini-pro] leaked markers/diagrams/text (유리문에 '고물상'이라는 식별 가능한 한글 텍스트가 노출됨)",
     "[gpt-high] 찰리 가슴의 UBIK 로고, 유리문의 한글 및 벽 포스터의 판독 가능한 글자가 노출되어 글자·로고 금지 조건을 위반합니다."
    ],
    "A": [
     "[gpt-high] 찰리 가슴에 판독 가능한 UBIK 표기가 남아 있어 글자·로고 금지 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 600
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "장소의 실제 구조를 유지하며 캐릭터 특징과 역동적인 당기기 자세를 물리적 오류 없이 훌륭하게 구현했으나, 출구를 화면 좌측 가장자리에 밀착시키라는 구도 지시에는 약간 미치지 못함.  ★위반: [gpt-high] 찰리 가슴에 판독 가능한 UBIK 표기가 남아 있어 글자·로고 금지 조건을 위반합니다."
   },
   {
    "label": "B",
    "score": 600,
    "verdict_ko": "프레이밍 지시는 잘 따랐으나, 이현우의 오른팔이 아래로 내려와 있음에도 등 뒤에서 기형적인 팔이 뻗어 나오는 치명적인 해부학적 오류와 금지된 텍스트가 포함됨.  ★위반: [gemini-pro] physically impossible anatomy (이현우의 오른팔이 몸통 옆으로 늘어져 있음에도 불구하고, 등 뒤에서 불가능한 구조의 또 다른 팔이 길게 뻗어 나와 찰리를 잡고 있음) / [gemini-pro] leaked markers/diagrams/text (유리문에 '고물상'이라는 식별 가능한 한글 텍스트가 노출됨) / [gpt-high] 찰리 가슴의 UBIK 로고, 유리문의 한글 및 벽 포스터의 판독 가능한 글자가 노출되어 글자·로고 금지 조건을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L25B04.png",
    "asset_id": "7433b9c2-c294-46c1-bd2c-18bf7fa04322",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1597480>",
    "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd8d-894f-7fda-9295-cccfd2283201",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S15sh11__bgfirst_bg.png",
   "bg_asset_id": "28bbef6d-c5d4-4371-888b-bea3ba16935d",
   "bg_record_key": "S15sh11::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S15sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:23:20.673701+00:00",
  "fingerprint": "90c2e0499ac235cc652f665381855c2a4f5269c190d8f89ce7b004befa489319",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S15sh11_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S15sh11_sel.png",
  "source_sha256": "142eb33e598df671ad4b4b94c882f8c8205b4187af13fc664c88590de9a4c158",
  "file": "S15sh11_cine.png",
  "staged_sha256": "9ec0428c9e57e48d3c21ed6e1ffc0710c51f2b94a4af25101fb2b894188ca731",
  "latency_ms": 11621
 },
 "S15sh13::signage": {
  "fp": "47c12f14b260aae9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S15sh13": {
  "input_fingerprint": "0180ab2b7205290b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다가 낡은 유선 전화기 수화기를 귓가에 바짝 가져다 댄 채 손으로 입을 가리고 멈춰선 초조한 표정.\n\nLOCATION (lock): At the telephone near the junk shop's repair and computer area, with the monitor glowing among piled electrical scrap. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Worn corded telephone receiver (Held tightly against 마츠다's ear during the call) — Its side is visible beside his cheek, separate from the hand covering his mouth; used as Identifies the covert call without obscuring the eyes or becoming an oversized foreground prop; Piled electrical appliances (Remaining inside the shop behind 마츠다); used as Soft background context confirms that the camera has returned into the shop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the daytime shop ambience with restrained facial contrast that reveals his anxious expression without introducing a sinister new color or light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop remains filled with appliances and scrap, and the reward article has not been cleared from the browser. Charlie has left the shop with his dirty exterior, worn coat and hat, and peeling chest logo unchanged. 마츠다: He remains inside the shop and is using a telephone.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 마츠다 right now, so 마츠다's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 마츠다: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다가 낡은 유선 전화기 수화기를 귓가에 바짝 가져다 댄 채 손으로 입을 가리고 멈춰선 초조한 표정.\n\nLOCATION (lock): At the telephone near the junk shop's repair and computer area, with the monitor glowing among piled electrical scrap. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Worn corded telephone receiver (Held tightly against 마츠다's ear during the call) — Its side is visible beside his cheek, separate from the hand covering his mouth; used as Identifies the covert call without obscuring the eyes or becoming an oversized foreground prop; Piled electrical appliances (Remaining inside the shop behind 마츠다); used as Soft background context confirms that the camera has returned into the shop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the daytime shop ambience with restrained facial contrast that reveals his anxious expression without introducing a sinister new color or light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop remains filled with appliances and scrap, and the reward article has not been cleared from the browser. Charlie has left the shop with his dirty exterior, worn coat and hat, and peeling chest logo unchanged. 마츠다: He remains inside the shop and is using a telephone.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 마츠다 right now, so 마츠다's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 마츠다: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다가 낡은 유선 전화기 수화기를 귓가에 바짝 가져다 댄 채 손으로 입을 가리고 멈춰선 초조한 표정.\n\nLOCATION (lock): At the telephone near the junk shop's repair and computer area, with the monitor glowing among piled electrical scrap. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Worn corded telephone receiver (Held tightly against 마츠다's ear during the call) — Its side is visible beside his cheek, separate from the hand covering his mouth; used as Identifies the covert call without obscuring the eyes or becoming an oversized foreground prop; Piled electrical appliances (Remaining inside the shop behind 마츠다); used as Soft background context confirms that the camera has returned into the shop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the daytime shop ambience with restrained facial contrast that reveals his anxious expression without introducing a sinister new color or light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop remains filled with appliances and scrap, and the reward article has not been cleared from the browser. Charlie has left the shop with his dirty exterior, worn coat and hat, and peeling chest logo unchanged. 마츠다: He remains inside the shop and is using a telephone.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 마츠다 right now, so 마츠다's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 마츠다: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "마츠다의 시선은 프레임 왼쪽 앞의 허공을 향하고 있으며, 수화기의 측면이 카메라를 향해 명확히 노출되어 있음.",
    "built_space": "이전 샷의 실내 공간 구조를 완벽히 재현함. 왼쪽의 출입문, 길게 늘어선 냉장고들, 중앙 배경의 책상과 켜진 모니터, 오른쪽의 TV 금속 선반이 모두 이전 샷과 동일한 위치 및 방향에 있음.",
    "entities": "마츠다(얼굴과 연령대는 일치하나 레퍼런스의 녹색 모자는 누락됨), 낡은 유선 전화기 수화기 및 본체, 배경의 전자기기 폐기물과 켜진 모니터.",
    "hard_violations": [],
    "physics": "오른손은 입을 가리고 얼굴에 자연스럽게 밀착되어 있으며, 왼손은 수화기를 안정적으로 쥐고 지탱함. 수화기에 연결된 선이 프레임 왼쪽 아래의 회색 플라스틱 상자 위에 놓인 전화기 본체로 이어짐."
   },
   {
    "label": "B",
    "direction": "마츠다의 시선은 정면 좌측 허공을 향하고 있음.",
    "built_space": "이전 샷의 공간과 전혀 다름. 레퍼런스에는 없는 커다란 다중 격자형 창문이 왼쪽에 렌더링되었으며, 냉장고와 선반 등 고정 요소들의 배치가 완전히 달라져 공간의 연속성이 붕괴됨.",
    "entities": "마츠다(얼굴, 연령대, 녹색 모자 등 복장 일치), 유선 전화기 수화기, 우측 하단의 모니터, 좌측 하단의 회로 기판 더미.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스의 지정된 위치(Location lock) 및 구조를 완전히 무시하고, 해당 시점에서 존재할 수 없는 전혀 새로운 다중 격자 유리창 구조와 실내 형태를 발명하여 렌더링함",
     "[gpt-high] 기준 점포의 왼쪽 가전 진열 벽을 대형 창벽으로 바꾸고 후면 창 구성을 추가하여, 고정된 장소의 건축 구조를 변경했다."
    ],
    "physics": "오른손이 입을 가리고 얼굴에 닿아 있으며, 왼손이 수화기를 쥐어 공중에 지탱하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "캐릭터의 모자가 누락된 점은 아쉬우나, 이전 샷의 복잡한 실내 구조와 조명, 가구 배치를 완벽하게 재현했으며 요구된 포즈와 해부학적 구조를 사실적으로 잘 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "마츠다의 복장(모자)은 잘 일치하지만, 레퍼런스에 명시된 공간 구조를 무시하고 엉뚱한 거대 창문을 추가하여 위치 고정(Location lock) 조건을 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "마츠다의 시선은 프레임 왼쪽 앞의 허공을 향하고 있으며, 수화기의 측면이 카메라를 향해 명확히 노출되어 있음.",
        "built_space": "이전 샷의 실내 공간 구조를 완벽히 재현함. 왼쪽의 출입문, 길게 늘어선 냉장고들, 중앙 배경의 책상과 켜진 모니터, 오른쪽의 TV 금속 선반이 모두 이전 샷과 동일한 위치 및 방향에 있음.",
        "entities": "마츠다(얼굴과 연령대는 일치하나 레퍼런스의 녹색 모자는 누락됨), 낡은 유선 전화기 수화기 및 본체, 배경의 전자기기 폐기물과 켜진 모니터.",
        "hard_violations": [],
        "physics": "오른손은 입을 가리고 얼굴에 자연스럽게 밀착되어 있으며, 왼손은 수화기를 안정적으로 쥐고 지탱함. 수화기에 연결된 선이 프레임 왼쪽 아래의 회색 플라스틱 상자 위에 놓인 전화기 본체로 이어짐."
       },
       {
        "label": "B",
        "direction": "마츠다의 시선은 정면 좌측 허공을 향하고 있음.",
        "built_space": "이전 샷의 공간과 전혀 다름. 레퍼런스에는 없는 커다란 다중 격자형 창문이 왼쪽에 렌더링되었으며, 냉장고와 선반 등 고정 요소들의 배치가 완전히 달라져 공간의 연속성이 붕괴됨.",
        "entities": "마츠다(얼굴, 연령대, 녹색 모자 등 복장 일치), 유선 전화기 수화기, 우측 하단의 모니터, 좌측 하단의 회로 기판 더미.",
        "hard_violations": [
         "레퍼런스의 지정된 위치(Location lock) 및 구조를 완전히 무시하고, 해당 시점에서 존재할 수 없는 전혀 새로운 다중 격자 유리창 구조와 실내 형태를 발명하여 렌더링함"
        ],
        "physics": "오른손이 입을 가리고 얼굴에 닿아 있으며, 왼손이 수화기를 쥐어 공중에 지탱하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "캐릭터의 모자가 누락된 점은 아쉬우나, 이전 샷의 복잡한 실내 구조와 조명, 가구 배치를 완벽하게 재현했으며 요구된 포즈와 해부학적 구조를 사실적으로 잘 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "마츠다의 복장(모자)은 잘 일치하지만, 레퍼런스에 명시된 공간 구조를 무시하고 엉뚱한 거대 창문을 추가하여 위치 고정(Location lock) 조건을 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "마츠다의 시선은 프레임 왼쪽 앞의 허공을 향하고 있으며, 수화기의 측면이 카메라를 향해 명확히 노출되어 있음.",
        "built_space": "이전 샷의 실내 공간 구조를 완벽히 재현함. 왼쪽의 출입문, 길게 늘어선 냉장고들, 중앙 배경의 책상과 켜진 모니터, 오른쪽의 TV 금속 선반이 모두 이전 샷과 동일한 위치 및 방향에 있음.",
        "entities": "마츠다(얼굴과 연령대는 일치하나 레퍼런스의 녹색 모자는 누락됨), 낡은 유선 전화기 수화기 및 본체, 배경의 전자기기 폐기물과 켜진 모니터.",
        "hard_violations": [],
        "physics": "오른손은 입을 가리고 얼굴에 자연스럽게 밀착되어 있으며, 왼손은 수화기를 안정적으로 쥐고 지탱함. 수화기에 연결된 선이 프레임 왼쪽 아래의 회색 플라스틱 상자 위에 놓인 전화기 본체로 이어짐."
       },
       {
        "label": "B",
        "direction": "마츠다의 시선은 정면 좌측 허공을 향하고 있음.",
        "built_space": "이전 샷의 공간과 전혀 다름. 레퍼런스에는 없는 커다란 다중 격자형 창문이 왼쪽에 렌더링되었으며, 냉장고와 선반 등 고정 요소들의 배치가 완전히 달라져 공간의 연속성이 붕괴됨.",
        "entities": "마츠다(얼굴, 연령대, 녹색 모자 등 복장 일치), 유선 전화기 수화기, 우측 하단의 모니터, 좌측 하단의 회로 기판 더미.",
        "hard_violations": [
         "레퍼런스의 지정된 위치(Location lock) 및 구조를 완전히 무시하고, 해당 시점에서 존재할 수 없는 전혀 새로운 다중 격자 유리창 구조와 실내 형태를 발명하여 렌더링함"
        ],
        "physics": "오른손이 입을 가리고 얼굴에 닿아 있으며, 왼손이 수화기를 쥐어 공중에 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "수화기와 입을 가리는 동작, 모자는 잘 맞지만, 창벽으로 바뀐 점포 구조가 장소 고정을 어기고 구도도 요구한 클로즈업보다 넓다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "기존 점포의 출입구·가전 선반·수리 공간을 유지하면서 초조한 비밀 통화를 구현했으나, 기준의 모자가 없고 클로즈업은 다소 느슨하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 카메라가 아닌 화면 왼쪽의 프레임 밖을 향하며, 특정 상대는 보이지 않는다. 수화기 위쪽은 귀에 밀착되고 아래쪽은 입 옆으로 향한다. 반대 손이 입을 가리며 수화기 측면과 눈은 별도로 보인다.",
        "built_space": "천장 형광등 두 개, 왼쪽의 넓은 다분할 창벽, 뒤쪽 창 두 군데, 오른쪽 가전 선반, 왼쪽 폐부품 더미와 오른쪽 아래 발광 모니터 한 대가 보인다. 기준 사진의 왼쪽 가전 진열 구역이 큰 창벽으로 바뀌어 동일 점포 구조로 보기 어렵다. 인물은 폐부품과 모니터 사이에 있으며, 상반신과 주변 공간까지 보여 얼굴 중심 클로즈업보다 넓다.",
        "entities": "인물은 노년의 동아시아계 남성 한 명이며 기준 마츠다의 얼굴, 올리브색 모자, 체크 셔츠, 낡은 황갈색 조끼와 대체로 맞는다. 일본 국적 자체는 외관만으로 확인할 수 없다. 보이는 팔에는 기준의 녹색 덧소매가 없다. 낡은 아이보리색 유선 수화기와 전기 폐품은 분명하다. 모니터는 녹색의 거의 빈 화면으로 보여 보상 기사 유지가 확인되지 않는다. 다른 사람이나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "기준 점포의 왼쪽 가전 진열 벽을 대형 창벽으로 바꾸고 후면 창 구성을 추가하여, 고정된 장소의 건축 구조를 변경했다."
        ],
        "physics": "한 손이 수화기를 확실히 잡고 다른 손의 손가락과 손바닥이 입에 닿아 있다. 전화선은 수화기 아래에서 중력 방향으로 늘어진다. 폐부품은 통과 작업면에, 모니터는 작업대 쪽에 놓여 있다. 하체는 프레임 밖이라 발의 접지는 확인할 수 없지만, 보이는 상체에 부유나 불가능한 자세는 없다."
       },
       {
        "label": "B",
        "direction": "눈은 화면 왼쪽 아래의 프레임 밖을 향하고 미간이 긴장되어 있다. 수화기의 청취부는 귀에 닿고 송화부는 가려진 입 옆에 위치한다. 한 손은 수화기를 잡고 다른 손은 입을 덮어, 전화기와 입을 가리는 손이 명확히 구분된다.",
        "built_space": "뒤쪽 왼편 유리 출입구 한 곳, 그 오른쪽 좁은 창 한 곳, 천장 형광등 세 개, 오른쪽 다단 금속 가전 선반과 선풍기 한 대, 중앙 뒤쪽 냉장고·세탁기 열이 보인다. 출입구 근처 작업대에는 켜진 모니터 한 대가 있다. 이 배치는 기준 점포와 잘 이어진다. 인물은 전경 전화 작업대 옆에 있으며, 전화기 본체는 왼쪽 아래에 놓여 있다. 머리 윗부분은 잘리지만 가슴과 넓은 통로까지 들어와 클로즈업으로는 다소 넓다.",
        "entities": "노년의 동아시아계 남성 한 명만 보이며, 얼굴의 연령감과 체격, 체크 셔츠와 낡은 황갈색 조끼는 기준에 대체로 부합한다. 일본 국적은 외관만으로 판별할 수 없다. 기준의 모자는 없고 회색 섞인 머리카락이 드러나며, 보이는 팔의 녹색 덧소매도 빠졌다. 낡은 유선 수화기, 전화기 본체, 전선, 가전제품과 발광 모니터가 보인다. 모니터에는 밝은 페이지 형태가 있으나 보상 기사인지는 식별되지 않는다. 추가 인물과 읽히는 글자는 없다.",
        "hard_violations": [],
        "physics": "수화기는 손가락으로 단단히 지지되고 귀에 밀착되어 있다. 반대 손은 자연스럽게 입과 볼에 접촉한다. 꼬인 전화선은 수화기에서 작업대 위 본체로 이어지며 아래로 처진다. 전화기 본체는 작업대에, 가전제품과 선풍기는 선반이나 바닥에 지지된다. 하체가 보이지 않으므로 발의 위치는 판단할 수 없지만, 상체와 팔의 정지 자세에는 물리적 모순이 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "수화기와 입을 가리는 동작, 모자는 잘 맞지만, 창벽으로 바뀐 점포 구조가 장소 고정을 어기고 구도도 요구한 클로즈업보다 넓다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "기존 점포의 출입구·가전 선반·수리 공간을 유지하면서 초조한 비밀 통화를 구현했으나, 기준의 모자가 없고 클로즈업은 다소 느슨하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "시선은 카메라가 아닌 화면 왼쪽의 프레임 밖을 향하며, 특정 상대는 보이지 않는다. 수화기 위쪽은 귀에 밀착되고 아래쪽은 입 옆으로 향한다. 반대 손이 입을 가리며 수화기 측면과 눈은 별도로 보인다.",
        "built_space": "천장 형광등 두 개, 왼쪽의 넓은 다분할 창벽, 뒤쪽 창 두 군데, 오른쪽 가전 선반, 왼쪽 폐부품 더미와 오른쪽 아래 발광 모니터 한 대가 보인다. 기준 사진의 왼쪽 가전 진열 구역이 큰 창벽으로 바뀌어 동일 점포 구조로 보기 어렵다. 인물은 폐부품과 모니터 사이에 있으며, 상반신과 주변 공간까지 보여 얼굴 중심 클로즈업보다 넓다.",
        "entities": "인물은 노년의 동아시아계 남성 한 명이며 기준 마츠다의 얼굴, 올리브색 모자, 체크 셔츠, 낡은 황갈색 조끼와 대체로 맞는다. 일본 국적 자체는 외관만으로 확인할 수 없다. 보이는 팔에는 기준의 녹색 덧소매가 없다. 낡은 아이보리색 유선 수화기와 전기 폐품은 분명하다. 모니터는 녹색의 거의 빈 화면으로 보여 보상 기사 유지가 확인되지 않는다. 다른 사람이나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "기준 점포의 왼쪽 가전 진열 벽을 대형 창벽으로 바꾸고 후면 창 구성을 추가하여, 고정된 장소의 건축 구조를 변경했다."
        ],
        "physics": "한 손이 수화기를 확실히 잡고 다른 손의 손가락과 손바닥이 입에 닿아 있다. 전화선은 수화기 아래에서 중력 방향으로 늘어진다. 폐부품은 통과 작업면에, 모니터는 작업대 쪽에 놓여 있다. 하체는 프레임 밖이라 발의 접지는 확인할 수 없지만, 보이는 상체에 부유나 불가능한 자세는 없다."
       },
       {
        "label": "A",
        "direction": "눈은 화면 왼쪽 아래의 프레임 밖을 향하고 미간이 긴장되어 있다. 수화기의 청취부는 귀에 닿고 송화부는 가려진 입 옆에 위치한다. 한 손은 수화기를 잡고 다른 손은 입을 덮어, 전화기와 입을 가리는 손이 명확히 구분된다.",
        "built_space": "뒤쪽 왼편 유리 출입구 한 곳, 그 오른쪽 좁은 창 한 곳, 천장 형광등 세 개, 오른쪽 다단 금속 가전 선반과 선풍기 한 대, 중앙 뒤쪽 냉장고·세탁기 열이 보인다. 출입구 근처 작업대에는 켜진 모니터 한 대가 있다. 이 배치는 기준 점포와 잘 이어진다. 인물은 전경 전화 작업대 옆에 있으며, 전화기 본체는 왼쪽 아래에 놓여 있다. 머리 윗부분은 잘리지만 가슴과 넓은 통로까지 들어와 클로즈업으로는 다소 넓다.",
        "entities": "노년의 동아시아계 남성 한 명만 보이며, 얼굴의 연령감과 체격, 체크 셔츠와 낡은 황갈색 조끼는 기준에 대체로 부합한다. 일본 국적은 외관만으로 판별할 수 없다. 기준의 모자는 없고 회색 섞인 머리카락이 드러나며, 보이는 팔의 녹색 덧소매도 빠졌다. 낡은 유선 수화기, 전화기 본체, 전선, 가전제품과 발광 모니터가 보인다. 모니터에는 밝은 페이지 형태가 있으나 보상 기사인지는 식별되지 않는다. 추가 인물과 읽히는 글자는 없다.",
        "hard_violations": [],
        "physics": "수화기는 손가락으로 단단히 지지되고 귀에 밀착되어 있다. 반대 손은 자연스럽게 입과 볼에 접촉한다. 꼬인 전화선은 수화기에서 작업대 위 본체로 이어지며 아래로 처진다. 전화기 본체는 작업대에, 가전제품과 선풍기는 선반이나 바닥에 지지된다. 하체가 보이지 않으므로 발의 위치는 판단할 수 없지만, 상체와 팔의 정지 자세에는 물리적 모순이 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.069
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.819
   },
   "violations": {
    "B": [
     "[gemini-pro] 레퍼런스의 지정된 위치(Location lock) 및 구조를 완전히 무시하고, 해당 시점에서 존재할 수 없는 전혀 새로운 다중 격자 유리창 구조와 실내 형태를 발명하여 렌더링함",
     "[gpt-high] 기준 점포의 왼쪽 가전 진열 벽을 대형 창벽으로 바꾸고 후면 창 구성을 추가하여, 고정된 장소의 건축 구조를 변경했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 819
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "캐릭터의 모자가 누락된 점은 아쉬우나, 이전 샷의 복잡한 실내 구조와 조명, 가구 배치를 완벽하게 재현했으며 요구된 포즈와 해부학적 구조를 사실적으로 잘 구현했습니다."
   },
   {
    "label": "B",
    "score": 819,
    "verdict_ko": "마츠다의 복장(모자)은 잘 일치하지만, 레퍼런스에 명시된 공간 구조를 무시하고 엉뚱한 거대 창문을 추가하여 위치 고정(Location lock) 조건을 크게 위반했습니다.  ★위반: [gemini-pro] 레퍼런스의 지정된 위치(Location lock) 및 구조를 완전히 무시하고, 해당 시점에서 존재할 수 없는 전혀 새로운 다중 격자 유리창 구조와 실내 형태를 발명하여 렌더링함 / [gpt-high] 기준 점포의 왼쪽 가전 진열 벽을 대형 창벽으로 바꾸고 후면 창 구성을 추가하여, 고정된 장소의 건축 구조를 변경했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S15sh11_sel.png",
    "asset_id": "3ab56e16-c1d2-4884-a8f5-16bd78df4008",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 마츠다: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876556>",
    "asset_id": "8e5adf95-2c7b-4499-9815-c7a393ac5668",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd8e-49f4-70e5-ad4f-08795a679f1e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S15sh11"
  }
 },
 "S15sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:23:32.631457+00:00",
  "fingerprint": "b06391e6a1385672a2aa6d578e194c1e5b3c884f47a85c48f6bd59f39f37dbe0",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S15sh13_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S15sh13_sel.png",
  "source_sha256": "2965e494b74e7eb11ad3827d5a38ddea38c042f8fd00d4a824e00667a4b2e02b",
  "file": "S15sh13_cine.png",
  "staged_sha256": "3d02a679c4731af1b0006838326085855131e26f3e8d43755072c958b79e9f31",
  "latency_ms": 11692
 },
 "S16sh2::signage": {
  "fp": "34bef1ee7dee9a80",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S16sh2": {
  "input_fingerprint": "fe156fd2cc1850a6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물줄기를 맞은 찰리가 입가 금속 부품을 위로 한껏 올린 채 두 팔을 좌우로 활짝 벌린 유쾌한 자태.\n\nLOCATION (lock): In the open washing area immediately outside a child's container home in the refugee settlement. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hose water stream (Striking 찰리 and washing dirt from his body); used as Enters from the left as a diagonal connecting the off-screen playmates to 찰리; 라울's container home (The washing takes place in front of it) — A partial exterior face appears behind 찰리; used as Grounds the playful action in its domestic setting without crowding either outstretched hand.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use daytime ambient light with controlled highlights on the water and 찰리's washed surfaces, keeping his joyful facial configuration clearly readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Water streams from the hose outside Raul's container, wetting Charlie and washing dirt from his metal exterior. His peeling UBIK chest logo and old coat and hat remain established; washing is not a repair or replacement.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물줄기를 맞은 찰리가 입가 금속 부품을 위로 한껏 올린 채 두 팔을 좌우로 활짝 벌린 유쾌한 자태.\n\nLOCATION (lock): In the open washing area immediately outside a child's container home in the refugee settlement. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hose water stream (Striking 찰리 and washing dirt from his body); used as Enters from the left as a diagonal connecting the off-screen playmates to 찰리; 라울's container home (The washing takes place in front of it) — A partial exterior face appears behind 찰리; used as Grounds the playful action in its domestic setting without crowding either outstretched hand.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use daytime ambient light with controlled highlights on the water and 찰리's washed surfaces, keeping his joyful facial configuration clearly readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Water streams from the hose outside Raul's container, wetting Charlie and washing dirt from his metal exterior. His peeling UBIK chest logo and old coat and hat remain established; washing is not a repair or replacement.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물줄기를 맞은 찰리가 입가 금속 부품을 위로 한껏 올린 채 두 팔을 좌우로 활짝 벌린 유쾌한 자태.\n\nLOCATION (lock): In the open washing area immediately outside a child's container home in the refugee settlement. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hose water stream (Striking 찰리 and washing dirt from his body); used as Enters from the left as a diagonal connecting the off-screen playmates to 찰리; 라울's container home (The washing takes place in front of it) — A partial exterior face appears behind 찰리; used as Grounds the playful action in its domestic setting without crowding either outstretched hand.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use daytime ambient light with controlled highlights on the water and 찰리's washed surfaces, keeping his joyful facial configuration clearly readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Water streams from the hose outside Raul's container, wetting Charlie and washing dirt from his metal exterior. His peeling UBIK chest logo and old coat and hat remain established; washing is not a repair or replacement.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "좌측에서 찰리의 가슴으로 물줄기가 분사됨. 시선은 정면을 향함.",
    "built_space": "찰리 바로 뒤에 컨테이너 주택 외벽이 위치함.",
    "entities": "코트, 모자, UBIK 로고는 있으나 치아가 묘사되어 캐릭터 설정에 어긋남. 프롬프트에 없는 배경 인물 다수 존재.",
    "hard_violations": [
     "[gemini-pro] 지문에 없는 배경 인물 추가 (invented people)",
     "[gpt-high] 화면 밖에 있어야 할 호스 사용자의 손이 화면 안에 들어왔다.",
     "[gpt-high] 찰리 외에 배경의 성인과 아이들을 다수 추가했다.",
     "[gpt-high] 가슴의 ‘UBIK’ 글자가 읽혀 글자 금지 조건을 위반한다.",
     "[gpt-high] 장난감 자동차, 물통, 빗자루, 빨래 등 허용된 장소 설명에 없는 소품을 다수 추가했다."
    ],
    "physics": "바닥에 서서 팔을 벌리고 있으며, 물이 가슴에 부딪혀 튀는 물리적 묘사가 정상적임."
   },
   {
    "label": "B",
    "direction": "좌측 상단에서 찰리의 흉부를 향해 물줄기가 분사됨.",
    "built_space": "찰리 뒤로 컨테이너 구조물 외벽이 넓게 배치됨.",
    "entities": "점 형태의 눈과 선 형태의 입을 가진 기계 얼굴을 정확히 구현. 프롬프트에 없는 배경 인물 2명 존재.",
    "hard_violations": [
     "[gemini-pro] 지문에 없는 배경 인물 추가 (invented people)",
     "[gpt-high] 찰리만 보여야 하는 장면에 배경 인물들이 추가되었다.",
     "[gpt-high] 가슴의 ‘UBIK’ 글자가 선명하게 읽혀 글자 금지 조건을 위반한다.",
     "[gpt-high] 장소 설명이 허용하지 않은 배경 용기와 상자 등의 소품을 추가했다."
    ],
    "physics": "지면에 안정적으로 서서 두 팔을 벌리고 있으며, 물방울의 물리적 튀김 효과가 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "캐릭터의 기계 얼굴(점 형태의 눈, 선 형태의 입)은 정확히 구현했으나, 지문에 명시되지 않은 배경 인물을 포함하여 엄격한 인물 제한 규칙을 위반함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "캐릭터 얼굴에 치아를 묘사하여 '선 형태의 입' 설정을 크게 위반했으며, 지문에 없는 인물들을 배경에 다수 추가함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "좌측에서 찰리의 가슴으로 물줄기가 분사됨. 시선은 정면을 향함.",
        "built_space": "찰리 바로 뒤에 컨테이너 주택 외벽이 위치함.",
        "entities": "코트, 모자, UBIK 로고는 있으나 치아가 묘사되어 캐릭터 설정에 어긋남. 프롬프트에 없는 배경 인물 다수 존재.",
        "hard_violations": [
         "지문에 없는 배경 인물 추가 (invented people)"
        ],
        "physics": "바닥에 서서 팔을 벌리고 있으며, 물이 가슴에 부딪혀 튀는 물리적 묘사가 정상적임."
       },
       {
        "label": "B",
        "direction": "좌측 상단에서 찰리의 흉부를 향해 물줄기가 분사됨.",
        "built_space": "찰리 뒤로 컨테이너 구조물 외벽이 넓게 배치됨.",
        "entities": "점 형태의 눈과 선 형태의 입을 가진 기계 얼굴을 정확히 구현. 프롬프트에 없는 배경 인물 2명 존재.",
        "hard_violations": [
         "지문에 없는 배경 인물 추가 (invented people)"
        ],
        "physics": "지면에 안정적으로 서서 두 팔을 벌리고 있으며, 물방울의 물리적 튀김 효과가 자연스러움."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "캐릭터의 기계 얼굴(점 형태의 눈, 선 형태의 입)은 정확히 구현했으나, 지문에 명시되지 않은 배경 인물을 포함하여 엄격한 인물 제한 규칙을 위반함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "캐릭터 얼굴에 치아를 묘사하여 '선 형태의 입' 설정을 크게 위반했으며, 지문에 없는 인물들을 배경에 다수 추가함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "좌측에서 찰리의 가슴으로 물줄기가 분사됨. 시선은 정면을 향함.",
        "built_space": "찰리 바로 뒤에 컨테이너 주택 외벽이 위치함.",
        "entities": "코트, 모자, UBIK 로고는 있으나 치아가 묘사되어 캐릭터 설정에 어긋남. 프롬프트에 없는 배경 인물 다수 존재.",
        "hard_violations": [
         "지문에 없는 배경 인물 추가 (invented people)"
        ],
        "physics": "바닥에 서서 팔을 벌리고 있으며, 물이 가슴에 부딪혀 튀는 물리적 묘사가 정상적임."
       },
       {
        "label": "B",
        "direction": "좌측 상단에서 찰리의 흉부를 향해 물줄기가 분사됨.",
        "built_space": "찰리 뒤로 컨테이너 구조물 외벽이 넓게 배치됨.",
        "entities": "점 형태의 눈과 선 형태의 입을 가진 기계 얼굴을 정확히 구현. 프롬프트에 없는 배경 인물 2명 존재.",
        "hard_violations": [
         "지문에 없는 배경 인물 추가 (invented people)"
        ],
        "physics": "지면에 안정적으로 서서 두 팔을 벌리고 있으며, 물방울의 물리적 튀김 효과가 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "물줄기의 명중과 양팔을 벌린 동작은 맞지만, 상반신 위주의 좁은 구도와 약한 웃음 표현에 더해 추가 인물과 읽히는 가슴 글자가 실격 사유다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "상대적으로 넓은 구도와 각진 기계 머리, 활짝 벌린 팔과 유쾌한 표정이 더 가깝지만, 추가 인물·호스 잡은 손·읽히는 글자 때문에 이 후보도 실격이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 위 호스에서 나온 물이 오른쪽 아래로 진행해 찰리의 화면 왼쪽 어깨와 가슴에 닿는다. 찰리는 거의 정면에서 약간 화면 왼쪽을 바라보며 두 팔과 손바닥을 좌우로 펼친다. 물의 목표는 맞지만 입가 금속 부품을 한껏 올린 웃음보다는 입을 조금 벌린 표정에 가깝다.",
        "built_space": "찰리 뒤와 오른쪽에 골이 진 컨테이너 외벽 일부가 있고, 왼쪽 어깨 뒤로 창 하나가 부분적으로 보인다. 왼쪽 배경에는 다른 컨테이너와 통로가 이어진다. 양손은 외벽 설비와 겹치지 않지만 찰리의 상반신이 화면을 크게 차지해 요청한 와이드 쇼트보다 좁다. 장소 참고 이미지는 없어 특정 건물과의 일치 여부는 확인할 수 없다.",
        "entities": "찰리 한 명 외에 왼쪽 통로에 적어도 세 명의 추가 인물이 보이고, 왼쪽 위에도 호스를 잡은 손 일부처럼 보이는 형태가 있다. 찰리는 고릴라형 금속 머리, 밝은 얼굴판, 낡은 모자와 코트, 벗겨진 가슴 도장을 갖췄다. 다만 머리는 둥글고 눈과 입은 점·선 형태보다 사람 눈과 벌어진 입에 가깝다. 가슴에는 읽을 수 있는 ‘UBIK’ 글자가 있다. 배경에는 요청하지 않은 여러 용기와 상자가 추가되어 있다. 낮의 야외 세척 장면과 젖고 낡은 금속 표면은 보인다.",
        "hard_violations": [
         "찰리만 보여야 하는 장면에 배경 인물들이 추가되었다.",
         "가슴의 ‘UBIK’ 글자가 선명하게 읽혀 글자 금지 조건을 위반한다.",
         "장소 설명이 허용하지 않은 배경 용기와 상자 등의 소품을 추가했다."
        ],
        "physics": "물줄기는 보이는 호스 출구에서 이어져 찰리의 어깨와 가슴에서 흩어지고 아래로 흐르므로 충돌과 중력 방향이 타당하다. 팔은 어깨에 연결되어 있고 모자는 머리에, 코트는 몸에 걸쳐 있다. 하체와 발은 프레임 밖이어서 지면 접촉은 확인할 수 없지만 공중에 떠 있다는 증거는 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽 아래 노즐에서 나온 물이 오른쪽 위로 향해 찰리의 화면 왼쪽 상완과 어깨·가슴 부근에 부딪힌다. 찰리의 얼굴과 시선은 물이 오는 화면 왼쪽을 향하고 양팔은 좌우로 크게 벌어져 있다. 올라간 입꼬리와 크게 열린 입이 즐거운 반응을 분명히 전달하지만, 선 형태의 입가 부품을 들어 올리는 모습보다는 치아가 드러나는 웃음으로 구현되었다.",
        "built_space": "뒤쪽 중앙에 컨테이너 전면이 있으며 창 하나, 인물에 일부 가려진 출입문 하나, 문 위 차양 하나가 보인다. 좌우에는 별도의 지붕과 주거 공간이 이어진다. 찰리는 컨테이너 앞 젖은 공터에 있으며 양손을 벌릴 공간은 확보되어 있다. A보다 주변 공간을 포함하지만 다리 아래는 잘렸고, 컨테이너 전면과 주변 살림이 크게 드러나 간결한 부분 외벽 배경보다 복잡하다.",
        "entities": "찰리는 각진 고릴라형 기계 머리와 흰 얼굴판, 낡은 모자와 코트, 벗겨진 가슴 도장을 갖추고 있다. 눈에는 눈동자가 표현되어 점 형태와 다르고 입에는 치아가 드러나 선 형태와 다르다. 가슴의 ‘UBIK’는 읽을 수 있다. 왼쪽 아래에는 다른 사람의 손이 노즐을 잡고 있으며 양쪽 배경에도 여러 성인과 아이가 있다. 추가된 물통 두 개, 장난감 자동차, 빗자루, 빨래와 생활용품도 뚜렷하다. 낮빛과 젖은 금속·직물·물웅덩이는 실물 재질로 표현되었다.",
        "hard_violations": [
         "화면 밖에 있어야 할 호스 사용자의 손이 화면 안에 들어왔다.",
         "찰리 외에 배경의 성인과 아이들을 다수 추가했다.",
         "가슴의 ‘UBIK’ 글자가 읽혀 글자 금지 조건을 위반한다.",
         "장난감 자동차, 물통, 빗자루, 빨래 등 허용된 장소 설명에 없는 소품을 다수 추가했다."
        ],
        "physics": "노즐은 화면 왼쪽 아래의 손이 잡고 있으며 물은 노즐에서 찰리까지 연속적으로 이어진다. 몸에 부딪힌 물방울이 퍼지고 아래로 떨어지는 모습도 타당하다. 양팔과 손의 연결, 머리에 얹힌 모자, 몸에 걸친 코트에 명백한 지지 오류는 없다. 발은 프레임 밖이므로 접지는 확인할 수 없지만 부유하거나 도약하는 자세는 아니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "물줄기의 명중과 양팔을 벌린 동작은 맞지만, 상반신 위주의 좁은 구도와 약한 웃음 표현에 더해 추가 인물과 읽히는 가슴 글자가 실격 사유다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "상대적으로 넓은 구도와 각진 기계 머리, 활짝 벌린 팔과 유쾌한 표정이 더 가깝지만, 추가 인물·호스 잡은 손·읽히는 글자 때문에 이 후보도 실격이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 위 호스에서 나온 물이 오른쪽 아래로 진행해 찰리의 화면 왼쪽 어깨와 가슴에 닿는다. 찰리는 거의 정면에서 약간 화면 왼쪽을 바라보며 두 팔과 손바닥을 좌우로 펼친다. 물의 목표는 맞지만 입가 금속 부품을 한껏 올린 웃음보다는 입을 조금 벌린 표정에 가깝다.",
        "built_space": "찰리 뒤와 오른쪽에 골이 진 컨테이너 외벽 일부가 있고, 왼쪽 어깨 뒤로 창 하나가 부분적으로 보인다. 왼쪽 배경에는 다른 컨테이너와 통로가 이어진다. 양손은 외벽 설비와 겹치지 않지만 찰리의 상반신이 화면을 크게 차지해 요청한 와이드 쇼트보다 좁다. 장소 참고 이미지는 없어 특정 건물과의 일치 여부는 확인할 수 없다.",
        "entities": "찰리 한 명 외에 왼쪽 통로에 적어도 세 명의 추가 인물이 보이고, 왼쪽 위에도 호스를 잡은 손 일부처럼 보이는 형태가 있다. 찰리는 고릴라형 금속 머리, 밝은 얼굴판, 낡은 모자와 코트, 벗겨진 가슴 도장을 갖췄다. 다만 머리는 둥글고 눈과 입은 점·선 형태보다 사람 눈과 벌어진 입에 가깝다. 가슴에는 읽을 수 있는 ‘UBIK’ 글자가 있다. 배경에는 요청하지 않은 여러 용기와 상자가 추가되어 있다. 낮의 야외 세척 장면과 젖고 낡은 금속 표면은 보인다.",
        "hard_violations": [
         "찰리만 보여야 하는 장면에 배경 인물들이 추가되었다.",
         "가슴의 ‘UBIK’ 글자가 선명하게 읽혀 글자 금지 조건을 위반한다.",
         "장소 설명이 허용하지 않은 배경 용기와 상자 등의 소품을 추가했다."
        ],
        "physics": "물줄기는 보이는 호스 출구에서 이어져 찰리의 어깨와 가슴에서 흩어지고 아래로 흐르므로 충돌과 중력 방향이 타당하다. 팔은 어깨에 연결되어 있고 모자는 머리에, 코트는 몸에 걸쳐 있다. 하체와 발은 프레임 밖이어서 지면 접촉은 확인할 수 없지만 공중에 떠 있다는 증거는 없다."
       },
       {
        "label": "A",
        "direction": "왼쪽 아래 노즐에서 나온 물이 오른쪽 위로 향해 찰리의 화면 왼쪽 상완과 어깨·가슴 부근에 부딪힌다. 찰리의 얼굴과 시선은 물이 오는 화면 왼쪽을 향하고 양팔은 좌우로 크게 벌어져 있다. 올라간 입꼬리와 크게 열린 입이 즐거운 반응을 분명히 전달하지만, 선 형태의 입가 부품을 들어 올리는 모습보다는 치아가 드러나는 웃음으로 구현되었다.",
        "built_space": "뒤쪽 중앙에 컨테이너 전면이 있으며 창 하나, 인물에 일부 가려진 출입문 하나, 문 위 차양 하나가 보인다. 좌우에는 별도의 지붕과 주거 공간이 이어진다. 찰리는 컨테이너 앞 젖은 공터에 있으며 양손을 벌릴 공간은 확보되어 있다. A보다 주변 공간을 포함하지만 다리 아래는 잘렸고, 컨테이너 전면과 주변 살림이 크게 드러나 간결한 부분 외벽 배경보다 복잡하다.",
        "entities": "찰리는 각진 고릴라형 기계 머리와 흰 얼굴판, 낡은 모자와 코트, 벗겨진 가슴 도장을 갖추고 있다. 눈에는 눈동자가 표현되어 점 형태와 다르고 입에는 치아가 드러나 선 형태와 다르다. 가슴의 ‘UBIK’는 읽을 수 있다. 왼쪽 아래에는 다른 사람의 손이 노즐을 잡고 있으며 양쪽 배경에도 여러 성인과 아이가 있다. 추가된 물통 두 개, 장난감 자동차, 빗자루, 빨래와 생활용품도 뚜렷하다. 낮빛과 젖은 금속·직물·물웅덩이는 실물 재질로 표현되었다.",
        "hard_violations": [
         "화면 밖에 있어야 할 호스 사용자의 손이 화면 안에 들어왔다.",
         "찰리 외에 배경의 성인과 아이들을 다수 추가했다.",
         "가슴의 ‘UBIK’ 글자가 읽혀 글자 금지 조건을 위반한다.",
         "장난감 자동차, 물통, 빗자루, 빨래 등 허용된 장소 설명에 없는 소품을 다수 추가했다."
        ],
        "physics": "노즐은 화면 왼쪽 아래의 손이 잡고 있으며 물은 노즐에서 찰리까지 연속적으로 이어진다. 몸에 부딪힌 물방울이 퍼지고 아래로 떨어지는 모습도 타당하다. 양팔과 손의 연결, 머리에 얹힌 모자, 몸에 걸친 코트에 명백한 지지 오류는 없다. 발은 프레임 밖이므로 접지는 확인할 수 없지만 부유하거나 도약하는 자세는 아니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.8,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.55,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 지문에 없는 배경 인물 추가 (invented people)",
     "[gpt-high] 화면 밖에 있어야 할 호스 사용자의 손이 화면 안에 들어왔다.",
     "[gpt-high] 찰리 외에 배경의 성인과 아이들을 다수 추가했다.",
     "[gpt-high] 가슴의 ‘UBIK’ 글자가 읽혀 글자 금지 조건을 위반한다.",
     "[gpt-high] 장난감 자동차, 물통, 빗자루, 빨래 등 허용된 장소 설명에 없는 소품을 다수 추가했다."
    ],
    "B": [
     "[gemini-pro] 지문에 없는 배경 인물 추가 (invented people)",
     "[gpt-high] 찰리만 보여야 하는 장면에 배경 인물들이 추가되었다.",
     "[gpt-high] 가슴의 ‘UBIK’ 글자가 선명하게 읽혀 글자 금지 조건을 위반한다.",
     "[gpt-high] 장소 설명이 허용하지 않은 배경 용기와 상자 등의 소품을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1417,
   "A": 1550
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "캐릭터의 기계 얼굴(점 형태의 눈, 선 형태의 입)은 정확히 구현했으나, 지문에 명시되지 않은 배경 인물을 포함하여 엄격한 인물 제한 규칙을 위반함.  ★위반: [gemini-pro] 지문에 없는 배경 인물 추가 (invented people) / [gpt-high] 찰리만 보여야 하는 장면에 배경 인물들이 추가되었다. / [gpt-high] 가슴의 ‘UBIK’ 글자가 선명하게 읽혀 글자 금지 조건을 위반한다. / [gpt-high] 장소 설명이 허용하지 않은 배경 용기와 상자 등의 소품을 추가했다."
   },
   {
    "label": "A",
    "score": 1550,
    "verdict_ko": "캐릭터 얼굴에 치아를 묘사하여 '선 형태의 입' 설정을 크게 위반했으며, 지문에 없는 인물들을 배경에 다수 추가함.  ★위반: [gemini-pro] 지문에 없는 배경 인물 추가 (invented people) / [gpt-high] 화면 밖에 있어야 할 호스 사용자의 손이 화면 안에 들어왔다. / [gpt-high] 찰리 외에 배경의 성인과 아이들을 다수 추가했다. / [gpt-high] 가슴의 ‘UBIK’ 글자가 읽혀 글자 금지 조건을 위반한다. / [gpt-high] 장난감 자동차, 물통, 빗자루, 빨래 등 허용된 장소 설명에 없는 소품을 다수 추가했다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd8f-0a7a-7072-8691-c8c0e62f176e",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S16sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:23:44.662303+00:00",
  "fingerprint": "ff6bce93517200858ad45e61f1217d1553cd59d5c932b4fccf46468f4536930d",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S16sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S16sh2_sel.png",
  "source_sha256": "f07a138844b49e016b88d41f81f7ca10e8ebfc48118df899f0a65860d23de119",
  "file": "S16sh2_cine.png",
  "staged_sha256": "eabb30d1e8386c972acfb04fc86c42523406c870ff80ab3ed66f3b6c4f1bfa9e",
  "latency_ms": 11706
 },
 "S16sh6::signage": {
  "fp": "ba1d6ca57e8df401",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S16sh6": {
  "input_fingerprint": "38d1b3d67400a4df",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈썹을 위로 치켜세운 채 감탄한 듯 입을 동그랗게 모은 페드로의 얼굴.\n\nLOCATION (lock): At an outdoor observation spot a short distance from the container home where the robot is being washed. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light gently separates the facial expression without adding a new lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hose-washing setup remains outside the container, and Charlie's exterior is wet with dirt being rinsed away. His old coat and hat and peeling UBIK logo have not been replaced. 페드로: He remains at a distance from the washing area.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈썹을 위로 치켜세운 채 감탄한 듯 입을 동그랗게 모은 페드로의 얼굴.\n\nLOCATION (lock): At an outdoor observation spot a short distance from the container home where the robot is being washed. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light gently separates the facial expression without adding a new lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hose-washing setup remains outside the container, and Charlie's exterior is wet with dirt being rinsed away. His old coat and hat and peeling UBIK logo have not been replaced. 페드로: He remains at a distance from the washing area.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈썹을 위로 치켜세운 채 감탄한 듯 입을 동그랗게 모은 페드로의 얼굴.\n\nLOCATION (lock): At an outdoor observation spot a short distance from the container home where the robot is being washed. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light gently separates the facial expression without adding a new lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hose-washing setup remains outside the container, and Charlie's exterior is wet with dirt being rinsed away. His old coat and hat and peeling UBIK logo have not been replaced. 페드로: He remains at a distance from the washing area.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 화면 밖 왼쪽 아래를 향하고 있습니다.",
    "built_space": "레퍼런스 이미지의 컨테이너 건물, 창문, 세발자전거, 흰색 물통 등의 배경 요소가 올바른 위치에 자연스럽게 배치되었습니다.",
    "entities": "페드로의 얼굴 특징과 의상(비니, 가죽 재킷)이 참조 이미지와 완벽히 일치하며, 눈썹을 올리고 입을 동그랗게 모은 표정이 잘 표현되었습니다.",
    "hard_violations": [
     "[gpt-high] 왼쪽 배경에 페드로가 아닌 인물이 남아 있어, 이 장면의 단독 인물 지정과 이전 장면 인물 재등장 금지를 위반했다."
    ],
    "physics": "땅을 딛고 자연스럽게 서 있는 자세입니다."
   },
   {
    "label": "B",
    "direction": "인물의 시선은 화면 밖 왼쪽 아래를 향하고 있습니다.",
    "built_space": "컨테이너 건물과 주변 사물들의 배치가 레퍼런스 이미지의 공간과 잘 일치합니다.",
    "entities": "페드로의 인상착의와 요구된 표정은 훌륭하게 구현되었으나, 프롬프트에 없는 인물이 배경 왼쪽(싱크대 부근)에 추가로 묘사되었습니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 인물이 배경에 창조되어 추가됨",
     "[gpt-high] 페드로만 보여야 하는 장면에 왼쪽 배경의 다른 사람들을 포함해, 이전 장면 인물의 재등장 금지와 추가 인물 금지를 위반했다."
    ],
    "physics": "땅을 딛고 자연스럽게 서 있는 자세입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 페드로의 표정(치켜세운 눈썹, 둥글게 모은 입)을 클로즈업으로 완벽히 구현했으며, 프롬프트의 지시대로 배경에 불필요한 인물을 배제하여 훌륭한 결과물을 만들었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "페드로의 표정과 외양 묘사는 우수하지만, 프롬프트에서 엄격히 금지한 '명시되지 않은 인물'이 배경 왼쪽에 뚜렷하게 등장하여 치명적인 지침 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 화면 밖 왼쪽 아래를 향하고 있습니다.",
        "built_space": "레퍼런스 이미지의 컨테이너 건물, 창문, 세발자전거, 흰색 물통 등의 배경 요소가 올바른 위치에 자연스럽게 배치되었습니다.",
        "entities": "페드로의 얼굴 특징과 의상(비니, 가죽 재킷)이 참조 이미지와 완벽히 일치하며, 눈썹을 올리고 입을 동그랗게 모은 표정이 잘 표현되었습니다.",
        "hard_violations": [],
        "physics": "땅을 딛고 자연스럽게 서 있는 자세입니다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 화면 밖 왼쪽 아래를 향하고 있습니다.",
        "built_space": "컨테이너 건물과 주변 사물들의 배치가 레퍼런스 이미지의 공간과 잘 일치합니다.",
        "entities": "페드로의 인상착의와 요구된 표정은 훌륭하게 구현되었으나, 프롬프트에 없는 인물이 배경 왼쪽(싱크대 부근)에 추가로 묘사되었습니다.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 인물이 배경에 창조되어 추가됨"
        ],
        "physics": "땅을 딛고 자연스럽게 서 있는 자세입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 페드로의 표정(치켜세운 눈썹, 둥글게 모은 입)을 클로즈업으로 완벽히 구현했으며, 프롬프트의 지시대로 배경에 불필요한 인물을 배제하여 훌륭한 결과물을 만들었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "페드로의 표정과 외양 묘사는 우수하지만, 프롬프트에서 엄격히 금지한 '명시되지 않은 인물'이 배경 왼쪽에 뚜렷하게 등장하여 치명적인 지침 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 화면 밖 왼쪽 아래를 향하고 있습니다.",
        "built_space": "레퍼런스 이미지의 컨테이너 건물, 창문, 세발자전거, 흰색 물통 등의 배경 요소가 올바른 위치에 자연스럽게 배치되었습니다.",
        "entities": "페드로의 얼굴 특징과 의상(비니, 가죽 재킷)이 참조 이미지와 완벽히 일치하며, 눈썹을 올리고 입을 동그랗게 모은 표정이 잘 표현되었습니다.",
        "hard_violations": [],
        "physics": "땅을 딛고 자연스럽게 서 있는 자세입니다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 화면 밖 왼쪽 아래를 향하고 있습니다.",
        "built_space": "컨테이너 건물과 주변 사물들의 배치가 레퍼런스 이미지의 공간과 잘 일치합니다.",
        "entities": "페드로의 인상착의와 요구된 표정은 훌륭하게 구현되었으나, 프롬프트에 없는 인물이 배경 왼쪽(싱크대 부근)에 추가로 묘사되었습니다.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 인물이 배경에 창조되어 추가됨"
        ],
        "physics": "땅을 딛고 자연스럽게 서 있는 자세입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴 클로즈업과 둥근 입은 맞지만, 눈썹이 감탄보다 찡그림에 가깝고 금지된 배경 인물들이 남아 있다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "A보다 치켜올린 눈썹과 동그랗게 모은 입, 부드러운 낮빛을 잘 구현했지만, 배경에 다른 인물이 남아 있어 사용 불가 조건에 걸린다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "페드로의 얼굴과 두 눈은 화면 왼쪽 바깥을 향한다. 바라보는 대상은 보이지 않아 찰리의 세척을 보고 있는지는 확인할 수 없지만, 렌즈를 응시하지 않는 관찰 반응으로 읽힌다. 무기나 방향성 소품은 없다.",
        "built_space": "얼굴이 화면 오른쪽을 크게 차지하는 클로즈업이다. 뒤에는 낡은 회청색 컨테이너 한 동, 일부 가려진 창문 한 개, 양옆 경사진 차양, 빨랫줄과 바닥의 흰 물통 두 개가 보인다. 참고 장소의 재료와 배치는 대체로 유지되며, 설비 중복이나 불가능한 반사는 보이지 않는다. 세척 구역과 페드로 사이의 정확한 거리는 이 화면만으로 확정하기 어렵다.",
        "entities": "검은 비니와 낡은 검은 가죽 재킷을 입은 젊은 남성으로, 참고 페드로의 짙은 눈썹과 얼굴 윤곽, 지정된 라틴계 혼혈 청년 외형에 대체로 부합한다. 입술은 둥글게 모았지만 눈썹 사이의 수축이 강해 감탄보다는 의아함에 가깝다. 왼쪽 배경에 서 있는 사람과 허리를 굽힌 사람 등 다른 인물이 보인다. 찰리와 호스는 화면 밖이며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "페드로만 보여야 하는 장면에 왼쪽 배경의 다른 사람들을 포함해, 이전 장면 인물의 재등장 금지와 추가 인물 금지를 위반했다."
        ],
        "physics": "페드로의 머리는 목과 어깨에 정상적으로 이어져 있고 비니와 재킷도 몸에 자연스럽게 걸쳐 있다. 하체는 클로즈업 밖이므로 발의 지지는 확인 대상이 아니다. 물통과 주변 물건은 바닥이나 작업대에 놓여 있으며, 떠 있는 신체나 무지지 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "페드로는 화면 왼쪽 바깥을 바라보며 입을 동그랗게 모으고 있다. 실제 시선 대상은 프레임 밖이라 특정할 수 없지만, 카메라가 아닌 외부 상황에 반응하는 방향이다. 겨누는 물체나 이동 중인 신체는 없다.",
        "built_space": "페드로의 얼굴과 어깨를 담은 클로즈업이며, 배경에는 회청색 컨테이너 한 동과 창문 한 개, 양옆 차양, 빨랫줄, 흰 물통 두 개와 작은 장난감 차량 한 개가 보인다. 젖은 흙바닥과 낡은 금속 외벽은 이전 장소와 잘 이어진다. 배경 물건이 전경 크기로 과장되지 않았고, 중복 설비나 반사 오류는 보이지 않는다.",
        "entities": "검은 비니, 낡은 가죽 재킷, 짙은 눈썹과 젊은 남성의 얼굴이 참고 페드로와 대체로 일치한다. 민족적 정체성을 외형만으로 확정할 수는 없으나 지정된 외형과 충돌하지 않는다. 눈썹 안쪽이 올라가고 입술이 작은 원을 이루어 A보다 요청된 감탄 표정에 가깝다. 왼쪽 배경에는 흐리지만 허리를 굽힌 다른 인물이 남아 있다. 찰리와 세척 호스는 보이지 않으며 판독 가능한 글자도 없다.",
        "hard_violations": [
         "왼쪽 배경에 페드로가 아닌 인물이 남아 있어, 이 장면의 단독 인물 지정과 이전 장면 인물 재등장 금지를 위반했다."
        ],
        "physics": "목과 어깨가 머리를 정상적으로 지지하며, 표정은 자연스러운 눈과 얼굴 근육의 연기로 구현되어 있다. 비니는 머리에, 재킷은 어깨에 걸쳐 있고 배경 물통과 장난감은 지면 위에 놓여 있다. 프레임 밖 하체의 자세는 판단할 수 없으며, 보이는 범위에는 부유하거나 지지가 없는 물체가 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "얼굴 클로즈업과 둥근 입은 맞지만, 눈썹이 감탄보다 찡그림에 가깝고 금지된 배경 인물들이 남아 있다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "A보다 치켜올린 눈썹과 동그랗게 모은 입, 부드러운 낮빛을 잘 구현했지만, 배경에 다른 인물이 남아 있어 사용 불가 조건에 걸린다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "페드로의 얼굴과 두 눈은 화면 왼쪽 바깥을 향한다. 바라보는 대상은 보이지 않아 찰리의 세척을 보고 있는지는 확인할 수 없지만, 렌즈를 응시하지 않는 관찰 반응으로 읽힌다. 무기나 방향성 소품은 없다.",
        "built_space": "얼굴이 화면 오른쪽을 크게 차지하는 클로즈업이다. 뒤에는 낡은 회청색 컨테이너 한 동, 일부 가려진 창문 한 개, 양옆 경사진 차양, 빨랫줄과 바닥의 흰 물통 두 개가 보인다. 참고 장소의 재료와 배치는 대체로 유지되며, 설비 중복이나 불가능한 반사는 보이지 않는다. 세척 구역과 페드로 사이의 정확한 거리는 이 화면만으로 확정하기 어렵다.",
        "entities": "검은 비니와 낡은 검은 가죽 재킷을 입은 젊은 남성으로, 참고 페드로의 짙은 눈썹과 얼굴 윤곽, 지정된 라틴계 혼혈 청년 외형에 대체로 부합한다. 입술은 둥글게 모았지만 눈썹 사이의 수축이 강해 감탄보다는 의아함에 가깝다. 왼쪽 배경에 서 있는 사람과 허리를 굽힌 사람 등 다른 인물이 보인다. 찰리와 호스는 화면 밖이며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "페드로만 보여야 하는 장면에 왼쪽 배경의 다른 사람들을 포함해, 이전 장면 인물의 재등장 금지와 추가 인물 금지를 위반했다."
        ],
        "physics": "페드로의 머리는 목과 어깨에 정상적으로 이어져 있고 비니와 재킷도 몸에 자연스럽게 걸쳐 있다. 하체는 클로즈업 밖이므로 발의 지지는 확인 대상이 아니다. 물통과 주변 물건은 바닥이나 작업대에 놓여 있으며, 떠 있는 신체나 무지지 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "페드로는 화면 왼쪽 바깥을 바라보며 입을 동그랗게 모으고 있다. 실제 시선 대상은 프레임 밖이라 특정할 수 없지만, 카메라가 아닌 외부 상황에 반응하는 방향이다. 겨누는 물체나 이동 중인 신체는 없다.",
        "built_space": "페드로의 얼굴과 어깨를 담은 클로즈업이며, 배경에는 회청색 컨테이너 한 동과 창문 한 개, 양옆 차양, 빨랫줄, 흰 물통 두 개와 작은 장난감 차량 한 개가 보인다. 젖은 흙바닥과 낡은 금속 외벽은 이전 장소와 잘 이어진다. 배경 물건이 전경 크기로 과장되지 않았고, 중복 설비나 반사 오류는 보이지 않는다.",
        "entities": "검은 비니, 낡은 가죽 재킷, 짙은 눈썹과 젊은 남성의 얼굴이 참고 페드로와 대체로 일치한다. 민족적 정체성을 외형만으로 확정할 수는 없으나 지정된 외형과 충돌하지 않는다. 눈썹 안쪽이 올라가고 입술이 작은 원을 이루어 A보다 요청된 감탄 표정에 가깝다. 왼쪽 배경에는 흐리지만 허리를 굽힌 다른 인물이 남아 있다. 찰리와 세척 호스는 보이지 않으며 판독 가능한 글자도 없다.",
        "hard_violations": [
         "왼쪽 배경에 페드로가 아닌 인물이 남아 있어, 이 장면의 단독 인물 지정과 이전 장면 인물 재등장 금지를 위반했다."
        ],
        "physics": "목과 어깨가 머리를 정상적으로 지지하며, 표정은 자연스러운 눈과 얼굴 근육의 연기로 구현되어 있다. 비니는 머리에, 재킷은 어깨에 걸쳐 있고 배경 물통과 장난감은 지면 위에 놓여 있다. 프레임 밖 하체의 자세는 판단할 수 없으며, 보이는 범위에는 부유하거나 지지가 없는 물체가 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.179
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.929
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트에 명시되지 않은 인물이 배경에 창조되어 추가됨",
     "[gpt-high] 페드로만 보여야 하는 장면에 왼쪽 배경의 다른 사람들을 포함해, 이전 장면 인물의 재등장 금지와 추가 인물 금지를 위반했다."
    ],
    "A": [
     "[gpt-high] 왼쪽 배경에 페드로가 아닌 인물이 남아 있어, 이 장면의 단독 인물 지정과 이전 장면 인물 재등장 금지를 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 929
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지시된 페드로의 표정(치켜세운 눈썹, 둥글게 모은 입)을 클로즈업으로 완벽히 구현했으며, 프롬프트의 지시대로 배경에 불필요한 인물을 배제하여 훌륭한 결과물을 만들었습니다.  ★위반: [gpt-high] 왼쪽 배경에 페드로가 아닌 인물이 남아 있어, 이 장면의 단독 인물 지정과 이전 장면 인물 재등장 금지를 위반했다."
   },
   {
    "label": "B",
    "score": 929,
    "verdict_ko": "페드로의 표정과 외양 묘사는 우수하지만, 프롬프트에서 엄격히 금지한 '명시되지 않은 인물'이 배경 왼쪽에 뚜렷하게 등장하여 치명적인 지침 위반이 발생했습니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 인물이 배경에 창조되어 추가됨 / [gpt-high] 페드로만 보여야 하는 장면에 왼쪽 배경의 다른 사람들을 포함해, 이전 장면 인물의 재등장 금지와 추가 인물 금지를 위반했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S16sh2_sel.png",
    "asset_id": "c73a00a4-6fb9-4b7d-97b6-971945d4101a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:826925>",
    "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd8f-cb41-757e-8d34-651683823012",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S16sh2"
  }
 },
 "S16sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:23:56.713278+00:00",
  "fingerprint": "de626da3044d04a1a2b6f17f164f32f815672c86b4f76ee5404cdb9ba8ed1feb",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S16sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S16sh6_sel.png",
  "source_sha256": "a3a560775c32fef54b20bc9b497e41d6d43e5f418f834df436327fcdda7c5b93",
  "file": "S16sh6_cine.png",
  "staged_sha256": "5322be34762eab4ce623168c555a08fb44930f06c7178b0932260adc1e9e91a5",
  "latency_ms": 11470
 },
 "S17sh3::signage": {
  "fp": "3bcf363465690e82",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S17sh3": {
  "input_fingerprint": "38dc202dbdac7b75",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다가 한쪽 손가락 세 개를 쫙 펴 보이며 비열한 미소를 짓는 찰나.\n\nLOCATION (lock): At the owner's repair desk inside the refugee-settlement junk shop, in daytime ambient light among old appliances. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Desk (In front of 마츠다 before he collapses onto it) — The visitor-facing edge runs obliquely across the bottom of the image; used as Maintains the dialogue axis and establishes the surface involved in the following action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime interior ambient light preserves clear separation between the raised fingers and the calculating smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Appliances and scrap remain piled in the shop, and a television is under repair. Charlie is not stored in the shop. 마츠다: He remains in the shop after interrupting his television repair.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다가 한쪽 손가락 세 개를 쫙 펴 보이며 비열한 미소를 짓는 찰나.\n\nLOCATION (lock): At the owner's repair desk inside the refugee-settlement junk shop, in daytime ambient light among old appliances. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Desk (In front of 마츠다 before he collapses onto it) — The visitor-facing edge runs obliquely across the bottom of the image; used as Maintains the dialogue axis and establishes the surface involved in the following action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime interior ambient light preserves clear separation between the raised fingers and the calculating smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Appliances and scrap remain piled in the shop, and a television is under repair. Charlie is not stored in the shop. 마츠다: He remains in the shop after interrupting his television repair.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다가 한쪽 손가락 세 개를 쫙 펴 보이며 비열한 미소를 짓는 찰나.\n\nLOCATION (lock): At the owner's repair desk inside the refugee-settlement junk shop, in daytime ambient light among old appliances. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Desk (In front of 마츠다 before he collapses onto it) — The visitor-facing edge runs obliquely across the bottom of the image; used as Maintains the dialogue axis and establishes the surface involved in the following action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime interior ambient light preserves clear separation between the raised fingers and the calculating smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Appliances and scrap remain piled in the shop, and a television is under repair. Charlie is not stored in the shop. 마츠다: He remains in the shop after interrupting his television repair.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "마츠다의 시선과 손 방향이 화면 우측 전경에 위치한 알 수 없는 인물을 향하고 있음.",
    "built_space": "고물상 내부. 좌측에 냉장고, 우측에 TV 선반, 후경에 출입문이 있으며 전경의 책상이 사선으로 배치되어 참조 이미지의 공간과 일치함.",
    "entities": "마츠다의 얼굴과 의상은 참조 이미지와 일치하고 손가락 세 개를 펴고 있으나, 샷 텍스트에 명시되지 않은 타인이 화면 우측에 추가됨.",
    "hard_violations": [
     "[gemini-pro] 발명된 인물 (우측 전경에 샷 텍스트에 없는 타인 등장)",
     "[gpt-high] 마츠다만 등장해야 하는 장면에 별도의 남성을 오른쪽 전경의 머리와 상체로 추가했습니다."
    ],
    "physics": "마츠다는 책상 앞에 앉아 체중을 지탱하고 있으나, 오른팔이 가슴을 무리하게 가로질러 뻗은 자세가 다소 부자연스러움."
   },
   {
    "label": "B",
    "direction": "마츠다의 시선과 올려진 오른손이 화면 좌측 전경에 위치한 알 수 없는 인물을 향함.",
    "built_space": "참조 이미지와 동일한 고물상 내부 구조(우측 TV 선반, 좌측 냉장고, 후경 출입문)를 보여주며, 전경에 사선으로 놓인 수리용 책상 배치도 적절함.",
    "entities": "마츠다의 외모와 의상이 참조와 일치하며 손가락 세 개를 펴고 비열한 미소를 짓고 있으나, 명시되지 않은 타인이 좌측에 추가됨.",
    "hard_violations": [
     "[gemini-pro] 발명된 인물 (좌측 전경에 샷 텍스트에 없는 타인 등장)",
     "[gpt-high] 마츠다만 보여야 하는 장면에 별도의 남성 방문객을 왼쪽 전경의 얼굴·상체·손으로 추가했습니다."
    ],
    "physics": "책상 앞의 자세와 들어 올린 오른팔의 움직임이 물리적으로 자연스럽고 안정적으로 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시된 비열한 미소와 손동작을 자연스럽게 연출했으나, 프롬프트에 없는 인물이 전경에 추가되어 치명적인 규정 위반이 발생했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오른팔이 몸을 가로지르는 어색한 자세를 취하고 있으며, 역시 프롬프트에 명시되지 않은 인물이 화면에 등장해 치명적 위반입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "마츠다의 시선과 손 방향이 화면 우측 전경에 위치한 알 수 없는 인물을 향하고 있음.",
        "built_space": "고물상 내부. 좌측에 냉장고, 우측에 TV 선반, 후경에 출입문이 있으며 전경의 책상이 사선으로 배치되어 참조 이미지의 공간과 일치함.",
        "entities": "마츠다의 얼굴과 의상은 참조 이미지와 일치하고 손가락 세 개를 펴고 있으나, 샷 텍스트에 명시되지 않은 타인이 화면 우측에 추가됨.",
        "hard_violations": [
         "발명된 인물 (우측 전경에 샷 텍스트에 없는 타인 등장)"
        ],
        "physics": "마츠다는 책상 앞에 앉아 체중을 지탱하고 있으나, 오른팔이 가슴을 무리하게 가로질러 뻗은 자세가 다소 부자연스러움."
       },
       {
        "label": "B",
        "direction": "마츠다의 시선과 올려진 오른손이 화면 좌측 전경에 위치한 알 수 없는 인물을 향함.",
        "built_space": "참조 이미지와 동일한 고물상 내부 구조(우측 TV 선반, 좌측 냉장고, 후경 출입문)를 보여주며, 전경에 사선으로 놓인 수리용 책상 배치도 적절함.",
        "entities": "마츠다의 외모와 의상이 참조와 일치하며 손가락 세 개를 펴고 비열한 미소를 짓고 있으나, 명시되지 않은 타인이 좌측에 추가됨.",
        "hard_violations": [
         "발명된 인물 (좌측 전경에 샷 텍스트에 없는 타인 등장)"
        ],
        "physics": "책상 앞의 자세와 들어 올린 오른팔의 움직임이 물리적으로 자연스럽고 안정적으로 지탱됨."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시된 비열한 미소와 손동작을 자연스럽게 연출했으나, 프롬프트에 없는 인물이 전경에 추가되어 치명적인 규정 위반이 발생했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오른팔이 몸을 가로지르는 어색한 자세를 취하고 있으며, 역시 프롬프트에 명시되지 않은 인물이 화면에 등장해 치명적 위반입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "마츠다의 시선과 손 방향이 화면 우측 전경에 위치한 알 수 없는 인물을 향하고 있음.",
        "built_space": "고물상 내부. 좌측에 냉장고, 우측에 TV 선반, 후경에 출입문이 있으며 전경의 책상이 사선으로 배치되어 참조 이미지의 공간과 일치함.",
        "entities": "마츠다의 얼굴과 의상은 참조 이미지와 일치하고 손가락 세 개를 펴고 있으나, 샷 텍스트에 명시되지 않은 타인이 화면 우측에 추가됨.",
        "hard_violations": [
         "발명된 인물 (우측 전경에 샷 텍스트에 없는 타인 등장)"
        ],
        "physics": "마츠다는 책상 앞에 앉아 체중을 지탱하고 있으나, 오른팔이 가슴을 무리하게 가로질러 뻗은 자세가 다소 부자연스러움."
       },
       {
        "label": "B",
        "direction": "마츠다의 시선과 올려진 오른손이 화면 좌측 전경에 위치한 알 수 없는 인물을 향함.",
        "built_space": "참조 이미지와 동일한 고물상 내부 구조(우측 TV 선반, 좌측 냉장고, 후경 출입문)를 보여주며, 전경에 사선으로 놓인 수리용 책상 배치도 적절함.",
        "entities": "마츠다의 외모와 의상이 참조와 일치하며 손가락 세 개를 펴고 비열한 미소를 짓고 있으나, 명시되지 않은 타인이 좌측에 추가됨.",
        "hard_violations": [
         "발명된 인물 (좌측 전경에 샷 텍스트에 없는 타인 등장)"
        ],
        "physics": "책상 앞의 자세와 들어 올린 오른팔의 움직임이 물리적으로 자연스럽고 안정적으로 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "세 손가락을 펼친 손짓, 계산적인 미소와 중간 샷은 더 충실하지만, 왼쪽 전경에 허용되지 않은 상대 인물을 추가해 탈락 사유가 있습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "세 손가락과 미소는 구현했지만, 오른쪽에 금지된 상대 인물을 추가했고 목재 작업대도 이전 장면의 금속성 작업면과 달라집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "마츠다는 화면 왼쪽 전경의 남성을 바라보며 웃습니다. 얼굴 옆에 세 손가락을 위로 펼치고 손바닥을 상대와 카메라 쪽으로 보여 줍니다. 손짓과 미소가 서로 가리지 않지만, 시선의 실제 대상인 남성은 이 장면에 등장하면 안 됩니다.",
        "built_space": "마츠다는 작업대 뒤에 허리 부근까지 보이는 중간 샷으로 배치됩니다. 금속성 작업대 하나가 화면 하단을 비스듬히 가로지릅니다. 뒤 중앙에는 유리 출입구 하나, 오른쪽에는 여러 층의 가전 선반과 식별 가능한 전면식 세탁기 한 대 및 선풍기 한 대가 있습니다. 왼쪽에도 오래된 가전이 쌓여 있어 이전 장면의 통로형 상점과 대체로 이어집니다. 다만 작업대 방문객 쪽에 추가 남성이 서 있습니다.",
        "entities": "노년의 동아시아계 남성 한 명이 마츠다로 보이며, 회색 섞인 짧은 머리와 얼굴, 체크 셔츠 및 때 묻은 갈색 조끼가 이전 장면과 잘 맞습니다. 일본 국적 자체는 외관으로 확인할 수 없습니다. 한 손에는 정확히 세 손가락이 펴져 있습니다. 별도의 성인 남성 얼굴과 상체, 손 일부가 왼쪽에 추가되었습니다. 낡은 텔레비전과 가전, 작업대 위 전선과 공구는 보이지만 수리 중인 텔레비전은 명확히 식별되지 않습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "마츠다만 보여야 하는 장면에 별도의 남성 방문객을 왼쪽 전경의 얼굴·상체·손으로 추가했습니다."
        ],
        "physics": "들어 올린 손은 손목과 굽힌 팔을 통해 몸에 자연스럽게 연결되고, 나머지 두 손가락을 접은 자세도 가능합니다. 하체는 작업대에 가려져 있지만 공중에 떠 있다는 징후는 없습니다. 전선과 공구는 작업대에, 가전은 선반이나 바닥에 놓여 있습니다. 추가 인물도 전경에서 잘린 서 있는 자세로 보이며, 부유 물체나 불가능한 관절은 확인되지 않습니다."
       },
       {
        "label": "B",
        "direction": "마츠다는 오른쪽 전경 남성의 얼굴을 바라보며 웃습니다. 들어 올린 손의 세 손가락은 오른쪽 위로 벌어져 상대를 향하고, 카메라에는 주로 손등이 보입니다. 세 손가락과 얼굴은 구분되지만 시선과 손짓의 대상인 남성은 허용되지 않은 인물입니다.",
        "built_space": "마츠다는 작업대 뒤에서 상체를 앞으로 기울인 중간 샷으로 보입니다. 하단에는 비스듬한 모서리를 가진 작업대 하나가 있으나 작업면은 거친 갈색 목재입니다. 오른쪽 벽에는 여러 층의 텔레비전 선반과 선풍기 한 대, 왼쪽 통로에는 냉장고류 및 전면식 세탁기 한 대가 보입니다. 낮빛이 들어오는 상점이라는 특징은 유지하지만 작업대 재질은 이전 장면의 금속성 작업면과 다릅니다. 추가 남성이 방문객 쪽 작업대 앞을 크게 차지합니다.",
        "entities": "마츠다는 회색 섞인 머리의 노년 동아시아계 남성으로, 체크 셔츠와 갈색 작업 조끼 및 얼굴의 전반적인 특징이 참조와 부합합니다. 일본 국적은 외관만으로 확인할 수 없습니다. 한 손에 세 손가락을 펼쳤고 다른 손은 작업대 위에 놓았습니다. 오른쪽에는 별도의 성인 남성 머리와 상체가 있습니다. 오래된 텔레비전과 가전, 전선과 공구는 있지만 특정 텔레비전의 수리 상태는 분명하지 않습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "마츠다만 등장해야 하는 장면에 별도의 남성을 오른쪽 전경의 머리와 상체로 추가했습니다."
        ],
        "physics": "마츠다의 한쪽 팔꿈치와 반대쪽 손이 작업대에 닿아 앞으로 기울인 상체를 지지합니다. 세 손가락을 펼친 손은 팔에 정상적으로 연결되어 있고 가능한 동작입니다. 공구와 전선은 작업대에 놓여 있으며 뒤쪽 가전도 선반이나 바닥의 지지를 받습니다. 하체는 가려져 있고, 지지 없이 떠 있는 몸이나 물체는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "세 손가락을 펼친 손짓, 계산적인 미소와 중간 샷은 더 충실하지만, 왼쪽 전경에 허용되지 않은 상대 인물을 추가해 탈락 사유가 있습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "세 손가락과 미소는 구현했지만, 오른쪽에 금지된 상대 인물을 추가했고 목재 작업대도 이전 장면의 금속성 작업면과 달라집니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "마츠다는 화면 왼쪽 전경의 남성을 바라보며 웃습니다. 얼굴 옆에 세 손가락을 위로 펼치고 손바닥을 상대와 카메라 쪽으로 보여 줍니다. 손짓과 미소가 서로 가리지 않지만, 시선의 실제 대상인 남성은 이 장면에 등장하면 안 됩니다.",
        "built_space": "마츠다는 작업대 뒤에 허리 부근까지 보이는 중간 샷으로 배치됩니다. 금속성 작업대 하나가 화면 하단을 비스듬히 가로지릅니다. 뒤 중앙에는 유리 출입구 하나, 오른쪽에는 여러 층의 가전 선반과 식별 가능한 전면식 세탁기 한 대 및 선풍기 한 대가 있습니다. 왼쪽에도 오래된 가전이 쌓여 있어 이전 장면의 통로형 상점과 대체로 이어집니다. 다만 작업대 방문객 쪽에 추가 남성이 서 있습니다.",
        "entities": "노년의 동아시아계 남성 한 명이 마츠다로 보이며, 회색 섞인 짧은 머리와 얼굴, 체크 셔츠 및 때 묻은 갈색 조끼가 이전 장면과 잘 맞습니다. 일본 국적 자체는 외관으로 확인할 수 없습니다. 한 손에는 정확히 세 손가락이 펴져 있습니다. 별도의 성인 남성 얼굴과 상체, 손 일부가 왼쪽에 추가되었습니다. 낡은 텔레비전과 가전, 작업대 위 전선과 공구는 보이지만 수리 중인 텔레비전은 명확히 식별되지 않습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "마츠다만 보여야 하는 장면에 별도의 남성 방문객을 왼쪽 전경의 얼굴·상체·손으로 추가했습니다."
        ],
        "physics": "들어 올린 손은 손목과 굽힌 팔을 통해 몸에 자연스럽게 연결되고, 나머지 두 손가락을 접은 자세도 가능합니다. 하체는 작업대에 가려져 있지만 공중에 떠 있다는 징후는 없습니다. 전선과 공구는 작업대에, 가전은 선반이나 바닥에 놓여 있습니다. 추가 인물도 전경에서 잘린 서 있는 자세로 보이며, 부유 물체나 불가능한 관절은 확인되지 않습니다."
       },
       {
        "label": "A",
        "direction": "마츠다는 오른쪽 전경 남성의 얼굴을 바라보며 웃습니다. 들어 올린 손의 세 손가락은 오른쪽 위로 벌어져 상대를 향하고, 카메라에는 주로 손등이 보입니다. 세 손가락과 얼굴은 구분되지만 시선과 손짓의 대상인 남성은 허용되지 않은 인물입니다.",
        "built_space": "마츠다는 작업대 뒤에서 상체를 앞으로 기울인 중간 샷으로 보입니다. 하단에는 비스듬한 모서리를 가진 작업대 하나가 있으나 작업면은 거친 갈색 목재입니다. 오른쪽 벽에는 여러 층의 텔레비전 선반과 선풍기 한 대, 왼쪽 통로에는 냉장고류 및 전면식 세탁기 한 대가 보입니다. 낮빛이 들어오는 상점이라는 특징은 유지하지만 작업대 재질은 이전 장면의 금속성 작업면과 다릅니다. 추가 남성이 방문객 쪽 작업대 앞을 크게 차지합니다.",
        "entities": "마츠다는 회색 섞인 머리의 노년 동아시아계 남성으로, 체크 셔츠와 갈색 작업 조끼 및 얼굴의 전반적인 특징이 참조와 부합합니다. 일본 국적은 외관만으로 확인할 수 없습니다. 한 손에 세 손가락을 펼쳤고 다른 손은 작업대 위에 놓았습니다. 오른쪽에는 별도의 성인 남성 머리와 상체가 있습니다. 오래된 텔레비전과 가전, 전선과 공구는 있지만 특정 텔레비전의 수리 상태는 분명하지 않습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "마츠다만 등장해야 하는 장면에 별도의 남성을 오른쪽 전경의 머리와 상체로 추가했습니다."
        ],
        "physics": "마츠다의 한쪽 팔꿈치와 반대쪽 손이 작업대에 닿아 앞으로 기울인 상체를 지지합니다. 세 손가락을 펼친 손은 팔에 정상적으로 연결되어 있고 가능한 동작입니다. 공구와 전선은 작업대에 놓여 있으며 뒤쪽 가전도 선반이나 바닥의 지지를 받습니다. 하체는 가려져 있고, 지지 없이 떠 있는 몸이나 물체는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.417,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.167,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 발명된 인물 (우측 전경에 샷 텍스트에 없는 타인 등장)",
     "[gpt-high] 마츠다만 등장해야 하는 장면에 별도의 남성을 오른쪽 전경의 머리와 상체로 추가했습니다."
    ],
    "B": [
     "[gemini-pro] 발명된 인물 (좌측 전경에 샷 텍스트에 없는 타인 등장)",
     "[gpt-high] 마츠다만 보여야 하는 장면에 별도의 남성 방문객을 왼쪽 전경의 얼굴·상체·손으로 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 1167
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지시된 비열한 미소와 손동작을 자연스럽게 연출했으나, 프롬프트에 없는 인물이 전경에 추가되어 치명적인 규정 위반이 발생했습니다.  ★위반: [gemini-pro] 발명된 인물 (좌측 전경에 샷 텍스트에 없는 타인 등장) / [gpt-high] 마츠다만 보여야 하는 장면에 별도의 남성 방문객을 왼쪽 전경의 얼굴·상체·손으로 추가했습니다."
   },
   {
    "label": "A",
    "score": 1167,
    "verdict_ko": "오른팔이 몸을 가로지르는 어색한 자세를 취하고 있으며, 역시 프롬프트에 명시되지 않은 인물이 화면에 등장해 치명적 위반입니다.  ★위반: [gemini-pro] 발명된 인물 (우측 전경에 샷 텍스트에 없는 타인 등장) / [gpt-high] 마츠다만 등장해야 하는 장면에 별도의 남성을 오른쪽 전경의 머리와 상체로 추가했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S15sh13_sel.png",
    "asset_id": "b3769162-2a5b-4eaa-ac42-62a2014c870f",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 마츠다: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876556>",
    "asset_id": "8e5adf95-2c7b-4499-9815-c7a393ac5668",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd90-8818-72a9-b082-62c16b785ad5",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S15sh13"
  }
 },
 "S17sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:24:08.517037+00:00",
  "fingerprint": "6e86b737e77aa2cc5f23b28f03e8c02043e6b19ac9612e5530af3e323273b9b8",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S17sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S17sh3_sel.png",
  "source_sha256": "5c10481a26f02042a47d504545d8b6d98dc948989492d9b46146a772812d5fae",
  "file": "S17sh3_cine.png",
  "staged_sha256": "3a2ab4afe9e96b680aa1edcf545a68a5f6d47c3ee4290b114d5984973d534d78",
  "latency_ms": 11919
 },
 "S17sh7::signage": {
  "fp": "eed5b5c8e42cd33f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S17sh7": {
  "input_fingerprint": "8e6ce6715a3b3aa1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다의 찢어진 가슴팍 옷자락에서 붉은 피가 공중에 튀어 오른 타격의 순간.\n\nLOCATION (lock): Beside the repair desk inside the junk shop, surrounded by old electrical goods in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Desk (Immediately beneath 마츠다's falling upper body) — A narrow portion of the top and visitor-facing edge is visible below his chest; used as Supplies the destination of the collapse and a stable spatial reference beneath the impact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding interior illumination, allowing the described red blood to register without a separate impact-light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The television under repair and the shop's piles of appliances and scrap remain in place. 마츠다: He has been shot and is collapsing onto the desk; the text does not establish a wound location or a specific blood pattern.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다의 찢어진 가슴팍 옷자락에서 붉은 피가 공중에 튀어 오른 타격의 순간.\n\nLOCATION (lock): Beside the repair desk inside the junk shop, surrounded by old electrical goods in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Desk (Immediately beneath 마츠다's falling upper body) — A narrow portion of the top and visitor-facing edge is visible below his chest; used as Supplies the destination of the collapse and a stable spatial reference beneath the impact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding interior illumination, allowing the described red blood to register without a separate impact-light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The television under repair and the shop's piles of appliances and scrap remain in place. 마츠다: He has been shot and is collapsing onto the desk; the text does not establish a wound location or a specific blood pattern.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다의 찢어진 가슴팍 옷자락에서 붉은 피가 공중에 튀어 오른 타격의 순간.\n\nLOCATION (lock): Beside the repair desk inside the junk shop, surrounded by old electrical goods in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Desk (Immediately beneath 마츠다's falling upper body) — A narrow portion of the top and visitor-facing edge is visible below his chest; used as Supplies the destination of the collapse and a stable spatial reference beneath the impact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding interior illumination, allowing the described red blood to register without a separate impact-light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The television under repair and the shop's piles of appliances and scrap remain in place. 마츠다: He has been shot and is collapsing onto the desk; the text does not establish a wound location or a specific blood pattern.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성 얼굴, 60대, 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 아래쪽 책상을 향하며, 피는 가슴에서 수평으로 뿜어짐.",
    "built_space": "고물상 내부. 전면에 책상과 공구, 후면에 가전제품 선반이 위치해 레퍼런스와 일치함.",
    "entities": "마츠다의 외형은 일치하나 의상이 찢어지지 않았으며, 왼쪽 벽에 읽을 수 있는 간판 텍스트가 있음.",
    "hard_violations": [
     "[gpt-high] 왼쪽 아래에 이전 스틸의 다른 사람으로 읽히는 어깨와 옷을 남겨, 마츠다 외 인물의 등장을 금지한 조건을 위반했습니다.",
     "[gpt-high] 왼쪽 위 벽 간판에 판독 가능한 한글 일부가 노출되어, 읽을 수 있는 글자를 전면 금지한 조건을 위반했습니다."
    ],
    "physics": "피는 공중에 흩뿌려진 상태이며, 상체는 앞으로 기울어져 프레임 밖의 다리로 지탱됨."
   },
   {
    "label": "B",
    "direction": "시선은 앞을 향하고, 핏방울은 가슴에서 위쪽으로 튀어 오름.",
    "built_space": "고물상 내부. 앞쪽에 책상과 공구, 뒤쪽에 가전제품 선반이 레퍼런스와 완벽히 일치하게 배치됨.",
    "entities": "마츠다의 외형이 일치하고 조끼와 셔츠 가슴팍이 찢어져 있으며, 배경 텍스트는 판독 불가능함.",
    "hard_violations": [
     "[gpt-high] 왼쪽 아래의 다른 사람 어깨와 화면 하단의 손을 남겨, 이전 스틸의 다른 인물을 이 장면에서 제외하라는 명시적 조건을 위반했습니다."
    ],
    "physics": "핏방울이 공중으로 솟아오르며, 상체는 충격으로 약간 뒤로 젖혀진 채 다리로 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 '찢어진 가슴팍 옷자락'과 '위로 튀어 오르는 피'를 정확히 묘사했으며, 배경의 텍스트를 판독 불가하게 처리하여 프롬프트를 훌륭하게 준수했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "옷자락이 찢어지지 않았고 피가 위로 튀어 오르지 않으며, 왼쪽 벽에 읽을 수 있는 한글 텍스트가 명확히 존재해 부정 프롬프트를 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 아래쪽 책상을 향하며, 피는 가슴에서 수평으로 뿜어짐.",
        "built_space": "고물상 내부. 전면에 책상과 공구, 후면에 가전제품 선반이 위치해 레퍼런스와 일치함.",
        "entities": "마츠다의 외형은 일치하나 의상이 찢어지지 않았으며, 왼쪽 벽에 읽을 수 있는 간판 텍스트가 있음.",
        "hard_violations": [],
        "physics": "피는 공중에 흩뿌려진 상태이며, 상체는 앞으로 기울어져 프레임 밖의 다리로 지탱됨."
       },
       {
        "label": "B",
        "direction": "시선은 앞을 향하고, 핏방울은 가슴에서 위쪽으로 튀어 오름.",
        "built_space": "고물상 내부. 앞쪽에 책상과 공구, 뒤쪽에 가전제품 선반이 레퍼런스와 완벽히 일치하게 배치됨.",
        "entities": "마츠다의 외형이 일치하고 조끼와 셔츠 가슴팍이 찢어져 있으며, 배경 텍스트는 판독 불가능함.",
        "hard_violations": [],
        "physics": "핏방울이 공중으로 솟아오르며, 상체는 충격으로 약간 뒤로 젖혀진 채 다리로 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 '찢어진 가슴팍 옷자락'과 '위로 튀어 오르는 피'를 정확히 묘사했으며, 배경의 텍스트를 판독 불가하게 처리하여 프롬프트를 훌륭하게 준수했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "옷자락이 찢어지지 않았고 피가 위로 튀어 오르지 않으며, 왼쪽 벽에 읽을 수 있는 한글 텍스트가 명확히 존재해 부정 프롬프트를 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 아래쪽 책상을 향하며, 피는 가슴에서 수평으로 뿜어짐.",
        "built_space": "고물상 내부. 전면에 책상과 공구, 후면에 가전제품 선반이 위치해 레퍼런스와 일치함.",
        "entities": "마츠다의 외형은 일치하나 의상이 찢어지지 않았으며, 왼쪽 벽에 읽을 수 있는 간판 텍스트가 있음.",
        "hard_violations": [],
        "physics": "피는 공중에 흩뿌려진 상태이며, 상체는 앞으로 기울어져 프레임 밖의 다리로 지탱됨."
       },
       {
        "label": "B",
        "direction": "시선은 앞을 향하고, 핏방울은 가슴에서 위쪽으로 튀어 오름.",
        "built_space": "고물상 내부. 앞쪽에 책상과 공구, 뒤쪽에 가전제품 선반이 레퍼런스와 완벽히 일치하게 배치됨.",
        "entities": "마츠다의 외형이 일치하고 조끼와 셔츠 가슴팍이 찢어져 있으며, 배경 텍스트는 판독 불가능함.",
        "hard_violations": [],
        "physics": "핏방울이 공중으로 솟아오르며, 상체는 충격으로 약간 뒤로 젖혀진 채 다리로 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "상승하는 혈액과 이전 의상은 맞지만, 금지된 다른 사람의 어깨와 손이 남았고 클로즈업 대신 넓은 상반신 구도로 책상 위 붕괴도 약합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "다른 사람의 어깨와 판독 가능한 간판 글자 때문에 부적격이지만, 두 후보 중 가슴의 타격과 책상으로 무너지는 상체의 관계는 더 충실합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "마츠다는 머리를 화면 오른쪽으로 기울이고 시선을 전방 아래로 둡니다. 상체는 아직 거의 직립하여 바로 아래 책상으로 쓰러지는 방향이 뚜렷하지 않습니다. 피는 가슴의 찢어진 부분에서 화면 왼쪽 위로 뻗어 올라가므로 공중으로 튀어 오르는 방향은 맞습니다. 총이나 조준 대상은 보이지 않습니다.",
        "built_space": "전경의 수리 책상 한 개, 뒤쪽 유리 출입구 한 곳, 오른쪽 냉장고 한 대와 나란한 전면 투입식 세탁기 두 대, 여러 층의 텔레비전 선반이 보입니다. 왼쪽에도 가전 적재대와 작은 작업대가 있어 이전 장소와 대체로 일치합니다. 다만 책상 상판과 주변 가게가 넓게 드러나며, 가슴 바로 아래 상판과 손님 쪽 가장자리만 좁게 보이는 클로즈업이 아닙니다. 왼쪽 아래에는 책상 반대편 사람의 어깨와 하단의 손 일부가 남아 있습니다.",
        "entities": "중심 인물은 짧은 회흑색 머리의 노년 동아시아계 남성으로, 이전 스틸의 마츠다 얼굴과 베이지 작업 조끼, 어두운 체크 셔츠를 대체로 유지합니다. 외관만으로 일본 국적을 확인할 수는 없습니다. 찢어진 가슴 옷과 붉은 혈액, 낡은 전자제품과 수리 부품이 보입니다. 별도 인물의 어깨와 손은 마츠다만 등장해야 한다는 조건과 맞지 않습니다. 수리 중인 특정 텔레비전의 동일성은 이 화면에서 확정하기 어렵습니다.",
        "hard_violations": [
         "왼쪽 아래의 다른 사람 어깨와 화면 하단의 손을 남겨, 이전 스틸의 다른 인물을 이 장면에서 제외하라는 명시적 조건을 위반했습니다."
        ],
        "physics": "하체는 책상 뒤에 가려져 발의 지지는 확인되지 않지만, 몸통이 화면 아래로 이어지는 직립 자세여서 공중에 떠 있는 인체는 아닙니다. 팔도 아래로 이어지며 손의 접촉점은 가려져 있습니다. 가슴에서 이어지는 혈액 줄기와 분리된 물방울은 충격으로 분출된 순간으로 해석 가능합니다. 책상은 붕괴를 받을 위치에 있으나 아직 상체나 손이 닿는 모습은 확인되지 않습니다. 전자제품과 공구는 선반 또는 상판에 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "마츠다의 얼굴과 시선은 화면 왼쪽 아래의 책상을 향하고, 상체도 앞으로 굽어 책상 쪽으로 내려갑니다. 따라서 붕괴의 목적지가 명확합니다. 피는 가슴에서 주로 화면 왼쪽과 약간 아래쪽으로 퍼져, 위로 솟는 방향은 A보다 약합니다. 무기나 조준 행동은 없습니다.",
        "built_space": "수리 책상 한 개가 전경에 있고 뒤쪽 유리 출입구 한 곳, 오른쪽 냉장고 한 대와 전면 투입식 세탁기 두 대, 다층 텔레비전 선반이 이전 장소의 배치를 이어갑니다. 마츠다의 가슴은 책상 위로 기울어 붕괴 위치가 맞습니다. 그러나 머리부터 허리 부근과 가게 양쪽까지 보여 여전히 요구한 클로즈업보다 넓고, 상판도 좁은 띠 이상으로 노출됩니다. 왼쪽 아래에는 다른 사람의 어깨가 있으며 왼쪽 위 벽에는 글자가 보이는 간판이 추가되어 있습니다.",
        "entities": "노년 동아시아계 남성 한 명의 얼굴, 짧은 회흑색 머리, 베이지 조끼와 체크 셔츠는 이전 스틸의 마츠다와 대체로 맞습니다. 일본인이라는 국적은 영상 외관만으로 검증할 수 없습니다. 가슴에는 찢어진 옷과 붉은 혈액 분출이 있고, 주변에는 낡은 텔레비전과 수리 부품이 있습니다. 왼쪽 아래의 별도 인물 어깨는 허용되지 않습니다. 왼쪽 위 간판은 한글 일부가 판독 가능한 크기로 드러나 무문자 조건도 어깁니다.",
        "hard_violations": [
         "왼쪽 아래에 이전 스틸의 다른 사람으로 읽히는 어깨와 옷을 남겨, 마츠다 외 인물의 등장을 금지한 조건을 위반했습니다.",
         "왼쪽 위 벽 간판에 판독 가능한 한글 일부가 노출되어, 읽을 수 있는 글자를 전면 금지한 조건을 위반했습니다."
        ],
        "physics": "상체가 허리에서 앞으로 꺾이고 하체는 화면 아래로 이어져, 총격 뒤 지지를 잃고 책상으로 넘어지는 중간 동작으로 가능합니다. 발과 손은 가려져 실제 접촉점은 확인되지 않지만 몸 전체가 떠 있는 자세는 아닙니다. 가슴과 책상 사이에 남은 간격도 접촉 직전 순간으로 자연스럽습니다. 혈액은 가슴의 분출 지점과 연결된 줄기 및 떨어져 나온 방울로 보여 충격에 의한 운동 근거가 있습니다. 가전제품과 공구는 선반과 책상에 지지되어 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "상승하는 혈액과 이전 의상은 맞지만, 금지된 다른 사람의 어깨와 손이 남았고 클로즈업 대신 넓은 상반신 구도로 책상 위 붕괴도 약합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "다른 사람의 어깨와 판독 가능한 간판 글자 때문에 부적격이지만, 두 후보 중 가슴의 타격과 책상으로 무너지는 상체의 관계는 더 충실합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "마츠다는 머리를 화면 오른쪽으로 기울이고 시선을 전방 아래로 둡니다. 상체는 아직 거의 직립하여 바로 아래 책상으로 쓰러지는 방향이 뚜렷하지 않습니다. 피는 가슴의 찢어진 부분에서 화면 왼쪽 위로 뻗어 올라가므로 공중으로 튀어 오르는 방향은 맞습니다. 총이나 조준 대상은 보이지 않습니다.",
        "built_space": "전경의 수리 책상 한 개, 뒤쪽 유리 출입구 한 곳, 오른쪽 냉장고 한 대와 나란한 전면 투입식 세탁기 두 대, 여러 층의 텔레비전 선반이 보입니다. 왼쪽에도 가전 적재대와 작은 작업대가 있어 이전 장소와 대체로 일치합니다. 다만 책상 상판과 주변 가게가 넓게 드러나며, 가슴 바로 아래 상판과 손님 쪽 가장자리만 좁게 보이는 클로즈업이 아닙니다. 왼쪽 아래에는 책상 반대편 사람의 어깨와 하단의 손 일부가 남아 있습니다.",
        "entities": "중심 인물은 짧은 회흑색 머리의 노년 동아시아계 남성으로, 이전 스틸의 마츠다 얼굴과 베이지 작업 조끼, 어두운 체크 셔츠를 대체로 유지합니다. 외관만으로 일본 국적을 확인할 수는 없습니다. 찢어진 가슴 옷과 붉은 혈액, 낡은 전자제품과 수리 부품이 보입니다. 별도 인물의 어깨와 손은 마츠다만 등장해야 한다는 조건과 맞지 않습니다. 수리 중인 특정 텔레비전의 동일성은 이 화면에서 확정하기 어렵습니다.",
        "hard_violations": [
         "왼쪽 아래의 다른 사람 어깨와 화면 하단의 손을 남겨, 이전 스틸의 다른 인물을 이 장면에서 제외하라는 명시적 조건을 위반했습니다."
        ],
        "physics": "하체는 책상 뒤에 가려져 발의 지지는 확인되지 않지만, 몸통이 화면 아래로 이어지는 직립 자세여서 공중에 떠 있는 인체는 아닙니다. 팔도 아래로 이어지며 손의 접촉점은 가려져 있습니다. 가슴에서 이어지는 혈액 줄기와 분리된 물방울은 충격으로 분출된 순간으로 해석 가능합니다. 책상은 붕괴를 받을 위치에 있으나 아직 상체나 손이 닿는 모습은 확인되지 않습니다. 전자제품과 공구는 선반 또는 상판에 놓여 있습니다."
       },
       {
        "label": "A",
        "direction": "마츠다의 얼굴과 시선은 화면 왼쪽 아래의 책상을 향하고, 상체도 앞으로 굽어 책상 쪽으로 내려갑니다. 따라서 붕괴의 목적지가 명확합니다. 피는 가슴에서 주로 화면 왼쪽과 약간 아래쪽으로 퍼져, 위로 솟는 방향은 A보다 약합니다. 무기나 조준 행동은 없습니다.",
        "built_space": "수리 책상 한 개가 전경에 있고 뒤쪽 유리 출입구 한 곳, 오른쪽 냉장고 한 대와 전면 투입식 세탁기 두 대, 다층 텔레비전 선반이 이전 장소의 배치를 이어갑니다. 마츠다의 가슴은 책상 위로 기울어 붕괴 위치가 맞습니다. 그러나 머리부터 허리 부근과 가게 양쪽까지 보여 여전히 요구한 클로즈업보다 넓고, 상판도 좁은 띠 이상으로 노출됩니다. 왼쪽 아래에는 다른 사람의 어깨가 있으며 왼쪽 위 벽에는 글자가 보이는 간판이 추가되어 있습니다.",
        "entities": "노년 동아시아계 남성 한 명의 얼굴, 짧은 회흑색 머리, 베이지 조끼와 체크 셔츠는 이전 스틸의 마츠다와 대체로 맞습니다. 일본인이라는 국적은 영상 외관만으로 검증할 수 없습니다. 가슴에는 찢어진 옷과 붉은 혈액 분출이 있고, 주변에는 낡은 텔레비전과 수리 부품이 있습니다. 왼쪽 아래의 별도 인물 어깨는 허용되지 않습니다. 왼쪽 위 간판은 한글 일부가 판독 가능한 크기로 드러나 무문자 조건도 어깁니다.",
        "hard_violations": [
         "왼쪽 아래에 이전 스틸의 다른 사람으로 읽히는 어깨와 옷을 남겨, 마츠다 외 인물의 등장을 금지한 조건을 위반했습니다.",
         "왼쪽 위 벽 간판에 판독 가능한 한글 일부가 노출되어, 읽을 수 있는 글자를 전면 금지한 조건을 위반했습니다."
        ],
        "physics": "상체가 허리에서 앞으로 꺾이고 하체는 화면 아래로 이어져, 총격 뒤 지지를 잃고 책상으로 넘어지는 중간 동작으로 가능합니다. 발과 손은 가려져 실제 접촉점은 확인되지 않지만 몸 전체가 떠 있는 자세는 아닙니다. 가슴과 책상 사이에 남은 간격도 접촉 직전 순간으로 자연스럽습니다. 혈액은 가슴의 분출 지점과 연결된 줄기 및 떨어져 나온 방울로 보여 충격에 의한 운동 근거가 있습니다. 가전제품과 공구는 선반과 책상에 지지되어 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.321,
    "B": 1.417
   },
   "violations": {
    "B": [
     "[gpt-high] 왼쪽 아래의 다른 사람 어깨와 화면 하단의 손을 남겨, 이전 스틸의 다른 인물을 이 장면에서 제외하라는 명시적 조건을 위반했습니다."
    ],
    "A": [
     "[gpt-high] 왼쪽 아래에 이전 스틸의 다른 사람으로 읽히는 어깨와 옷을 남겨, 마츠다 외 인물의 등장을 금지한 조건을 위반했습니다.",
     "[gpt-high] 왼쪽 위 벽 간판에 판독 가능한 한글 일부가 노출되어, 읽을 수 있는 글자를 전면 금지한 조건을 위반했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1417,
   "A": 1321
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "지시된 '찢어진 가슴팍 옷자락'과 '위로 튀어 오르는 피'를 정확히 묘사했으며, 배경의 텍스트를 판독 불가하게 처리하여 프롬프트를 훌륭하게 준수했습니다.  ★위반: [gpt-high] 왼쪽 아래의 다른 사람 어깨와 화면 하단의 손을 남겨, 이전 스틸의 다른 인물을 이 장면에서 제외하라는 명시적 조건을 위반했습니다."
   },
   {
    "label": "A",
    "score": 1321,
    "verdict_ko": "옷자락이 찢어지지 않았고 피가 위로 튀어 오르지 않으며, 왼쪽 벽에 읽을 수 있는 한글 텍스트가 명확히 존재해 부정 프롬프트를 위반했습니다.  ★위반: [gpt-high] 왼쪽 아래에 이전 스틸의 다른 사람으로 읽히는 어깨와 옷을 남겨, 마츠다 외 인물의 등장을 금지한 조건을 위반했습니다. / [gpt-high] 왼쪽 위 벽 간판에 판독 가능한 한글 일부가 노출되어, 읽을 수 있는 글자를 전면 금지한 조건을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S17sh3_sel.png",
    "asset_id": "595bbc64-1d91-45e5-8434-1e429e68f268",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 마츠다: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1290375>",
    "asset_id": "08af16fa-c815-43f6-9599-371c64d66165",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd91-4c07-72e2-88a5-49e7277256e7",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S17sh3"
  }
 },
 "S17sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:24:20.762647+00:00",
  "fingerprint": "d5aa30740f9a7093d574cfcb6c772710520e4aa7e471cee3907434e8ba7f7559",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S17sh7_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S17sh7_sel.png",
  "source_sha256": "474880bcd6375063e9d97826baf9f73c9f86a6cce017bbfdfb75843eb1424257",
  "file": "S17sh7_cine.png",
  "staged_sha256": "9b437b04bf8afc8ec4f6c5599e34b8331563a5e1cd5355730f76ffe5957263bb",
  "latency_ms": 11092
 },
 "S17sh9::signage": {
  "fp": "5a2cac2e2df03d17",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S17sh9": {
  "input_fingerprint": "56bfa1f62f2cd1c1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진의 수하 1이 한 손으로 검은 무전기를 입가에 바짝 가져다 댄 채 무표정하게 앞을 응시하는 구도.\n\nLOCATION (lock): In the junk shop's repair area near the owner's desk, with daytime ambient light over the piled appliances. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Black radio (Held close to the mouth during the report) — Seen obliquely beside his lower face, with its speaking end directed toward his mouth; used as Makes the reporting action explicit while remaining small relative to his upper body; Desk (Remains beside the reporting man) — The visitor-facing corner appears at the lower left; used as Retains a modest spatial link to the preceding collapse without showing the firing action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same restrained daytime interior illumination, keeping his expression and the radio legible without a dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The television remains under repair amid the shop's appliances and scrap. 박철진의 수하 1: He holds his radio up to speak into it.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진의 수하 1 right now, so 박철진의 수하 1's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진의 수하 1: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진의 수하 1 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진의 수하 1이 한 손으로 검은 무전기를 입가에 바짝 가져다 댄 채 무표정하게 앞을 응시하는 구도.\n\nLOCATION (lock): In the junk shop's repair area near the owner's desk, with daytime ambient light over the piled appliances. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Black radio (Held close to the mouth during the report) — Seen obliquely beside his lower face, with its speaking end directed toward his mouth; used as Makes the reporting action explicit while remaining small relative to his upper body; Desk (Remains beside the reporting man) — The visitor-facing corner appears at the lower left; used as Retains a modest spatial link to the preceding collapse without showing the firing action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same restrained daytime interior illumination, keeping his expression and the radio legible without a dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The television remains under repair amid the shop's appliances and scrap. 박철진의 수하 1: He holds his radio up to speak into it.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진의 수하 1 right now, so 박철진의 수하 1's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진의 수하 1: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진의 수하 1 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진의 수하 1이 한 손으로 검은 무전기를 입가에 바짝 가져다 댄 채 무표정하게 앞을 응시하는 구도.\n\nLOCATION (lock): In the junk shop's repair area near the owner's desk, with daytime ambient light over the piled appliances. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Black radio (Held close to the mouth during the report) — Seen obliquely beside his lower face, with its speaking end directed toward his mouth; used as Makes the reporting action explicit while remaining small relative to his upper body; Desk (Remains beside the reporting man) — The visitor-facing corner appears at the lower left; used as Retains a modest spatial link to the preceding collapse without showing the firing action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same restrained daytime interior illumination, keeping his expression and the radio legible without a dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The television remains under repair amid the shop's appliances and scrap. 박철진의 수하 1: He holds his radio up to speak into it.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진의 수하 1 right now, so 박철진의 수하 1's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진의 수하 1: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진의 수하 1 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "인물은 무표정하게 앞(카메라의 약간 왼쪽)을 응시하고 있습니다. 검은 무전기를 쥔 손이 입가로 올라와 있으며, 무전기의 송화 부분이 입을 향하고 있습니다.",
    "built_space": "고물상 수리 공간. 왼쪽 아래에 책상의 모서리가 보이며, 책상 위 인두기, 전선 등의 공구 배치가 레퍼런스와 완벽하게 일치합니다. 뒤쪽의 가전제품 선반과 냉장고의 위치도 록(lock)된 장소와 동일합니다.",
    "entities": "박철진의 수하 1로 지정된 인물의 얼굴, 모자, 복장이 레퍼런스와 일치합니다. 검은 무전기가 묘사되었습니다.",
    "hard_violations": [
     "[gemini-pro] 신체 구조 불가능 및 복제: 오른쪽 팔뚝이 책상 위에 가로로 놓여 있는데, 동시에 또 다른 오른쪽 팔이 수직으로 올라와 무전기를 쥐고 있어 팔이 세 개입니다.",
     "[gemini-pro] 신체 구조 불가능: 무전기를 쥔 손의 엄지손가락과 손가락 관절(knuckles)이 무전기의 같은 쪽(왼쪽)에 위치하여 해부학적으로 불가능합니다.",
     "[gemini-pro] 텍스트 노출 금지 위반: 왼쪽 팔의 패치에 '대테러 존어'라는 글씨가 선명하게 읽힙니다.",
     "[gpt-high] 왼팔 완장의 한글 일부가 판독 가능하여, 모든 글자를 읽을 수 없게 처리하라는 조건을 위반한다."
    ],
    "physics": "책상 위에 놓인 팔은 표면에 자연스럽게 닿아 있으나, 무전기를 든 손은 세 번째 팔로 뻗어나와 있어 물리적 모순을 일으킵니다."
   },
   {
    "label": "B",
    "direction": "인물은 무표정하게 앞을 응시하고 있습니다. 무전기는 입가에 바짝 대어 있으며 송화 부분이 입을 향합니다.",
    "built_space": "고물상 수리 공간. 왼쪽 아래에 책상이 보이나, 레퍼런스에서 텅 비어 있던 중앙 통로 자리에 이전 샷에 없던 나무 테이블과 분해된 브라운관 TV가 가로막듯 배치되어 장소의 고정된 형태를 무시했습니다.",
    "entities": "박철진의 수하 1로 지정된 인물의 얼굴과 복장이 일치합니다. 검은색 무전기가 묘사되었습니다.",
    "hard_violations": [
     "[gemini-pro] 신체 구조 불가능: 무전기를 쥔 손의 엄지손가락과 손가락 관절(knuckles)이 무전기의 같은 쪽(오른쪽)에 겹쳐 있어 해부학적으로 불가능한 손가락 구조를 보입니다.",
     "[gemini-pro] 텍스트 노출 금지 위반: 왼쪽 팔의 패치에 '대하 존여'라는 글씨가 명확하게 읽힙니다.",
     "[gemini-pro] 발명된 사물 및 장소 왜곡: 레퍼런스(이전 샷)에서 아무것도 없던 열린 통로 중앙에 테이블과 대형 TV를 임의로 생성하여 공간 구조를 파괴했습니다.",
     "[gpt-high] 왼팔 완장에 판독 가능한 한글이 노출되어, 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
    ],
    "physics": "무전기를 쥐고 있는 손 자체는 공중에 떠 있지 않고 팔과 연결되어 있으나, 손가락 뼈의 구조와 꺾임이 물리적으로 성립하지 않습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "이전 샷의 배경과 소품 배치를 완벽하게 유지했으나, 오른쪽 팔이 두 개로 복제되고 손 구조가 불가능하며 팔의 패치 글씨가 그대로 읽히는 치명적인 위반이 있습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "손의 해부학적 구조가 불가능하고 패치의 글씨가 읽히며, 무엇보다 이전 샷에서 비어 있던 통로 중앙에 테이블과 TV를 임의로 생성하여 장소의 고정된 특징을 심각하게 훼손했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 무표정하게 앞(카메라의 약간 왼쪽)을 응시하고 있습니다. 검은 무전기를 쥔 손이 입가로 올라와 있으며, 무전기의 송화 부분이 입을 향하고 있습니다.",
        "built_space": "고물상 수리 공간. 왼쪽 아래에 책상의 모서리가 보이며, 책상 위 인두기, 전선 등의 공구 배치가 레퍼런스와 완벽하게 일치합니다. 뒤쪽의 가전제품 선반과 냉장고의 위치도 록(lock)된 장소와 동일합니다.",
        "entities": "박철진의 수하 1로 지정된 인물의 얼굴, 모자, 복장이 레퍼런스와 일치합니다. 검은 무전기가 묘사되었습니다.",
        "hard_violations": [
         "신체 구조 불가능 및 복제: 오른쪽 팔뚝이 책상 위에 가로로 놓여 있는데, 동시에 또 다른 오른쪽 팔이 수직으로 올라와 무전기를 쥐고 있어 팔이 세 개입니다.",
         "신체 구조 불가능: 무전기를 쥔 손의 엄지손가락과 손가락 관절(knuckles)이 무전기의 같은 쪽(왼쪽)에 위치하여 해부학적으로 불가능합니다.",
         "텍스트 노출 금지 위반: 왼쪽 팔의 패치에 '대테러 존어'라는 글씨가 선명하게 읽힙니다."
        ],
        "physics": "책상 위에 놓인 팔은 표면에 자연스럽게 닿아 있으나, 무전기를 든 손은 세 번째 팔로 뻗어나와 있어 물리적 모순을 일으킵니다."
       },
       {
        "label": "B",
        "direction": "인물은 무표정하게 앞을 응시하고 있습니다. 무전기는 입가에 바짝 대어 있으며 송화 부분이 입을 향합니다.",
        "built_space": "고물상 수리 공간. 왼쪽 아래에 책상이 보이나, 레퍼런스에서 텅 비어 있던 중앙 통로 자리에 이전 샷에 없던 나무 테이블과 분해된 브라운관 TV가 가로막듯 배치되어 장소의 고정된 형태를 무시했습니다.",
        "entities": "박철진의 수하 1로 지정된 인물의 얼굴과 복장이 일치합니다. 검은색 무전기가 묘사되었습니다.",
        "hard_violations": [
         "신체 구조 불가능: 무전기를 쥔 손의 엄지손가락과 손가락 관절(knuckles)이 무전기의 같은 쪽(오른쪽)에 겹쳐 있어 해부학적으로 불가능한 손가락 구조를 보입니다.",
         "텍스트 노출 금지 위반: 왼쪽 팔의 패치에 '대하 존여'라는 글씨가 명확하게 읽힙니다.",
         "발명된 사물 및 장소 왜곡: 레퍼런스(이전 샷)에서 아무것도 없던 열린 통로 중앙에 테이블과 대형 TV를 임의로 생성하여 공간 구조를 파괴했습니다."
        ],
        "physics": "무전기를 쥐고 있는 손 자체는 공중에 떠 있지 않고 팔과 연결되어 있으나, 손가락 뼈의 구조와 꺾임이 물리적으로 성립하지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "이전 샷의 배경과 소품 배치를 완벽하게 유지했으나, 오른쪽 팔이 두 개로 복제되고 손 구조가 불가능하며 팔의 패치 글씨가 그대로 읽히는 치명적인 위반이 있습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "손의 해부학적 구조가 불가능하고 패치의 글씨가 읽히며, 무엇보다 이전 샷에서 비어 있던 통로 중앙에 테이블과 TV를 임의로 생성하여 장소의 고정된 특징을 심각하게 훼손했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 무표정하게 앞(카메라의 약간 왼쪽)을 응시하고 있습니다. 검은 무전기를 쥔 손이 입가로 올라와 있으며, 무전기의 송화 부분이 입을 향하고 있습니다.",
        "built_space": "고물상 수리 공간. 왼쪽 아래에 책상의 모서리가 보이며, 책상 위 인두기, 전선 등의 공구 배치가 레퍼런스와 완벽하게 일치합니다. 뒤쪽의 가전제품 선반과 냉장고의 위치도 록(lock)된 장소와 동일합니다.",
        "entities": "박철진의 수하 1로 지정된 인물의 얼굴, 모자, 복장이 레퍼런스와 일치합니다. 검은 무전기가 묘사되었습니다.",
        "hard_violations": [
         "신체 구조 불가능 및 복제: 오른쪽 팔뚝이 책상 위에 가로로 놓여 있는데, 동시에 또 다른 오른쪽 팔이 수직으로 올라와 무전기를 쥐고 있어 팔이 세 개입니다.",
         "신체 구조 불가능: 무전기를 쥔 손의 엄지손가락과 손가락 관절(knuckles)이 무전기의 같은 쪽(왼쪽)에 위치하여 해부학적으로 불가능합니다.",
         "텍스트 노출 금지 위반: 왼쪽 팔의 패치에 '대테러 존어'라는 글씨가 선명하게 읽힙니다."
        ],
        "physics": "책상 위에 놓인 팔은 표면에 자연스럽게 닿아 있으나, 무전기를 든 손은 세 번째 팔로 뻗어나와 있어 물리적 모순을 일으킵니다."
       },
       {
        "label": "B",
        "direction": "인물은 무표정하게 앞을 응시하고 있습니다. 무전기는 입가에 바짝 대어 있으며 송화 부분이 입을 향합니다.",
        "built_space": "고물상 수리 공간. 왼쪽 아래에 책상이 보이나, 레퍼런스에서 텅 비어 있던 중앙 통로 자리에 이전 샷에 없던 나무 테이블과 분해된 브라운관 TV가 가로막듯 배치되어 장소의 고정된 형태를 무시했습니다.",
        "entities": "박철진의 수하 1로 지정된 인물의 얼굴과 복장이 일치합니다. 검은색 무전기가 묘사되었습니다.",
        "hard_violations": [
         "신체 구조 불가능: 무전기를 쥔 손의 엄지손가락과 손가락 관절(knuckles)이 무전기의 같은 쪽(오른쪽)에 겹쳐 있어 해부학적으로 불가능한 손가락 구조를 보입니다.",
         "텍스트 노출 금지 위반: 왼쪽 팔의 패치에 '대하 존여'라는 글씨가 명확하게 읽힙니다.",
         "발명된 사물 및 장소 왜곡: 레퍼런스(이전 샷)에서 아무것도 없던 열린 통로 중앙에 테이블과 대형 TV를 임의로 생성하여 공간 구조를 파괴했습니다."
        ],
        "physics": "무전기를 쥐고 있는 손 자체는 공중에 떠 있지 않고 팔과 연결되어 있으나, 손가락 뼈의 구조와 꺾임이 물리적으로 성립하지 않습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "입가에 밀착한 무전기, 무표정한 전방 응시, 왼쪽 아래 책상 모서리가 지정 구도에 더 충실하지만, 완장의 읽히는 한글 때문에 실격이다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물과 수리점의 연속성은 좋지만 책상이 화면 하단 전체를 차지하고 무전기가 입에서 조금 떨어져 있으며, 완장의 읽히는 한글 때문에 실격이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남자는 무표정하게 카메라 축에 가까운 전방을 응시한다. 오른손의 검은 무전기는 아래 얼굴 옆에서 비스듬히 보이고, 상단 발화부가 입술 바로 곁으로 향한다. 안테나는 왼쪽 위로 뻗으며 발화 방향과 혼동되지 않는다.",
        "built_space": "뒤쪽 유리 출입구 하나, 오른쪽 다단 금속 선반과 그 위의 여러 텔레비전·전자레인지, 나란한 전면 투입식 세탁기 두 대, 오른쪽 냉장고 한 대와 왼쪽의 적층 가전이 보인다. 왼쪽 아래에는 공구와 전선이 놓인 금속 작업대 한 개의 모서리가 들어온다. 그 뒤 별도 작업면에는 뒤판이 열린 수리 중 텔레비전 한 대가 있다. 인물은 작업대 옆에 서 있으며, 기존 장소의 낡은 재질과 낮빛은 대체로 유지된다. 불가능한 반사는 보이지 않는다.",
        "entities": "보이는 사람은 성인 남성 한 명뿐이다. 짧은 검은 머리, 짙은 눈썹, 수염 자국, 얼굴 윤곽과 체격이 인물 참조에 가깝다. 사진만으로 국적이나 민족을 확정할 수는 없으며, 참조의 외형을 다른 외형으로 바꾼 흔적은 없다. 올리브색 모자와 제복, 검은 완장도 일치한다. 검은 휴대용 무전기, 책상, 수리 중인 텔레비전이 확인된다. 다만 완장의 금색 한글 일부가 명확하게 읽혀 글자 금지 조건을 어긴다.",
        "hard_violations": [
         "왼팔 완장에 판독 가능한 한글이 노출되어, 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "무전기는 오른손 손가락으로 감싸 쥐고 있으며 손목과 굽힌 팔이 무게를 받는다. 손의 피부와 제복 소매는 해당 남성과 자연스럽게 이어진다. 발은 미디엄 숏 밖이지만 상체에 부유나 불가능한 자세의 징후는 없다. 공구는 작업대에, 열린 텔레비전은 뒤쪽 작업면에, 가전은 바닥 또는 선반에 지지되어 있다."
       },
       {
        "label": "B",
        "direction": "남자의 얼굴과 시선은 카메라보다 화면 왼쪽의 화면 밖 전방을 향한다. 표정은 무덤덤하다. 오른손의 무전기는 아래 얼굴 옆으로 기울어 있고 상단은 입 쪽에 놓였지만, 입술과는 몇 센티미터 간격이 있어 바짝 가져다 댄 동작은 A보다 약하다.",
        "built_space": "뒤쪽 유리 출입구 하나, 오른쪽 다단 가전 선반, 나란한 전면 투입식 세탁기 두 대, 그 뒤의 냉장고와 왼쪽 적층 가전이 이전 장소와 유사하게 배치된다. 왼쪽에는 수리 장비와 작은 목재 책상이 있고, 금속 작업대 한 개가 전경 하단을 거의 가로지른다. 인물은 그 작업대 뒤에서 팔꿈치를 얹고 있다. 미디엄 숏이지만 책상이 왼쪽 아래 모서리로만 남는 지정 구도보다 전경 비중이 크다. 불가능한 반사는 보이지 않는다.",
        "entities": "성인 남성 한 명만 보이며 얼굴 윤곽, 짧은 검은 머리, 수염 자국과 체격은 인물 참조와 가깝다. 민족이나 국적은 외모만으로 단정할 수 없다. 올리브색 제복과 모자, 검은 완장이 유지되고, 이전 장면의 노인은 등장하지 않는다. 검은 무전기와 작업대가 있으며 왼쪽에는 열린 수리용 기기와 부품이 보이나, 수리 중인 텔레비전의 구체적 형태는 A보다 덜 명확하다. 완장에는 읽을 수 있는 금색 한글이 있다.",
        "hard_violations": [
         "왼팔 완장의 한글 일부가 판독 가능하여, 모든 글자를 읽을 수 없게 처리하라는 조건을 위반한다."
        ],
        "physics": "오른손이 무전기를 실제로 움켜쥐며 굽힌 팔꿈치는 금속 작업대 위에 닿아 지지된다. 손·손목·팔의 연결과 기기의 무게 지지가 자연스럽다. 하체는 화면 밖이지만 상체가 공중에 떠 있다는 징후는 없다. 전경 공구와 부품은 작업대에 놓이고 배경 가전은 선반이나 바닥에 받쳐져 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "입가에 밀착한 무전기, 무표정한 전방 응시, 왼쪽 아래 책상 모서리가 지정 구도에 더 충실하지만, 완장의 읽히는 한글 때문에 실격이다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물과 수리점의 연속성은 좋지만 책상이 화면 하단 전체를 차지하고 무전기가 입에서 조금 떨어져 있으며, 완장의 읽히는 한글 때문에 실격이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "남자는 무표정하게 카메라 축에 가까운 전방을 응시한다. 오른손의 검은 무전기는 아래 얼굴 옆에서 비스듬히 보이고, 상단 발화부가 입술 바로 곁으로 향한다. 안테나는 왼쪽 위로 뻗으며 발화 방향과 혼동되지 않는다.",
        "built_space": "뒤쪽 유리 출입구 하나, 오른쪽 다단 금속 선반과 그 위의 여러 텔레비전·전자레인지, 나란한 전면 투입식 세탁기 두 대, 오른쪽 냉장고 한 대와 왼쪽의 적층 가전이 보인다. 왼쪽 아래에는 공구와 전선이 놓인 금속 작업대 한 개의 모서리가 들어온다. 그 뒤 별도 작업면에는 뒤판이 열린 수리 중 텔레비전 한 대가 있다. 인물은 작업대 옆에 서 있으며, 기존 장소의 낡은 재질과 낮빛은 대체로 유지된다. 불가능한 반사는 보이지 않는다.",
        "entities": "보이는 사람은 성인 남성 한 명뿐이다. 짧은 검은 머리, 짙은 눈썹, 수염 자국, 얼굴 윤곽과 체격이 인물 참조에 가깝다. 사진만으로 국적이나 민족을 확정할 수는 없으며, 참조의 외형을 다른 외형으로 바꾼 흔적은 없다. 올리브색 모자와 제복, 검은 완장도 일치한다. 검은 휴대용 무전기, 책상, 수리 중인 텔레비전이 확인된다. 다만 완장의 금색 한글 일부가 명확하게 읽혀 글자 금지 조건을 어긴다.",
        "hard_violations": [
         "왼팔 완장에 판독 가능한 한글이 노출되어, 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "무전기는 오른손 손가락으로 감싸 쥐고 있으며 손목과 굽힌 팔이 무게를 받는다. 손의 피부와 제복 소매는 해당 남성과 자연스럽게 이어진다. 발은 미디엄 숏 밖이지만 상체에 부유나 불가능한 자세의 징후는 없다. 공구는 작업대에, 열린 텔레비전은 뒤쪽 작업면에, 가전은 바닥 또는 선반에 지지되어 있다."
       },
       {
        "label": "A",
        "direction": "남자의 얼굴과 시선은 카메라보다 화면 왼쪽의 화면 밖 전방을 향한다. 표정은 무덤덤하다. 오른손의 무전기는 아래 얼굴 옆으로 기울어 있고 상단은 입 쪽에 놓였지만, 입술과는 몇 센티미터 간격이 있어 바짝 가져다 댄 동작은 A보다 약하다.",
        "built_space": "뒤쪽 유리 출입구 하나, 오른쪽 다단 가전 선반, 나란한 전면 투입식 세탁기 두 대, 그 뒤의 냉장고와 왼쪽 적층 가전이 이전 장소와 유사하게 배치된다. 왼쪽에는 수리 장비와 작은 목재 책상이 있고, 금속 작업대 한 개가 전경 하단을 거의 가로지른다. 인물은 그 작업대 뒤에서 팔꿈치를 얹고 있다. 미디엄 숏이지만 책상이 왼쪽 아래 모서리로만 남는 지정 구도보다 전경 비중이 크다. 불가능한 반사는 보이지 않는다.",
        "entities": "성인 남성 한 명만 보이며 얼굴 윤곽, 짧은 검은 머리, 수염 자국과 체격은 인물 참조와 가깝다. 민족이나 국적은 외모만으로 단정할 수 없다. 올리브색 제복과 모자, 검은 완장이 유지되고, 이전 장면의 노인은 등장하지 않는다. 검은 무전기와 작업대가 있으며 왼쪽에는 열린 수리용 기기와 부품이 보이나, 수리 중인 텔레비전의 구체적 형태는 A보다 덜 명확하다. 완장에는 읽을 수 있는 금색 한글이 있다.",
        "hard_violations": [
         "왼팔 완장의 한글 일부가 판독 가능하여, 모든 글자를 읽을 수 없게 처리하라는 조건을 위반한다."
        ],
        "physics": "오른손이 무전기를 실제로 움켜쥐며 굽힌 팔꿈치는 금속 작업대 위에 닿아 지지된다. 손·손목·팔의 연결과 기기의 무게 지지가 자연스럽다. 하체는 화면 밖이지만 상체가 공중에 떠 있다는 징후는 없다. 전경 공구와 부품은 작업대에 놓이고 배경 가전은 선반이나 바닥에 받쳐져 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 신체 구조 불가능 및 복제: 오른쪽 팔뚝이 책상 위에 가로로 놓여 있는데, 동시에 또 다른 오른쪽 팔이 수직으로 올라와 무전기를 쥐고 있어 팔이 세 개입니다.",
     "[gemini-pro] 신체 구조 불가능: 무전기를 쥔 손의 엄지손가락과 손가락 관절(knuckles)이 무전기의 같은 쪽(왼쪽)에 위치하여 해부학적으로 불가능합니다.",
     "[gemini-pro] 텍스트 노출 금지 위반: 왼쪽 팔의 패치에 '대테러 존어'라는 글씨가 선명하게 읽힙니다.",
     "[gpt-high] 왼팔 완장의 한글 일부가 판독 가능하여, 모든 글자를 읽을 수 없게 처리하라는 조건을 위반한다."
    ],
    "B": [
     "[gemini-pro] 신체 구조 불가능: 무전기를 쥔 손의 엄지손가락과 손가락 관절(knuckles)이 무전기의 같은 쪽(오른쪽)에 겹쳐 있어 해부학적으로 불가능한 손가락 구조를 보입니다.",
     "[gemini-pro] 텍스트 노출 금지 위반: 왼쪽 팔의 패치에 '대하 존여'라는 글씨가 명확하게 읽힙니다.",
     "[gemini-pro] 발명된 사물 및 장소 왜곡: 레퍼런스(이전 샷)에서 아무것도 없던 열린 통로 중앙에 테이블과 대형 TV를 임의로 생성하여 공간 구조를 파괴했습니다.",
     "[gpt-high] 왼팔 완장에 판독 가능한 한글이 노출되어, 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1500,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "이전 샷의 배경과 소품 배치를 완벽하게 유지했으나, 오른쪽 팔이 두 개로 복제되고 손 구조가 불가능하며 팔의 패치 글씨가 그대로 읽히는 치명적인 위반이 있습니다.  ★위반: [gemini-pro] 신체 구조 불가능 및 복제: 오른쪽 팔뚝이 책상 위에 가로로 놓여 있는데, 동시에 또 다른 오른쪽 팔이 수직으로 올라와 무전기를 쥐고 있어 팔이 세 개입니다. / [gemini-pro] 신체 구조 불가능: 무전기를 쥔 손의 엄지손가락과 손가락 관절(knuckles)이 무전기의 같은 쪽(왼쪽)에 위치하여 해부학적으로 불가능합니다. / [gemini-pro] 텍스트 노출 금지 위반: 왼쪽 팔의 패치에 '대테러 존어'라는 글씨가 선명하게 읽힙니다. / [gpt-high] 왼팔 완장의 한글 일부가 판독 가능하여, 모든 글자를 읽을 수 없게 처리하라는 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "손의 해부학적 구조가 불가능하고 패치의 글씨가 읽히며, 무엇보다 이전 샷에서 비어 있던 통로 중앙에 테이블과 TV를 임의로 생성하여 장소의 고정된 특징을 심각하게 훼손했습니다.  ★위반: [gemini-pro] 신체 구조 불가능: 무전기를 쥔 손의 엄지손가락과 손가락 관절(knuckles)이 무전기의 같은 쪽(오른쪽)에 겹쳐 있어 해부학적으로 불가능한 손가락 구조를 보입니다. / [gemini-pro] 텍스트 노출 금지 위반: 왼쪽 팔의 패치에 '대하 존여'라는 글씨가 명확하게 읽힙니다. / [gemini-pro] 발명된 사물 및 장소 왜곡: 레퍼런스(이전 샷)에서 아무것도 없던 열린 통로 중앙에 테이블과 대형 TV를 임의로 생성하여 공간 구조를 파괴했습니다. / [gpt-high] 왼팔 완장에 판독 가능한 한글이 노출되어, 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S17sh7_sel.png",
    "asset_id": "5d91c022-17da-4ece-9527-28530e747fde",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진의 수하 1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:967362>",
    "asset_id": "49efd036-7f14-493a-b65e-cec03d0dc53c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd92-02c9-72cc-b837-4c22dc9b98c2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S17sh7"
  }
 },
 "S17sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:24:32.183751+00:00",
  "fingerprint": "ea26608188a4e4af7e50513c542b86d0c25027992e98cef18808ecee3cc72b65",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S17sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S17sh9_sel.png",
  "source_sha256": "364bba56b2d6eac4d5efd59d21660f324d83d76a15efcb46faaea022cc929b49",
  "file": "S17sh9_cine.png",
  "staged_sha256": "36b9588b98a8b5e4da5a614b2b573d794a16d43df2db53431eec022bca61f1ba",
  "latency_ms": 19546
 },
 "S18sh6::signage": {
  "fp": "993576daa871b2d6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::1aed9e1a8f70eab7": {
  "subjects": [],
  "subject_text": "박철진의 민병대 사무실\n책상과 스피커폰이 놓인 사무실. 책상 위에 큰 지도가 펼쳐져 있고 벽 쪽에는 다트용 공간이 있다.",
  "identity": "canonical",
  "scope_id": "L28",
  "scope_role": "location_interior",
  "scope_sha": "8a546e8f9d3f1669"
 },
 "S18sh6::bgfirst_bg": {
  "input_fingerprint": "5ed2784792ff07f6",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 단도 끝이 윤성찬의 얼굴 옆 나무 벽면에 깊숙이 박힌 채 나무 파편이 허공에 정지해 있는 타격의 순간.\n\nLOCATION (lock): At the wooden wall behind the visitor's standing position inside a militia command office, in daytime interior light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wooden wall impact point immediately beside 윤성찬's face in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Wooden wall beside 윤성찬 (Struck by the knife, with fragments airborne at the entry point) — Viewed at a shallow oblique angle beside the visible edge of his face; used as Places the impact immediately beside the threatened face without flattening their separation; Knife (Its tip is deeply embedded in the wall) — The exposed portion projects away from the wall and is seen from the side at an oblique angle; used as Forms the sharp impact detail while the adjacent face supplies human scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Steady daytime interior illumination reveals the wood fragments and 윤성찬's flushed reaction without an added flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 단도 끝이 윤성찬의 얼굴 옆 나무 벽면에 깊숙이 박힌 채 나무 파편이 허공에 정지해 있는 타격의 순간.\n\nLOCATION (lock): At the wooden wall behind the visitor's standing position inside a militia command office, in daytime interior light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wooden wall impact point immediately beside 윤성찬's face in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Wooden wall beside 윤성찬 (Struck by the knife, with fragments airborne at the entry point) — Viewed at a shallow oblique angle beside the visible edge of his face; used as Places the impact immediately beside the threatened face without flattening their separation; Knife (Its tip is deeply embedded in the wall) — The exposed portion projects away from the wall and is seen from the side at an oblique angle; used as Forms the sharp impact detail while the adjacent face supplies human scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Steady daytime interior illumination reveals the wood fragments and 윤성찬's flushed reaction without an added flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S18sh6__bgfirst_bg.png",
  "asset_id": "82d63152-24ea-49ff-af27-dbd647387e04",
  "input_asset_ids": [
   "4827fcb4-47f2-4df7-b3e2-68e4ae5f0ce9",
   "32b99c56-4af6-4df1-b932-68c40cd54ef2"
  ]
 },
 "S18sh6": {
  "input_fingerprint": "c1dfc54a967119d6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 단도 끝이 윤성찬의 얼굴 옆 나무 벽면에 깊숙이 박힌 채 나무 파편이 허공에 정지해 있는 타격의 순간.\n\nLOCATION (lock): At the wooden wall behind the visitor's standing position inside a militia command office, in daytime interior light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wooden wall impact point immediately beside 윤성찬's face in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Wooden wall beside 윤성찬 (Struck by the knife, with fragments airborne at the entry point) — Viewed at a shallow oblique angle beside the visible edge of his face; used as Places the impact immediately beside the threatened face without flattening their separation; Knife (Its tip is deeply embedded in the wall) — The exposed portion projects away from the wall and is seen from the side at an oblique angle; used as Forms the sharp impact detail while the adjacent face supplies human scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Steady daytime interior illumination reveals the wood fragments and 윤성찬's flushed reaction without an added flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The thrown knife is now at the office wall and must not reappear as a handheld prop without retrieval. The speakerphone remains in the office. 윤성찬: He remains standing with his handkerchief, his face flushed from the sudden fright.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 단도 끝이 윤성찬의 얼굴 옆 나무 벽면에 깊숙이 박힌 채 나무 파편이 허공에 정지해 있는 타격의 순간.\n\nLOCATION (lock): At the wooden wall behind the visitor's standing position inside a militia command office, in daytime interior light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wooden wall impact point immediately beside 윤성찬's face in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Wooden wall beside 윤성찬 (Struck by the knife, with fragments airborne at the entry point) — Viewed at a shallow oblique angle beside the visible edge of his face; used as Places the impact immediately beside the threatened face without flattening their separation; Knife (Its tip is deeply embedded in the wall) — The exposed portion projects away from the wall and is seen from the side at an oblique angle; used as Forms the sharp impact detail while the adjacent face supplies human scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Steady daytime interior illumination reveals the wood fragments and 윤성찬's flushed reaction without an added flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The thrown knife is now at the office wall and must not reappear as a handheld prop without retrieval. The speakerphone remains in the office. 윤성찬: He remains standing with his handkerchief, his face flushed from the sudden fright.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 단도 끝이 윤성찬의 얼굴 옆 나무 벽면에 깊숙이 박힌 채 나무 파편이 허공에 정지해 있는 타격의 순간.\n\nLOCATION (lock): At the wooden wall behind the visitor's standing position inside a militia command office, in daytime interior light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wooden wall impact point immediately beside 윤성찬's face in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Wooden wall beside 윤성찬 (Struck by the knife, with fragments airborne at the entry point) — Viewed at a shallow oblique angle beside the visible edge of his face; used as Places the impact immediately beside the threatened face without flattening their separation; Knife (Its tip is deeply embedded in the wall) — The exposed portion projects away from the wall and is seen from the side at an oblique angle; used as Forms the sharp impact detail while the adjacent face supplies human scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Steady daytime interior illumination reveals the wood fragments and 윤성찬's flushed reaction without an added flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The thrown knife is now at the office wall and must not reappear as a handheld prop without retrieval. The speakerphone remains in the office. 윤성찬: He remains standing with his handkerchief, his face flushed from the sudden fright.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S18sh6__bgfirst_bg.png",
     "asset_id": "82d63152-24ea-49ff-af27-dbd647387e04",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S18sh6.png",
     "asset_id": "4827fcb4-47f2-4df7-b3e2-68e4ae5f0ce9",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:781789>",
     "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L28B02.png",
     "asset_id": "32b99c56-4af6-4df1-b932-68c40cd54ef2",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:781789>",
     "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 우측 벽에 박힌 단도를 향함.",
    "built_space": "지휘소 내부. 열린 문, 벽면의 지도, 의자가 참조 사진과 일치하게 배치됨.",
    "entities": "남색 정장을 입은 윤성찬, 하얀 손수건, 벽에 박힌 단도와 나무 파편이 모두 존재함.",
    "hard_violations": [],
    "physics": "단도는 나무 벽에 박혀 지탱되며, 허공의 파편들은 타격 충격으로 발사된 상태임. 인물은 바닥에 서 있음."
   },
   {
    "label": "B",
    "direction": "인물의 시선은 좌측 단도를 향함.",
    "built_space": "타격점이 있는 나무 벽면의 좁은 영역만 보임.",
    "entities": "윤성찬의 붉어진 얼굴과 단도가 확인되나 손수건은 없음.",
    "hard_violations": [],
    "physics": "단도는 벽에 지탱되며 파편들은 타격에 의해 발사됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 프레임 우측의 타격점 레이아웃, 인물의 손수건 소지 상태, 배경 장소의 디테일을 지시문과 정확히 일치하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "타격점 위치가 좌측으로 반전되어 레이아웃 지시를 어겼으며, 손수건이 누락되었고 배경 맥락이 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 우측 벽에 박힌 단도를 향함.",
        "built_space": "지휘소 내부. 열린 문, 벽면의 지도, 의자가 참조 사진과 일치하게 배치됨.",
        "entities": "남색 정장을 입은 윤성찬, 하얀 손수건, 벽에 박힌 단도와 나무 파편이 모두 존재함.",
        "hard_violations": [],
        "physics": "단도는 나무 벽에 박혀 지탱되며, 허공의 파편들은 타격 충격으로 발사된 상태임. 인물은 바닥에 서 있음."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 좌측 단도를 향함.",
        "built_space": "타격점이 있는 나무 벽면의 좁은 영역만 보임.",
        "entities": "윤성찬의 붉어진 얼굴과 단도가 확인되나 손수건은 없음.",
        "hard_violations": [],
        "physics": "단도는 벽에 지탱되며 파편들은 타격에 의해 발사됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 프레임 우측의 타격점 레이아웃, 인물의 손수건 소지 상태, 배경 장소의 디테일을 지시문과 정확히 일치하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "타격점 위치가 좌측으로 반전되어 레이아웃 지시를 어겼으며, 손수건이 누락되었고 배경 맥락이 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 우측 벽에 박힌 단도를 향함.",
        "built_space": "지휘소 내부. 열린 문, 벽면의 지도, 의자가 참조 사진과 일치하게 배치됨.",
        "entities": "남색 정장을 입은 윤성찬, 하얀 손수건, 벽에 박힌 단도와 나무 파편이 모두 존재함.",
        "hard_violations": [],
        "physics": "단도는 나무 벽에 박혀 지탱되며, 허공의 파편들은 타격 충격으로 발사된 상태임. 인물은 바닥에 서 있음."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 좌측 단도를 향함.",
        "built_space": "타격점이 있는 나무 벽면의 좁은 영역만 보임.",
        "entities": "윤성찬의 붉어진 얼굴과 단도가 확인되나 손수건은 없음.",
        "hard_violations": [],
        "physics": "단도는 벽에 지탱되며 파편들은 타격에 의해 발사됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "얼굴 바로 옆에 박힌 단도와 비산하는 파편을 밀착 클로즈업으로 구현해 우세하지만, 타격점이 지정된 중앙 오른쪽이 아니라 왼쪽에 있다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "벽에 깊이 박힌 단도와 장소는 잘 맞지만, 상반신과 실내까지 넓혀 얼굴과 타격점의 밀착 클로즈업을 약화했고 얼굴·셔츠·넥타이도 인물 참조와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남성의 눈은 화면 왼쪽 칼과 벽의 타격점 쪽으로 돌아가 있다. 단도는 오른쪽에서 왼쪽으로 향하며 끝이 나무 벽에 들어가 있고, 손잡이 쪽은 얼굴 가장자리 뒤로 가려진다. 칼이 얼굴 자체를 찌른 모습은 아니다.",
        "built_space": "갈색 세로 목재 벽 한 면과 타격점 하나가 보인다. 비스듬한 벽면과 그 앞 얼굴의 깊이 차이가 읽힌다. 창·문·선반은 클로즈업 밖이므로 수량을 확인할 수 없으며, 중복 설비나 반사는 없다. 얼굴은 오른쪽에 있지만 타격점은 중앙 왼쪽에 있어 지정 배치와 차이가 난다.",
        "entities": "회색 머리와 깊은 주름을 가진 고령 남성 한 명으로, 참조 인물의 얼굴 윤곽과 머리 모양에 가깝다. 짙은 정장과 흰 셔츠 일부가 보이고 얼굴은 붉게 상기되어 있다. 금속 단도 한 자루, 손상된 목재 벽, 나무 파편이 식별된다. 칼날이 상당 부분 노출되어 깊이 박힌 정도는 덜 명확하다. 손수건과 스피커폰은 프레임 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "단도 끝은 파손된 목재 진입점에 물려 벽이 칼을 지탱한다. 허공의 파편은 그 진입점에서 충격으로 튀어나온 순간으로 설명된다. 얼굴과 어깨의 연결은 자연스럽고, 하체가 잘려 서 있는 자세의 발 지지는 확인할 수 없지만 몸이 공중에 떠 있다는 징후는 없다."
       },
       {
        "label": "B",
        "direction": "남성은 고개와 눈을 오른쪽의 단도 및 타격점으로 돌렸다. 단도 손잡이는 왼쪽 앞으로 돌출되고 칼끝은 오른쪽 목재 벽 안으로 향한다. 위협 대상인 얼굴과 칼 사이에는 A보다 큰 화면상 간격이 있다.",
        "built_space": "오른쪽 목재 벽 한 면에 타격점 하나가 있고, 왼쪽에는 창 하나, 열린 문 하나, 의자 하나가 보인다. 벽의 지도와 사진 배치는 장소 참조와 대체로 대응한다. 남성은 벽 앞에 몸을 틀고 있으며 의자에 앉아 있는 접촉은 보이지 않는다. 타격점은 오른쪽에 있지만 얼굴은 왼쪽이고 상반신과 실내까지 포함해 요구된 밀착 구도보다 넓다.",
        "entities": "회색 머리의 고령 남성 한 명이 보이나 참조보다 얼굴이 둥글고 볼과 턱의 형태가 다르다. 짙은 재킷은 유사하지만 셔츠는 밝은 청회색의 열린 깃이며 참조의 넥타이가 없다. 손에는 흰 손수건이 실제로 쥐어져 있다. 단도 한 자루와 나무 파편, 목재 벽이 명확하며 스피커폰은 프레임 밖이다. 배경 종이에는 식별 가능한 문구가 보이지 않는다.",
        "hard_violations": [],
        "physics": "단도는 짧게 드러난 칼날 끝이 벽에 깊게 박혀 지지되고, 손잡이가 벽 바깥으로 돌출된 상태가 물리적으로 가능하다. 파편은 칼 진입점의 충격으로 방사상 비산한 모습이다. 손수건은 손가락으로 움켜쥐고 있으며 몸통의 비틀림도 가능하다. 발은 프레임 밖이므로 지면 접촉은 확인할 수 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "얼굴 바로 옆에 박힌 단도와 비산하는 파편을 밀착 클로즈업으로 구현해 우세하지만, 타격점이 지정된 중앙 오른쪽이 아니라 왼쪽에 있다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "벽에 깊이 박힌 단도와 장소는 잘 맞지만, 상반신과 실내까지 넓혀 얼굴과 타격점의 밀착 클로즈업을 약화했고 얼굴·셔츠·넥타이도 인물 참조와 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "남성의 눈은 화면 왼쪽 칼과 벽의 타격점 쪽으로 돌아가 있다. 단도는 오른쪽에서 왼쪽으로 향하며 끝이 나무 벽에 들어가 있고, 손잡이 쪽은 얼굴 가장자리 뒤로 가려진다. 칼이 얼굴 자체를 찌른 모습은 아니다.",
        "built_space": "갈색 세로 목재 벽 한 면과 타격점 하나가 보인다. 비스듬한 벽면과 그 앞 얼굴의 깊이 차이가 읽힌다. 창·문·선반은 클로즈업 밖이므로 수량을 확인할 수 없으며, 중복 설비나 반사는 없다. 얼굴은 오른쪽에 있지만 타격점은 중앙 왼쪽에 있어 지정 배치와 차이가 난다.",
        "entities": "회색 머리와 깊은 주름을 가진 고령 남성 한 명으로, 참조 인물의 얼굴 윤곽과 머리 모양에 가깝다. 짙은 정장과 흰 셔츠 일부가 보이고 얼굴은 붉게 상기되어 있다. 금속 단도 한 자루, 손상된 목재 벽, 나무 파편이 식별된다. 칼날이 상당 부분 노출되어 깊이 박힌 정도는 덜 명확하다. 손수건과 스피커폰은 프레임 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "단도 끝은 파손된 목재 진입점에 물려 벽이 칼을 지탱한다. 허공의 파편은 그 진입점에서 충격으로 튀어나온 순간으로 설명된다. 얼굴과 어깨의 연결은 자연스럽고, 하체가 잘려 서 있는 자세의 발 지지는 확인할 수 없지만 몸이 공중에 떠 있다는 징후는 없다."
       },
       {
        "label": "A",
        "direction": "남성은 고개와 눈을 오른쪽의 단도 및 타격점으로 돌렸다. 단도 손잡이는 왼쪽 앞으로 돌출되고 칼끝은 오른쪽 목재 벽 안으로 향한다. 위협 대상인 얼굴과 칼 사이에는 A보다 큰 화면상 간격이 있다.",
        "built_space": "오른쪽 목재 벽 한 면에 타격점 하나가 있고, 왼쪽에는 창 하나, 열린 문 하나, 의자 하나가 보인다. 벽의 지도와 사진 배치는 장소 참조와 대체로 대응한다. 남성은 벽 앞에 몸을 틀고 있으며 의자에 앉아 있는 접촉은 보이지 않는다. 타격점은 오른쪽에 있지만 얼굴은 왼쪽이고 상반신과 실내까지 포함해 요구된 밀착 구도보다 넓다.",
        "entities": "회색 머리의 고령 남성 한 명이 보이나 참조보다 얼굴이 둥글고 볼과 턱의 형태가 다르다. 짙은 재킷은 유사하지만 셔츠는 밝은 청회색의 열린 깃이며 참조의 넥타이가 없다. 손에는 흰 손수건이 실제로 쥐어져 있다. 단도 한 자루와 나무 파편, 목재 벽이 명확하며 스피커폰은 프레임 밖이다. 배경 종이에는 식별 가능한 문구가 보이지 않는다.",
        "hard_violations": [],
        "physics": "단도는 짧게 드러난 칼날 끝이 벽에 깊게 박혀 지지되고, 손잡이가 벽 바깥으로 돌출된 상태가 물리적으로 가능하다. 파편은 칼 진입점의 충격으로 방사상 비산한 모습이다. 손수건은 손가락으로 움켜쥐고 있으며 몸통의 비틀림도 가능하다. 발은 프레임 밖이므로 지면 접촉은 확인할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.429
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1750,
   "B": 1429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "요구된 프레임 우측의 타격점 레이아웃, 인물의 손수건 소지 상태, 배경 장소의 디테일을 지시문과 정확히 일치하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1429,
    "verdict_ko": "타격점 위치가 좌측으로 반전되어 레이아웃 지시를 어겼으며, 손수건이 누락되었고 배경 맥락이 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L28B02.png",
    "asset_id": "32b99c56-4af6-4df1-b932-68c40cd54ef2",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:781789>",
    "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd93-40c0-7ef1-913a-4e3ca56aa8f4",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S18sh6__bgfirst_bg.png",
   "bg_asset_id": "82d63152-24ea-49ff-af27-dbd647387e04",
   "bg_record_key": "S18sh6::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S18sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:24:52.132422+00:00",
  "fingerprint": "9e51736ecaf8bd27d6e52f734bd71b8e1b10b25ef395e74ed503cb588a87aa29",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S18sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S18sh6_sel.png",
  "source_sha256": "bc49421d85534b6f1363a16324e7d48cbd97c7687395e4b375ace449036cbe1b",
  "file": "S18sh6_cine.png",
  "staged_sha256": "d510daf4f88c47b29e51a43f6a7c7d726d5cb9b8069d29d147840db43b45bd61",
  "latency_ms": 10951
 },
 "S18sh8::signage": {
  "fp": "003072fe81dfad14",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S18sh8": {
  "input_fingerprint": "7e6c80376e0f0f49",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진이 두 팔을 양옆으로 벌린 채 거만한 미소를 지으며 윤성찬을 똑바로 응시하는 위압적인 구도.\n\nLOCATION (lock): In the speakerphone conversation area inside the militia command office, facing the visitor in daytime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the office's daytime ambient illumination, letting the low framing and unequal postures carry the intimidation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The knife remains at the wall after the throw; no retrieval is described. The speakerphone remains present. 박철진: He is no longer holding the knife.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진이 두 팔을 양옆으로 벌린 채 거만한 미소를 지으며 윤성찬을 똑바로 응시하는 위압적인 구도.\n\nLOCATION (lock): In the speakerphone conversation area inside the militia command office, facing the visitor in daytime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the office's daytime ambient illumination, letting the low framing and unequal postures carry the intimidation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The knife remains at the wall after the throw; no retrieval is described. The speakerphone remains present. 박철진: He is no longer holding the knife.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진이 두 팔을 양옆으로 벌린 채 거만한 미소를 지으며 윤성찬을 똑바로 응시하는 위압적인 구도.\n\nLOCATION (lock): In the speakerphone conversation area inside the militia command office, facing the visitor in daytime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the office's daytime ambient illumination, letting the low framing and unequal postures carry the intimidation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The knife remains at the wall after the throw; no retrieval is described. The speakerphone remains present. 박철진: He is no longer holding the knife.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라 렌즈를 똑바로 응시하며 양팔을 벌리고 거만한 미소를 지음.",
    "built_space": "배경 구조와 사물은 이전 컷과 일치하나, 프롬프트에서 유지하라고 지시한 스피커폰이 보이지 않음.",
    "entities": "캐릭터 레퍼런스의 인물이 아닌, 이전 컷에 등장했던 인물(윤성찬)의 얼굴을 군복에 잘못 적용함.",
    "hard_violations": [
     "[gemini-pro] 이전 컷 인물의 얼굴을 현재 컷 캐릭터에 잘못 적용 (신원 위반)",
     "[gpt-high] 완장에 판독 가능한 한글과 영문자 형태가 노출되어, 읽을 수 있는 글자 금지 조건을 위반한다."
    ],
    "physics": "칼은 벽에 안정적으로 박혀 있고, 인물의 자세도 중력에 맞게 서 있음."
   },
   {
    "label": "B",
    "direction": "카메라를 향해 시선을 고정하고 거만한 미소를 지으며 양팔을 벌림.",
    "built_space": "이전 컷의 배경 구조가 정확히 유지되었으며, 전경에 요구된 스피커폰이 놓여 있음.",
    "entities": "제시된 캐릭터 레퍼런스('박철진')의 얼굴, 체형, 복장을 정확하게 재현함.",
    "hard_violations": [
     "[gpt-high] 완장의 흰색 한글이 읽을 수 있는 형태로 노출되어, 화면 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반한다."
    ],
    "physics": "칼이 나무 벽에 자연스럽게 박혀 있으며, 인물의 체중 분산과 팔의 지지가 안정적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "캐릭터 레퍼런스의 외모를 정확히 반영했으며, 양팔을 벌린 위압적인 자세와 배경 연속성(스피커폰, 벽에 박힌 칼)을 성공적으로 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 캐릭터가 아닌 이전 컷 스틸의 인물 얼굴을 잘못 가져와 사용하여 치명적인 신원 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라 렌즈를 똑바로 응시하며 양팔을 벌리고 거만한 미소를 지음.",
        "built_space": "배경 구조와 사물은 이전 컷과 일치하나, 프롬프트에서 유지하라고 지시한 스피커폰이 보이지 않음.",
        "entities": "캐릭터 레퍼런스의 인물이 아닌, 이전 컷에 등장했던 인물(윤성찬)의 얼굴을 군복에 잘못 적용함.",
        "hard_violations": [
         "이전 컷 인물의 얼굴을 현재 컷 캐릭터에 잘못 적용 (신원 위반)"
        ],
        "physics": "칼은 벽에 안정적으로 박혀 있고, 인물의 자세도 중력에 맞게 서 있음."
       },
       {
        "label": "B",
        "direction": "카메라를 향해 시선을 고정하고 거만한 미소를 지으며 양팔을 벌림.",
        "built_space": "이전 컷의 배경 구조가 정확히 유지되었으며, 전경에 요구된 스피커폰이 놓여 있음.",
        "entities": "제시된 캐릭터 레퍼런스('박철진')의 얼굴, 체형, 복장을 정확하게 재현함.",
        "hard_violations": [],
        "physics": "칼이 나무 벽에 자연스럽게 박혀 있으며, 인물의 체중 분산과 팔의 지지가 안정적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "캐릭터 레퍼런스의 외모를 정확히 반영했으며, 양팔을 벌린 위압적인 자세와 배경 연속성(스피커폰, 벽에 박힌 칼)을 성공적으로 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 캐릭터가 아닌 이전 컷 스틸의 인물 얼굴을 잘못 가져와 사용하여 치명적인 신원 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라 렌즈를 똑바로 응시하며 양팔을 벌리고 거만한 미소를 지음.",
        "built_space": "배경 구조와 사물은 이전 컷과 일치하나, 프롬프트에서 유지하라고 지시한 스피커폰이 보이지 않음.",
        "entities": "캐릭터 레퍼런스의 인물이 아닌, 이전 컷에 등장했던 인물(윤성찬)의 얼굴을 군복에 잘못 적용함.",
        "hard_violations": [
         "이전 컷 인물의 얼굴을 현재 컷 캐릭터에 잘못 적용 (신원 위반)"
        ],
        "physics": "칼은 벽에 안정적으로 박혀 있고, 인물의 자세도 중력에 맞게 서 있음."
       },
       {
        "label": "B",
        "direction": "카메라를 향해 시선을 고정하고 거만한 미소를 지으며 양팔을 벌림.",
        "built_space": "이전 컷의 배경 구조가 정확히 유지되었으며, 전경에 요구된 스피커폰이 놓여 있음.",
        "entities": "제시된 캐릭터 레퍼런스('박철진')의 얼굴, 체형, 복장을 정확하게 재현함.",
        "hard_violations": [],
        "physics": "칼이 나무 벽에 자연스럽게 박혀 있으며, 인물의 체중 분산과 팔의 지지가 안정적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "양팔을 벌린 미디엄 숏과 거만한 미소, 인물 참조에 가까운 얼굴은 더 충실하지만, 완장의 판독 가능한 글자가 명시적 금지 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "낮은 시점과 양팔을 벌린 구도는 맞지만, 박철진 대신 이전 숏 노인의 얼굴을 옮긴 듯한 외모이며 완장 글자도 노출된다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남자는 양팔을 화면 좌우로 벌리고 손바닥을 앞쪽으로 펼친 채 카메라 쪽을 똑바로 본다. 화면 밖 방문자 윤성찬의 위치를 카메라 축에 둔 연출로 해석할 수 있지만, 상대가 보이지 않아 실제 눈맞춤까지 확인되지는 않는다. 오른쪽 칼은 칼끝이 목재 벽 안으로 들어가고 손잡이가 실내 쪽으로 돌출되어 있다.",
        "built_space": "왼쪽에 두 칸으로 나뉜 창 하나, 중앙 뒤에 출입구 하나, 창 아래 의자 두 개와 인물 뒤 의자 등받이 하나가 보인다. 오른쪽 목재 벽에는 지도와 종이 게시물, 칼 하나와 파손 흔적이 있다. 하단 탁자 위에는 스피커폰 하나가 놓였다. 참조의 창·출입구·목재 벽 관계와 낮의 자연광을 대체로 유지한다. 참조에서 보인 의자는 하나지만 전체 의자 수가 확정된 것은 아니므로 추가로 보이는 의자를 중복 설비로 단정할 수 없다.",
        "entities": "한 명의 중년 남성이 보이며 짧은 머리와 회색 섞인 수염, 얼굴 윤곽은 박철진 인물 참조에 비교적 가깝다. 참조의 올리브색 군복과 검은 완장은 유지하지만 군모는 없다. 민족적 배경은 외모만으로 확정할 수 없으며, 참조 인물과의 외형적 일치가 B보다 높다. 칼은 손이 아니라 벽에 남아 있고 스피커폰도 보인다. 완장에는 판독 가능한 흰색 한글이 있다.",
        "hard_violations": [
         "완장의 흰색 한글이 읽을 수 있는 형태로 노출되어, 화면 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "상체는 곧게 선 자세이고 두 팔은 어깨와 팔꿈치로 자연스럽게 지탱된다. 발은 미디엄 숏 밖이므로 접지는 확인할 수 없지만 공중에 떠 있다는 징후는 없다. 빈손의 손가락 구조도 대체로 자연스럽다. 칼은 벽에 박힌 칼날로 고정되고, 스피커폰은 탁자 면에 받쳐져 있다."
       },
       {
        "label": "B",
        "direction": "남자는 양팔과 손바닥을 좌우로 펼치고 카메라 쪽을 응시한다. 카메라 축의 화면 밖 방문자를 향한 시선으로 성립하지만 윤성찬은 직접 보이지 않는다. 오른쪽 칼은 벽을 향해 박혀 있고 손잡이는 왼쪽 실내 방향으로 나온다.",
        "built_space": "왼쪽에 창 하나와 의자 하나, 그 옆에 열린 문 하나가 있으며 문 너머로 상자 두 개가 포개져 있다. 오른쪽에는 목재 벽, 지도와 게시물, 칼 하나와 파손 자국이 보인다. 왼쪽 아래에는 전화 장치 일부가 잘려 보인다. 낮은 카메라가 천장까지 담으며 참조의 창·문·벽 배치와 주간 조명을 대체로 유지한다. 불가능한 반사나 중복된 문·칼은 보이지 않는다.",
        "entities": "한 명의 남성이 올리브색 군복과 검은 완장을 착용했지만, 흰 머리를 뒤로 넘긴 무수염의 고령 얼굴이다. 이는 수염이 있는 중년 박철진 참조보다 이전 숏의 노인 얼굴에 매우 가깝고, 지정된 44세 인물의 정체성과 맞지 않는다. 군모도 없다. 칼은 벽에 남아 있고 양손은 비어 있다. 왼쪽 아래 전화 장치는 일부만 보여 지정 스피커폰인지 확실히 확인하기 어렵다. 완장에는 판독 가능한 흰색 글자가 있다.",
        "hard_violations": [
         "완장에 판독 가능한 한글과 영문자 형태가 노출되어, 읽을 수 있는 글자 금지 조건을 위반한다."
        ],
        "physics": "서 있는 몸통에서 양팔을 벌린 자세이며 어깨·팔꿈치·손목의 연결은 물리적으로 가능하다. 하체와 발은 프레임 밖이어서 접지는 직접 확인되지 않지만 부유한 자세는 아니다. 칼은 벽에 박혀 지지되고 문 너머 상자들은 서로와 바닥에 받쳐져 있다. 왼쪽 아래 전화 장치는 일부와 받침 주변이 잘려 있어 지지면을 확정할 수 없지만 공중에 떠 있다고 볼 근거도 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "양팔을 벌린 미디엄 숏과 거만한 미소, 인물 참조에 가까운 얼굴은 더 충실하지만, 완장의 판독 가능한 글자가 명시적 금지 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "낮은 시점과 양팔을 벌린 구도는 맞지만, 박철진 대신 이전 숏 노인의 얼굴을 옮긴 듯한 외모이며 완장 글자도 노출된다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "남자는 양팔을 화면 좌우로 벌리고 손바닥을 앞쪽으로 펼친 채 카메라 쪽을 똑바로 본다. 화면 밖 방문자 윤성찬의 위치를 카메라 축에 둔 연출로 해석할 수 있지만, 상대가 보이지 않아 실제 눈맞춤까지 확인되지는 않는다. 오른쪽 칼은 칼끝이 목재 벽 안으로 들어가고 손잡이가 실내 쪽으로 돌출되어 있다.",
        "built_space": "왼쪽에 두 칸으로 나뉜 창 하나, 중앙 뒤에 출입구 하나, 창 아래 의자 두 개와 인물 뒤 의자 등받이 하나가 보인다. 오른쪽 목재 벽에는 지도와 종이 게시물, 칼 하나와 파손 흔적이 있다. 하단 탁자 위에는 스피커폰 하나가 놓였다. 참조의 창·출입구·목재 벽 관계와 낮의 자연광을 대체로 유지한다. 참조에서 보인 의자는 하나지만 전체 의자 수가 확정된 것은 아니므로 추가로 보이는 의자를 중복 설비로 단정할 수 없다.",
        "entities": "한 명의 중년 남성이 보이며 짧은 머리와 회색 섞인 수염, 얼굴 윤곽은 박철진 인물 참조에 비교적 가깝다. 참조의 올리브색 군복과 검은 완장은 유지하지만 군모는 없다. 민족적 배경은 외모만으로 확정할 수 없으며, 참조 인물과의 외형적 일치가 B보다 높다. 칼은 손이 아니라 벽에 남아 있고 스피커폰도 보인다. 완장에는 판독 가능한 흰색 한글이 있다.",
        "hard_violations": [
         "완장의 흰색 한글이 읽을 수 있는 형태로 노출되어, 화면 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "상체는 곧게 선 자세이고 두 팔은 어깨와 팔꿈치로 자연스럽게 지탱된다. 발은 미디엄 숏 밖이므로 접지는 확인할 수 없지만 공중에 떠 있다는 징후는 없다. 빈손의 손가락 구조도 대체로 자연스럽다. 칼은 벽에 박힌 칼날로 고정되고, 스피커폰은 탁자 면에 받쳐져 있다."
       },
       {
        "label": "A",
        "direction": "남자는 양팔과 손바닥을 좌우로 펼치고 카메라 쪽을 응시한다. 카메라 축의 화면 밖 방문자를 향한 시선으로 성립하지만 윤성찬은 직접 보이지 않는다. 오른쪽 칼은 벽을 향해 박혀 있고 손잡이는 왼쪽 실내 방향으로 나온다.",
        "built_space": "왼쪽에 창 하나와 의자 하나, 그 옆에 열린 문 하나가 있으며 문 너머로 상자 두 개가 포개져 있다. 오른쪽에는 목재 벽, 지도와 게시물, 칼 하나와 파손 자국이 보인다. 왼쪽 아래에는 전화 장치 일부가 잘려 보인다. 낮은 카메라가 천장까지 담으며 참조의 창·문·벽 배치와 주간 조명을 대체로 유지한다. 불가능한 반사나 중복된 문·칼은 보이지 않는다.",
        "entities": "한 명의 남성이 올리브색 군복과 검은 완장을 착용했지만, 흰 머리를 뒤로 넘긴 무수염의 고령 얼굴이다. 이는 수염이 있는 중년 박철진 참조보다 이전 숏의 노인 얼굴에 매우 가깝고, 지정된 44세 인물의 정체성과 맞지 않는다. 군모도 없다. 칼은 벽에 남아 있고 양손은 비어 있다. 왼쪽 아래 전화 장치는 일부만 보여 지정 스피커폰인지 확실히 확인하기 어렵다. 완장에는 판독 가능한 흰색 글자가 있다.",
        "hard_violations": [
         "완장에 판독 가능한 한글과 영문자 형태가 노출되어, 읽을 수 있는 글자 금지 조건을 위반한다."
        ],
        "physics": "서 있는 몸통에서 양팔을 벌린 자세이며 어깨·팔꿈치·손목의 연결은 물리적으로 가능하다. 하체와 발은 프레임 밖이어서 접지는 직접 확인되지 않지만 부유한 자세는 아니다. 칼은 벽에 박혀 지지되고 문 너머 상자들은 서로와 바닥에 받쳐져 있다. 왼쪽 아래 전화 장치는 일부와 받침 주변이 잘려 있어 지지면을 확정할 수 없지만 공중에 떠 있다고 볼 근거도 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.929,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.679,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 컷 인물의 얼굴을 현재 컷 캐릭터에 잘못 적용 (신원 위반)",
     "[gpt-high] 완장에 판독 가능한 한글과 영문자 형태가 노출되어, 읽을 수 있는 글자 금지 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 완장의 흰색 한글이 읽을 수 있는 형태로 노출되어, 화면 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 679
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "캐릭터 레퍼런스의 외모를 정확히 반영했으며, 양팔을 벌린 위압적인 자세와 배경 연속성(스피커폰, 벽에 박힌 칼)을 성공적으로 구현했습니다.  ★위반: [gpt-high] 완장의 흰색 한글이 읽을 수 있는 형태로 노출되어, 화면 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 679,
    "verdict_ko": "지정된 캐릭터가 아닌 이전 컷 스틸의 인물 얼굴을 잘못 가져와 사용하여 치명적인 신원 위반이 발생했습니다.  ★위반: [gemini-pro] 이전 컷 인물의 얼굴을 현재 컷 캐릭터에 잘못 적용 (신원 위반) / [gpt-high] 완장에 판독 가능한 한글과 영문자 형태가 노출되어, 읽을 수 있는 글자 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S18sh6_sel.png",
    "asset_id": "4edf9b21-a1e6-4015-9ee8-5cf4a0f47ad0",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd93-f69d-7ebd-9b0f-91280db80924",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S18sh6"
  }
 },
 "S18sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:25:03.424627+00:00",
  "fingerprint": "a336c76889fd14961dc3488b6438554313a8c4ed9b7576993a1c7f32d1b4cae7",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S18sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S18sh8_sel.png",
  "source_sha256": "c4e3abb41e33d368c19305368412f6024e86ef5b2f1da7dd2d007a9f639d21b3",
  "file": "S18sh8_cine.png",
  "staged_sha256": "6fe74ff1a500e4830c6978a6f78c93741b4d97189084f22a8c8d6b520b0df465",
  "latency_ms": 12620
 },
 "S19sh1::confined_fp_apt": {
  "applies": true,
  "reason_ko": "차량 내부(차량 캐빈)를 배경으로 하는 숏으로, 뒷좌석에 앉은 인물이 앞좌석 쪽으로 몸을 기울이는 구도를 정확히 표현하려면 앞뒤 좌석의 공간적 배치와 방향성을 명확히 해야 하므로 평면도 레이아웃 보조가 필요합니다.",
  "input_fingerprint": "ab1fd32e39876220"
 },
 "S19sh1::signage": {
  "fp": "e3c7418f69435a63",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::fccca8a535c3": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/confinedfp_base_fccca8a535c3.png",
  "place_text": "Inside the rear passenger compartment of a luxury car, facing the chauffeur's front seat through the compact cabin in daylight.",
  "input_fingerprint": "6f293ca47cf09864"
 },
 "S19sh1::confined_fp": {
  "reads": {
   "controls": "A steering wheel is located at the front, associated with the Chauffeur seat.",
   "mirrors": "No mirrors are indicated in the diagram.",
   "camera": "Positioned over the Center Console in the rear, pointing straight forward towards the Chauffeur seat.",
   "occupants": "윤성찬 is located in the Rear Left Passenger seat."
  },
  "mismatches": [
   "The diagram shows the camera pointing forward from the rear center console, which would only capture the back of the front seat, whereas the text requires a medium shot showing Yoon Sung-chan's fierce gaze as he leans forward (which would necessitate a backward-facing camera from the front cabin).",
   "The text mentions a front passenger seat providing foreground framing at frame left, contradicting the diagram's forward-facing angle where the Chauffeur seat occupies the center forward view."
  ],
  "scene_description_en": "The camera is situated above the rear center console, pointing directly forward toward the front cabin. In the center of the frame, the backrest of the Chauffeur's seat is prominently visible. Beyond the Chauffeur's seat, the steering wheel is seen in the distance. To the immediate left of the camera's forward view, Yoon Sung-chan occupies the Rear Left Passenger seat, positioned largely beside the camera rather than in front of it. To the right of the camera, the Rear Right Passenger seat is unoccupied. There are no mirrors or reflective surfaces depicted in this layout.",
  "fixed": true,
  "input_fingerprint": "8dfd607ef6da03f7"
 },
 "era_assess::63f75864458f32b7": {
  "subjects": [],
  "subject_text": "윤성찬의 승용차 내부\n앞좌석과 뒷좌석이 구분된 고급 세단 실내. 전방 운전대와 좌석 등받이가 보이고 측면 창문이 길게 나 있다.",
  "identity": "canonical",
  "scope_id": "L29",
  "scope_role": "location_interior",
  "scope_sha": "5cf76da2ee045fd2"
 },
 "S19sh1": {
  "input_fingerprint": "0e47ecac54691e1b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고급 승용차 뒷좌석에 앉은 윤성찬이 앞좌석 쪽으로 몸을 바짝 기울인 채 매서운 눈빛을 뿜는 구도.\n\nLOCATION (lock): Inside the rear passenger compartment of a luxury car, facing the chauffeur's front seat through the compact cabin in daylight. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Front passenger seat (Its inner edge lies beside the camera's view into the rear cabin) — Only a narrow rear-facing edge is visible at frame left; used as Provides restrained foreground framing without obscuring the forward lean; Rear seat (Occupied by 윤성찬 as he leans away from its back) — The backrest is seen obliquely behind his upper body; used as Clarifies that the command is delivered from the rear passenger position.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the cabin keeps his eyes readable without specifying an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬: He is seated in the car and retains his handkerchief from the office visit.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is situated above the rear center console, pointing directly forward toward the front cabin. In the center of the frame, the backrest of the Chauffeur's seat is prominently visible. Beyond the Chauffeur's seat, the steering wheel is seen in the distance. To the immediate left of the camera's forward view, Yoon Sung-chan occupies the Rear Left Passenger seat, positioned largely beside the camera rather than in front of it. To the right of the camera, the Rear Right Passenger seat is unoccupied. There are no mirrors or reflective surfaces depicted in this layout.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고급 승용차 뒷좌석에 앉은 윤성찬이 앞좌석 쪽으로 몸을 바짝 기울인 채 매서운 눈빛을 뿜는 구도.\n\nLOCATION (lock): Inside the rear passenger compartment of a luxury car, facing the chauffeur's front seat through the compact cabin in daylight. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the cabin keeps his eyes readable without specifying an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬: He is seated in the car and retains his handkerchief from the office visit.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is situated above the rear center console, pointing directly forward toward the front cabin. In the center of the frame, the backrest of the Chauffeur's seat is prominently visible. Beyond the Chauffeur's seat, the steering wheel is seen in the distance. To the immediate left of the camera's forward view, Yoon Sung-chan occupies the Rear Left Passenger seat, positioned largely beside the camera rather than in front of it. To the right of the camera, the Rear Right Passenger seat is unoccupied. There are no mirrors or reflective surfaces depicted in this layout.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고급 승용차 뒷좌석에 앉은 윤성찬이 앞좌석 쪽으로 몸을 바짝 기울인 채 매서운 눈빛을 뿜는 구도.\n\nLOCATION (lock): Inside the rear passenger compartment of a luxury car, facing the chauffeur's front seat through the compact cabin in daylight. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the cabin keeps his eyes readable without specifying an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬: He is seated in the car and retains his handkerchief from the office visit.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S19sh1_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "윤성찬",
     "path": "<bytes:781789>",
     "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S19sh1_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "윤성찬",
     "path": "<bytes:781789>",
     "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "시선이 앞좌석이 아닌 카메라 렌즈를 직접 향하고 있음.",
    "built_space": "카메라가 뒤에서 앞을 향하는 구도이나, 도면(좌측 뒷좌석)과 달리 인물이 우측 뒷좌석에 잘못 배치됨.",
    "entities": "윤성찬의 인상착의는 일치하나, 프롬프트에 명시된 손수건 소품이 누락됨.",
    "hard_violations": [
     "[gemini-pro] 도면에 명시된 인물의 좌석 위치(Rear Left)를 무시하고 우측에 배치함",
     "[gemini-pro] 하반신이 완전히 생략되고 상체가 시트 쿠션에 묻혀 융합된 물리적으로 불가능한 신체 구조",
     "[gpt-high] 윤성찬이 지정된 왼쪽 뒷좌석이 아니라 오른쪽의 앞선 별도 좌석 구역에 배치되어 있다.",
     "[gpt-high] 왼쪽 좌석 위에 요청되지 않은 별도 쿠션과 분리된 머리받침을 추가했다.",
     "[gpt-high] 운전대 중앙에 자동차 로고가 노출되어 로고 금지 조건을 위반한다."
    ],
    "physics": "인물의 하반신을 지지하는 바닥이나 다리가 없으며, 몸통이 둥둥 뜬 채로 시트에 녹아들어 있음."
   },
   {
    "label": "B",
    "direction": "앞좌석 방향을 향해 매섭게 쏘아보는 시선이 잘 표현됨.",
    "built_space": "카메라가 뒤에서 앞을 향하는 구도. 도면 지시대로 인물이 좌측 뒷좌석에 정확히 위치함.",
    "entities": "윤성찬의 외모와 의상이 일치하며, 오른손에 손수건을 쥐고 있음.",
    "hard_violations": [
     "[gemini-pro] 인물의 얼굴 뒤쪽 운전석 시트에 소속을 알 수 없는 세 번째 손(여분의 신체 부위)이 생성됨"
    ],
    "physics": "앞으로 몸을 기울인 자세는 자연스러우나, 얼굴 뒤의 여분의 손은 물리적으로 설명되지 않고 공중에 떠 있듯 시트를 잡고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지정된 좌석 배치(좌측)와 손수건 소품은 잘 반영했으나, 인물 뒤편에 정체불명의 세 번째 손이 생성되는 치명적인 오류로 인해 사용 불가함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "도면의 좌석 위치를 위반(우측 배치)하였고, 하반신이 시트와 융합된 형태의 심각한 신체 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "앞좌석 방향을 향해 매섭게 쏘아보는 시선이 잘 표현됨.",
        "built_space": "카메라가 뒤에서 앞을 향하는 구도. 도면 지시대로 인물이 좌측 뒷좌석에 정확히 위치함.",
        "entities": "윤성찬의 외모와 의상이 일치하며, 오른손에 손수건을 쥐고 있음.",
        "hard_violations": [
         "인물의 얼굴 뒤쪽 운전석 시트에 소속을 알 수 없는 세 번째 손(여분의 신체 부위)이 생성됨"
        ],
        "physics": "앞으로 몸을 기울인 자세는 자연스러우나, 얼굴 뒤의 여분의 손은 물리적으로 설명되지 않고 공중에 떠 있듯 시트를 잡고 있음."
       },
       {
        "label": "A",
        "direction": "시선이 앞좌석이 아닌 카메라 렌즈를 직접 향하고 있음.",
        "built_space": "카메라가 뒤에서 앞을 향하는 구도이나, 도면(좌측 뒷좌석)과 달리 인물이 우측 뒷좌석에 잘못 배치됨.",
        "entities": "윤성찬의 인상착의는 일치하나, 프롬프트에 명시된 손수건 소품이 누락됨.",
        "hard_violations": [
         "도면에 명시된 인물의 좌석 위치(Rear Left)를 무시하고 우측에 배치함",
         "하반신이 완전히 생략되고 상체가 시트 쿠션에 묻혀 융합된 물리적으로 불가능한 신체 구조"
        ],
        "physics": "인물의 하반신을 지지하는 바닥이나 다리가 없으며, 몸통이 둥둥 뜬 채로 시트에 녹아들어 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지정된 좌석 배치(좌측)와 손수건 소품은 잘 반영했으나, 인물 뒤편에 정체불명의 세 번째 손이 생성되는 치명적인 오류로 인해 사용 불가함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "도면의 좌석 위치를 위반(우측 배치)하였고, 하반신이 시트와 융합된 형태의 심각한 신체 오류가 발생함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앞좌석 방향을 향해 매섭게 쏘아보는 시선이 잘 표현됨.",
        "built_space": "카메라가 뒤에서 앞을 향하는 구도. 도면 지시대로 인물이 좌측 뒷좌석에 정확히 위치함.",
        "entities": "윤성찬의 외모와 의상이 일치하며, 오른손에 손수건을 쥐고 있음.",
        "hard_violations": [
         "인물의 얼굴 뒤쪽 운전석 시트에 소속을 알 수 없는 세 번째 손(여분의 신체 부위)이 생성됨"
        ],
        "physics": "앞으로 몸을 기울인 자세는 자연스러우나, 얼굴 뒤의 여분의 손은 물리적으로 설명되지 않고 공중에 떠 있듯 시트를 잡고 있음."
       },
       {
        "label": "A",
        "direction": "시선이 앞좌석이 아닌 카메라 렌즈를 직접 향하고 있음.",
        "built_space": "카메라가 뒤에서 앞을 향하는 구도이나, 도면(좌측 뒷좌석)과 달리 인물이 우측 뒷좌석에 잘못 배치됨.",
        "entities": "윤성찬의 인상착의는 일치하나, 프롬프트에 명시된 손수건 소품이 누락됨.",
        "hard_violations": [
         "도면에 명시된 인물의 좌석 위치(Rear Left)를 무시하고 우측에 배치함",
         "하반신이 완전히 생략되고 상체가 시트 쿠션에 묻혀 융합된 물리적으로 불가능한 신체 구조"
        ],
        "physics": "인물의 하반신을 지지하는 바닥이나 다리가 없으며, 몸통이 둥둥 뜬 채로 시트에 녹아들어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "뒷좌석 왼쪽에서 몸을 기울이고 손수건을 쥔 모습은 더 가깝지만, 시선이 앞좌석이 아닌 후방 카메라를 향하며 지정된 미디엄 숏과 좌석 배경 구성을 놓쳤다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "인물이 오른쪽의 별도 좌석에 놓이고 중앙 등받이가 화면을 지배하며, 불필요한 분리 머리받침과 로고까지 보여 지정된 좌석 배치와 전방을 향한 동작을 크게 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "상체는 왼쪽 앞좌석과 중앙 콘솔 쪽으로 기울지만, 얼굴과 두 눈은 오른쪽 뒤편의 카메라 쪽을 향한다. 매서운 표정은 있으나 시선의 목표가 전방 운전석은 아니다. 무기나 이동 물체는 없다.",
        "built_space": "앞좌석 등받이와 머리받침이 좌우에 각각 하나씩 있고, 중앙 콘솔 하나, 대시보드 하나, 오른쪽 운전대 하나, 실내 거울 하나가 보인다. 윤성찬은 왼쪽 뒷좌석 쪽에 있다. 카메라는 뒤에서 전방을 넓게 보므로, 왼쪽에 앞좌석의 좁은 가장자리만 두고 인물 뒤로 뒷좌석 등받이를 비스듬히 보이라는 구성이 아니다. 특히 오른쪽 앞좌석 등받이가 크게 노출된다. 거울에 불가능한 인물 반사는 없다.",
        "entities": "보이는 사람은 회색 머리의 고령 남성 한 명뿐이다. 주름진 얼굴, 머리 모양, 짙은 남색 정장과 흰 셔츠, 푸른 넥타이는 인물 참조와 대체로 맞는다. 외모만으로 국적이나 민족을 확정할 수는 없다. 오른손에는 흰 손수건이 있다. 가죽과 목재 마감은 고급 승용차로 읽히며 창밖은 낮이다. 판독 가능한 글자나 도식은 보이지 않는다.",
        "hard_violations": [],
        "physics": "하체는 왼쪽 뒷좌석 쪽에 놓이고, 오른손과 팔뚝은 중앙 팔걸이에 기대며 손수건을 누르고 있다. 다른 손은 앞좌석의 안쪽 가장자리에 걸쳐 상체를 지탱한다. 앉은 상태에서 몸을 앞으로 기울이는 동작으로 가능한 자세이며, 지지 없이 떠 있는 신체나 물건은 없다."
       },
       {
        "label": "B",
        "direction": "남성은 중앙 등받이의 오른쪽에서 몸을 안쪽으로 틀고, 얼굴과 눈을 뒤쪽 카메라로 돌린다. 눈빛은 날카롭지만 전방 운전석을 향하지 않는다. 몸의 회전 역시 앞좌석으로 바짝 기울여 말하는 순간보다 뒤를 돌아보는 모습에 가깝다.",
        "built_space": "중앙에 큰 등받이와 머리받침 한 세트가 있고, 그 왼쪽에는 별도 좌방석 위에 쿠션과 분리된 머리받침이 놓여 있다. 오른쪽에는 남성이 앉은 별도 좌방석이 있으며, 화면 아래에는 다시 후열 좌석 가장자리와 중앙 콘솔이 보인다. 운전대는 왼쪽이다. 인물은 도면의 왼쪽 뒷좌석이 아니라 오른쪽의 더 앞선 좌석 구역에 놓인다. 중앙 등받이가 상체 일부를 가리고, 요구된 왼쪽의 좁은 앞좌석 가장자리와 인물 뒤 뒷좌석 등받이 구성도 성립하지 않는다.",
        "entities": "고령의 회색 머리 남성 한 명이며 얼굴과 남색 정장, 흰 셔츠는 참조와 대체로 유사하다. 손수건처럼 보이는 흰 천은 가슴 주머니에 꽂혀 있으나 손에 보유한 상태는 확인되지 않는다. 가죽·목재 실내와 주간 채광은 맞는다. 왼쪽 좌석 위의 별도 쿠션과 분리 머리받침은 요청되지 않은 소품이며, 운전대 중앙에는 식별 가능한 타원형 자동차 로고가 보인다.",
        "hard_violations": [
         "윤성찬이 지정된 왼쪽 뒷좌석이 아니라 오른쪽의 앞선 별도 좌석 구역에 배치되어 있다.",
         "왼쪽 좌석 위에 요청되지 않은 별도 쿠션과 분리된 머리받침을 추가했다.",
         "운전대 중앙에 자동차 로고가 노출되어 로고 금지 조건을 위반한다."
        ],
        "physics": "남성의 엉덩이는 오른쪽 좌방석에 닿아 있어 몸 자체가 떠 있지는 않다. 팔은 몸 앞쪽으로 굽혀져 있고 손은 대부분 가려진다. 왼쪽의 쿠션과 분리 머리받침도 좌방석 위에 놓여 지지된다. 따라서 무지지 부유 문제보다는 좌석 배치와 몸을 돌린 방향이 핵심 문제다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "뒷좌석 왼쪽에서 몸을 기울이고 손수건을 쥔 모습은 더 가깝지만, 시선이 앞좌석이 아닌 후방 카메라를 향하며 지정된 미디엄 숏과 좌석 배경 구성을 놓쳤다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "인물이 오른쪽의 별도 좌석에 놓이고 중앙 등받이가 화면을 지배하며, 불필요한 분리 머리받침과 로고까지 보여 지정된 좌석 배치와 전방을 향한 동작을 크게 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "상체는 왼쪽 앞좌석과 중앙 콘솔 쪽으로 기울지만, 얼굴과 두 눈은 오른쪽 뒤편의 카메라 쪽을 향한다. 매서운 표정은 있으나 시선의 목표가 전방 운전석은 아니다. 무기나 이동 물체는 없다.",
        "built_space": "앞좌석 등받이와 머리받침이 좌우에 각각 하나씩 있고, 중앙 콘솔 하나, 대시보드 하나, 오른쪽 운전대 하나, 실내 거울 하나가 보인다. 윤성찬은 왼쪽 뒷좌석 쪽에 있다. 카메라는 뒤에서 전방을 넓게 보므로, 왼쪽에 앞좌석의 좁은 가장자리만 두고 인물 뒤로 뒷좌석 등받이를 비스듬히 보이라는 구성이 아니다. 특히 오른쪽 앞좌석 등받이가 크게 노출된다. 거울에 불가능한 인물 반사는 없다.",
        "entities": "보이는 사람은 회색 머리의 고령 남성 한 명뿐이다. 주름진 얼굴, 머리 모양, 짙은 남색 정장과 흰 셔츠, 푸른 넥타이는 인물 참조와 대체로 맞는다. 외모만으로 국적이나 민족을 확정할 수는 없다. 오른손에는 흰 손수건이 있다. 가죽과 목재 마감은 고급 승용차로 읽히며 창밖은 낮이다. 판독 가능한 글자나 도식은 보이지 않는다.",
        "hard_violations": [],
        "physics": "하체는 왼쪽 뒷좌석 쪽에 놓이고, 오른손과 팔뚝은 중앙 팔걸이에 기대며 손수건을 누르고 있다. 다른 손은 앞좌석의 안쪽 가장자리에 걸쳐 상체를 지탱한다. 앉은 상태에서 몸을 앞으로 기울이는 동작으로 가능한 자세이며, 지지 없이 떠 있는 신체나 물건은 없다."
       },
       {
        "label": "A",
        "direction": "남성은 중앙 등받이의 오른쪽에서 몸을 안쪽으로 틀고, 얼굴과 눈을 뒤쪽 카메라로 돌린다. 눈빛은 날카롭지만 전방 운전석을 향하지 않는다. 몸의 회전 역시 앞좌석으로 바짝 기울여 말하는 순간보다 뒤를 돌아보는 모습에 가깝다.",
        "built_space": "중앙에 큰 등받이와 머리받침 한 세트가 있고, 그 왼쪽에는 별도 좌방석 위에 쿠션과 분리된 머리받침이 놓여 있다. 오른쪽에는 남성이 앉은 별도 좌방석이 있으며, 화면 아래에는 다시 후열 좌석 가장자리와 중앙 콘솔이 보인다. 운전대는 왼쪽이다. 인물은 도면의 왼쪽 뒷좌석이 아니라 오른쪽의 더 앞선 좌석 구역에 놓인다. 중앙 등받이가 상체 일부를 가리고, 요구된 왼쪽의 좁은 앞좌석 가장자리와 인물 뒤 뒷좌석 등받이 구성도 성립하지 않는다.",
        "entities": "고령의 회색 머리 남성 한 명이며 얼굴과 남색 정장, 흰 셔츠는 참조와 대체로 유사하다. 손수건처럼 보이는 흰 천은 가슴 주머니에 꽂혀 있으나 손에 보유한 상태는 확인되지 않는다. 가죽·목재 실내와 주간 채광은 맞는다. 왼쪽 좌석 위의 별도 쿠션과 분리 머리받침은 요청되지 않은 소품이며, 운전대 중앙에는 식별 가능한 타원형 자동차 로고가 보인다.",
        "hard_violations": [
         "윤성찬이 지정된 왼쪽 뒷좌석이 아니라 오른쪽의 앞선 별도 좌석 구역에 배치되어 있다.",
         "왼쪽 좌석 위에 요청되지 않은 별도 쿠션과 분리된 머리받침을 추가했다.",
         "운전대 중앙에 자동차 로고가 노출되어 로고 금지 조건을 위반한다."
        ],
        "physics": "남성의 엉덩이는 오른쪽 좌방석에 닿아 있어 몸 자체가 떠 있지는 않다. 팔은 몸 앞쪽으로 굽혀져 있고 손은 대부분 가려진다. 왼쪽의 쿠션과 분리 머리받침도 좌방석 위에 놓여 지지된다. 따라서 무지지 부유 문제보다는 좌석 배치와 몸을 돌린 방향이 핵심 문제다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.75,
    "B": 1.75
   },
   "violations": {
    "B": [
     "[gemini-pro] 인물의 얼굴 뒤쪽 운전석 시트에 소속을 알 수 없는 세 번째 손(여분의 신체 부위)이 생성됨"
    ],
    "A": [
     "[gemini-pro] 도면에 명시된 인물의 좌석 위치(Rear Left)를 무시하고 우측에 배치함",
     "[gemini-pro] 하반신이 완전히 생략되고 상체가 시트 쿠션에 묻혀 융합된 물리적으로 불가능한 신체 구조",
     "[gpt-high] 윤성찬이 지정된 왼쪽 뒷좌석이 아니라 오른쪽의 앞선 별도 좌석 구역에 배치되어 있다.",
     "[gpt-high] 왼쪽 좌석 위에 요청되지 않은 별도 쿠션과 분리된 머리받침을 추가했다.",
     "[gpt-high] 운전대 중앙에 자동차 로고가 노출되어 로고 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지정된 좌석 배치(좌측)와 손수건 소품은 잘 반영했으나, 인물 뒤편에 정체불명의 세 번째 손이 생성되는 치명적인 오류로 인해 사용 불가함.  ★위반: [gemini-pro] 인물의 얼굴 뒤쪽 운전석 시트에 소속을 알 수 없는 세 번째 손(여분의 신체 부위)이 생성됨"
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "도면의 좌석 위치를 위반(우측 배치)하였고, 하반신이 시트와 융합된 형태의 심각한 신체 오류가 발생함.  ★위반: [gemini-pro] 도면에 명시된 인물의 좌석 위치(Rear Left)를 무시하고 우측에 배치함 / [gemini-pro] 하반신이 완전히 생략되고 상체가 시트 쿠션에 묻혀 융합된 물리적으로 불가능한 신체 구조 / [gpt-high] 윤성찬이 지정된 왼쪽 뒷좌석이 아니라 오른쪽의 앞선 별도 좌석 구역에 배치되어 있다. / [gpt-high] 왼쪽 좌석 위에 요청되지 않은 별도 쿠션과 분리된 머리받침을 추가했다. / [gpt-high] 운전대 중앙에 자동차 로고가 노출되어 로고 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S19sh1_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "윤성찬",
    "path": "<bytes:781789>",
    "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd94-c5bb-78c3-abfe-e1451af0d817",
  "confined_fp": {
   "base_key": "confinedfp::fccca8a535c3",
   "apt_reason": "차량 내부(차량 캐빈)를 배경으로 하는 숏으로, 뒷좌석에 앉은 인물이 앞좌석 쪽으로 몸을 기울이는 구도를 정확히 표현하려면 앞뒤 좌석의 공간적 배치와 방향성을 명확히 해야 하므로 평면도 레이아웃 보조가 필요합니다.",
   "fixed": true,
   "mismatches": [
    "The diagram shows the camera pointing forward from the rear center console, which would only capture the back of the front seat, whereas the text requires a medium shot showing Yoon Sung-chan's fierce gaze as he leans forward (which would necessitate a backward-facing camera from the front cabin).",
    "The text mentions a front passenger seat providing foreground framing at frame left, contradicting the diagram's forward-facing angle where the Chauffeur seat occupies the center forward view."
   ]
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S19sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:25:16.437358+00:00",
  "fingerprint": "dd834fd6f5bbd427cc8089d671e36ba28505f32bf4c6962cc60314605c10cebc",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S19sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S19sh1_sel.png",
  "source_sha256": "4d4e8a45624c10d00e4d33ce357750558ae985da0631478354aad7051e6e90d0",
  "file": "S19sh1_cine.png",
  "staged_sha256": "f9d9e8f73adb94a6fbabbdd2d323208ec7063bf0d93f87af339692ed3dc92c09",
  "latency_ms": 11983
 },
 "S20sh5::signage": {
  "fp": "2ffd08dc5ba8e0ff",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::89be1a99f37d12f3": {
  "subjects": [],
  "subject_text": "인천 난민촌 거리와 주거 골목\n컨테이너 주택이 밀집한 거리와 좁은 골목. 도로변에 작은 상점들이 이어지고 가로등과 방송용 스피커가 설치되어 있다.",
  "identity": "canonical",
  "scope_id": "L12",
  "scope_role": "location_exterior",
  "scope_sha": "06eda36f926c7bbf"
 },
 "groupbg::home_approach_alley": {
  "input_fingerprint": "868ec7ee9fd55424",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "home_approach_alley",
    "tags": [
     "S20sh5",
     "S20sh7"
    ]
   },
   "context_sig": "688085c837652031"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At an outdoor residential alley junction near the searched container home in the refugee settlement.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리와 주거 골목: 녹슨 컨테이너 박스들이 빽빽하게 들어서 있고 낡은 가로등과 확성기가 달린 비좁은 슬럼가. (특징: 녹슨 외벽의 컨테이너 가건물들; 수선 흔적이 있는 낡은 옷을 입은 사람들; 어두워지면 일제히 꺼지는 낡은 가로등)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우와 페드로, 실실 웃으며 왔다가 박철진 수하들이 현우 컨테이너 안을 뒤지는 것을 보고 놀라 숨는다.\n- 현우, 페드로와 함께 멀어지려는데, 순간 앞을 가로막는 경찰들.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At an outdoor residential alley junction near the searched container home in the refugee settlement.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리와 주거 골목: 녹슨 컨테이너 박스들이 빽빽하게 들어서 있고 낡은 가로등과 확성기가 달린 비좁은 슬럼가. (특징: 녹슨 외벽의 컨테이너 가건물들; 수선 흔적이 있는 낡은 옷을 입은 사람들; 어두워지면 일제히 꺼지는 낡은 가로등)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우와 페드로, 실실 웃으며 왔다가 박철진 수하들이 현우 컨테이너 안을 뒤지는 것을 보고 놀라 숨는다.\n- 현우, 페드로와 함께 멀어지려는데, 순간 앞을 가로막는 경찰들.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_home_approach_alley_f0e441.png",
  "asset_id": "0f733de5-e006-4bc6-b386-b0ef150b4b0d",
  "input_asset_ids": [
   "7331e487-c728-4126-88f2-40fa9a6a780b"
  ],
  "origin_tag": "S20sh5",
  "place_text": "At an outdoor residential alley junction near the searched container home in the refugee settlement.",
  "origin_inputs": {
   "place_text": "At an outdoor residential alley junction near the searched container home in the refugee settlement.",
   "time_of_day_en": "day",
   "conti_asset_id": "7331e487-c728-4126-88f2-40fa9a6a780b"
  }
 },
 "S20sh5::bgfirst_bg": {
  "input_fingerprint": "c6e71f96341501fb",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 다른 골목을 빠져나가려는 이현우의 앞을 건장한 아시아계 사복경찰들이 일렬로 막아선 위압적인 구도.\n\nLOCATION (lock): At an outdoor residential alley junction near the searched container home in the refugee settlement.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Alley retreat route (Blocked across its width by the plainclothes officers) — The route recedes obliquely between the foreground pair and the police line; used as Shows that the officers occupy the available exit rather than merely standing nearby.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime ambient light keeps the obstructed route and the opposing bodies readable without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 다른 골목을 빠져나가려는 이현우의 앞을 건장한 아시아계 사복경찰들이 일렬로 막아선 위압적인 구도.\n\nLOCATION (lock): At an outdoor residential alley junction near the searched container home in the refugee settlement.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Alley retreat route (Blocked across its width by the plainclothes officers) — The route recedes obliquely between the foreground pair and the police line; used as Shows that the officers occupy the available exit rather than merely standing nearby.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime ambient light keeps the obstructed route and the opposing bodies readable without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S20sh5__bgfirst_bg.png",
  "asset_id": "82d1593b-39c7-491f-8a91-9cdcd9d94969",
  "input_asset_ids": [
   "7331e487-c728-4126-88f2-40fa9a6a780b",
   "0f733de5-e006-4bc6-b386-b0ef150b4b0d"
  ]
 },
 "S20sh5": {
  "input_fingerprint": "81dc30d1f49cab3b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다른 골목을 빠져나가려는 이현우의 앞을 건장한 아시아계 사복경찰들이 일렬로 막아선 위압적인 구도.\n\nLOCATION (lock): At an outdoor residential alley junction near the searched container home in the refugee settlement. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Alley retreat route (Blocked across its width by the plainclothes officers) — The route recedes obliquely between the foreground pair and the police line; used as Shows that the officers occupy the available exit rather than merely standing nearby.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime ambient light keeps the obstructed route and the opposing bodies readable without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container home 7-31 is being searched, with the resulting disorder remaining in place. 이현우: He is in the alley, still bearing his earlier facial injuries and dog-bitten leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다른 골목을 빠져나가려는 이현우의 앞을 건장한 아시아계 사복경찰들이 일렬로 막아선 위압적인 구도.\n\nLOCATION (lock): At an outdoor residential alley junction near the searched container home in the refugee settlement. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Alley retreat route (Blocked across its width by the plainclothes officers) — The route recedes obliquely between the foreground pair and the police line; used as Shows that the officers occupy the available exit rather than merely standing nearby.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime ambient light keeps the obstructed route and the opposing bodies readable without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container home 7-31 is being searched, with the resulting disorder remaining in place. 이현우: He is in the alley, still bearing his earlier facial injuries and dog-bitten leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다른 골목을 빠져나가려는 이현우의 앞을 건장한 아시아계 사복경찰들이 일렬로 막아선 위압적인 구도.\n\nLOCATION (lock): At an outdoor residential alley junction near the searched container home in the refugee settlement. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Alley retreat route (Blocked across its width by the plainclothes officers) — The route recedes obliquely between the foreground pair and the police line; used as Shows that the officers occupy the available exit rather than merely standing nearby.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime ambient light keeps the obstructed route and the opposing bodies readable without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container home 7-31 is being searched, with the resulting disorder remaining in place. 이현우: He is in the alley, still bearing his earlier facial injuries and dog-bitten leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S20sh5__bgfirst_bg.png",
     "asset_id": "82d1593b-39c7-491f-8a91-9cdcd9d94969",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S20sh5.png",
     "asset_id": "7331e487-c728-4126-88f2-40fa9a6a780b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_home_approach_alley_f0e441.png",
     "asset_id": "0f733de5-e006-4bc6-b386-b0ef150b4b0d",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "중앙의 경찰이 이현우를 향해 손을 뻗어 제지하고 있으며, 이현우는 경찰들을 마주 보고 있음. 우측 전경의 인물은 상황을 지켜봄.",
    "built_space": "레퍼런스와 일치하는 컨테이너 주택 골목. 인물들이 골목의 폭을 가로막고 적절히 배치됨.",
    "entities": "이현우(얼굴 상처, 개에 물린 다리 붕대 일치함), 사복경찰 5명. 샷 텍스트에 없는 우측 전경의 모자 쓴 남성 1명.",
    "hard_violations": [
     "[gemini-pro] 프롬프트의 샷 텍스트에 명시되지 않은 인물(우측 전경의 모자 쓴 남성)이 임의로 추가됨."
    ],
    "physics": "모든 인물이 지면에 두 발을 딛고 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "이현우는 화면 좌측을 향해 시선을 두고 있으며, 뒤편에 일렬로 선 경찰들은 주로 카메라 정면 방향을 주시함.",
    "built_space": "레퍼런스와 일치하는 컨테이너 골목이며 인물들이 골목을 가로막고 있음. 원경에 고층 건물들이 보임.",
    "entities": "이현우(얼굴 상처, 다리 붕대 구현됨), 전술 조끼를 착용한 경찰 6명. 추가로 개입된 인물 없음.",
    "hard_violations": [
     "[gpt-high] 경찰 조끼에 판독 가능한 경찰 글자 표식이 있어 읽을 수 있는 글자 금지 조건을 위반한다.",
     "[gpt-high] 장면에 지정되지 않은 권총 세 정을 경찰들의 손에 추가했다."
    ],
    "physics": "모든 인물이 바닥에 발을 딛고 자연스럽게 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷 텍스트에 없는 임의의 인물을 우측 전경에 추가하여 등장인물 제한 규칙을 심각하게 위반했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 인물들만 정확히 배치했으며, 이현우의 부상 상태와 골목을 막아선 경찰들의 구도를 무난하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙의 경찰이 이현우를 향해 손을 뻗어 제지하고 있으며, 이현우는 경찰들을 마주 보고 있음. 우측 전경의 인물은 상황을 지켜봄.",
        "built_space": "레퍼런스와 일치하는 컨테이너 주택 골목. 인물들이 골목의 폭을 가로막고 적절히 배치됨.",
        "entities": "이현우(얼굴 상처, 개에 물린 다리 붕대 일치함), 사복경찰 5명. 샷 텍스트에 없는 우측 전경의 모자 쓴 남성 1명.",
        "hard_violations": [
         "프롬프트의 샷 텍스트에 명시되지 않은 인물(우측 전경의 모자 쓴 남성)이 임의로 추가됨."
        ],
        "physics": "모든 인물이 지면에 두 발을 딛고 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "이현우는 화면 좌측을 향해 시선을 두고 있으며, 뒤편에 일렬로 선 경찰들은 주로 카메라 정면 방향을 주시함.",
        "built_space": "레퍼런스와 일치하는 컨테이너 골목이며 인물들이 골목을 가로막고 있음. 원경에 고층 건물들이 보임.",
        "entities": "이현우(얼굴 상처, 다리 붕대 구현됨), 전술 조끼를 착용한 경찰 6명. 추가로 개입된 인물 없음.",
        "hard_violations": [],
        "physics": "모든 인물이 바닥에 발을 딛고 자연스럽게 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷 텍스트에 없는 임의의 인물을 우측 전경에 추가하여 등장인물 제한 규칙을 심각하게 위반했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 인물들만 정확히 배치했으며, 이현우의 부상 상태와 골목을 막아선 경찰들의 구도를 무난하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "중앙의 경찰이 이현우를 향해 손을 뻗어 제지하고 있으며, 이현우는 경찰들을 마주 보고 있음. 우측 전경의 인물은 상황을 지켜봄.",
        "built_space": "레퍼런스와 일치하는 컨테이너 주택 골목. 인물들이 골목의 폭을 가로막고 적절히 배치됨.",
        "entities": "이현우(얼굴 상처, 개에 물린 다리 붕대 일치함), 사복경찰 5명. 샷 텍스트에 없는 우측 전경의 모자 쓴 남성 1명.",
        "hard_violations": [
         "프롬프트의 샷 텍스트에 명시되지 않은 인물(우측 전경의 모자 쓴 남성)이 임의로 추가됨."
        ],
        "physics": "모든 인물이 지면에 두 발을 딛고 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "이현우는 화면 좌측을 향해 시선을 두고 있으며, 뒤편에 일렬로 선 경찰들은 주로 카메라 정면 방향을 주시함.",
        "built_space": "레퍼런스와 일치하는 컨테이너 골목이며 인물들이 골목을 가로막고 있음. 원경에 고층 건물들이 보임.",
        "entities": "이현우(얼굴 상처, 다리 붕대 구현됨), 전술 조끼를 착용한 경찰 6명. 추가로 개입된 인물 없음.",
        "hard_violations": [],
        "physics": "모든 인물이 바닥에 발을 딛고 자연스럽게 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이현우와 경찰들이 같은 전경 방향을 바라봐 앞길을 막는 대치가 약하며, 판독 가능한 경찰 표식과 추가된 권총이 금지 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "이현우를 향한 정지 손짓과 골목 폭을 가로막는 사복경찰들의 대치가 명확하고 장소도 충실하나, 경찰들의 깊이가 달라 정확한 일렬 배치는 덜 선명하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 전경의 이현우는 경찰들에게 등을 일부 돌린 채 화면 앞쪽 왼편을 바라본다. 뒤쪽 경찰 여섯 명도 대체로 카메라 쪽을 보고 있어 서로 마주한 차단보다는 추적하는 배치로 읽힌다. 보이는 권총 세 정의 총구는 아래쪽 노면을 향하며 이현우를 직접 겨누지는 않는다.",
        "built_space": "녹슨 컨테이너 양쪽 벽 사이에 포장 골목이 이어진다. 왼쪽 가까운 창 하나, 오른쪽 문 두 개와 벽등 하나, 전봇대의 확성기와 가로등, 양쪽 실외기·상자·기댄 판재가 보인다. 재료와 주요 시설은 참조와 유사하지만 먼 배경이 대형 고층건물 군으로 바뀌었다. 경찰 여섯 명은 골목 폭에 걸쳐 있으나 앞뒤로 분산되어 있고, 이현우 앞의 전경 통로는 비어 있다. 요청한 비스듬한 후퇴로보다는 정면으로 깊어지는 골목이다.",
        "entities": "이현우 한 명과 성인 아시아계 남성 경찰 여섯 명이 보인다. 이현우의 젊은 얼굴, 헝클어진 검은 머리, 낡은 짙은 후드와 찢어진 청바지는 참조에 대체로 부합하며 얼굴 상처와 피 묻은 다리 붕대도 있다. 경찰들은 건장하지만 일반 사복 위에 경찰 표식이 있는 전술조끼를 입었다. 조끼에는 판독 가능한 경찰 글자 표식이 있으며, 요청하지 않은 권총 세 정이 추가되었다. 주택 수색의 구체적인 흔적은 확인되지 않는다.",
        "hard_violations": [
         "경찰 조끼에 판독 가능한 경찰 글자 표식이 있어 읽을 수 있는 글자 금지 조건을 위반한다.",
         "장면에 지정되지 않은 권총 세 정을 경찰들의 손에 추가했다."
        ],
        "physics": "경찰들의 보이는 발은 노면에 닿아 있고 벌린 다리로 체중을 지탱한다. 이현우는 발이 프레임 밖이지만 두 다리가 노면 방향으로 자연스럽게 이어져 있어 부유한다고 볼 근거는 없다. 권총은 각각 손으로 잡고 있으며 붕대는 다리에 감겨 있다. 상자와 판재는 바닥 또는 벽에 지지되어 있고 불가능한 신체 자세는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 전경의 이현우는 오른쪽 안쪽의 경찰들을 향해 몸과 얼굴을 돌리고 있다. 중앙 경찰은 이현우 쪽을 보면서 손바닥을 그에게 내밀어 정지를 요구한다. 나머지 경찰들도 대체로 이현우 쪽에 주의를 두며, 오른쪽 전경의 모자 쓴 남성 역시 대치 방향을 향한다. 무기나 총구는 보이지 않는다.",
        "built_space": "왼쪽의 녹슨 컨테이너와 창 하나, 오른쪽 문 두 개와 벽등 하나, 검은 차양, 확성기 달린 전봇대, 가로등과 전선이 참조의 골목을 잘 유지한다. 왼쪽에는 덮개 씌운 물건과 기댄 판재·갈색 통·흰 용기, 오른쪽에는 실외기와 상자가 보인다. 중경 경찰 다섯 명이 양쪽 컨테이너 사이 통행 폭을 가로막고, 전경 양쪽 인물 사이로 골목이 이어진다. 다만 중앙 경찰이 앞으로 나와 있어 엄밀한 일렬은 아니며 골목의 사선 후퇴도 강하지 않다.",
        "entities": "이현우 한 명, 중경의 건장한 성인 아시아계 남성 경찰 다섯 명, 오른쪽 전경의 모자 쓴 성인 남성 한 명이 보인다. 전경 남성은 같은 사복경찰 무리로 해석 가능하지만 별도의 신원 표시는 없다. 이현우의 검은 헝클어진 머리와 젊은 옆얼굴, 낡은 후드와 청바지는 참조에 대체로 맞고 얼굴 상처와 피 묻은 다리 붕대가 보인다. 경찰들은 재킷과 셔츠 등 사복을 입었으며 읽을 수 있는 글자나 노출된 무기는 없다. 특정 주택의 수색 흔적은 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "중경 경찰들의 발은 노면에 닿아 있고 벌린 다리와 체중 이동으로 안정적으로 서 있다. 중앙 경찰의 정지 손짓은 어깨와 팔꿈치에서 자연스럽게 연결된다. 이현우의 하체는 프레임 아래로 이어지며 서 있는 자세가 성립한다. 오른쪽 전경 남성은 몸을 낮추고 무릎을 굽힌 자세로 보이고, 접지 부위는 잘려 있어 정확한 지지점은 확인되지 않지만 공중에 떠 있다는 증거는 없다. 주변 물건도 바닥과 벽에 지지되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이현우와 경찰들이 같은 전경 방향을 바라봐 앞길을 막는 대치가 약하며, 판독 가능한 경찰 표식과 추가된 권총이 금지 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "이현우를 향한 정지 손짓과 골목 폭을 가로막는 사복경찰들의 대치가 명확하고 장소도 충실하나, 경찰들의 깊이가 달라 정확한 일렬 배치는 덜 선명하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 전경의 이현우는 경찰들에게 등을 일부 돌린 채 화면 앞쪽 왼편을 바라본다. 뒤쪽 경찰 여섯 명도 대체로 카메라 쪽을 보고 있어 서로 마주한 차단보다는 추적하는 배치로 읽힌다. 보이는 권총 세 정의 총구는 아래쪽 노면을 향하며 이현우를 직접 겨누지는 않는다.",
        "built_space": "녹슨 컨테이너 양쪽 벽 사이에 포장 골목이 이어진다. 왼쪽 가까운 창 하나, 오른쪽 문 두 개와 벽등 하나, 전봇대의 확성기와 가로등, 양쪽 실외기·상자·기댄 판재가 보인다. 재료와 주요 시설은 참조와 유사하지만 먼 배경이 대형 고층건물 군으로 바뀌었다. 경찰 여섯 명은 골목 폭에 걸쳐 있으나 앞뒤로 분산되어 있고, 이현우 앞의 전경 통로는 비어 있다. 요청한 비스듬한 후퇴로보다는 정면으로 깊어지는 골목이다.",
        "entities": "이현우 한 명과 성인 아시아계 남성 경찰 여섯 명이 보인다. 이현우의 젊은 얼굴, 헝클어진 검은 머리, 낡은 짙은 후드와 찢어진 청바지는 참조에 대체로 부합하며 얼굴 상처와 피 묻은 다리 붕대도 있다. 경찰들은 건장하지만 일반 사복 위에 경찰 표식이 있는 전술조끼를 입었다. 조끼에는 판독 가능한 경찰 글자 표식이 있으며, 요청하지 않은 권총 세 정이 추가되었다. 주택 수색의 구체적인 흔적은 확인되지 않는다.",
        "hard_violations": [
         "경찰 조끼에 판독 가능한 경찰 글자 표식이 있어 읽을 수 있는 글자 금지 조건을 위반한다.",
         "장면에 지정되지 않은 권총 세 정을 경찰들의 손에 추가했다."
        ],
        "physics": "경찰들의 보이는 발은 노면에 닿아 있고 벌린 다리로 체중을 지탱한다. 이현우는 발이 프레임 밖이지만 두 다리가 노면 방향으로 자연스럽게 이어져 있어 부유한다고 볼 근거는 없다. 권총은 각각 손으로 잡고 있으며 붕대는 다리에 감겨 있다. 상자와 판재는 바닥 또는 벽에 지지되어 있고 불가능한 신체 자세는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 전경의 이현우는 오른쪽 안쪽의 경찰들을 향해 몸과 얼굴을 돌리고 있다. 중앙 경찰은 이현우 쪽을 보면서 손바닥을 그에게 내밀어 정지를 요구한다. 나머지 경찰들도 대체로 이현우 쪽에 주의를 두며, 오른쪽 전경의 모자 쓴 남성 역시 대치 방향을 향한다. 무기나 총구는 보이지 않는다.",
        "built_space": "왼쪽의 녹슨 컨테이너와 창 하나, 오른쪽 문 두 개와 벽등 하나, 검은 차양, 확성기 달린 전봇대, 가로등과 전선이 참조의 골목을 잘 유지한다. 왼쪽에는 덮개 씌운 물건과 기댄 판재·갈색 통·흰 용기, 오른쪽에는 실외기와 상자가 보인다. 중경 경찰 다섯 명이 양쪽 컨테이너 사이 통행 폭을 가로막고, 전경 양쪽 인물 사이로 골목이 이어진다. 다만 중앙 경찰이 앞으로 나와 있어 엄밀한 일렬은 아니며 골목의 사선 후퇴도 강하지 않다.",
        "entities": "이현우 한 명, 중경의 건장한 성인 아시아계 남성 경찰 다섯 명, 오른쪽 전경의 모자 쓴 성인 남성 한 명이 보인다. 전경 남성은 같은 사복경찰 무리로 해석 가능하지만 별도의 신원 표시는 없다. 이현우의 검은 헝클어진 머리와 젊은 옆얼굴, 낡은 후드와 청바지는 참조에 대체로 맞고 얼굴 상처와 피 묻은 다리 붕대가 보인다. 경찰들은 재킷과 셔츠 등 사복을 입었으며 읽을 수 있는 글자나 노출된 무기는 없다. 특정 주택의 수색 흔적은 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "중경 경찰들의 발은 노면에 닿아 있고 벌린 다리와 체중 이동으로 안정적으로 서 있다. 중앙 경찰의 정지 손짓은 어깨와 팔꿈치에서 자연스럽게 연결된다. 이현우의 하체는 프레임 아래로 이어지며 서 있는 자세가 성립한다. 오른쪽 전경 남성은 몸을 낮추고 무릎을 굽힌 자세로 보이고, 접지 부위는 잘려 있어 정확한 지지점은 확인되지 않지만 공중에 떠 있다는 증거는 없다. 주변 물건도 바닥과 벽에 지지되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.375
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.125
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트의 샷 텍스트에 명시되지 않은 인물(우측 전경의 모자 쓴 남성)이 임의로 추가됨."
    ],
    "B": [
     "[gpt-high] 경찰 조끼에 판독 가능한 경찰 글자 표식이 있어 읽을 수 있는 글자 금지 조건을 위반한다.",
     "[gpt-high] 장면에 지정되지 않은 권총 세 정을 경찰들의 손에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "A": 1179,
   "B": 1125
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "샷 텍스트에 없는 임의의 인물을 우측 전경에 추가하여 등장인물 제한 규칙을 심각하게 위반했습니다.  ★위반: [gemini-pro] 프롬프트의 샷 텍스트에 명시되지 않은 인물(우측 전경의 모자 쓴 남성)이 임의로 추가됨."
   },
   {
    "label": "B",
    "score": 1125,
    "verdict_ko": "지시된 인물들만 정확히 배치했으며, 이현우의 부상 상태와 골목을 막아선 경찰들의 구도를 무난하게 구현했습니다.  ★위반: [gpt-high] 경찰 조끼에 판독 가능한 경찰 글자 표식이 있어 읽을 수 있는 글자 금지 조건을 위반한다. / [gpt-high] 장면에 지정되지 않은 권총 세 정을 경찰들의 손에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_home_approach_alley_f0e441.png",
    "asset_id": "0f733de5-e006-4bc6-b386-b0ef150b4b0d",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd95-8c02-781b-b57d-2166b7a2d360",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S20sh5__bgfirst_bg.png",
   "bg_asset_id": "82d1593b-39c7-491f-8a91-9cdcd9d94969",
   "bg_record_key": "S20sh5::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "home_approach_alley",
   "groupbg_asset_id": "0f733de5-e006-4bc6-b386-b0ef150b4b0d"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S20sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:25:28.904444+00:00",
  "fingerprint": "769a461ca0cfc91b7534c765b93cd0b74361792557f845a06d5400cc150fa652",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S20sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S20sh5_sel.png",
  "source_sha256": "e6974ccc30ba7795b98471c2eb2c705591fa82cedaacdcb7ca7396b73951a307",
  "file": "S20sh5_cine.png",
  "staged_sha256": "c5816c3960e95669d4384afacfc363b4c0372b6ae6aff44de551980dff7b1e19",
  "latency_ms": 9746
 },
 "S20sh7::signage": {
  "fp": "15dab62afac7c3dd",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S20sh7": {
  "input_fingerprint": "01a220927a115a81",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 페드로의 두 손바닥이 앞의 사복경찰 가슴팍에 닿아 맹렬하게 밀어내는 힘에 의해 경찰의 상체가 뒤로 젖혀진 타격의 순간.\n\nLOCATION (lock): In the refugee settlement's residential alley where plainclothes police block the escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Escape opening beyond the recoiling officer in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Opening in the blocked alley (Becoming available as the officer is pushed backward) — Visible beyond the recoiling officer on the right side of the composition; used as Connects the physical shove to the possibility of 이현우's escape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight balance so the contact and recoil read through posture rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the alley walls, ground surfaces, and daylight appearance from the reference. EXCLUDE fallen market boxes and pursuit debris, which belong to later locations.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed by the ongoing search. 페드로: He is pitched forward in a forceful pushing motion.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 페드로의 두 손바닥이 앞의 사복경찰 가슴팍에 닿아 맹렬하게 밀어내는 힘에 의해 경찰의 상체가 뒤로 젖혀진 타격의 순간.\n\nLOCATION (lock): In the refugee settlement's residential alley where plainclothes police block the escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Escape opening beyond the recoiling officer in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Opening in the blocked alley (Becoming available as the officer is pushed backward) — Visible beyond the recoiling officer on the right side of the composition; used as Connects the physical shove to the possibility of 이현우's escape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight balance so the contact and recoil read through posture rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the alley walls, ground surfaces, and daylight appearance from the reference. EXCLUDE fallen market boxes and pursuit debris, which belong to later locations.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed by the ongoing search. 페드로: He is pitched forward in a forceful pushing motion.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 페드로의 두 손바닥이 앞의 사복경찰 가슴팍에 닿아 맹렬하게 밀어내는 힘에 의해 경찰의 상체가 뒤로 젖혀진 타격의 순간.\n\nLOCATION (lock): In the refugee settlement's residential alley where plainclothes police block the escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Escape opening beyond the recoiling officer in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Opening in the blocked alley (Becoming available as the officer is pushed backward) — Visible beyond the recoiling officer on the right side of the composition; used as Connects the physical shove to the possibility of 이현우's escape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight balance so the contact and recoil read through posture rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the alley walls, ground surfaces, and daylight appearance from the reference. EXCLUDE fallen market boxes and pursuit debris, which belong to later locations.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed by the ongoing search. 페드로: He is pitched forward in a forceful pushing motion.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "페드로의 시선과 두 손이 앞선 사복경찰의 가슴을 정확히 향하여 닿아 있음.",
    "built_space": "컨테이너가 늘어선 주거지 골목. 프롬프트가 요구한 우측 중간 배경의 '도주로 개방 공간'이 닫힌 문과 벽으로 막혀 있어 제대로 묘사되지 않음.",
    "entities": "페드로는 참조 이미지의 인상착의(비니, 스터드 가죽 재킷, 체인 등)와 정확히 일치함. 사복경찰 1명만 등장하며 이전 샷의 인물과 겹치지 않게 새로 디자인됨.",
    "hard_violations": [
     "[gpt-high] 골목 먼 배경에 장면이 요구하지 않은 추가 인물 한 명이 보입니다."
    ],
    "physics": "페드로가 두 발을 안정적으로 딛고 체중을 실어 미는 힘과, 그 타격에 의해 경찰의 상체가 뒤로 젖혀지는 물리적 작용이 매우 자연스러움."
   },
   {
    "label": "B",
    "direction": "페드로가 사복경찰의 가슴팍을 양손으로 밀어내며 시선을 맞추고 있음.",
    "built_space": "컨테이너 골목길이며, 우측 컨테이너에 열린 문이 있어 프롬프트가 요구한 도주로 개방 공간을 구도에 맞게 구현함.",
    "entities": "페드로의 인상착의는 일치함. 그러나 타격받는 경찰과 배경의 인물들이 이전 샷 참조 이미지의 인물들(얼굴 및 체크 셔츠 등 복장)과 완벽히 동일함.",
    "hard_violations": [
     "[gemini-pro] 이전 샷 참조 이미지의 인물 외모 및 복장 복사 (등장 금지 지시 정면 위반)",
     "[gemini-pro] 대본에 없는 추가 인물 다수 등장 (Invented people / Extra bodies)",
     "[gpt-high] 장면에 필요한 페드로와 밀리는 경찰 외에 남성 세 명을 추가했습니다."
    ],
    "physics": "페드로가 힘껏 밀어내는 런지 자세와 경찰이 크게 중심을 잃고 밀려나는 동작이 올바르게 연결되어 지지받고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "우측 도주로 배경 요소가 누락되었으나, 미디엄 샷 구도와 이전 샷 인물 배제 지시를 엄격히 준수하며 밀어내는 타격 액션을 물리적으로 사실감 있게 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "우측 배경에 열린 문을 잘 배치했으나, 등장해서는 안 될 이전 샷의 인물들을 그대로 복사하고 대본에 없는 인원들을 다수 추가하여 치명적으로 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "페드로의 시선과 두 손이 앞선 사복경찰의 가슴을 정확히 향하여 닿아 있음.",
        "built_space": "컨테이너가 늘어선 주거지 골목. 프롬프트가 요구한 우측 중간 배경의 '도주로 개방 공간'이 닫힌 문과 벽으로 막혀 있어 제대로 묘사되지 않음.",
        "entities": "페드로는 참조 이미지의 인상착의(비니, 스터드 가죽 재킷, 체인 등)와 정확히 일치함. 사복경찰 1명만 등장하며 이전 샷의 인물과 겹치지 않게 새로 디자인됨.",
        "hard_violations": [],
        "physics": "페드로가 두 발을 안정적으로 딛고 체중을 실어 미는 힘과, 그 타격에 의해 경찰의 상체가 뒤로 젖혀지는 물리적 작용이 매우 자연스러움."
       },
       {
        "label": "B",
        "direction": "페드로가 사복경찰의 가슴팍을 양손으로 밀어내며 시선을 맞추고 있음.",
        "built_space": "컨테이너 골목길이며, 우측 컨테이너에 열린 문이 있어 프롬프트가 요구한 도주로 개방 공간을 구도에 맞게 구현함.",
        "entities": "페드로의 인상착의는 일치함. 그러나 타격받는 경찰과 배경의 인물들이 이전 샷 참조 이미지의 인물들(얼굴 및 체크 셔츠 등 복장)과 완벽히 동일함.",
        "hard_violations": [
         "이전 샷 참조 이미지의 인물 외모 및 복장 복사 (등장 금지 지시 정면 위반)",
         "대본에 없는 추가 인물 다수 등장 (Invented people / Extra bodies)"
        ],
        "physics": "페드로가 힘껏 밀어내는 런지 자세와 경찰이 크게 중심을 잃고 밀려나는 동작이 올바르게 연결되어 지지받고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "우측 도주로 배경 요소가 누락되었으나, 미디엄 샷 구도와 이전 샷 인물 배제 지시를 엄격히 준수하며 밀어내는 타격 액션을 물리적으로 사실감 있게 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "우측 배경에 열린 문을 잘 배치했으나, 등장해서는 안 될 이전 샷의 인물들을 그대로 복사하고 대본에 없는 인원들을 다수 추가하여 치명적으로 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "페드로의 시선과 두 손이 앞선 사복경찰의 가슴을 정확히 향하여 닿아 있음.",
        "built_space": "컨테이너가 늘어선 주거지 골목. 프롬프트가 요구한 우측 중간 배경의 '도주로 개방 공간'이 닫힌 문과 벽으로 막혀 있어 제대로 묘사되지 않음.",
        "entities": "페드로는 참조 이미지의 인상착의(비니, 스터드 가죽 재킷, 체인 등)와 정확히 일치함. 사복경찰 1명만 등장하며 이전 샷의 인물과 겹치지 않게 새로 디자인됨.",
        "hard_violations": [],
        "physics": "페드로가 두 발을 안정적으로 딛고 체중을 실어 미는 힘과, 그 타격에 의해 경찰의 상체가 뒤로 젖혀지는 물리적 작용이 매우 자연스러움."
       },
       {
        "label": "B",
        "direction": "페드로가 사복경찰의 가슴팍을 양손으로 밀어내며 시선을 맞추고 있음.",
        "built_space": "컨테이너 골목길이며, 우측 컨테이너에 열린 문이 있어 프롬프트가 요구한 도주로 개방 공간을 구도에 맞게 구현함.",
        "entities": "페드로의 인상착의는 일치함. 그러나 타격받는 경찰과 배경의 인물들이 이전 샷 참조 이미지의 인물들(얼굴 및 체크 셔츠 등 복장)과 완벽히 동일함.",
        "hard_violations": [
         "이전 샷 참조 이미지의 인물 외모 및 복장 복사 (등장 금지 지시 정면 위반)",
         "대본에 없는 추가 인물 다수 등장 (Invented people / Extra bodies)"
        ],
        "physics": "페드로가 힘껏 밀어내는 런지 자세와 경찰이 크게 중심을 잃고 밀려나는 동작이 올바르게 연결되어 지지받고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "두 손바닥의 가슴 접촉과 경찰의 반동은 보이지만, 불필요한 인물 세 명을 추가하고 미디엄 숏을 전신 구도로 넓혀 핵심 구도 지시를 위반합니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "미디엄 숏에서 페드로의 전방 기울임과 양손 밀치기, 경찰의 상체 반동을 더 충실하게 담았지만, 골목 먼 배경의 추가 인물 때문에 완전한 합격은 아닙니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "페드로는 오른쪽 경찰을 바라보며 두 손바닥을 경찰의 윗가슴에 대고 오른쪽 뒤로 밀고 있습니다. 경찰은 상체와 얼굴이 뒤로 젖혀져 위쪽을 향합니다. 뒤쪽 두 남성은 접촉 장면을 보고 있으며, 오른쪽 전경 남성도 그쪽으로 몸을 돌리고 있습니다. 무기나 겨냥하는 소품은 없습니다.",
        "built_space": "녹슨 골판 컨테이너 벽 두 줄과 균열 난 포장 골목, 왼쪽 창문 하나와 배관 하나, 기대 놓은 판재, 덮개 씌운 적치물, 갈색 통, 파란 상자, 흰 용기가 보입니다. 오른쪽에는 열린 밝은 문짝 하나와 검은 차양이 있습니다. 재질과 낮빛은 장소 참조에 가깝습니다. 그러나 주인공 두 명의 발끝까지 보여 미디엄 숏보다 넓고, 경찰 너머의 통로에는 다른 남성들이 서 있으며 오른쪽 전경 인물이 탈출 방향을 크게 가립니다.",
        "entities": "총 다섯 명으로, 페드로와 밀리는 사복경찰 외에 배경 남성 두 명과 오른쪽 전경 남성 한 명이 있습니다. 페드로는 젊은 남성 얼굴, 검은 비니, 낡은 가죽 재킷, 팔 장식, 체인과 찢어진 검은 바지로 인물 참조의 주요 특징을 따릅니다. 밀리는 인물은 어두운 재킷과 체크 셔츠를 입은 성인 남성으로 사복경찰 역할에 맞지만 이전 스틸 중앙 경찰의 외형과 의상을 강하게 반복합니다. 판독 가능한 글자는 보이지 않습니다.",
        "hard_violations": [
         "장면에 필요한 페드로와 밀리는 경찰 외에 남성 세 명을 추가했습니다."
        ],
        "physics": "페드로는 벌린 두 다리와 바닥에 닿은 부츠로 버티며 양팔을 뻗습니다. 경찰도 발을 노면에 둔 채 무릎을 굽히고 상체만 뒤로 젖혀져 있어 밀침에 따른 균형 상실로 설명됩니다. 두 손의 가슴 접촉이 보이고, 들어 올린 경찰 팔은 어깨에 자연스럽게 연결됩니다. 배경 남성들은 서 있으며 전경 남성의 하체는 화면 밖입니다. 근거 없이 공중에 뜬 몸이나 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "페드로의 시선과 뻗은 양팔이 바로 앞 경찰의 윗가슴을 향하고, 두 손바닥이 그 부위에 닿습니다. 경찰의 상체는 오른쪽 뒤로 젖혀지고 턱과 시선은 위쪽으로 향합니다. 밀어내는 방향과 반동 방향이 서로 맞습니다. 먼 배경 인물의 시선은 판별하기 어렵습니다.",
        "built_space": "두 줄의 녹슨 컨테이너와 갈라진 포장면이 이어지고, 왼쪽에는 창문 하나가 부분적으로 보입니다. 오른쪽에는 닫힌 문 하나, 벽등 하나, 검은 차양, 실외기 두 대, 기대 놓은 판재와 상자 두 개가 보입니다. 전선과 전봇대도 유지되어 장소의 재질과 낮빛은 참조에 가깝습니다. 두 인물을 허벅지 부근에서 자른 미디엄 구도이며, 경찰 오른쪽 아래로 빈 통로 일부가 드러납니다. 다만 중간 오른쪽 배경의 탈출 개구부는 경찰 몸에 상당 부분 가려져 선명하지 않습니다.",
        "entities": "페드로와 성인 남성 사복경찰 외에, 골목 소실점 부근에서 작은 사람의 머리와 상체 일부가 보입니다. 페드로는 참조에 가까운 젊은 얼굴과 검은 비니, 낡은 가죽 재킷, 팔목 장식과 바지 체인을 갖췄습니다. 경찰은 남색 일상복 재킷과 청바지를 입어 사복 차림으로 읽힙니다. 두 주요 인물 외의 먼 배경 인물은 장면에 요구되지 않았습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "골목 먼 배경에 장면이 요구하지 않은 추가 인물 한 명이 보입니다."
        ],
        "physics": "페드로는 골반보다 어깨를 앞으로 내밀어 체중을 싣고, 양손으로 경찰 가슴을 누릅니다. 경찰은 골반과 허벅지가 아래로 이어진 상태에서 상체가 뒤로 기울어 밀침의 반동으로 가능한 자세입니다. 두 사람의 발과 지면 접점은 프레임 밖이므로 직접 확인할 수 없지만, 몸 전체가 공중에 떠 있다는 증거는 없습니다. 재킷과 체인은 몸에 걸려 있고 실외기와 상자는 바닥에 놓여 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "두 손바닥의 가슴 접촉과 경찰의 반동은 보이지만, 불필요한 인물 세 명을 추가하고 미디엄 숏을 전신 구도로 넓혀 핵심 구도 지시를 위반합니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "미디엄 숏에서 페드로의 전방 기울임과 양손 밀치기, 경찰의 상체 반동을 더 충실하게 담았지만, 골목 먼 배경의 추가 인물 때문에 완전한 합격은 아닙니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "페드로는 오른쪽 경찰을 바라보며 두 손바닥을 경찰의 윗가슴에 대고 오른쪽 뒤로 밀고 있습니다. 경찰은 상체와 얼굴이 뒤로 젖혀져 위쪽을 향합니다. 뒤쪽 두 남성은 접촉 장면을 보고 있으며, 오른쪽 전경 남성도 그쪽으로 몸을 돌리고 있습니다. 무기나 겨냥하는 소품은 없습니다.",
        "built_space": "녹슨 골판 컨테이너 벽 두 줄과 균열 난 포장 골목, 왼쪽 창문 하나와 배관 하나, 기대 놓은 판재, 덮개 씌운 적치물, 갈색 통, 파란 상자, 흰 용기가 보입니다. 오른쪽에는 열린 밝은 문짝 하나와 검은 차양이 있습니다. 재질과 낮빛은 장소 참조에 가깝습니다. 그러나 주인공 두 명의 발끝까지 보여 미디엄 숏보다 넓고, 경찰 너머의 통로에는 다른 남성들이 서 있으며 오른쪽 전경 인물이 탈출 방향을 크게 가립니다.",
        "entities": "총 다섯 명으로, 페드로와 밀리는 사복경찰 외에 배경 남성 두 명과 오른쪽 전경 남성 한 명이 있습니다. 페드로는 젊은 남성 얼굴, 검은 비니, 낡은 가죽 재킷, 팔 장식, 체인과 찢어진 검은 바지로 인물 참조의 주요 특징을 따릅니다. 밀리는 인물은 어두운 재킷과 체크 셔츠를 입은 성인 남성으로 사복경찰 역할에 맞지만 이전 스틸 중앙 경찰의 외형과 의상을 강하게 반복합니다. 판독 가능한 글자는 보이지 않습니다.",
        "hard_violations": [
         "장면에 필요한 페드로와 밀리는 경찰 외에 남성 세 명을 추가했습니다."
        ],
        "physics": "페드로는 벌린 두 다리와 바닥에 닿은 부츠로 버티며 양팔을 뻗습니다. 경찰도 발을 노면에 둔 채 무릎을 굽히고 상체만 뒤로 젖혀져 있어 밀침에 따른 균형 상실로 설명됩니다. 두 손의 가슴 접촉이 보이고, 들어 올린 경찰 팔은 어깨에 자연스럽게 연결됩니다. 배경 남성들은 서 있으며 전경 남성의 하체는 화면 밖입니다. 근거 없이 공중에 뜬 몸이나 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "페드로의 시선과 뻗은 양팔이 바로 앞 경찰의 윗가슴을 향하고, 두 손바닥이 그 부위에 닿습니다. 경찰의 상체는 오른쪽 뒤로 젖혀지고 턱과 시선은 위쪽으로 향합니다. 밀어내는 방향과 반동 방향이 서로 맞습니다. 먼 배경 인물의 시선은 판별하기 어렵습니다.",
        "built_space": "두 줄의 녹슨 컨테이너와 갈라진 포장면이 이어지고, 왼쪽에는 창문 하나가 부분적으로 보입니다. 오른쪽에는 닫힌 문 하나, 벽등 하나, 검은 차양, 실외기 두 대, 기대 놓은 판재와 상자 두 개가 보입니다. 전선과 전봇대도 유지되어 장소의 재질과 낮빛은 참조에 가깝습니다. 두 인물을 허벅지 부근에서 자른 미디엄 구도이며, 경찰 오른쪽 아래로 빈 통로 일부가 드러납니다. 다만 중간 오른쪽 배경의 탈출 개구부는 경찰 몸에 상당 부분 가려져 선명하지 않습니다.",
        "entities": "페드로와 성인 남성 사복경찰 외에, 골목 소실점 부근에서 작은 사람의 머리와 상체 일부가 보입니다. 페드로는 참조에 가까운 젊은 얼굴과 검은 비니, 낡은 가죽 재킷, 팔목 장식과 바지 체인을 갖췄습니다. 경찰은 남색 일상복 재킷과 청바지를 입어 사복 차림으로 읽힙니다. 두 주요 인물 외의 먼 배경 인물은 장면에 요구되지 않았습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "골목 먼 배경에 장면이 요구하지 않은 추가 인물 한 명이 보입니다."
        ],
        "physics": "페드로는 골반보다 어깨를 앞으로 내밀어 체중을 싣고, 양손으로 경찰 가슴을 누릅니다. 경찰은 골반과 허벅지가 아래로 이어진 상태에서 상체가 뒤로 기울어 밀침의 반동으로 가능한 자세입니다. 두 사람의 발과 지면 접점은 프레임 밖이므로 직접 확인할 수 없지만, 몸 전체가 공중에 떠 있다는 증거는 없습니다. 재킷과 체인은 몸에 걸려 있고 실외기와 상자는 바닥에 놓여 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.929
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.679
   },
   "violations": {
    "B": [
     "[gemini-pro] 이전 샷 참조 이미지의 인물 외모 및 복장 복사 (등장 금지 지시 정면 위반)",
     "[gemini-pro] 대본에 없는 추가 인물 다수 등장 (Invented people / Extra bodies)",
     "[gpt-high] 장면에 필요한 페드로와 밀리는 경찰 외에 남성 세 명을 추가했습니다."
    ],
    "A": [
     "[gpt-high] 골목 먼 배경에 장면이 요구하지 않은 추가 인물 한 명이 보입니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 679
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "우측 도주로 배경 요소가 누락되었으나, 미디엄 샷 구도와 이전 샷 인물 배제 지시를 엄격히 준수하며 밀어내는 타격 액션을 물리적으로 사실감 있게 구현함.  ★위반: [gpt-high] 골목 먼 배경에 장면이 요구하지 않은 추가 인물 한 명이 보입니다."
   },
   {
    "label": "B",
    "score": 679,
    "verdict_ko": "우측 배경에 열린 문을 잘 배치했으나, 등장해서는 안 될 이전 샷의 인물들을 그대로 복사하고 대본에 없는 인원들을 다수 추가하여 치명적으로 위반함.  ★위반: [gemini-pro] 이전 샷 참조 이미지의 인물 외모 및 복장 복사 (등장 금지 지시 정면 위반) / [gemini-pro] 대본에 없는 추가 인물 다수 등장 (Invented people / Extra bodies) / [gpt-high] 장면에 필요한 페드로와 밀리는 경찰 외에 남성 세 명을 추가했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S20sh5_sel.png",
    "asset_id": "a0314a1f-6782-446f-903d-e7a0ddb779ad",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:826925>",
    "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd96-2fb5-70f4-a8a2-922ab9e9e40a",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S20sh5"
  }
 },
 "S20sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:25:38.990932+00:00",
  "fingerprint": "32f4480ea6b2f3a103dd1385e6907f8776ff9394c8b96fd5896e3fa9481a81a0",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S20sh7_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S20sh7_sel.png",
  "source_sha256": "b0803a57c7f4d9ea6ff4f4d7363cf9cfd61c5f6d43ffa02f701ba803549d7d4c",
  "file": "S20sh7_cine.png",
  "staged_sha256": "2940baf95ae78962f9fb2278e212a269f21b47bb30a391b845911445a086ad88",
  "latency_ms": 13212
 },
 "S20sh12::signage": {
  "fp": "f6cfb748408c42a4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::market_capture_alley": {
  "input_fingerprint": "bed50cc9c3fa08ab",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "market_capture_alley",
    "tags": [
     "S20sh12"
    ]
   },
   "context_sig": "4ab53133e771b5bd"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the entrance to a narrower side alley branching off the refugee settlement's market lanes.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리와 주거 골목: 녹슨 컨테이너 박스들이 빽빽하게 들어서 있고 낡은 가로등과 확성기가 달린 비좁은 슬럼가. (특징: 녹슨 외벽의 컨테이너 가건물들; 수선 흔적이 있는 낡은 옷을 입은 사람들; 어두워지면 일제히 꺼지는 낡은 가로등)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 더욱 좁은 골목길로 도망가려던 찰나-! 퍽!!\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the entrance to a narrower side alley branching off the refugee settlement's market lanes.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리와 주거 골목: 녹슨 컨테이너 박스들이 빽빽하게 들어서 있고 낡은 가로등과 확성기가 달린 비좁은 슬럼가. (특징: 녹슨 외벽의 컨테이너 가건물들; 수선 흔적이 있는 낡은 옷을 입은 사람들; 어두워지면 일제히 꺼지는 낡은 가로등)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 더욱 좁은 골목길로 도망가려던 찰나-! 퍽!!\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_market_capture_alley_13d54d.png",
  "asset_id": "b72ff740-1baa-4a51-b31a-135364ea551f",
  "input_asset_ids": [
   "c419904c-288b-4c09-87c9-7533ec42de5a"
  ],
  "origin_tag": "S20sh12",
  "place_text": "At the entrance to a narrower side alley branching off the refugee settlement's market lanes.",
  "origin_inputs": {
   "place_text": "At the entrance to a narrower side alley branching off the refugee settlement's market lanes.",
   "time_of_day_en": "day",
   "conti_asset_id": "c419904c-288b-4c09-87c9-7533ec42de5a"
  }
 },
 "S20sh12::bgfirst_bg": {
  "input_fingerprint": "3f0e6b77418ac40d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 맹렬히 휘둘러진 개머리판이 이현우의 광대뼈에 정통으로 꽂히며 얼굴이 옆으로 확 꺾인 타격의 찰나.\n\nLOCATION (lock): At the entrance to a narrower side alley branching off the refugee settlement's market lanes.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rifle butt (In contact with 이현우's cheek during the strike) — Enters laterally from frame right with its striking end against the cheek; the wielder remains outside the crop; used as Makes the interception readable without letting the weapon overwhelm the human reaction; Narrow alley entrance (The route 이현우 was attempting to enter when intercepted) — A limited oblique view remains behind his collapsing shoulder; used as Retains the spatial context of the interrupted escape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged daytime ambient illumination keeps the cheek contact and facial reaction clear without inventing additional injury detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 맹렬히 휘둘러진 개머리판이 이현우의 광대뼈에 정통으로 꽂히며 얼굴이 옆으로 확 꺾인 타격의 찰나.\n\nLOCATION (lock): At the entrance to a narrower side alley branching off the refugee settlement's market lanes.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rifle butt (In contact with 이현우's cheek during the strike) — Enters laterally from frame right with its striking end against the cheek; the wielder remains outside the crop; used as Makes the interception readable without letting the weapon overwhelm the human reaction; Narrow alley entrance (The route 이현우 was attempting to enter when intercepted) — A limited oblique view remains behind his collapsing shoulder; used as Retains the spatial context of the interrupted escape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged daytime ambient illumination keeps the cheek contact and facial reaction clear without inventing additional injury detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S20sh12__bgfirst_bg.png",
  "asset_id": "d2cafdfa-96ab-4e94-a85e-b019a99054d1",
  "input_asset_ids": [
   "c419904c-288b-4c09-87c9-7533ec42de5a",
   "b72ff740-1baa-4a51-b31a-135364ea551f"
  ]
 },
 "S20sh12": {
  "input_fingerprint": "007552e36dea95bb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 맹렬히 휘둘러진 개머리판이 이현우의 광대뼈에 정통으로 꽂히며 얼굴이 옆으로 확 꺾인 타격의 찰나.\n\nLOCATION (lock): At the entrance to a narrower side alley branching off the refugee settlement's market lanes. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rifle butt (In contact with 이현우's cheek during the strike) — Enters laterally from frame right with its striking end against the cheek; the wielder remains outside the crop; used as Makes the interception readable without letting the weapon overwhelm the human reaction; Narrow alley entrance (The route 이현우 was attempting to enter when intercepted) — A limited oblique view remains behind his collapsing shoulder; used as Retains the spatial context of the interrupted escape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged daytime ambient illumination keeps the cheek contact and facial reaction clear without inventing additional injury detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market-alley boxes knocked down during the escape remain scattered. Container 7-31 remains disturbed by the search. 이현우: He retains his dog-bitten leg and earlier facial injuries while taking the rifle-butt blow. The resulting facial bruising carries into detention.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 맹렬히 휘둘러진 개머리판이 이현우의 광대뼈에 정통으로 꽂히며 얼굴이 옆으로 확 꺾인 타격의 찰나.\n\nLOCATION (lock): At the entrance to a narrower side alley branching off the refugee settlement's market lanes. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rifle butt (In contact with 이현우's cheek during the strike) — Enters laterally from frame right with its striking end against the cheek; the wielder remains outside the crop; used as Makes the interception readable without letting the weapon overwhelm the human reaction; Narrow alley entrance (The route 이현우 was attempting to enter when intercepted) — A limited oblique view remains behind his collapsing shoulder; used as Retains the spatial context of the interrupted escape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged daytime ambient illumination keeps the cheek contact and facial reaction clear without inventing additional injury detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market-alley boxes knocked down during the escape remain scattered. Container 7-31 remains disturbed by the search. 이현우: He retains his dog-bitten leg and earlier facial injuries while taking the rifle-butt blow. The resulting facial bruising carries into detention.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 맹렬히 휘둘러진 개머리판이 이현우의 광대뼈에 정통으로 꽂히며 얼굴이 옆으로 확 꺾인 타격의 찰나.\n\nLOCATION (lock): At the entrance to a narrower side alley branching off the refugee settlement's market lanes. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rifle butt (In contact with 이현우's cheek during the strike) — Enters laterally from frame right with its striking end against the cheek; the wielder remains outside the crop; used as Makes the interception readable without letting the weapon overwhelm the human reaction; Narrow alley entrance (The route 이현우 was attempting to enter when intercepted) — A limited oblique view remains behind his collapsing shoulder; used as Retains the spatial context of the interrupted escape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged daytime ambient illumination keeps the cheek contact and facial reaction clear without inventing additional injury detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market-alley boxes knocked down during the escape remain scattered. Container 7-31 remains disturbed by the search. 이현우: He retains his dog-bitten leg and earlier facial injuries while taking the rifle-butt blow. The resulting facial bruising carries into detention.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S20sh12__bgfirst_bg.png",
     "asset_id": "d2cafdfa-96ab-4e94-a85e-b019a99054d1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S20sh12.png",
     "asset_id": "c419904c-288b-4c09-87c9-7533ec42de5a",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_market_capture_alley_13d54d.png",
     "asset_id": "b72ff740-1baa-4a51-b31a-135364ea551f",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "개머리판이 왼쪽을 향해 인물의 목과 턱을 타격하고 있으며, 인물의 고개는 뒤로 젖혀져 있음.",
    "built_space": "배경에 참고 이미지와 일치하는 컨테이너와 골목길이 보이며, 좁은 골목길 입구의 구조가 적절히 배치됨.",
    "entities": "이현우(참고 이미지의 외모와 일치, 후드티 착용, 얼굴 상처 있음), 소총(나무 개머리판).",
    "hard_violations": [
     "[gemini-pro] 이미지 위에 검은색 화살표들이 그려져 있음 (leaked markers/diagrams)",
     "[gpt-high] 실사 화면 위에 방향을 표시하는 화살표 세 개가 삽입되어 도식·오버레이 금지 조건을 위반한다.",
     "[gpt-high] 화면 밖에 있어야 할 공격자의 손과 팔을 오른쪽 프레임 안에 노출한다."
    ],
    "physics": "인물은 두 발로 서 있고, 총기는 화면 오른쪽에 보이는 손에 의해 지지됨."
   },
   {
    "label": "B",
    "direction": "개머리판이 프레임 오른쪽에서 들어와 인물의 광대뼈에 정통으로 꽂혔으며, 타격 방향으로 고개가 심하게 꺾임.",
    "built_space": "인물의 어깨 너머로 참고 이미지의 좁은 골목길 입구와 컨테이너가 제한적으로 보임.",
    "entities": "이현우(참고 이미지의 외모와 일치, 후드티 착용, 얼굴 상처 및 멍 있음), 소총(검은색 개머리판 및 손잡이).",
    "hard_violations": [],
    "physics": "인물의 몸은 바닥에 지탱되어 있고, 소총은 화면 밖으로 이어져 보이지 않는 프레임 밖의 손에 의해 지지되며 타격 충격으로 입에서 침이 흩뿌려짐."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 샷과 타격 부위(광대뼈)를 정확히 구현했으며, 무기를 휘두르는 인물을 프레임 밖으로 배치하는 지침을 잘 따랐습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화면에 지시되지 않은 검은색 화살표(다이어그램)가 유출된 치명적 위반이 있으며, 무기를 쥔 손이 프레임에 포함되어 지침에 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "개머리판이 왼쪽을 향해 인물의 목과 턱을 타격하고 있으며, 인물의 고개는 뒤로 젖혀져 있음.",
        "built_space": "배경에 참고 이미지와 일치하는 컨테이너와 골목길이 보이며, 좁은 골목길 입구의 구조가 적절히 배치됨.",
        "entities": "이현우(참고 이미지의 외모와 일치, 후드티 착용, 얼굴 상처 있음), 소총(나무 개머리판).",
        "hard_violations": [
         "이미지 위에 검은색 화살표들이 그려져 있음 (leaked markers/diagrams)"
        ],
        "physics": "인물은 두 발로 서 있고, 총기는 화면 오른쪽에 보이는 손에 의해 지지됨."
       },
       {
        "label": "B",
        "direction": "개머리판이 프레임 오른쪽에서 들어와 인물의 광대뼈에 정통으로 꽂혔으며, 타격 방향으로 고개가 심하게 꺾임.",
        "built_space": "인물의 어깨 너머로 참고 이미지의 좁은 골목길 입구와 컨테이너가 제한적으로 보임.",
        "entities": "이현우(참고 이미지의 외모와 일치, 후드티 착용, 얼굴 상처 및 멍 있음), 소총(검은색 개머리판 및 손잡이).",
        "hard_violations": [],
        "physics": "인물의 몸은 바닥에 지탱되어 있고, 소총은 화면 밖으로 이어져 보이지 않는 프레임 밖의 손에 의해 지지되며 타격 충격으로 입에서 침이 흩뿌려짐."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 샷과 타격 부위(광대뼈)를 정확히 구현했으며, 무기를 휘두르는 인물을 프레임 밖으로 배치하는 지침을 잘 따랐습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화면에 지시되지 않은 검은색 화살표(다이어그램)가 유출된 치명적 위반이 있으며, 무기를 쥔 손이 프레임에 포함되어 지침에 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "개머리판이 왼쪽을 향해 인물의 목과 턱을 타격하고 있으며, 인물의 고개는 뒤로 젖혀져 있음.",
        "built_space": "배경에 참고 이미지와 일치하는 컨테이너와 골목길이 보이며, 좁은 골목길 입구의 구조가 적절히 배치됨.",
        "entities": "이현우(참고 이미지의 외모와 일치, 후드티 착용, 얼굴 상처 있음), 소총(나무 개머리판).",
        "hard_violations": [
         "이미지 위에 검은색 화살표들이 그려져 있음 (leaked markers/diagrams)"
        ],
        "physics": "인물은 두 발로 서 있고, 총기는 화면 오른쪽에 보이는 손에 의해 지지됨."
       },
       {
        "label": "B",
        "direction": "개머리판이 프레임 오른쪽에서 들어와 인물의 광대뼈에 정통으로 꽂혔으며, 타격 방향으로 고개가 심하게 꺾임.",
        "built_space": "인물의 어깨 너머로 참고 이미지의 좁은 골목길 입구와 컨테이너가 제한적으로 보임.",
        "entities": "이현우(참고 이미지의 외모와 일치, 후드티 착용, 얼굴 상처 및 멍 있음), 소총(검은색 개머리판 및 손잡이).",
        "hard_violations": [],
        "physics": "인물의 몸은 바닥에 지탱되어 있고, 소총은 화면 밖으로 이어져 보이지 않는 프레임 밖의 손에 의해 지지되며 타격 충격으로 입에서 침이 흩뿌려짐."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "오른쪽에서 들어오는 개머리판과 얼굴 중심 클로즈업, 화면 밖 공격자는 잘 지켰지만 접촉점이 광대뼈보다 낮고 얼굴이 확 꺾이는 반응은 약하다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "화살표 도식과 화면 안 공격자의 손·팔이 명시적 금지 조건을 위반하며, 개머리판도 광대뼈 대신 귀 아래 목 옆에 닿는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "개머리판은 화면 오른쪽에서 왼쪽으로 들어와 이현우의 보이는 볼 뒤쪽과 턱선 부근에 닿는다. 접촉면 상단도 광대뼈 중심보다 낮아 정통 광대뼈 타격은 불명확하다. 얼굴은 화면 왼쪽으로 기울고 눈은 거의 감겨 있으며, 특정 대상을 바라보지 않는다. 총구는 화면 밖이다.",
        "built_space": "오른쪽에는 녹슨 골판 금속 벽, 왼쪽에는 천막 한 구역과 지지 기둥들, 쌓인 상자와 용기들, 바닥에 비스듬히 넘어진 나무 상자 하나가 보인다. 인물 뒤 왼쪽에 제한된 통로가 남아 있어 골목 입구의 맥락을 유지한다. 참고 장소의 재료와 시장 시설에 부합하며, 중복된 고정 시설이나 불가능한 반사는 보이지 않는다.",
        "entities": "젊은 동아시아계 남성 한 명으로, 검은 헝클어진 머리와 해진 짙은 회색 후드가 이현우 참고의 외형에 대체로 맞는다. 눈썹과 볼, 입 주변에 기존 상처로 읽히는 흔적이 있다. 검은 개머리판과 소총 몸체 일부가 보이고 공격자는 보이지 않는다. 다리의 물린 상처와 컨테이너 내부 상태는 크롭상 확인할 수 없다. 읽을 수 있는 글자나 도식은 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 이어진 상체가 기울어 있으며 공중에 떠 있는 전신은 아니다. 하체의 지지는 크롭 밖이다. 소총은 오른쪽 프레임 밖까지 이어져 있어 화면 밖 공격자가 쥔 연출로 성립하며, 손이 안 보인다는 이유만으로 부유 물체라 할 수 없다. 개머리판과 볼의 접촉은 있지만 얼굴 회전과 피부 압박은 비교적 약해 맹렬한 타격의 찰나보다 밀어붙이는 순간에 가깝다."
       },
       {
        "label": "B",
        "direction": "소총 개머리판은 오른쪽에서 왼쪽을 향하지만 넓은 끝면이 이현우의 귀 아래와 목 옆에 닿고 광대뼈에는 닿지 않는다. 얼굴은 화면 왼쪽으로 돌아가고 눈은 감겨 있다. 화면에 추가된 화살표 세 개도 왼쪽 방향을 가리킨다. 총구는 화면 밖이다.",
        "built_space": "왼쪽 천막과 상자 더미, 중앙으로 깊게 이어지는 시장 통로, 오른쪽 금속 벽과 드럼통 하나가 보인다. 오른쪽 위에는 가로등 하나와 확성기 하나가 있어 참고 장소의 시설 구성이 잘 남아 있다. 다만 넓은 시장 통로를 정면으로 많이 드러내어, 어깨 뒤에 좁은 샛골목 입구만 제한적으로 남기는 구도와 다르다.",
        "entities": "검은 헝클어진 머리와 얼굴 상처가 있는 젊은 동아시아계 남성은 이현우의 외형에 대체로 맞지만, 상체를 더 넓게 보여 주고 후드의 소매 구조도 참고와 다르게 보인다. 나무 개머리판 소총 외에 별도 인물의 손과 녹색 소매 팔이 오른쪽에 등장한다. 다리 상처와 컨테이너 수색 상태는 확인할 수 없다. 읽을 수 있는 글자는 없으나 명백한 화살표 도식이 있다.",
        "hard_violations": [
         "실사 화면 위에 방향을 표시하는 화살표 세 개가 삽입되어 도식·오버레이 금지 조건을 위반한다.",
         "화면 밖에 있어야 할 공격자의 손과 팔을 오른쪽 프레임 안에 노출한다."
        ],
        "physics": "오른쪽 손이 소총을 실제로 잡고 있어 무기의 지지는 보인다. 이현우의 목과 상체는 연결되어 있고, 어깨를 움츠리며 몸을 돌리는 자세 자체는 가능하다. 다만 목 옆의 접촉으로 머리가 돌아가는 장면이어서, 요청한 광대뼈 직접 타격의 물리적 관계는 성립하지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "오른쪽에서 들어오는 개머리판과 얼굴 중심 클로즈업, 화면 밖 공격자는 잘 지켰지만 접촉점이 광대뼈보다 낮고 얼굴이 확 꺾이는 반응은 약하다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "화살표 도식과 화면 안 공격자의 손·팔이 명시적 금지 조건을 위반하며, 개머리판도 광대뼈 대신 귀 아래 목 옆에 닿는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "개머리판은 화면 오른쪽에서 왼쪽으로 들어와 이현우의 보이는 볼 뒤쪽과 턱선 부근에 닿는다. 접촉면 상단도 광대뼈 중심보다 낮아 정통 광대뼈 타격은 불명확하다. 얼굴은 화면 왼쪽으로 기울고 눈은 거의 감겨 있으며, 특정 대상을 바라보지 않는다. 총구는 화면 밖이다.",
        "built_space": "오른쪽에는 녹슨 골판 금속 벽, 왼쪽에는 천막 한 구역과 지지 기둥들, 쌓인 상자와 용기들, 바닥에 비스듬히 넘어진 나무 상자 하나가 보인다. 인물 뒤 왼쪽에 제한된 통로가 남아 있어 골목 입구의 맥락을 유지한다. 참고 장소의 재료와 시장 시설에 부합하며, 중복된 고정 시설이나 불가능한 반사는 보이지 않는다.",
        "entities": "젊은 동아시아계 남성 한 명으로, 검은 헝클어진 머리와 해진 짙은 회색 후드가 이현우 참고의 외형에 대체로 맞는다. 눈썹과 볼, 입 주변에 기존 상처로 읽히는 흔적이 있다. 검은 개머리판과 소총 몸체 일부가 보이고 공격자는 보이지 않는다. 다리의 물린 상처와 컨테이너 내부 상태는 크롭상 확인할 수 없다. 읽을 수 있는 글자나 도식은 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 이어진 상체가 기울어 있으며 공중에 떠 있는 전신은 아니다. 하체의 지지는 크롭 밖이다. 소총은 오른쪽 프레임 밖까지 이어져 있어 화면 밖 공격자가 쥔 연출로 성립하며, 손이 안 보인다는 이유만으로 부유 물체라 할 수 없다. 개머리판과 볼의 접촉은 있지만 얼굴 회전과 피부 압박은 비교적 약해 맹렬한 타격의 찰나보다 밀어붙이는 순간에 가깝다."
       },
       {
        "label": "A",
        "direction": "소총 개머리판은 오른쪽에서 왼쪽을 향하지만 넓은 끝면이 이현우의 귀 아래와 목 옆에 닿고 광대뼈에는 닿지 않는다. 얼굴은 화면 왼쪽으로 돌아가고 눈은 감겨 있다. 화면에 추가된 화살표 세 개도 왼쪽 방향을 가리킨다. 총구는 화면 밖이다.",
        "built_space": "왼쪽 천막과 상자 더미, 중앙으로 깊게 이어지는 시장 통로, 오른쪽 금속 벽과 드럼통 하나가 보인다. 오른쪽 위에는 가로등 하나와 확성기 하나가 있어 참고 장소의 시설 구성이 잘 남아 있다. 다만 넓은 시장 통로를 정면으로 많이 드러내어, 어깨 뒤에 좁은 샛골목 입구만 제한적으로 남기는 구도와 다르다.",
        "entities": "검은 헝클어진 머리와 얼굴 상처가 있는 젊은 동아시아계 남성은 이현우의 외형에 대체로 맞지만, 상체를 더 넓게 보여 주고 후드의 소매 구조도 참고와 다르게 보인다. 나무 개머리판 소총 외에 별도 인물의 손과 녹색 소매 팔이 오른쪽에 등장한다. 다리 상처와 컨테이너 수색 상태는 확인할 수 없다. 읽을 수 있는 글자는 없으나 명백한 화살표 도식이 있다.",
        "hard_violations": [
         "실사 화면 위에 방향을 표시하는 화살표 세 개가 삽입되어 도식·오버레이 금지 조건을 위반한다.",
         "화면 밖에 있어야 할 공격자의 손과 팔을 오른쪽 프레임 안에 노출한다."
        ],
        "physics": "오른쪽 손이 소총을 실제로 잡고 있어 무기의 지지는 보인다. 이현우의 목과 상체는 연결되어 있고, 어깨를 움츠리며 몸을 돌리는 자세 자체는 가능하다. 다만 목 옆의 접촉으로 머리가 돌아가는 장면이어서, 요청한 광대뼈 직접 타격의 물리적 관계는 성립하지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.571,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.321,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 이미지 위에 검은색 화살표들이 그려져 있음 (leaked markers/diagrams)",
     "[gpt-high] 실사 화면 위에 방향을 표시하는 화살표 세 개가 삽입되어 도식·오버레이 금지 조건을 위반한다.",
     "[gpt-high] 화면 밖에 있어야 할 공격자의 손과 팔을 오른쪽 프레임 안에 노출한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 321
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 클로즈업 샷과 타격 부위(광대뼈)를 정확히 구현했으며, 무기를 휘두르는 인물을 프레임 밖으로 배치하는 지침을 잘 따랐습니다."
   },
   {
    "label": "A",
    "score": 321,
    "verdict_ko": "화면에 지시되지 않은 검은색 화살표(다이어그램)가 유출된 치명적 위반이 있으며, 무기를 쥔 손이 프레임에 포함되어 지침에 어긋납니다.  ★위반: [gemini-pro] 이미지 위에 검은색 화살표들이 그려져 있음 (leaked markers/diagrams) / [gpt-high] 실사 화면 위에 방향을 표시하는 화살표 세 개가 삽입되어 도식·오버레이 금지 조건을 위반한다. / [gpt-high] 화면 밖에 있어야 할 공격자의 손과 팔을 오른쪽 프레임 안에 노출한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_market_capture_alley_13d54d.png",
    "asset_id": "b72ff740-1baa-4a51-b31a-135364ea551f",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd97-0888-70d8-91c4-2b4c69e576e9",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S20sh12__bgfirst_bg.png",
   "bg_asset_id": "d2cafdfa-96ab-4e94-a85e-b019a99054d1",
   "bg_record_key": "S20sh12::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "market_capture_alley",
   "groupbg_asset_id": "b72ff740-1baa-4a51-b31a-135364ea551f"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S20sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:25:52.685874+00:00",
  "fingerprint": "97de213ab08d304299eba446cdc492d6e263d4a12c6a35cbc96b1a231ce6c840",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S20sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S20sh12_sel.png",
  "source_sha256": "16996a931ab65a1bdd8b9df8c10c7f29e7c6398f86ac5dde9059c55ce9f43fe0",
  "file": "S20sh12_cine.png",
  "staged_sha256": "9804eee9988bcf7213fe23b9515e5c0e8846e65943a29a5692942885c7e66125",
  "latency_ms": 18355
 },
 "S21sh1::signage": {
  "fp": "92a2d7bc26dd7e33",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::32987604ed1eed30": {
  "subjects": [],
  "subject_text": "이현우의 컨테이너 내부 앰버의 방\n주거용 컨테이너 안에서 방문으로 분리된 작은 방. 공용 생활공간에 인접한 단출한 실내 공간이다.",
  "identity": "canonical",
  "scope_id": "L23",
  "scope_role": "location_interior",
  "scope_sha": "48706fb757ded1e0"
 },
 "S21sh1::bgfirst_bg": {
  "input_fingerprint": "c14ef20dea02b35e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 낡은 침상 위에 누운 앰버가 옅은 미소를 띤 채 평온하게 두 눈을 감고 있는 얼굴.\n\nLOCATION (lock): On a worn bed in a child's container home used as a temporary resting place, in soft daytime ambient light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Worn bed (Occupied by the sleeping 앰버) — A cropped portion of its upper surface appears beneath her head and shoulder; used as A narrow contextual border anchors the face in the sleeping space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate daytime ambient light and gentle facial contrast preserve the quiet, directly observed intimacy.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 낡은 침상 위에 누운 앰버가 옅은 미소를 띤 채 평온하게 두 눈을 감고 있는 얼굴.\n\nLOCATION (lock): On a worn bed in a child's container home used as a temporary resting place, in soft daytime ambient light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Worn bed (Occupied by the sleeping 앰버) — A cropped portion of its upper surface appears beneath her head and shoulder; used as A narrow contextual border anchors the face in the sleeping space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate daytime ambient light and gentle facial contrast preserve the quiet, directly observed intimacy.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S21sh1__bgfirst_bg.png",
  "asset_id": "85709c0f-1b56-4d8a-a0fb-6c009eca849c",
  "input_asset_ids": [
   "1e4cb78a-a709-4c36-bd49-7e384185700d",
   "026786ff-2cf4-4ab0-b904-9fb5c074c9a2"
  ]
 },
 "S21sh1": {
  "input_fingerprint": "392cd434bf347f71",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낡은 침상 위에 누운 앰버가 옅은 미소를 띤 채 평온하게 두 눈을 감고 있는 얼굴.\n\nLOCATION (lock): On a worn bed in a child's container home used as a temporary resting place, in soft daytime ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Worn bed (Occupied by the sleeping 앰버) — A cropped portion of its upper surface appears beneath her head and shoulder; used as A narrow contextual border anchors the face in the sleeping space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate daytime ambient light and gentle facial contrast preserve the quiet, directly observed intimacy.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep, reclining on the bed with both eyes closed and her face relaxed. The positions of her arms and legs, the angle of her head, and whether her torso faces upward or sideways are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 앰버: She is asleep inside Raul's container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낡은 침상 위에 누운 앰버가 옅은 미소를 띤 채 평온하게 두 눈을 감고 있는 얼굴.\n\nLOCATION (lock): On a worn bed in a child's container home used as a temporary resting place, in soft daytime ambient light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Worn bed (Occupied by the sleeping 앰버) — A cropped portion of its upper surface appears beneath her head and shoulder; used as A narrow contextual border anchors the face in the sleeping space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate daytime ambient light and gentle facial contrast preserve the quiet, directly observed intimacy.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep, reclining on the bed with both eyes closed and her face relaxed. The positions of her arms and legs, the angle of her head, and whether her torso faces upward or sideways are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 앰버: She is asleep inside Raul's container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낡은 침상 위에 누운 앰버가 옅은 미소를 띤 채 평온하게 두 눈을 감고 있는 얼굴.\n\nLOCATION (lock): On a worn bed in a child's container home used as a temporary resting place, in soft daytime ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Worn bed (Occupied by the sleeping 앰버) — A cropped portion of its upper surface appears beneath her head and shoulder; used as A narrow contextual border anchors the face in the sleeping space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate daytime ambient light and gentle facial contrast preserve the quiet, directly observed intimacy.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep, reclining on the bed with both eyes closed and her face relaxed. The positions of her arms and legs, the angle of her head, and whether her torso faces upward or sideways are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 앰버: She is asleep inside Raul's container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S21sh1__bgfirst_bg.png",
     "asset_id": "85709c0f-1b56-4d8a-a0fb-6c009eca849c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S21sh1.png",
     "asset_id": "1e4cb78a-a709-4c36-bd49-7e384185700d",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L23B02.png",
     "asset_id": "026786ff-2cf4-4ab0-b904-9fb5c074c9a2",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 눈을 감고 정면을 향해 누워 있음.",
    "built_space": "침대, 회색 격자무늬 베개, 나무 헤드보드, 창문, 선반, 책, 식물 등이 방 안에 배치되어 있음.",
    "entities": "레퍼런스와 일치하는 금발 소녀(앰버)가 옅은 미소를 띠고 수면 중임.",
    "hard_violations": [],
    "physics": "머리와 상체가 침대와 베개에 의해 자연스럽게 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "두 눈을 감고 누워 있음.",
    "built_space": "나무 헤드보드와 체크무늬 베개 등 침대 윗부분만 화면에 좁게 나타남.",
    "entities": "레퍼런스와 일치하는 금발 소녀(앰버)가 옅은 미소를 띠고 수면 중임.",
    "hard_violations": [],
    "physics": "머리가 베개 위에 안정적으로 놓여 중력에 맞게 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 샷 크기와 얼굴을 감싸는 좁은 배경 테두리 프레이밍을 정확히 구현하여 가장 우수한 결과를 보여줍니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "장소의 디테일은 좋으나, 클로즈업 지시를 어기고 방의 배경을 너무 넓게 노출하여 프레이밍 조건에서 크게 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈을 감고 정면을 향해 누워 있음.",
        "built_space": "침대, 회색 격자무늬 베개, 나무 헤드보드, 창문, 선반, 책, 식물 등이 방 안에 배치되어 있음.",
        "entities": "레퍼런스와 일치하는 금발 소녀(앰버)가 옅은 미소를 띠고 수면 중임.",
        "hard_violations": [],
        "physics": "머리와 상체가 침대와 베개에 의해 자연스럽게 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "두 눈을 감고 누워 있음.",
        "built_space": "나무 헤드보드와 체크무늬 베개 등 침대 윗부분만 화면에 좁게 나타남.",
        "entities": "레퍼런스와 일치하는 금발 소녀(앰버)가 옅은 미소를 띠고 수면 중임.",
        "hard_violations": [],
        "physics": "머리가 베개 위에 안정적으로 놓여 중력에 맞게 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 샷 크기와 얼굴을 감싸는 좁은 배경 테두리 프레이밍을 정확히 구현하여 가장 우수한 결과를 보여줍니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "장소의 디테일은 좋으나, 클로즈업 지시를 어기고 방의 배경을 너무 넓게 노출하여 프레이밍 조건에서 크게 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 눈을 감고 정면을 향해 누워 있음.",
        "built_space": "침대, 회색 격자무늬 베개, 나무 헤드보드, 창문, 선반, 책, 식물 등이 방 안에 배치되어 있음.",
        "entities": "레퍼런스와 일치하는 금발 소녀(앰버)가 옅은 미소를 띠고 수면 중임.",
        "hard_violations": [],
        "physics": "머리와 상체가 침대와 베개에 의해 자연스럽게 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "두 눈을 감고 누워 있음.",
        "built_space": "나무 헤드보드와 체크무늬 베개 등 침대 윗부분만 화면에 좁게 나타남.",
        "entities": "레퍼런스와 일치하는 금발 소녀(앰버)가 옅은 미소를 띠고 수면 중임.",
        "hard_violations": [],
        "physics": "머리가 베개 위에 안정적으로 놓여 중력에 맞게 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "얼굴을 크게 담고 침구를 주변 맥락으로 제한해, 두 눈을 감고 옅게 미소 짓는 앰버의 얼굴 클로즈업이라는 핵심 지시를 가장 충실하게 구현했다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "수면 자세와 장소의 특징은 잘 맞지만, 상체와 방의 설비까지 넓게 보여 얼굴 중심의 밀착 구도와 좁은 침구 테두리 지시에서 벗어났다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 화면을 향해 비스듬히 돌아가 있고 두 눈은 완전히 감겨 있어 응시 대상이 없다. 입가에는 옅은 미소가 보이며, 이동하거나 무엇을 겨누는 동작은 없다.",
        "built_space": "얼굴이 화면 대부분을 차지한다. 머리 아래 베개와 체크무늬 침구, 왼쪽 위 목재 머리판 일부가 보이며 침대는 하나로 읽힌다. 창과 책장은 이 클로즈업 바깥에 있어 누락으로 볼 수 없다. 목재와 침구의 색감은 장소 참조와 이어지지만, 얼굴 바로 아래의 굵은 체크무늬는 참조의 회색 가는 격자 베개와 다르다. 반사나 중복 설비는 없다.",
        "entities": "금발의 약 10세 여자아이 한 명만 보인다. 참조와 머리색, 어린 얼굴의 비례, 체격이 대체로 부합한다. 한국계·백인 혼혈이라는 설정은 외관만으로 확정할 수 없고, 큰 눈 여부도 눈을 감아 확인할 수 없다. 회색 옷의 목과 어깨 부분이 보이며 참조의 회색 안쪽 옷과 색감이 비슷하다. 얼굴을 드러내라는 장면에 맞게 마스크는 없다. 읽을 수 있는 글자나 로고는 없다.",
        "hard_violations": [],
        "physics": "머리와 뺨이 베개에 내려앉아 지지되고, 보이는 어깨와 몸통은 침상에 누워 있다. 머리카락은 베개와 옷 위에 자연스럽게 펼쳐진다. 공중에 들린 신체나 지지 없는 물체는 없으며, 편안하게 잠든 자세로 성립한다."
       },
       {
        "label": "B",
        "direction": "얼굴은 위쪽과 카메라 쪽으로 약간 돌아가 있고 두 눈은 감겨 있다. 응시 대상은 없으며 입가는 부드럽게 올라가 있다. 몸은 침대 길이 방향으로 누워 있고 이동 동작이나 방향성 소품은 없다.",
        "built_space": "침대 하나의 목재 머리판, 회색 격자 베개 하나, 체크무늬 이불이 보인다. 뒤에는 창 하나, 커튼 한 폭, 오른쪽 책장 하나와 덩굴식물이 있으며 머리판 선반에는 시계 하나와 책·작은 용기가 놓여 있다. 주요 배치는 장소 참조와 대체로 맞고 중복 설비나 불가능한 반사는 없다. 다만 방의 배경과 덮인 상체가 상당 부분을 차지해 침구가 좁은 맥락 테두리에 머물지 않는다.",
        "entities": "금발의 어린 여자아이 한 명이며 약 10세라는 설정에 부합한다. 참조와 머리색과 어린 체격은 맞지만 얼굴의 정확한 동일성은 다소 불확실하다. 혼혈 배경은 외관만으로 확정할 수 없고 감긴 눈의 크기도 판별할 수 없다. 보이는 옷은 회색 둥근 목 티셔츠처럼 읽혀 참조의 회색 원피스와 완전히 같지는 않다. 침구와 목재 침대는 참조의 특징에 가깝다. 배경 시계와 책에는 읽을 수 있는 글자가 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 목은 베개에 받쳐지고 어깨와 몸통은 침상에 놓여 있다. 이불은 몸 위에 얹혀 주름지고 머리카락은 베개와 어깨 위에 내려앉는다. 들린 손발이나 지지 없는 물체는 보이지 않아 수면 중 중력에 따른 자세가 자연스럽다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "얼굴을 크게 담고 침구를 주변 맥락으로 제한해, 두 눈을 감고 옅게 미소 짓는 앰버의 얼굴 클로즈업이라는 핵심 지시를 가장 충실하게 구현했다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "수면 자세와 장소의 특징은 잘 맞지만, 상체와 방의 설비까지 넓게 보여 얼굴 중심의 밀착 구도와 좁은 침구 테두리 지시에서 벗어났다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 화면을 향해 비스듬히 돌아가 있고 두 눈은 완전히 감겨 있어 응시 대상이 없다. 입가에는 옅은 미소가 보이며, 이동하거나 무엇을 겨누는 동작은 없다.",
        "built_space": "얼굴이 화면 대부분을 차지한다. 머리 아래 베개와 체크무늬 침구, 왼쪽 위 목재 머리판 일부가 보이며 침대는 하나로 읽힌다. 창과 책장은 이 클로즈업 바깥에 있어 누락으로 볼 수 없다. 목재와 침구의 색감은 장소 참조와 이어지지만, 얼굴 바로 아래의 굵은 체크무늬는 참조의 회색 가는 격자 베개와 다르다. 반사나 중복 설비는 없다.",
        "entities": "금발의 약 10세 여자아이 한 명만 보인다. 참조와 머리색, 어린 얼굴의 비례, 체격이 대체로 부합한다. 한국계·백인 혼혈이라는 설정은 외관만으로 확정할 수 없고, 큰 눈 여부도 눈을 감아 확인할 수 없다. 회색 옷의 목과 어깨 부분이 보이며 참조의 회색 안쪽 옷과 색감이 비슷하다. 얼굴을 드러내라는 장면에 맞게 마스크는 없다. 읽을 수 있는 글자나 로고는 없다.",
        "hard_violations": [],
        "physics": "머리와 뺨이 베개에 내려앉아 지지되고, 보이는 어깨와 몸통은 침상에 누워 있다. 머리카락은 베개와 옷 위에 자연스럽게 펼쳐진다. 공중에 들린 신체나 지지 없는 물체는 없으며, 편안하게 잠든 자세로 성립한다."
       },
       {
        "label": "A",
        "direction": "얼굴은 위쪽과 카메라 쪽으로 약간 돌아가 있고 두 눈은 감겨 있다. 응시 대상은 없으며 입가는 부드럽게 올라가 있다. 몸은 침대 길이 방향으로 누워 있고 이동 동작이나 방향성 소품은 없다.",
        "built_space": "침대 하나의 목재 머리판, 회색 격자 베개 하나, 체크무늬 이불이 보인다. 뒤에는 창 하나, 커튼 한 폭, 오른쪽 책장 하나와 덩굴식물이 있으며 머리판 선반에는 시계 하나와 책·작은 용기가 놓여 있다. 주요 배치는 장소 참조와 대체로 맞고 중복 설비나 불가능한 반사는 없다. 다만 방의 배경과 덮인 상체가 상당 부분을 차지해 침구가 좁은 맥락 테두리에 머물지 않는다.",
        "entities": "금발의 어린 여자아이 한 명이며 약 10세라는 설정에 부합한다. 참조와 머리색과 어린 체격은 맞지만 얼굴의 정확한 동일성은 다소 불확실하다. 혼혈 배경은 외관만으로 확정할 수 없고 감긴 눈의 크기도 판별할 수 없다. 보이는 옷은 회색 둥근 목 티셔츠처럼 읽혀 참조의 회색 원피스와 완전히 같지는 않다. 침구와 목재 침대는 참조의 특징에 가깝다. 배경 시계와 책에는 읽을 수 있는 글자가 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 목은 베개에 받쳐지고 어깨와 몸통은 침상에 놓여 있다. 이불은 몸 위에 얹혀 주름지고 머리카락은 베개와 어깨 위에 내려앉는다. 들린 손발이나 지지 없는 물체는 보이지 않아 수면 중 중력에 따른 자세가 자연스럽다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.349,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.349,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1349
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 클로즈업 샷 크기와 얼굴을 감싸는 좁은 배경 테두리 프레이밍을 정확히 구현하여 가장 우수한 결과를 보여줍니다."
   },
   {
    "label": "A",
    "score": 1349,
    "verdict_ko": "장소의 디테일은 좋으나, 클로즈업 지시를 어기고 방의 배경을 너무 넓게 노출하여 프레이밍 조건에서 크게 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L23B02.png",
    "asset_id": "026786ff-2cf4-4ab0-b904-9fb5c074c9a2",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd98-3608-781e-bbc4-595cdb77d6f6",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S21sh1__bgfirst_bg.png",
   "bg_asset_id": "85709c0f-1b56-4d8a-a0fb-6c009eca849c",
   "bg_record_key": "S21sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S21sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:26:11.469485+00:00",
  "fingerprint": "4f51600874df3cd40d5237807ec1f5043cffcc6dcd072f0b3e47f21709025f6b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S21sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S21sh1_sel.png",
  "source_sha256": "17e44ed6687d6064c5c652a04accd5ea9909f6a3f319b4c8064a6bc193b1ef37",
  "file": "S21sh1_cine.png",
  "staged_sha256": "48fa8f5c726c8859d1395630894ddc2cd139ca9d8134973d0d2af003257a2075",
  "latency_ms": 18534
 },
 "S21sh4::signage": {
  "fp": "c89786a254e44ce9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S21sh4": {
  "input_fingerprint": "7b7cd12953ec6d17",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 라울이 입을 크게 벌린 채 손가락으로 열린 문밖을 가리키고 있는 다급한 구도.\n\nLOCATION (lock): Inside the sleeping area of the child's container home, beside the open entrance admitting daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open doorway beyond 라울's pointing finger in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open entrance door (Open after 라울's abrupt entrance) — The door is seen obliquely beside the opening toward which 라울 points; used as Places the source of his warning behind his pointing hand without showing the reported destruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain consistent daytime ambient illumination, using restrained contrast to make the interruption urgent without changing the room's lighting character.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The container door remains open after the abrupt entrance. 라울: He is inside the container after entering through the door.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 라울이 입을 크게 벌린 채 손가락으로 열린 문밖을 가리키고 있는 다급한 구도.\n\nLOCATION (lock): Inside the sleeping area of the child's container home, beside the open entrance admitting daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open doorway beyond 라울's pointing finger in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open entrance door (Open after 라울's abrupt entrance) — The door is seen obliquely beside the opening toward which 라울 points; used as Places the source of his warning behind his pointing hand without showing the reported destruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain consistent daytime ambient illumination, using restrained contrast to make the interruption urgent without changing the room's lighting character.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The container door remains open after the abrupt entrance. 라울: He is inside the container after entering through the door.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 라울이 입을 크게 벌린 채 손가락으로 열린 문밖을 가리키고 있는 다급한 구도.\n\nLOCATION (lock): Inside the sleeping area of the child's container home, beside the open entrance admitting daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open doorway beyond 라울's pointing finger in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open entrance door (Open after 라울's abrupt entrance) — The door is seen obliquely beside the opening toward which 라울 points; used as Places the source of his warning behind his pointing hand without showing the reported destruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain consistent daytime ambient illumination, using restrained contrast to make the interruption urgent without changing the room's lighting character.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The container door remains open after the abrupt entrance. 라울: He is inside the container after entering through the door.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "B",
    "direction": "라울의 눈은 왼쪽 전경의 금발 인물을 향한다. 뻗은 검지는 화면 오른쪽을 거의 수평으로 가리키며, 연장 방향은 밝은 출입구를 통과하기보다 오른쪽 문설주와 벽 쪽으로 이어진다. 열린 문은 손 뒤에 있지만 문밖을 지목하는 방향은 부정확하다.",
    "built_space": "금속 골판 벽 안에 열린 출입구 하나와 비스듬한 문짝 하나가 보인다. 출입구 양옆에는 목재 머리판을 가진 침대가 각각 하나씩, 총 두 개 있으며 체크무늬 베개도 두 개 보인다. 참고 사진에서 확인되는 목재와 체크 침구는 이어지지만, 두 번째 침대까지 같은 장소의 고정 배치인지는 확인할 수 없다. 라울은 침대 사이에 있고 금발 인물은 왼쪽 전경을 차지한다. 벽과 문에는 여러 안내 종이가 추가되어 있다.",
    "entities": "라울은 참고와 유사한 짙은 피부, 어린 얼굴, 뒤로 묶은 곱슬머리와 낡고 구멍 난 올리브색 티셔츠를 가진 남자아이로 표현됐다. 입을 크게 벌린 표정은 맞는다. 그러나 이전 장면의 인물을 연상시키는 금발 인물이 왼쪽에 추가됐고, 문밖에도 최소 세 사람의 형체가 보인다. 라울만 허용한 인물 조건과 명백히 충돌한다. 일부 안내문에는 식별 가능한 한글 글자도 드러난다.",
    "hard_violations": [
     "허용되지 않은 금발 인물을 왼쪽 전경에 추가했다.",
     "출입구 밖에 허용되지 않은 여러 사람을 추가했다.",
     "벽의 안내문에 읽을 수 있는 한글 글자가 노출되어 무문자 조건을 위반했다."
    ],
    "physics": "라울의 발은 프레임 밖이지만 몸통은 자연스러운 직립 자세이고, 가리키는 팔은 어깨와 팔꿈치에 정상적으로 연결되어 있다. 공중에 떠 있다고 볼 징후는 없다. 침구는 침대에 놓여 있고 문짝은 문틀의 경첩으로 지지되는 배치다. 보이는 인체 동작 자체에 불가능한 지지는 없다."
   },
   {
    "label": "A",
    "direction": "라울의 검지는 오른쪽 위로 뻗어 있으며 손끝 너머에 밝게 열린 출입구가 놓인다. 손가락의 방향이 문짝 옆 틈을 통해 문밖으로 이어진다. 눈도 오른쪽 바깥을 향하고 있어 경고하는 대상 방향과 일치한다.",
    "built_space": "컨테이너의 골판 금속 벽, 목재 머리판 하나, 체크무늬 침구가 있는 침대 하나와 베개 하나가 보인다. 오른쪽 배경에는 출입구 하나와 비스듬히 열린 금속 문짝 하나가 있다. 라울은 출입구 안쪽 침대 가장자리에 앉아 있고, 문밖의 낮빛이 실내로 들어온다. 참고 사진의 목재·체크 침구와 부드러운 주간 조명을 유지하며, 반사나 중복된 출입문은 없다.",
    "entities": "보이는 사람은 라울 한 명뿐이다. 약 10세 남자아이의 어린 얼굴과 체격, 참고의 라틴계·흑인 혼혈 설정에 부합하는 외형, 뒤로 묶은 곱슬 꽁지머리를 갖췄다. 낡은 올리브색 구멍 난 티셔츠와 해진 데님 반바지가 참고와 맞는다. 입을 크게 벌리고 외치는 표정이 명확하다. 신발은 프레임 밖이며, 추가 인물이나 읽을 수 있는 글자는 보이지 않는다.",
    "hard_violations": [],
    "physics": "엉덩이와 허벅지가 침대 매트리스에 닿아 앉은 몸을 지지한다. 아래로 내린 팔도 침대 쪽으로 이어지지만 손의 접촉점은 하단에서 잘려 있다. 가리키는 팔은 어깨에서 자연스럽게 들어 올려져 있고 손가락 관절도 가능한 자세다. 문짝은 문틀에 연결되어 있으며 침구는 매트리스 위에 놓여 있다. 지지 없이 떠 있는 몸이나 물체는 없다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "라울만 등장하며, 입을 크게 벌리고 우상단의 열린 문밖을 정확히 가리키는 미디엄 숏과 침실의 재질·낮빛을 충실히 구현했다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "금발 인물과 문밖의 여러 사람을 무단으로 추가했고, 손가락도 열린 출입구보다 오른쪽 문설주·벽 쪽을 향해 핵심 지시를 어겼다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "라울의 눈은 왼쪽 전경의 금발 인물을 향한다. 뻗은 검지는 화면 오른쪽을 거의 수평으로 가리키며, 연장 방향은 밝은 출입구를 통과하기보다 오른쪽 문설주와 벽 쪽으로 이어진다. 열린 문은 손 뒤에 있지만 문밖을 지목하는 방향은 부정확하다.",
        "built_space": "금속 골판 벽 안에 열린 출입구 하나와 비스듬한 문짝 하나가 보인다. 출입구 양옆에는 목재 머리판을 가진 침대가 각각 하나씩, 총 두 개 있으며 체크무늬 베개도 두 개 보인다. 참고 사진에서 확인되는 목재와 체크 침구는 이어지지만, 두 번째 침대까지 같은 장소의 고정 배치인지는 확인할 수 없다. 라울은 침대 사이에 있고 금발 인물은 왼쪽 전경을 차지한다. 벽과 문에는 여러 안내 종이가 추가되어 있다.",
        "entities": "라울은 참고와 유사한 짙은 피부, 어린 얼굴, 뒤로 묶은 곱슬머리와 낡고 구멍 난 올리브색 티셔츠를 가진 남자아이로 표현됐다. 입을 크게 벌린 표정은 맞는다. 그러나 이전 장면의 인물을 연상시키는 금발 인물이 왼쪽에 추가됐고, 문밖에도 최소 세 사람의 형체가 보인다. 라울만 허용한 인물 조건과 명백히 충돌한다. 일부 안내문에는 식별 가능한 한글 글자도 드러난다.",
        "hard_violations": [
         "허용되지 않은 금발 인물을 왼쪽 전경에 추가했다.",
         "출입구 밖에 허용되지 않은 여러 사람을 추가했다.",
         "벽의 안내문에 읽을 수 있는 한글 글자가 노출되어 무문자 조건을 위반했다."
        ],
        "physics": "라울의 발은 프레임 밖이지만 몸통은 자연스러운 직립 자세이고, 가리키는 팔은 어깨와 팔꿈치에 정상적으로 연결되어 있다. 공중에 떠 있다고 볼 징후는 없다. 침구는 침대에 놓여 있고 문짝은 문틀의 경첩으로 지지되는 배치다. 보이는 인체 동작 자체에 불가능한 지지는 없다."
       },
       {
        "label": "B",
        "direction": "라울의 검지는 오른쪽 위로 뻗어 있으며 손끝 너머에 밝게 열린 출입구가 놓인다. 손가락의 방향이 문짝 옆 틈을 통해 문밖으로 이어진다. 눈도 오른쪽 바깥을 향하고 있어 경고하는 대상 방향과 일치한다.",
        "built_space": "컨테이너의 골판 금속 벽, 목재 머리판 하나, 체크무늬 침구가 있는 침대 하나와 베개 하나가 보인다. 오른쪽 배경에는 출입구 하나와 비스듬히 열린 금속 문짝 하나가 있다. 라울은 출입구 안쪽 침대 가장자리에 앉아 있고, 문밖의 낮빛이 실내로 들어온다. 참고 사진의 목재·체크 침구와 부드러운 주간 조명을 유지하며, 반사나 중복된 출입문은 없다.",
        "entities": "보이는 사람은 라울 한 명뿐이다. 약 10세 남자아이의 어린 얼굴과 체격, 참고의 라틴계·흑인 혼혈 설정에 부합하는 외형, 뒤로 묶은 곱슬 꽁지머리를 갖췄다. 낡은 올리브색 구멍 난 티셔츠와 해진 데님 반바지가 참고와 맞는다. 입을 크게 벌리고 외치는 표정이 명확하다. 신발은 프레임 밖이며, 추가 인물이나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "엉덩이와 허벅지가 침대 매트리스에 닿아 앉은 몸을 지지한다. 아래로 내린 팔도 침대 쪽으로 이어지지만 손의 접촉점은 하단에서 잘려 있다. 가리키는 팔은 어깨에서 자연스럽게 들어 올려져 있고 손가락 관절도 가능한 자세다. 문짝은 문틀에 연결되어 있으며 침구는 매트리스 위에 놓여 있다. 지지 없이 떠 있는 몸이나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "라울만 등장하며, 입을 크게 벌리고 우상단의 열린 문밖을 정확히 가리키는 미디엄 숏과 침실의 재질·낮빛을 충실히 구현했다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "금발 인물과 문밖의 여러 사람을 무단으로 추가했고, 손가락도 열린 출입구보다 오른쪽 문설주·벽 쪽을 향해 핵심 지시를 어겼다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "라울의 눈은 왼쪽 전경의 금발 인물을 향한다. 뻗은 검지는 화면 오른쪽을 거의 수평으로 가리키며, 연장 방향은 밝은 출입구를 통과하기보다 오른쪽 문설주와 벽 쪽으로 이어진다. 열린 문은 손 뒤에 있지만 문밖을 지목하는 방향은 부정확하다.",
        "built_space": "금속 골판 벽 안에 열린 출입구 하나와 비스듬한 문짝 하나가 보인다. 출입구 양옆에는 목재 머리판을 가진 침대가 각각 하나씩, 총 두 개 있으며 체크무늬 베개도 두 개 보인다. 참고 사진에서 확인되는 목재와 체크 침구는 이어지지만, 두 번째 침대까지 같은 장소의 고정 배치인지는 확인할 수 없다. 라울은 침대 사이에 있고 금발 인물은 왼쪽 전경을 차지한다. 벽과 문에는 여러 안내 종이가 추가되어 있다.",
        "entities": "라울은 참고와 유사한 짙은 피부, 어린 얼굴, 뒤로 묶은 곱슬머리와 낡고 구멍 난 올리브색 티셔츠를 가진 남자아이로 표현됐다. 입을 크게 벌린 표정은 맞는다. 그러나 이전 장면의 인물을 연상시키는 금발 인물이 왼쪽에 추가됐고, 문밖에도 최소 세 사람의 형체가 보인다. 라울만 허용한 인물 조건과 명백히 충돌한다. 일부 안내문에는 식별 가능한 한글 글자도 드러난다.",
        "hard_violations": [
         "허용되지 않은 금발 인물을 왼쪽 전경에 추가했다.",
         "출입구 밖에 허용되지 않은 여러 사람을 추가했다.",
         "벽의 안내문에 읽을 수 있는 한글 글자가 노출되어 무문자 조건을 위반했다."
        ],
        "physics": "라울의 발은 프레임 밖이지만 몸통은 자연스러운 직립 자세이고, 가리키는 팔은 어깨와 팔꿈치에 정상적으로 연결되어 있다. 공중에 떠 있다고 볼 징후는 없다. 침구는 침대에 놓여 있고 문짝은 문틀의 경첩으로 지지되는 배치다. 보이는 인체 동작 자체에 불가능한 지지는 없다."
       },
       {
        "label": "A",
        "direction": "라울의 검지는 오른쪽 위로 뻗어 있으며 손끝 너머에 밝게 열린 출입구가 놓인다. 손가락의 방향이 문짝 옆 틈을 통해 문밖으로 이어진다. 눈도 오른쪽 바깥을 향하고 있어 경고하는 대상 방향과 일치한다.",
        "built_space": "컨테이너의 골판 금속 벽, 목재 머리판 하나, 체크무늬 침구가 있는 침대 하나와 베개 하나가 보인다. 오른쪽 배경에는 출입구 하나와 비스듬히 열린 금속 문짝 하나가 있다. 라울은 출입구 안쪽 침대 가장자리에 앉아 있고, 문밖의 낮빛이 실내로 들어온다. 참고 사진의 목재·체크 침구와 부드러운 주간 조명을 유지하며, 반사나 중복된 출입문은 없다.",
        "entities": "보이는 사람은 라울 한 명뿐이다. 약 10세 남자아이의 어린 얼굴과 체격, 참고의 라틴계·흑인 혼혈 설정에 부합하는 외형, 뒤로 묶은 곱슬 꽁지머리를 갖췄다. 낡은 올리브색 구멍 난 티셔츠와 해진 데님 반바지가 참고와 맞는다. 입을 크게 벌리고 외치는 표정이 명확하다. 신발은 프레임 밖이며, 추가 인물이나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "엉덩이와 허벅지가 침대 매트리스에 닿아 앉은 몸을 지지한다. 아래로 내린 팔도 침대 쪽으로 이어지지만 손의 접촉점은 하단에서 잘려 있다. 가리키는 팔은 어깨에서 자연스럽게 들어 올려져 있고 손가락 관절도 가능한 자세다. 문짝은 문틀에 연결되어 있으며 침구는 매트리스 위에 놓여 있다. 지지 없이 떠 있는 몸이나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "totals": {
   "A": 9,
   "B": 1
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 9,
    "verdict_ko": "라울만 등장하며, 입을 크게 벌리고 우상단의 열린 문밖을 정확히 가리키는 미디엄 숏과 침실의 재질·낮빛을 충실히 구현했다."
   },
   {
    "label": "B",
    "score": 1,
    "verdict_ko": "금발 인물과 문밖의 여러 사람을 무단으로 추가했고, 손가락도 열린 출입구보다 오른쪽 문설주·벽 쪽을 향해 핵심 지시를 어겼다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S21sh1_sel.png",
    "asset_id": "ca7795b4-5731-4f38-b98f-b8ce7594d369",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:943979>",
    "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd99-653f-710e-83e8-5f724c373fab",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S21sh1"
  }
 },
 "S21sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:26:30.337605+00:00",
  "fingerprint": "f66971bf5b2e0a876561a2a10d3bd6345ad364fe28990c549f294ae734036edf",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S21sh4_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S21sh4_sel.png",
  "source_sha256": "2079b11192aa7900824850e8dcd0437504bf68bdd9f4d61704285a23e15c1e13",
  "file": "S21sh4_cine.png",
  "staged_sha256": "da9ffc5b01b7f0060451af27545601647df26b635b13123fec035ee15df65278",
  "latency_ms": 10614
 },
 "S22sh1::signage": {
  "fp": "cd7425867639743f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S22sh1": {
  "input_fingerprint": "f6730025d9fb6078",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복잡한 시장거리에 우뚝 선 거대한 덩치의 찰리가 진열된 온갖 잡동사니들을 향해 두 눈의 푸른빛을 빛내며 바라보는 외관.\n\nLOCATION (lock): In an open-air market lane inside the refugee settlement, beside stalls displaying assorted goods. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Market goods displays (Assorted goods laid out along the market street) — The displays recede diagonally to the right, with the goods receiving 찰리's attention continuing beyond the frame; used as Connects his partial profile to the objects of his curiosity without enlarging any single item; Market aisle and crowd (Occupied by numerous people moving through the market) — The aisle extends diagonally behind 찰리; used as Human-scale figures establish his large size; varied strides, head turns, and spacing prevent a duplicated crowd pattern.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps the market legible while the explicitly blue light of 찰리's eyes provides a selective color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market includes a banana stall. Charlie retains his weathered gorilla-shaped body, blue-lit eyes, worn coat and hat, and peeling UBIK chest logo; the earlier accumulated dirt has been washed away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복잡한 시장거리에 우뚝 선 거대한 덩치의 찰리가 진열된 온갖 잡동사니들을 향해 두 눈의 푸른빛을 빛내며 바라보는 외관.\n\nLOCATION (lock): In an open-air market lane inside the refugee settlement, beside stalls displaying assorted goods. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Market goods displays (Assorted goods laid out along the market street) — The displays recede diagonally to the right, with the goods receiving 찰리's attention continuing beyond the frame; used as Connects his partial profile to the objects of his curiosity without enlarging any single item; Market aisle and crowd (Occupied by numerous people moving through the market) — The aisle extends diagonally behind 찰리; used as Human-scale figures establish his large size; varied strides, head turns, and spacing prevent a duplicated crowd pattern.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps the market legible while the explicitly blue light of 찰리's eyes provides a selective color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market includes a banana stall. Charlie retains his weathered gorilla-shaped body, blue-lit eyes, worn coat and hat, and peeling UBIK chest logo; the earlier accumulated dirt has been washed away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복잡한 시장거리에 우뚝 선 거대한 덩치의 찰리가 진열된 온갖 잡동사니들을 향해 두 눈의 푸른빛을 빛내며 바라보는 외관.\n\nLOCATION (lock): In an open-air market lane inside the refugee settlement, beside stalls displaying assorted goods. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Market goods displays (Assorted goods laid out along the market street) — The displays recede diagonally to the right, with the goods receiving 찰리's attention continuing beyond the frame; used as Connects his partial profile to the objects of his curiosity without enlarging any single item; Market aisle and crowd (Occupied by numerous people moving through the market) — The aisle extends diagonally behind 찰리; used as Human-scale figures establish his large size; varied strides, head turns, and spacing prevent a duplicated crowd pattern.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps the market legible while the explicitly blue light of 찰리's eyes provides a selective color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market includes a banana stall. Charlie retains his weathered gorilla-shaped body, blue-lit eyes, worn coat and hat, and peeling UBIK chest logo; the earlier accumulated dirt has been washed away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리가 우측 가장자리에 진열된 잡동사니들을 향해 고개를 약간 숙여 바라보고 있습니다.",
    "built_space": "참조 사진의 골목 구조를 완벽히 유지하여 우측에 석재로 된 'Birreria' 상점과 상단 노란 간판, 중앙 아케이드 지붕이 정확히 배치되었습니다.",
    "entities": "찰리는 모자, 낡은 코트, 벗겨진 UBIK 로고, 푸른빛의 점 형태 눈을 갖췄으나 텍스트가 요구한 흰 얼굴판은 아닙니다. 흑인을 포함한 다민족 군중이 존재하나 바나나 가판대는 보이지 않습니다.",
    "hard_violations": [
     "[gpt-high] 가슴의 'UBIK', 시장·상점 간판 및 일부 상품에 판독 가능한 문자가 남아 있어 문자 및 로고 금지 조건을 위반한다."
    ],
    "physics": "찰리와 군중 모두 바닥에 안정적으로 발을 딛고 서 있습니다."
   },
   {
    "label": "B",
    "direction": "찰리가 우측 아래의 진열대와 물건들을 향해 시선을 고정하고 있습니다.",
    "built_space": "부감 구도로 촬영되었으나, 참조 사진에서 우측에 있어야 할 'Birreria' 건물과 노란 간판이 좌측으로 이동하는 등 골목의 뼈대가 완전히 뒤바뀌었습니다.",
    "entities": "찰리는 낡은 외투와 UBIK 로고, 푸른 눈을 가졌으나 흰 얼굴판이 아닙니다. 바나나 가판대는 등장하지만, 군중은 단일 인종에 가깝습니다.",
    "hard_violations": [
     "[gpt-high] 가슴의 'UBIK'와 여러 상점 간판에 읽을 수 있는 글자가 있어 판독 가능한 문자 및 로고 금지 조건을 위반한다."
    ],
    "physics": "등장인물과 사물 모두 중력에 맞게 지면과 진열대 위에 놓여 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "참조 사진의 건축물 구조와 다민족 군중을 정확히 구현했으나, 텍스트 배제 지침을 어겼고 흰 얼굴판이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "바나나 가판대는 묘사되었으나, 참조 사진의 건물 좌우 위치를 마음대로 뒤바꿔 장소 고정 지침을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리가 우측 가장자리에 진열된 잡동사니들을 향해 고개를 약간 숙여 바라보고 있습니다.",
        "built_space": "참조 사진의 골목 구조를 완벽히 유지하여 우측에 석재로 된 'Birreria' 상점과 상단 노란 간판, 중앙 아케이드 지붕이 정확히 배치되었습니다.",
        "entities": "찰리는 모자, 낡은 코트, 벗겨진 UBIK 로고, 푸른빛의 점 형태 눈을 갖췄으나 텍스트가 요구한 흰 얼굴판은 아닙니다. 흑인을 포함한 다민족 군중이 존재하나 바나나 가판대는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "찰리와 군중 모두 바닥에 안정적으로 발을 딛고 서 있습니다."
       },
       {
        "label": "B",
        "direction": "찰리가 우측 아래의 진열대와 물건들을 향해 시선을 고정하고 있습니다.",
        "built_space": "부감 구도로 촬영되었으나, 참조 사진에서 우측에 있어야 할 'Birreria' 건물과 노란 간판이 좌측으로 이동하는 등 골목의 뼈대가 완전히 뒤바뀌었습니다.",
        "entities": "찰리는 낡은 외투와 UBIK 로고, 푸른 눈을 가졌으나 흰 얼굴판이 아닙니다. 바나나 가판대는 등장하지만, 군중은 단일 인종에 가깝습니다.",
        "hard_violations": [],
        "physics": "등장인물과 사물 모두 중력에 맞게 지면과 진열대 위에 놓여 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "참조 사진의 건축물 구조와 다민족 군중을 정확히 구현했으나, 텍스트 배제 지침을 어겼고 흰 얼굴판이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "바나나 가판대는 묘사되었으나, 참조 사진의 건물 좌우 위치를 마음대로 뒤바꿔 장소 고정 지침을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리가 우측 가장자리에 진열된 잡동사니들을 향해 고개를 약간 숙여 바라보고 있습니다.",
        "built_space": "참조 사진의 골목 구조를 완벽히 유지하여 우측에 석재로 된 'Birreria' 상점과 상단 노란 간판, 중앙 아케이드 지붕이 정확히 배치되었습니다.",
        "entities": "찰리는 모자, 낡은 코트, 벗겨진 UBIK 로고, 푸른빛의 점 형태 눈을 갖췄으나 텍스트가 요구한 흰 얼굴판은 아닙니다. 흑인을 포함한 다민족 군중이 존재하나 바나나 가판대는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "찰리와 군중 모두 바닥에 안정적으로 발을 딛고 서 있습니다."
       },
       {
        "label": "B",
        "direction": "찰리가 우측 아래의 진열대와 물건들을 향해 시선을 고정하고 있습니다.",
        "built_space": "부감 구도로 촬영되었으나, 참조 사진에서 우측에 있어야 할 'Birreria' 건물과 노란 간판이 좌측으로 이동하는 등 골목의 뼈대가 완전히 뒤바뀌었습니다.",
        "entities": "찰리는 낡은 외투와 UBIK 로고, 푸른 눈을 가졌으나 흰 얼굴판이 아닙니다. 바나나 가판대는 등장하지만, 군중은 단일 인종에 가깝습니다.",
        "hard_violations": [],
        "physics": "등장인물과 사물 모두 중력에 맞게 지면과 진열대 위에 놓여 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오른쪽 잡동사니를 내려다보는 방향은 맞지만, 찰리가 화면을 과도하게 채워 와이드 숏의 공간 관계가 약하고 얼굴도 유기적인 고릴라에 가까우며 판독 가능한 글자가 남아 있다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "와이드 숏에서 찰리의 부분 측면과 오른쪽 진열대, 군중을 통한 크기 대비 및 기준 장소의 구조를 더 충실히 구현했지만, 읽을 수 있는 간판과 가슴 글자로 인해 최종 사용 조건은 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴과 푸른 두 눈은 화면 오른쪽 아래의 소품 진열대 쪽으로 향한다. 특정 물건을 응시하는지는 분명하지 않지만 관심 방향은 진열대와 연결된다. 진열대와 통로는 오른쪽 위로 후퇴하며 군중은 통로 양방향으로 걷거나 주변을 본다. 무기나 겨냥하는 도구는 없다.",
        "built_space": "높은 사선 시점에서 찰리가 중앙과 왼쪽 대부분을 차지하고 다리 아래가 잘린다. 왼쪽에는 석재 상점 전면, 유리 출입구, 노란 돌출 간판 하나와 그 아래 구형 조명 하나가 보인다. 왼쪽 앞에는 바나나 매대 하나, 오른쪽에는 상자와 바구니를 놓은 연속 진열대가 있으며 통로는 찰리 뒤로 이어진다. 기준 사진의 석재와 노란 간판은 재현하지만 아치형 시장 입구와 전신주 등 장소를 확인할 주요 구조는 이 구도에서 보이지 않는다. 불가능한 반사는 확인되지 않는다.",
        "entities": "찰리는 거대한 고릴라형 몸, 낡은 모자와 코트, 푸른 두 눈, 벗겨진 가슴 도장을 갖고 있다. 그러나 얼굴은 흰 얼굴판과 각진 기계 외장보다는 갈색의 유기적인 고릴라 주둥이에 가깝고 입도 단순한 선 형태가 아니다. 코트에는 짙은 얼룩이 남아 있어 세척 후 상태가 명확하지 않다. 바나나와 여러 잡화가 실제 매대에 보인다. 배경에는 성인 남녀로 보이는 군중이 있으며 정확한 국적은 식별할 수 없다. 군중 자체는 카메라 지시에서 요구한 요소다. 가슴의 'UBIK', 노란 간판의 '포차', 상점명이 판독된다.",
        "hard_violations": [
         "가슴의 'UBIK'와 여러 상점 간판에 읽을 수 있는 글자가 있어 판독 가능한 문자 및 로고 금지 조건을 위반한다."
        ],
        "physics": "찰리는 몸통을 세운 직립 자세이며 다리는 화면 아래로 이어진다. 발이 잘려 지면 접촉은 직접 확인되지 않지만 공중에 떠 있다는 증거는 없다. 군중의 보이는 발은 통로 바닥을 딛는다. 잡화는 상자와 진열대 위에 놓여 있고 바나나는 매대에 쌓이거나 가로 지지대에 매달려 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리는 머리를 화면 오른쪽 아래로 돌리고 숙여 오른쪽의 냄비·공구 진열대를 바라본다. 부분 측면 얼굴과 두 푸른 눈의 방향이 관심 대상에 더 명확히 연결된다. 앞쪽 남성은 등을 보인 채 오른쪽으로 고개를 돌리고, 왼쪽 손님과 상인은 매대 물건을 내려다본다. 나머지 군중은 뒤쪽 통로에 서로 다른 방향으로 서 있거나 이동한다.",
        "built_space": "지면 가까운 와이드 숏으로 찰리의 모자부터 바닥에 닿은 발까지 보이며, 앞쪽 인물이 하체 일부를 가린다. 오른쪽에는 기준 사진과 같은 석재 상점, 노란 돌출 간판 하나와 가로 간판 일부, 구형 조명 하나가 보인다. 중앙 뒤에는 반원형 시장 입구 하나와 녹색 간판, 그 왼쪽에는 전신주 하나와 전선들이 있다. 왼쪽의 낮은 처마와 배수관, 골목 포장 및 노란 가장자리 선도 장소 기준과 연결된다. 양쪽 매대 사이에 사람이 설 통로가 확보되어 있다. 오른쪽 진열 물품은 화면 밖으로 이어지지만 후퇴 방향은 화면 오른쪽 대각선보다는 중앙 쪽 깊이에 가깝다. 불가능한 반사는 보이지 않는다.",
        "entities": "찰리는 거대한 고릴라형 기계 몸, 분절된 머리 외장과 손, 낡은 모자와 코트, 푸른 점 형태의 두 눈, 벗겨진 가슴 도장을 갖는다. 얼굴판은 흰색보다는 짙은 회색이고 주둥이 형태가 강하지만 A보다 기계적 외장이 뚜렷하다. 전자 잡화, 옷, 냄비와 공구가 진열되어 있다. 바나나 매대는 이 프레임에서 확인되지 않는다. 성인 남녀와 여러 피부색의 군중이 보이지만 국적은 단정할 수 없다. 가슴의 'UBIK', 시장과 상점 간판, 일부 상품 표기가 읽힌다.",
        "hard_violations": [
         "가슴의 'UBIK', 시장·상점 간판 및 일부 상품에 판독 가능한 문자가 남아 있어 문자 및 로고 금지 조건을 위반한다."
        ],
        "physics": "찰리의 보이는 기계 발은 포장 바닥을 딛고 있으며 직립한 몸을 지탱한다. 앞쪽 인물에 가려진 부분을 제외하면 관절과 체중 지지가 자연스럽다. 사람들은 발로 바닥을 딛고 있고 상인이 다루는 물건에는 손이 닿아 있다. 냄비와 공구는 탁자나 상자에 놓이고 옷은 옷걸이와 가로 봉에 걸려 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "오른쪽 잡동사니를 내려다보는 방향은 맞지만, 찰리가 화면을 과도하게 채워 와이드 숏의 공간 관계가 약하고 얼굴도 유기적인 고릴라에 가까우며 판독 가능한 글자가 남아 있다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "와이드 숏에서 찰리의 부분 측면과 오른쪽 진열대, 군중을 통한 크기 대비 및 기준 장소의 구조를 더 충실히 구현했지만, 읽을 수 있는 간판과 가슴 글자로 인해 최종 사용 조건은 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴과 푸른 두 눈은 화면 오른쪽 아래의 소품 진열대 쪽으로 향한다. 특정 물건을 응시하는지는 분명하지 않지만 관심 방향은 진열대와 연결된다. 진열대와 통로는 오른쪽 위로 후퇴하며 군중은 통로 양방향으로 걷거나 주변을 본다. 무기나 겨냥하는 도구는 없다.",
        "built_space": "높은 사선 시점에서 찰리가 중앙과 왼쪽 대부분을 차지하고 다리 아래가 잘린다. 왼쪽에는 석재 상점 전면, 유리 출입구, 노란 돌출 간판 하나와 그 아래 구형 조명 하나가 보인다. 왼쪽 앞에는 바나나 매대 하나, 오른쪽에는 상자와 바구니를 놓은 연속 진열대가 있으며 통로는 찰리 뒤로 이어진다. 기준 사진의 석재와 노란 간판은 재현하지만 아치형 시장 입구와 전신주 등 장소를 확인할 주요 구조는 이 구도에서 보이지 않는다. 불가능한 반사는 확인되지 않는다.",
        "entities": "찰리는 거대한 고릴라형 몸, 낡은 모자와 코트, 푸른 두 눈, 벗겨진 가슴 도장을 갖고 있다. 그러나 얼굴은 흰 얼굴판과 각진 기계 외장보다는 갈색의 유기적인 고릴라 주둥이에 가깝고 입도 단순한 선 형태가 아니다. 코트에는 짙은 얼룩이 남아 있어 세척 후 상태가 명확하지 않다. 바나나와 여러 잡화가 실제 매대에 보인다. 배경에는 성인 남녀로 보이는 군중이 있으며 정확한 국적은 식별할 수 없다. 군중 자체는 카메라 지시에서 요구한 요소다. 가슴의 'UBIK', 노란 간판의 '포차', 상점명이 판독된다.",
        "hard_violations": [
         "가슴의 'UBIK'와 여러 상점 간판에 읽을 수 있는 글자가 있어 판독 가능한 문자 및 로고 금지 조건을 위반한다."
        ],
        "physics": "찰리는 몸통을 세운 직립 자세이며 다리는 화면 아래로 이어진다. 발이 잘려 지면 접촉은 직접 확인되지 않지만 공중에 떠 있다는 증거는 없다. 군중의 보이는 발은 통로 바닥을 딛는다. 잡화는 상자와 진열대 위에 놓여 있고 바나나는 매대에 쌓이거나 가로 지지대에 매달려 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리는 머리를 화면 오른쪽 아래로 돌리고 숙여 오른쪽의 냄비·공구 진열대를 바라본다. 부분 측면 얼굴과 두 푸른 눈의 방향이 관심 대상에 더 명확히 연결된다. 앞쪽 남성은 등을 보인 채 오른쪽으로 고개를 돌리고, 왼쪽 손님과 상인은 매대 물건을 내려다본다. 나머지 군중은 뒤쪽 통로에 서로 다른 방향으로 서 있거나 이동한다.",
        "built_space": "지면 가까운 와이드 숏으로 찰리의 모자부터 바닥에 닿은 발까지 보이며, 앞쪽 인물이 하체 일부를 가린다. 오른쪽에는 기준 사진과 같은 석재 상점, 노란 돌출 간판 하나와 가로 간판 일부, 구형 조명 하나가 보인다. 중앙 뒤에는 반원형 시장 입구 하나와 녹색 간판, 그 왼쪽에는 전신주 하나와 전선들이 있다. 왼쪽의 낮은 처마와 배수관, 골목 포장 및 노란 가장자리 선도 장소 기준과 연결된다. 양쪽 매대 사이에 사람이 설 통로가 확보되어 있다. 오른쪽 진열 물품은 화면 밖으로 이어지지만 후퇴 방향은 화면 오른쪽 대각선보다는 중앙 쪽 깊이에 가깝다. 불가능한 반사는 보이지 않는다.",
        "entities": "찰리는 거대한 고릴라형 기계 몸, 분절된 머리 외장과 손, 낡은 모자와 코트, 푸른 점 형태의 두 눈, 벗겨진 가슴 도장을 갖는다. 얼굴판은 흰색보다는 짙은 회색이고 주둥이 형태가 강하지만 A보다 기계적 외장이 뚜렷하다. 전자 잡화, 옷, 냄비와 공구가 진열되어 있다. 바나나 매대는 이 프레임에서 확인되지 않는다. 성인 남녀와 여러 피부색의 군중이 보이지만 국적은 단정할 수 없다. 가슴의 'UBIK', 시장과 상점 간판, 일부 상품 표기가 읽힌다.",
        "hard_violations": [
         "가슴의 'UBIK', 시장·상점 간판 및 일부 상품에 판독 가능한 문자가 남아 있어 문자 및 로고 금지 조건을 위반한다."
        ],
        "physics": "찰리의 보이는 기계 발은 포장 바닥을 딛고 있으며 직립한 몸을 지탱한다. 앞쪽 인물에 가려진 부분을 제외하면 관절과 체중 지지가 자연스럽다. 사람들은 발로 바닥을 딛고 있고 상인이 다루는 물건에는 손이 닿아 있다. 냄비와 공구는 탁자나 상자에 놓이고 옷은 옷걸이와 가로 봉에 걸려 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.267
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.017
   },
   "violations": {
    "B": [
     "[gpt-high] 가슴의 'UBIK'와 여러 상점 간판에 읽을 수 있는 글자가 있어 판독 가능한 문자 및 로고 금지 조건을 위반한다."
    ],
    "A": [
     "[gpt-high] 가슴의 'UBIK', 시장·상점 간판 및 일부 상품에 판독 가능한 문자가 남아 있어 문자 및 로고 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1017
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "참조 사진의 건축물 구조와 다민족 군중을 정확히 구현했으나, 텍스트 배제 지침을 어겼고 흰 얼굴판이 누락되었습니다.  ★위반: [gpt-high] 가슴의 'UBIK', 시장·상점 간판 및 일부 상품에 판독 가능한 문자가 남아 있어 문자 및 로고 금지 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1017,
    "verdict_ko": "바나나 가판대는 묘사되었으나, 참조 사진의 건물 좌우 위치를 마음대로 뒤바꿔 장소 고정 지침을 심각하게 위반했습니다.  ★위반: [gpt-high] 가슴의 'UBIK'와 여러 상점 간판에 읽을 수 있는 글자가 있어 판독 가능한 문자 및 로고 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_refugee_market_street_sel.png",
    "asset_id": "557ad3c7-175c-42e9-9709-36354572bb5a",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd9a-146b-7356-87ed-71f41c3c4dbe",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S22sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:26:41.287113+00:00",
  "fingerprint": "89fae95a133b2065cfaca55f79eaa9211f006b43f77c2d2ba58451583a589042",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S22sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S22sh1_sel.png",
  "source_sha256": "a5680952b927fe2c2b57d0633aa3b702c72f5a3c0fcde8ea4e8e79d4bb2c1fa3",
  "file": "S22sh1_cine.png",
  "staged_sha256": "c24799aa8975b8582dccfcf93022f13f3121a14b021d5ae5a6e7099f31b41c0c",
  "latency_ms": 13038
 },
 "S22sh6::signage": {
  "fp": "e711c372c504ba46",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S22sh6": {
  "input_fingerprint": "2655df7bce1eb1d6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수많은 인파 속에서 찰리의 거대한 뒷모습이 복잡한 시장 텐트 틈새로 상체를 반쯤 들이밀고 한 발을 내디딘 외관.\n\nLOCATION (lock): In a crowded passage between the refugee settlement market's tented stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Crowd passage receiving 찰리 in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Market tents (Flanking the gap 찰리 is entering) — Their sides border the route obliquely on either side; used as Separate narrow side shapes funnel the eye toward his disappearing upper back without any single tent overwhelming the image; Crowd passage (A gap among the numerous market pedestrians admits 찰리) — The passage recedes toward the upper center; used as Staggered bodies partially conceal him, with differing step phases, shoulder rotations, and natural intervals preserving individual variation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established daytime ambient balance unchanged as overlapping bodies and tents create the sense of enclosure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains stocked; no banana has been taken. Charlie retains his old coat and hat, cleaned but weathered metal exterior, blue-lit eyes, and peeling chest logo as he moves into the crowd.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수많은 인파 속에서 찰리의 거대한 뒷모습이 복잡한 시장 텐트 틈새로 상체를 반쯤 들이밀고 한 발을 내디딘 외관.\n\nLOCATION (lock): In a crowded passage between the refugee settlement market's tented stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Crowd passage receiving 찰리 in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Market tents (Flanking the gap 찰리 is entering) — Their sides border the route obliquely on either side; used as Separate narrow side shapes funnel the eye toward his disappearing upper back without any single tent overwhelming the image; Crowd passage (A gap among the numerous market pedestrians admits 찰리) — The passage recedes toward the upper center; used as Staggered bodies partially conceal him, with differing step phases, shoulder rotations, and natural intervals preserving individual variation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established daytime ambient balance unchanged as overlapping bodies and tents create the sense of enclosure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains stocked; no banana has been taken. Charlie retains his old coat and hat, cleaned but weathered metal exterior, blue-lit eyes, and peeling chest logo as he moves into the crowd.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수많은 인파 속에서 찰리의 거대한 뒷모습이 복잡한 시장 텐트 틈새로 상체를 반쯤 들이밀고 한 발을 내디딘 외관.\n\nLOCATION (lock): In a crowded passage between the refugee settlement market's tented stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Crowd passage receiving 찰리 in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Market tents (Flanking the gap 찰리 is entering) — Their sides border the route obliquely on either side; used as Separate narrow side shapes funnel the eye toward his disappearing upper back without any single tent overwhelming the image; Crowd passage (A gap among the numerous market pedestrians admits 찰리) — The passage recedes toward the upper center; used as Staggered bodies partially conceal him, with differing step phases, shoulder rotations, and natural intervals preserving individual variation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established daytime ambient balance unchanged as overlapping bodies and tents create the sense of enclosure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains stocked; no banana has been taken. Charlie retains his old coat and hat, cleaned but weathered metal exterior, blue-lit eyes, and peeling chest logo as he moves into the crowd.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리가 군중이 모인 시장 골목 안쪽을 향해 전진하고 있음.",
    "built_space": "양옆으로 텐트가 사선으로 늘어서 있고 통로가 화면 중앙 상단으로 깊어지는 구조가 잘 구현됨.",
    "entities": "코트와 모자를 쓴 찰리의 거대한 뒷모습이 묘사됨. 이전 샷의 인물들이 등장하지 않음.",
    "hard_violations": [
     "[gpt-high] 배경 및 좌측 세로 간판에 읽을 수 있는 문자가 남아 있어, 모든 문자를 판독 불가능하게 하라는 조건을 위반합니다.",
     "[gpt-high] 이전 사진의 남색 조끼·자주색 체크 셔츠 남성과 회색 셔츠·짙은 재킷 남성의 외형·의상을 재사용하여, 찰리 외 인물의 재등장 금지 조건을 위반합니다."
    ],
    "physics": "찰리와 주변 상인, 보행자들의 발 접지와 자세가 자연스럽고 안정적임."
   },
   {
    "label": "B",
    "direction": "찰리가 군중을 헤치며 골목 깊숙한 곳으로 이동하고 있음.",
    "built_space": "양쪽에 천막과 바나나 가판대가 배치된 시장 통로가 묘사됨.",
    "entities": "찰리의 코트 등판에 가슴에 있어야 할 'UBIK' 로고가 그려져 있음.",
    "hard_violations": [
     "[gemini-pro] 이전 샷에 등장한 붉은 조끼 입은 남성이 좌측 배경에 그대로 복사되어 나타남 (인물 중복 금지 위반).",
     "[gemini-pro] 금속 맨몸 가슴에 위치한 로고가 겉옷(코트)의 등판에 나타나는 논리적 오류.",
     "[gpt-high] 찰리 등에 선명한 'UBIK' 문자 표식이 추가되고 배경 간판에도 읽히는 문자가 있어, 문자와 로고 금지 조건을 위반합니다.",
     "[gpt-high] 이전 사진의 남색 조끼·자주색 체크 셔츠 남성과 붉은 조끼 판매원의 외형·의상을 다시 사용하여, 찰리 외 인물의 재등장 금지 조건을 위반합니다."
    ],
    "physics": "찰리가 오른발을 들고 내딛는 걷기 동작을 취하고 있으나 지지 구조가 다소 어색함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트의 지시대로 찰리의 뒷모습을 논리적으로 연출했으며, 이전 샷의 인물을 철저히 배제하여 요구사항을 훌륭히 충족했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이전 샷의 특정 인물을 그대로 복사해 오고, 가슴에 있어야 할 로고를 겉옷 등판에 그리는 등 치명적인 지시 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리가 군중이 모인 시장 골목 안쪽을 향해 전진하고 있음.",
        "built_space": "양옆으로 텐트가 사선으로 늘어서 있고 통로가 화면 중앙 상단으로 깊어지는 구조가 잘 구현됨.",
        "entities": "코트와 모자를 쓴 찰리의 거대한 뒷모습이 묘사됨. 이전 샷의 인물들이 등장하지 않음.",
        "hard_violations": [],
        "physics": "찰리와 주변 상인, 보행자들의 발 접지와 자세가 자연스럽고 안정적임."
       },
       {
        "label": "B",
        "direction": "찰리가 군중을 헤치며 골목 깊숙한 곳으로 이동하고 있음.",
        "built_space": "양쪽에 천막과 바나나 가판대가 배치된 시장 통로가 묘사됨.",
        "entities": "찰리의 코트 등판에 가슴에 있어야 할 'UBIK' 로고가 그려져 있음.",
        "hard_violations": [
         "이전 샷에 등장한 붉은 조끼 입은 남성이 좌측 배경에 그대로 복사되어 나타남 (인물 중복 금지 위반).",
         "금속 맨몸 가슴에 위치한 로고가 겉옷(코트)의 등판에 나타나는 논리적 오류."
        ],
        "physics": "찰리가 오른발을 들고 내딛는 걷기 동작을 취하고 있으나 지지 구조가 다소 어색함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트의 지시대로 찰리의 뒷모습을 논리적으로 연출했으며, 이전 샷의 인물을 철저히 배제하여 요구사항을 훌륭히 충족했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이전 샷의 특정 인물을 그대로 복사해 오고, 가슴에 있어야 할 로고를 겉옷 등판에 그리는 등 치명적인 지시 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리가 군중이 모인 시장 골목 안쪽을 향해 전진하고 있음.",
        "built_space": "양옆으로 텐트가 사선으로 늘어서 있고 통로가 화면 중앙 상단으로 깊어지는 구조가 잘 구현됨.",
        "entities": "코트와 모자를 쓴 찰리의 거대한 뒷모습이 묘사됨. 이전 샷의 인물들이 등장하지 않음.",
        "hard_violations": [],
        "physics": "찰리와 주변 상인, 보행자들의 발 접지와 자세가 자연스럽고 안정적임."
       },
       {
        "label": "B",
        "direction": "찰리가 군중을 헤치며 골목 깊숙한 곳으로 이동하고 있음.",
        "built_space": "양쪽에 천막과 바나나 가판대가 배치된 시장 통로가 묘사됨.",
        "entities": "찰리의 코트 등판에 가슴에 있어야 할 'UBIK' 로고가 그려져 있음.",
        "hard_violations": [
         "이전 샷에 등장한 붉은 조끼 입은 남성이 좌측 배경에 그대로 복사되어 나타남 (인물 중복 금지 위반).",
         "금속 맨몸 가슴에 위치한 로고가 겉옷(코트)의 등판에 나타나는 논리적 오류."
        ],
        "physics": "찰리가 오른발을 들고 내딛는 걷기 동작을 취하고 있으나 지지 구조가 다소 어색함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "찰리의 전진하는 뒷모습은 보이지만 몸이 화면을 지나치게 크게 차지하고, 등에 추가된 선명한 문자와 이전 장면 인물·의상 재사용이 명시적 금지 조건을 위반합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인파가 찰리의 하체를 가리고 양옆 텐트가 통로를 만드는 구도는 더 충실하지만, 읽히는 간판과 이전 장면 인물·의상 재사용 때문에 최종 사용에는 부적합합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 카메라에 등을 보이고 화면 위쪽 중앙의 시장 통로로 향합니다. 얼굴과 눈은 보이지 않아 시선 자체는 확인할 수 없습니다. 들어 올린 발바닥이 카메라 쪽으로 보여 앞으로 걷는 방향과 맞습니다. 주변 사람들은 좌우의 상품이나 이동 방향을 보고 있습니다. 다만 찰리의 상체가 텐트 사이로 반쯤 사라지기보다 거의 전부 드러나 있습니다.",
        "built_space": "중앙 포장 통로 양쪽에 판매대가 있고, 왼쪽에는 큰 흰 텐트 한 동과 그 뒤 작은 푸른 차양, 오른쪽에는 흰색·청색 텐트 지붕 세 부분이 보입니다. 상부에는 짙은 방수포가 걸려 있습니다. 오른쪽 바나나 판매대는 충분히 채워져 있습니다. 통로의 원근은 성립하지만 찰리가 중경을 크게 점유하고 텐트도 넓은 면적으로 보여, 배경의 찰리와 좁은 측면 텐트라는 구성에서 벗어납니다. 이전 사진의 주황 배수관, 왼쪽 유리 상점, 오른쪽 석재 상점 등 장소 식별 요소는 확인하기 어렵습니다.",
        "entities": "찰리는 낡은 회갈색 코트와 모자를 쓴 거대한 금속 신체로 표현되어 보이는 범위의 기본 정체성은 맞습니다. 얼굴판·푸른 눈·가슴 표식은 뒤돌아 있어 평가 대상이 아닙니다. 대신 등에 참조에 없는 큰 'UBIK' 문자 표식이 생겼습니다. 군중은 다양한 피부색의 성인 남녀로 보이며, 군중 자체는 장면에 요구됩니다. 그러나 왼쪽 전경의 남색 조끼·자주색 체크 셔츠 남성과 왼쪽 판매대의 붉은 조끼 남성은 이전 사진의 인물과 의상을 재사용한 것으로 보입니다. 배경 간판에도 읽을 수 있는 한글이 있습니다.",
        "hard_violations": [
         "찰리 등에 선명한 'UBIK' 문자 표식이 추가되고 배경 간판에도 읽히는 문자가 있어, 문자와 로고 금지 조건을 위반합니다.",
         "이전 사진의 남색 조끼·자주색 체크 셔츠 남성과 붉은 조끼 판매원의 외형·의상을 다시 사용하여, 찰리 외 인물의 재등장 금지 조건을 위반합니다."
        ],
        "physics": "찰리의 한쪽 발은 발바닥을 드러내며 들려 있고 다른 다리의 접지 부분은 군중에 가려져 있습니다. 골반과 다리 배치는 보행으로 해석 가능하며, 떠 있다고 단정할 근거는 없습니다. 상자를 든 남성은 양손으로 상자를 받치고 있습니다. 상품은 판매대와 상자 위에 놓여 있고, 텐트는 기둥과 연결된 천 구조로 지지됩니다."
       },
       {
        "label": "B",
        "direction": "찰리는 등을 보인 채 위쪽 중앙으로 이어지는 통로를 향합니다. 얼굴이 가려져 눈의 방향은 확인되지 않습니다. 군중은 카메라 쪽으로 오거나 안쪽으로 들어가며, 일부는 옆 판매대를 봅니다. 찰리의 진행 목표는 분명하지만 상체를 틈으로 반쯤 밀어 넣는 굽힘은 약하고, 한 발을 내딛는 순간은 가려져 확인되지 않습니다.",
        "built_space": "약간 높은 시점에서 중앙 통로와 양쪽의 연속된 천막 판매대를 내려다봅니다. 왼쪽에는 전경의 잘린 천막, 푸른 방수포 천막, 흰 지붕, 주황 차양과 뒤쪽 푸른 지붕이 이어지고, 오른쪽에는 전경의 노랑·초록 천막과 여러 청색·흰색·분홍색 지붕이 겹칩니다. 왼쪽 바나나 판매대와 오른쪽 과일 판매대가 통로를 둘러쌉니다. 인파가 찰리의 하체를 단계적으로 가려 A보다 배경으로 들어가는 구도에 가깝습니다. 다만 찰리의 등이 여전히 크게 노출되고, 이전 사진의 고유 상점 외관은 뚜렷하게 유지되지 않습니다. 시점은 프롬프트가 허용한 범위이며 불가능한 반사는 없습니다.",
        "entities": "낡은 회갈색 코트, 중절모, 금속 목과 손이 보이는 찰리 한 명이 있습니다. 숨겨진 얼굴·눈·가슴 표식은 확인할 수 없으며 결함으로 보지 않습니다. 양쪽 판매대에는 바나나와 다른 과일이 충분히 남아 있습니다. 군중은 여러 피부색의 성인 남녀로 구성됩니다. 그러나 중앙의 남색 조끼·자주색 체크 셔츠 남성과 왼쪽 중경의 회색 셔츠·짙은 재킷 남성은 이전 사진 인물의 외형과 의상을 재사용한 것으로 보입니다. 배경과 좌측 세로 간판에는 식별 가능한 한글이 남아 있습니다.",
        "hard_violations": [
         "배경 및 좌측 세로 간판에 읽을 수 있는 문자가 남아 있어, 모든 문자를 판독 불가능하게 하라는 조건을 위반합니다.",
         "이전 사진의 남색 조끼·자주색 체크 셔츠 남성과 회색 셔츠·짙은 재킷 남성의 외형·의상을 재사용하여, 찰리 외 인물의 재등장 금지 조건을 위반합니다."
        ],
        "physics": "찰리의 다리와 발은 앞의 사람들에 가려져 접지를 직접 확인할 수 없습니다. 보이는 몸통과 팔에는 공중 부양이나 불가능한 관절 배치가 없으며, 발이 가려졌다는 이유만으로 무지지 상태라고 판단하지 않습니다. 군중은 통로에 서 있거나 걷고 있고, 판매원들의 손은 상품 쪽으로 뻗어 있습니다. 과일은 판매대가 받치며, 매달린 노란 상품은 오른쪽 기둥 부근에 고정되어 있습니다. 천막은 기둥과 프레임으로 지지됩니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "찰리의 전진하는 뒷모습은 보이지만 몸이 화면을 지나치게 크게 차지하고, 등에 추가된 선명한 문자와 이전 장면 인물·의상 재사용이 명시적 금지 조건을 위반합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인파가 찰리의 하체를 가리고 양옆 텐트가 통로를 만드는 구도는 더 충실하지만, 읽히는 간판과 이전 장면 인물·의상 재사용 때문에 최종 사용에는 부적합합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 카메라에 등을 보이고 화면 위쪽 중앙의 시장 통로로 향합니다. 얼굴과 눈은 보이지 않아 시선 자체는 확인할 수 없습니다. 들어 올린 발바닥이 카메라 쪽으로 보여 앞으로 걷는 방향과 맞습니다. 주변 사람들은 좌우의 상품이나 이동 방향을 보고 있습니다. 다만 찰리의 상체가 텐트 사이로 반쯤 사라지기보다 거의 전부 드러나 있습니다.",
        "built_space": "중앙 포장 통로 양쪽에 판매대가 있고, 왼쪽에는 큰 흰 텐트 한 동과 그 뒤 작은 푸른 차양, 오른쪽에는 흰색·청색 텐트 지붕 세 부분이 보입니다. 상부에는 짙은 방수포가 걸려 있습니다. 오른쪽 바나나 판매대는 충분히 채워져 있습니다. 통로의 원근은 성립하지만 찰리가 중경을 크게 점유하고 텐트도 넓은 면적으로 보여, 배경의 찰리와 좁은 측면 텐트라는 구성에서 벗어납니다. 이전 사진의 주황 배수관, 왼쪽 유리 상점, 오른쪽 석재 상점 등 장소 식별 요소는 확인하기 어렵습니다.",
        "entities": "찰리는 낡은 회갈색 코트와 모자를 쓴 거대한 금속 신체로 표현되어 보이는 범위의 기본 정체성은 맞습니다. 얼굴판·푸른 눈·가슴 표식은 뒤돌아 있어 평가 대상이 아닙니다. 대신 등에 참조에 없는 큰 'UBIK' 문자 표식이 생겼습니다. 군중은 다양한 피부색의 성인 남녀로 보이며, 군중 자체는 장면에 요구됩니다. 그러나 왼쪽 전경의 남색 조끼·자주색 체크 셔츠 남성과 왼쪽 판매대의 붉은 조끼 남성은 이전 사진의 인물과 의상을 재사용한 것으로 보입니다. 배경 간판에도 읽을 수 있는 한글이 있습니다.",
        "hard_violations": [
         "찰리 등에 선명한 'UBIK' 문자 표식이 추가되고 배경 간판에도 읽히는 문자가 있어, 문자와 로고 금지 조건을 위반합니다.",
         "이전 사진의 남색 조끼·자주색 체크 셔츠 남성과 붉은 조끼 판매원의 외형·의상을 다시 사용하여, 찰리 외 인물의 재등장 금지 조건을 위반합니다."
        ],
        "physics": "찰리의 한쪽 발은 발바닥을 드러내며 들려 있고 다른 다리의 접지 부분은 군중에 가려져 있습니다. 골반과 다리 배치는 보행으로 해석 가능하며, 떠 있다고 단정할 근거는 없습니다. 상자를 든 남성은 양손으로 상자를 받치고 있습니다. 상품은 판매대와 상자 위에 놓여 있고, 텐트는 기둥과 연결된 천 구조로 지지됩니다."
       },
       {
        "label": "A",
        "direction": "찰리는 등을 보인 채 위쪽 중앙으로 이어지는 통로를 향합니다. 얼굴이 가려져 눈의 방향은 확인되지 않습니다. 군중은 카메라 쪽으로 오거나 안쪽으로 들어가며, 일부는 옆 판매대를 봅니다. 찰리의 진행 목표는 분명하지만 상체를 틈으로 반쯤 밀어 넣는 굽힘은 약하고, 한 발을 내딛는 순간은 가려져 확인되지 않습니다.",
        "built_space": "약간 높은 시점에서 중앙 통로와 양쪽의 연속된 천막 판매대를 내려다봅니다. 왼쪽에는 전경의 잘린 천막, 푸른 방수포 천막, 흰 지붕, 주황 차양과 뒤쪽 푸른 지붕이 이어지고, 오른쪽에는 전경의 노랑·초록 천막과 여러 청색·흰색·분홍색 지붕이 겹칩니다. 왼쪽 바나나 판매대와 오른쪽 과일 판매대가 통로를 둘러쌉니다. 인파가 찰리의 하체를 단계적으로 가려 A보다 배경으로 들어가는 구도에 가깝습니다. 다만 찰리의 등이 여전히 크게 노출되고, 이전 사진의 고유 상점 외관은 뚜렷하게 유지되지 않습니다. 시점은 프롬프트가 허용한 범위이며 불가능한 반사는 없습니다.",
        "entities": "낡은 회갈색 코트, 중절모, 금속 목과 손이 보이는 찰리 한 명이 있습니다. 숨겨진 얼굴·눈·가슴 표식은 확인할 수 없으며 결함으로 보지 않습니다. 양쪽 판매대에는 바나나와 다른 과일이 충분히 남아 있습니다. 군중은 여러 피부색의 성인 남녀로 구성됩니다. 그러나 중앙의 남색 조끼·자주색 체크 셔츠 남성과 왼쪽 중경의 회색 셔츠·짙은 재킷 남성은 이전 사진 인물의 외형과 의상을 재사용한 것으로 보입니다. 배경과 좌측 세로 간판에는 식별 가능한 한글이 남아 있습니다.",
        "hard_violations": [
         "배경 및 좌측 세로 간판에 읽을 수 있는 문자가 남아 있어, 모든 문자를 판독 불가능하게 하라는 조건을 위반합니다.",
         "이전 사진의 남색 조끼·자주색 체크 셔츠 남성과 회색 셔츠·짙은 재킷 남성의 외형·의상을 재사용하여, 찰리 외 인물의 재등장 금지 조건을 위반합니다."
        ],
        "physics": "찰리의 다리와 발은 앞의 사람들에 가려져 접지를 직접 확인할 수 없습니다. 보이는 몸통과 팔에는 공중 부양이나 불가능한 관절 배치가 없으며, 발이 가려졌다는 이유만으로 무지지 상태라고 판단하지 않습니다. 군중은 통로에 서 있거나 걷고 있고, 판매원들의 손은 상품 쪽으로 뻗어 있습니다. 과일은 판매대가 받치며, 매달린 노란 상품은 오른쪽 기둥 부근에 고정되어 있습니다. 천막은 기둥과 프레임으로 지지됩니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.095
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.845
   },
   "violations": {
    "B": [
     "[gemini-pro] 이전 샷에 등장한 붉은 조끼 입은 남성이 좌측 배경에 그대로 복사되어 나타남 (인물 중복 금지 위반).",
     "[gemini-pro] 금속 맨몸 가슴에 위치한 로고가 겉옷(코트)의 등판에 나타나는 논리적 오류.",
     "[gpt-high] 찰리 등에 선명한 'UBIK' 문자 표식이 추가되고 배경 간판에도 읽히는 문자가 있어, 문자와 로고 금지 조건을 위반합니다.",
     "[gpt-high] 이전 사진의 남색 조끼·자주색 체크 셔츠 남성과 붉은 조끼 판매원의 외형·의상을 다시 사용하여, 찰리 외 인물의 재등장 금지 조건을 위반합니다."
    ],
    "A": [
     "[gpt-high] 배경 및 좌측 세로 간판에 읽을 수 있는 문자가 남아 있어, 모든 문자를 판독 불가능하게 하라는 조건을 위반합니다.",
     "[gpt-high] 이전 사진의 남색 조끼·자주색 체크 셔츠 남성과 회색 셔츠·짙은 재킷 남성의 외형·의상을 재사용하여, 찰리 외 인물의 재등장 금지 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 845
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "프롬프트의 지시대로 찰리의 뒷모습을 논리적으로 연출했으며, 이전 샷의 인물을 철저히 배제하여 요구사항을 훌륭히 충족했습니다.  ★위반: [gpt-high] 배경 및 좌측 세로 간판에 읽을 수 있는 문자가 남아 있어, 모든 문자를 판독 불가능하게 하라는 조건을 위반합니다. / [gpt-high] 이전 사진의 남색 조끼·자주색 체크 셔츠 남성과 회색 셔츠·짙은 재킷 남성의 외형·의상을 재사용하여, 찰리 외 인물의 재등장 금지 조건을 위반합니다."
   },
   {
    "label": "B",
    "score": 845,
    "verdict_ko": "이전 샷의 특정 인물을 그대로 복사해 오고, 가슴에 있어야 할 로고를 겉옷 등판에 그리는 등 치명적인 지시 위반이 발생했습니다.  ★위반: [gemini-pro] 이전 샷에 등장한 붉은 조끼 입은 남성이 좌측 배경에 그대로 복사되어 나타남 (인물 중복 금지 위반). / [gemini-pro] 금속 맨몸 가슴에 위치한 로고가 겉옷(코트)의 등판에 나타나는 논리적 오류. / [gpt-high] 찰리 등에 선명한 'UBIK' 문자 표식이 추가되고 배경 간판에도 읽히는 문자가 있어, 문자와 로고 금지 조건을 위반합니다. / [gpt-high] 이전 사진의 남색 조끼·자주색 체크 셔츠 남성과 붉은 조끼 판매원의 외형·의상을 다시 사용하여, 찰리 외 인물의 재등장 금지 조건을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S22sh1_sel.png",
    "asset_id": "f12fe40a-c94c-486b-b503-e304478012a2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1597480>",
    "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd9a-ea17-7093-9b98-783d6b084740",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S22sh1"
  }
 },
 "S22sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:26:54.640901+00:00",
  "fingerprint": "0f77c366b45150e8019c62895f11d957396df7b7ee4f9e935240ea70616fa516",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S22sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S22sh6_sel.png",
  "source_sha256": "8a5e8af13ee9a9c3cec3fb13161b717d24348cacdeb4c095c60516363b75672f",
  "file": "S22sh6_cine.png",
  "staged_sha256": "96998d60e808f1e63eb7af5cefc30a64f0b12855f42925cbfdc95a00c4122059",
  "latency_ms": 13000
 },
 "S22sh8::signage": {
  "fp": "ff3416d47a7c0729",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S22sh8": {
  "input_fingerprint": "4da6878af1bcf1fe",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 손가락을 뻗어 찰리가 사라진 인파 저편을 정확히 가리키며 눈을 부릅뜬 긴박한 표정.\n\nLOCATION (lock): At the crowded market-lane junction where the robot has just disappeared between stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Camera-left continuation of 찰리's exit route in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Near edge of the crowd route (찰리 has disappeared into the crowd beyond the camera-side edge) — A visible sliver of the route reaches the lower-left frame edge and continues out of view; used as Connects 앰버's finger and eyeline to a spatially plausible off-frame destination; nearby pedestrians retain varied strides and spacing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the market's daytime ambient continuity, with clear facial contrast supporting urgency rather than a stylized lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the market tents, stall fixtures, ground, and daylight colors from the reference. EXCLUDE the departing robot, which has already disappeared into the crowd.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains in place. Charlie is within the market crowd, retaining his old coat and hat, cleaned but worn metal body, blue-lit eyes, and peeling UBIK logo. 앰버: She has reached the banana-stall area and is ready to continue the search.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 손가락을 뻗어 찰리가 사라진 인파 저편을 정확히 가리키며 눈을 부릅뜬 긴박한 표정.\n\nLOCATION (lock): At the crowded market-lane junction where the robot has just disappeared between stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Camera-left continuation of 찰리's exit route in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Near edge of the crowd route (찰리 has disappeared into the crowd beyond the camera-side edge) — A visible sliver of the route reaches the lower-left frame edge and continues out of view; used as Connects 앰버's finger and eyeline to a spatially plausible off-frame destination; nearby pedestrians retain varied strides and spacing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the market's daytime ambient continuity, with clear facial contrast supporting urgency rather than a stylized lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the market tents, stall fixtures, ground, and daylight colors from the reference. EXCLUDE the departing robot, which has already disappeared into the crowd.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains in place. Charlie is within the market crowd, retaining his old coat and hat, cleaned but worn metal body, blue-lit eyes, and peeling UBIK logo. 앰버: She has reached the banana-stall area and is ready to continue the search.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 손가락을 뻗어 찰리가 사라진 인파 저편을 정확히 가리키며 눈을 부릅뜬 긴박한 표정.\n\nLOCATION (lock): At the crowded market-lane junction where the robot has just disappeared between stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Camera-left continuation of 찰리's exit route in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Near edge of the crowd route (찰리 has disappeared into the crowd beyond the camera-side edge) — A visible sliver of the route reaches the lower-left frame edge and continues out of view; used as Connects 앰버's finger and eyeline to a spatially plausible off-frame destination; nearby pedestrians retain varied strides and spacing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the market's daytime ambient continuity, with clear facial contrast supporting urgency rather than a stylized lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the market tents, stall fixtures, ground, and daylight colors from the reference. EXCLUDE the departing robot, which has already disappeared into the crowd.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains in place. Charlie is within the market crowd, retaining his old coat and hat, cleaned but worn metal body, blue-lit eyes, and peeling UBIK logo. 앰버: She has reached the banana-stall area and is ready to continue the search.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "앰버가 요구된 좌측 찰리의 퇴장 경로가 아닌 우측 상단을 가리키며 응시함.",
    "built_space": "시장 텐트와 가판대의 형태 및 위치는 레퍼런스를 따름.",
    "entities": "앰버는 금발 혼혈 소녀의 특징을 가지나, 배경 군중은 프롬프트에서 금지한 이전 샷 인물들로 복제됨.",
    "hard_violations": [
     "[gemini-pro] physically impossible staging (거대한 스케일과 지지점 없이 부양된 합성 오류)",
     "[gemini-pro] invented people (금지된 이전 샷 인물들의 무단 복제)",
     "[gpt-high] 이전 숏에서 제외하라고 명시한 군중의 얼굴·몸·의상을 다수 그대로 유지했다.",
     "[gpt-high] 상단 간판과 시장 표지에 읽을 수 있는 글자를 남겨 글자 금지 조건을 위반한다."
    ],
    "physics": "앰버의 다리가 보이지 않고 배경에 합성된 듯 공중에 떠 있으며, 주변 인물 대비 스케일이 불가능함."
   },
   {
    "label": "B",
    "direction": "손가락은 카메라 좌측을 향하나 시선이 정면 렌즈에 고정되어 목표 대상과 연결되지 않음.",
    "built_space": "시장 구조물과 가판대 수량이 일치하며, 앰버가 전경에 무리 없이 배치됨.",
    "entities": "앰버의 외모는 레퍼런스와 일치하나, 배경 군중이 이전 샷 이미지에서 그대로 복제됨.",
    "hard_violations": [
     "[gemini-pro] invented people (금지된 이전 샷 인물들의 무단 복제)",
     "[gpt-high] 이전 숏에서 제외하라고 명시한 여러 사람의 얼굴과 의상을 그대로 재사용했다.",
     "[gpt-high] 배경의 노란 세로 간판 등에 식별 가능한 글자가 남아 있어 읽을 수 있는 글자 금지 조건을 위반한다."
    ],
    "physics": "전경 바닥에 체중을 지지하고 서 있으며 팔과 손의 동작이 물리적으로 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "시선이 목표가 아닌 렌즈를 향하고 있으며, 이전 샷의 인물들을 그대로 복제하는 치명적인 지시 위반이 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "금지된 레퍼런스 인물 복제와 더불어 앰버의 방향 오류, 비정상적인 스케일 등 심각한 물리적 연출 실패가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버가 요구된 좌측 찰리의 퇴장 경로가 아닌 우측 상단을 가리키며 응시함.",
        "built_space": "시장 텐트와 가판대의 형태 및 위치는 레퍼런스를 따름.",
        "entities": "앰버는 금발 혼혈 소녀의 특징을 가지나, 배경 군중은 프롬프트에서 금지한 이전 샷 인물들로 복제됨.",
        "hard_violations": [
         "physically impossible staging (거대한 스케일과 지지점 없이 부양된 합성 오류)",
         "invented people (금지된 이전 샷 인물들의 무단 복제)"
        ],
        "physics": "앰버의 다리가 보이지 않고 배경에 합성된 듯 공중에 떠 있으며, 주변 인물 대비 스케일이 불가능함."
       },
       {
        "label": "B",
        "direction": "손가락은 카메라 좌측을 향하나 시선이 정면 렌즈에 고정되어 목표 대상과 연결되지 않음.",
        "built_space": "시장 구조물과 가판대 수량이 일치하며, 앰버가 전경에 무리 없이 배치됨.",
        "entities": "앰버의 외모는 레퍼런스와 일치하나, 배경 군중이 이전 샷 이미지에서 그대로 복제됨.",
        "hard_violations": [
         "invented people (금지된 이전 샷 인물들의 무단 복제)"
        ],
        "physics": "전경 바닥에 체중을 지지하고 서 있으며 팔과 손의 동작이 물리적으로 자연스러움."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "시선이 목표가 아닌 렌즈를 향하고 있으며, 이전 샷의 인물들을 그대로 복제하는 치명적인 지시 위반이 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "금지된 레퍼런스 인물 복제와 더불어 앰버의 방향 오류, 비정상적인 스케일 등 심각한 물리적 연출 실패가 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "앰버가 요구된 좌측 찰리의 퇴장 경로가 아닌 우측 상단을 가리키며 응시함.",
        "built_space": "시장 텐트와 가판대의 형태 및 위치는 레퍼런스를 따름.",
        "entities": "앰버는 금발 혼혈 소녀의 특징을 가지나, 배경 군중은 프롬프트에서 금지한 이전 샷 인물들로 복제됨.",
        "hard_violations": [
         "physically impossible staging (거대한 스케일과 지지점 없이 부양된 합성 오류)",
         "invented people (금지된 이전 샷 인물들의 무단 복제)"
        ],
        "physics": "앰버의 다리가 보이지 않고 배경에 합성된 듯 공중에 떠 있으며, 주변 인물 대비 스케일이 불가능함."
       },
       {
        "label": "B",
        "direction": "손가락은 카메라 좌측을 향하나 시선이 정면 렌즈에 고정되어 목표 대상과 연결되지 않음.",
        "built_space": "시장 구조물과 가판대 수량이 일치하며, 앰버가 전경에 무리 없이 배치됨.",
        "entities": "앰버의 외모는 레퍼런스와 일치하나, 배경 군중이 이전 샷 이미지에서 그대로 복제됨.",
        "hard_violations": [
         "invented people (금지된 이전 샷 인물들의 무단 복제)"
        ],
        "physics": "전경 바닥에 체중을 지지하고 서 있으며 팔과 손의 동작이 물리적으로 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오른쪽에 배치된 앰버의 미디엄 숏과 왼쪽 통로는 상대적으로 가깝지만, 손끝과 시선이 탈출 경로에 모이지 않고 이전 숏 인물과 읽히는 글자를 재사용했다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "손끝과 시선이 요구된 왼쪽 아래가 아닌 오른쪽을 향하며, 이전 숏의 인물·구도를 그대로 유지하고 앰버를 주변 사람보다 과대하게 배치했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 검지는 화면 왼쪽 약간 위를 향하며, 화면상 회색 티셔츠 남성의 어깨 쪽에 걸린다. 눈은 손끝 방향이 아니라 거의 카메라 쪽을 본다. 왼쪽 아래로 이어지는 통로는 있지만 손끝·시선이 그 통로 밖 찰리의 목적지로 연결되지 않는다. 주변 보행자는 카메라 쪽으로 오거나 골목 안으로 들어간다.",
        "built_space": "양쪽에 청색·백색 천막이 연속되고, 왼쪽에는 바나나 판매대 하나와 전경 생선 판매대 하나, 오른쪽에는 과일 판매대 하나와 전경 상품대 하나가 보인다. 상자와 천막 기둥은 지면에 놓여 있으며 중앙 통로가 왼쪽 아래까지 열린다. 앰버는 오른쪽 전경에서 허리 부근까지 보이는 미디엄 숏이다. 시장 재료와 낮빛은 참조에 가깝지만 기존 군중까지 재사용했다.",
        "entities": "금발의 어린 여자아이, 큰 눈, 회녹색 재킷과 얼룩진 회색 원피스는 앰버 참조에 대체로 부합한다. 혼혈 정체성을 외형만으로 확정할 수는 없으며 얼굴의 정확한 일치도는 제한적이다. 참조의 흰 마스크는 없다. 바나나는 있고 찰리 로봇은 보이지 않아 퇴장 조건은 맞는다. 반면 보라색 옷의 바나나 상인, 검은 재킷과 흰 장갑의 남성, 회색 티셔츠 남성, 오른쪽 파란 작업복 남성 등은 이전 숏 인물을 반복한다.",
        "hard_violations": [
         "이전 숏에서 제외하라고 명시한 여러 사람의 얼굴과 의상을 그대로 재사용했다.",
         "배경의 노란 세로 간판 등에 식별 가능한 글자가 남아 있어 읽을 수 있는 글자 금지 조건을 위반한다."
        ],
        "physics": "뻗은 팔과 검지는 어깨·팔꿈치·손목으로 자연스럽게 연결되어 지시 동작이 가능하다. 앰버의 발은 프레임 밖이므로 접지를 직접 확인할 수 없지만 공중에 떠 있다는 증거는 없다. 전경 보행자들은 발로 바닥을 딛거나 한쪽 발을 내딛으며, 판매대는 다리와 상자 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "앰버의 검지와 눈은 모두 화면 오른쪽 위의 프레임 밖을 향한다. 두 방향은 비교적 일치하지만 목표가 요구된 왼쪽 아래 카메라 쪽 탈출 경로와 반대다. 검지의 연장선은 오른쪽 천막 쪽을 지나며, 찰리가 사라진 왼쪽 통로를 가리키지 않는다.",
        "built_space": "왼쪽 바나나 판매대 하나, 오른쪽 과일 판매대 하나와 매달린 바나나 한 묶음, 양쪽의 청색·백색 천막 열이 이전 숏과 거의 같은 위치에 있다. 카메라는 이전 숏의 높은 골목 조망을 사실상 유지한다. 앰버는 중앙에서 허벅지 부근까지 크게 보이고 주변 성인과 판매대에 비해 과대하게 읽힌다. 왼쪽 아래는 사람과 판매대로 막혀 요구된 출구 통로의 가느다란 연결이 드러나지 않는다.",
        "entities": "금발의 어린 여자아이와 회녹색 재킷·회색 원피스는 참조의 앰버에 대체로 맞지만 흰 마스크가 빠졌다. 눈은 정상적인 인간 눈이며 긴박한 표정은 보인다. 바나나 판매대는 유지되고 찰리 로봇은 없다. 그러나 왼쪽 보라색 옷의 상인, 중앙의 회색 티셔츠 남성, 아래쪽 청색 셔츠와 검은 조끼의 인물, 오른쪽 파란 작업복 남성 등 이전 숏의 사람들이 동일한 의상과 위치로 남아 있다.",
        "hard_violations": [
         "이전 숏에서 제외하라고 명시한 군중의 얼굴·몸·의상을 다수 그대로 유지했다.",
         "상단 간판과 시장 표지에 읽을 수 있는 글자를 남겨 글자 금지 조건을 위반한다."
        ],
        "physics": "지시하는 팔은 몸에 연결되어 있고 손가락 자세도 실제로 가능한 동작이다. 앰버의 다리 아래와 발은 잘려 있어 지지면을 직접 확인할 수 없으며, 이것만으로 부유라고 단정할 수는 없다. 다만 성인 군중과의 크기·깊이 관계가 부자연스럽다. 판매대와 상자는 바닥에 놓여 있고 매달린 바나나는 위쪽 고정부에 연결되어 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "오른쪽에 배치된 앰버의 미디엄 숏과 왼쪽 통로는 상대적으로 가깝지만, 손끝과 시선이 탈출 경로에 모이지 않고 이전 숏 인물과 읽히는 글자를 재사용했다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "손끝과 시선이 요구된 왼쪽 아래가 아닌 오른쪽을 향하며, 이전 숏의 인물·구도를 그대로 유지하고 앰버를 주변 사람보다 과대하게 배치했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 검지는 화면 왼쪽 약간 위를 향하며, 화면상 회색 티셔츠 남성의 어깨 쪽에 걸린다. 눈은 손끝 방향이 아니라 거의 카메라 쪽을 본다. 왼쪽 아래로 이어지는 통로는 있지만 손끝·시선이 그 통로 밖 찰리의 목적지로 연결되지 않는다. 주변 보행자는 카메라 쪽으로 오거나 골목 안으로 들어간다.",
        "built_space": "양쪽에 청색·백색 천막이 연속되고, 왼쪽에는 바나나 판매대 하나와 전경 생선 판매대 하나, 오른쪽에는 과일 판매대 하나와 전경 상품대 하나가 보인다. 상자와 천막 기둥은 지면에 놓여 있으며 중앙 통로가 왼쪽 아래까지 열린다. 앰버는 오른쪽 전경에서 허리 부근까지 보이는 미디엄 숏이다. 시장 재료와 낮빛은 참조에 가깝지만 기존 군중까지 재사용했다.",
        "entities": "금발의 어린 여자아이, 큰 눈, 회녹색 재킷과 얼룩진 회색 원피스는 앰버 참조에 대체로 부합한다. 혼혈 정체성을 외형만으로 확정할 수는 없으며 얼굴의 정확한 일치도는 제한적이다. 참조의 흰 마스크는 없다. 바나나는 있고 찰리 로봇은 보이지 않아 퇴장 조건은 맞는다. 반면 보라색 옷의 바나나 상인, 검은 재킷과 흰 장갑의 남성, 회색 티셔츠 남성, 오른쪽 파란 작업복 남성 등은 이전 숏 인물을 반복한다.",
        "hard_violations": [
         "이전 숏에서 제외하라고 명시한 여러 사람의 얼굴과 의상을 그대로 재사용했다.",
         "배경의 노란 세로 간판 등에 식별 가능한 글자가 남아 있어 읽을 수 있는 글자 금지 조건을 위반한다."
        ],
        "physics": "뻗은 팔과 검지는 어깨·팔꿈치·손목으로 자연스럽게 연결되어 지시 동작이 가능하다. 앰버의 발은 프레임 밖이므로 접지를 직접 확인할 수 없지만 공중에 떠 있다는 증거는 없다. 전경 보행자들은 발로 바닥을 딛거나 한쪽 발을 내딛으며, 판매대는 다리와 상자 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "앰버의 검지와 눈은 모두 화면 오른쪽 위의 프레임 밖을 향한다. 두 방향은 비교적 일치하지만 목표가 요구된 왼쪽 아래 카메라 쪽 탈출 경로와 반대다. 검지의 연장선은 오른쪽 천막 쪽을 지나며, 찰리가 사라진 왼쪽 통로를 가리키지 않는다.",
        "built_space": "왼쪽 바나나 판매대 하나, 오른쪽 과일 판매대 하나와 매달린 바나나 한 묶음, 양쪽의 청색·백색 천막 열이 이전 숏과 거의 같은 위치에 있다. 카메라는 이전 숏의 높은 골목 조망을 사실상 유지한다. 앰버는 중앙에서 허벅지 부근까지 크게 보이고 주변 성인과 판매대에 비해 과대하게 읽힌다. 왼쪽 아래는 사람과 판매대로 막혀 요구된 출구 통로의 가느다란 연결이 드러나지 않는다.",
        "entities": "금발의 어린 여자아이와 회녹색 재킷·회색 원피스는 참조의 앰버에 대체로 맞지만 흰 마스크가 빠졌다. 눈은 정상적인 인간 눈이며 긴박한 표정은 보인다. 바나나 판매대는 유지되고 찰리 로봇은 없다. 그러나 왼쪽 보라색 옷의 상인, 중앙의 회색 티셔츠 남성, 아래쪽 청색 셔츠와 검은 조끼의 인물, 오른쪽 파란 작업복 남성 등 이전 숏의 사람들이 동일한 의상과 위치로 남아 있다.",
        "hard_violations": [
         "이전 숏에서 제외하라고 명시한 군중의 얼굴·몸·의상을 다수 그대로 유지했다.",
         "상단 간판과 시장 표지에 읽을 수 있는 글자를 남겨 글자 금지 조건을 위반한다."
        ],
        "physics": "지시하는 팔은 몸에 연결되어 있고 손가락 자세도 실제로 가능한 동작이다. 앰버의 다리 아래와 발은 잘려 있어 지지면을 직접 확인할 수 없으며, 이것만으로 부유라고 단정할 수는 없다. 다만 성인 군중과의 크기·깊이 관계가 부자연스럽다. 판매대와 상자는 바닥에 놓여 있고 매달린 바나나는 위쪽 고정부에 연결되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.083,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.833,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] physically impossible staging (거대한 스케일과 지지점 없이 부양된 합성 오류)",
     "[gemini-pro] invented people (금지된 이전 샷 인물들의 무단 복제)",
     "[gpt-high] 이전 숏에서 제외하라고 명시한 군중의 얼굴·몸·의상을 다수 그대로 유지했다.",
     "[gpt-high] 상단 간판과 시장 표지에 읽을 수 있는 글자를 남겨 글자 금지 조건을 위반한다."
    ],
    "B": [
     "[gemini-pro] invented people (금지된 이전 샷 인물들의 무단 복제)",
     "[gpt-high] 이전 숏에서 제외하라고 명시한 여러 사람의 얼굴과 의상을 그대로 재사용했다.",
     "[gpt-high] 배경의 노란 세로 간판 등에 식별 가능한 글자가 남아 있어 읽을 수 있는 글자 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 833
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "시선이 목표가 아닌 렌즈를 향하고 있으며, 이전 샷의 인물들을 그대로 복제하는 치명적인 지시 위반이 있습니다.  ★위반: [gemini-pro] invented people (금지된 이전 샷 인물들의 무단 복제) / [gpt-high] 이전 숏에서 제외하라고 명시한 여러 사람의 얼굴과 의상을 그대로 재사용했다. / [gpt-high] 배경의 노란 세로 간판 등에 식별 가능한 글자가 남아 있어 읽을 수 있는 글자 금지 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 833,
    "verdict_ko": "금지된 레퍼런스 인물 복제와 더불어 앰버의 방향 오류, 비정상적인 스케일 등 심각한 물리적 연출 실패가 발생했습니다.  ★위반: [gemini-pro] physically impossible staging (거대한 스케일과 지지점 없이 부양된 합성 오류) / [gemini-pro] invented people (금지된 이전 샷 인물들의 무단 복제) / [gpt-high] 이전 숏에서 제외하라고 명시한 군중의 얼굴·몸·의상을 다수 그대로 유지했다. / [gpt-high] 상단 간판과 시장 표지에 읽을 수 있는 글자를 남겨 글자 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S22sh6_sel.png",
    "asset_id": "85a428eb-5b27-47eb-8873-31605b1ec51e",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd9b-bf7d-787b-9b1e-596b19792df7",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S22sh6"
  }
 },
 "S22sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:27:07.978140+00:00",
  "fingerprint": "231a1c3037613c467cf5e14f9acc99684049c4a1df4fe48a4d8137eba1f32515",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S22sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S22sh8_sel.png",
  "source_sha256": "644143f3ea780131b448b8cb20dbb5192c56d9ed1697e062a387440af29d2851",
  "file": "S22sh8_cine.png",
  "staged_sha256": "5acd2fe8efeb0e7be780f8189868ef966ba35c4c5d5ef2e82de79d3ea2a2144f",
  "latency_ms": 11725
 },
 "S23sh3::signage": {
  "fp": "7390fe2416cb6b1b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::20c2413da7891cf9": {
  "subjects": [],
  "subject_text": "구치소 면회실\n넓고 단출한 면회실. 실내 한가운데 테이블이 놓여 있으며 양쪽에 마주 앉을 수 있는 좌석과 넓은 빈 바닥이 있다.",
  "identity": "canonical",
  "scope_id": "L30",
  "scope_role": "location_interior",
  "scope_sha": "9db460b0e8d7ab74"
 },
 "S23sh3::bgfirst_bg": {
  "input_fingerprint": "519b63495bc98f97",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앞으로 모은 수갑 찬 두 손바닥을 바짝 밀착한 자세의 이현우가 눈물 맺힌 눈으로 윤성찬을 애절하게 올려다보는 시선.\n\nLOCATION (lock): At the central table inside a spacious detention-center visiting room, in ordinary daytime interior light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Visitation table (Positioned between 이현우 and 윤성찬) — A narrow oblique strip of tabletop appears beneath the joined hands; used as Maintains the shared dialogue axis and the physical distance underlying the plea; Handcuffs (Still fastening 이현우's wrists before their removal) — Visible around the wrists beneath his pressed palms; used as A small lower-frame detail makes the plea's constrained physical posture legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate daytime ambient illumination preserves the bruising and moisture in 이현우's eyes with restrained, humane contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앞으로 모은 수갑 찬 두 손바닥을 바짝 밀착한 자세의 이현우가 눈물 맺힌 눈으로 윤성찬을 애절하게 올려다보는 시선.\n\nLOCATION (lock): At the central table inside a spacious detention-center visiting room, in ordinary daytime interior light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Visitation table (Positioned between 이현우 and 윤성찬) — A narrow oblique strip of tabletop appears beneath the joined hands; used as Maintains the shared dialogue axis and the physical distance underlying the plea; Handcuffs (Still fastening 이현우's wrists before their removal) — Visible around the wrists beneath his pressed palms; used as A small lower-frame detail makes the plea's constrained physical posture legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate daytime ambient illumination preserves the bruising and moisture in 이현우's eyes with restrained, humane contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S23sh3__bgfirst_bg.png",
  "asset_id": "198fa0af-9de1-4845-afac-576ce5d2bbb4",
  "input_asset_ids": [
   "fef5f11b-4b93-49e9-b87e-547e83090ee5",
   "63113696-c93a-4d23-a087-83bee2a865c6"
  ]
 },
 "S23sh3": {
  "input_fingerprint": "a87c61de654b7b0c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앞으로 모은 수갑 찬 두 손바닥을 바짝 밀착한 자세의 이현우가 눈물 맺힌 눈으로 윤성찬을 애절하게 올려다보는 시선.\n\nLOCATION (lock): At the central table inside a spacious detention-center visiting room, in ordinary daytime interior light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Visitation table (Positioned between 이현우 and 윤성찬) — A narrow oblique strip of tabletop appears beneath the joined hands; used as Maintains the shared dialogue axis and the physical distance underlying the plea; Handcuffs (Still fastening 이현우's wrists before their removal) — Visible around the wrists beneath his pressed palms; used as A small lower-frame detail makes the plea's constrained physical posture legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate daytime ambient illumination preserves the bruising and moisture in 이현우's eyes with restrained, humane contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A table stands in the middle of the spacious visiting room. 이현우: He is seated at the central table with his hands cuffed. His face is bruised, and the dog-bite injury to his leg remains untreated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앞으로 모은 수갑 찬 두 손바닥을 바짝 밀착한 자세의 이현우가 눈물 맺힌 눈으로 윤성찬을 애절하게 올려다보는 시선.\n\nLOCATION (lock): At the central table inside a spacious detention-center visiting room, in ordinary daytime interior light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Visitation table (Positioned between 이현우 and 윤성찬) — A narrow oblique strip of tabletop appears beneath the joined hands; used as Maintains the shared dialogue axis and the physical distance underlying the plea; Handcuffs (Still fastening 이현우's wrists before their removal) — Visible around the wrists beneath his pressed palms; used as A small lower-frame detail makes the plea's constrained physical posture legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate daytime ambient illumination preserves the bruising and moisture in 이현우's eyes with restrained, humane contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A table stands in the middle of the spacious visiting room. 이현우: He is seated at the central table with his hands cuffed. His face is bruised, and the dog-bite injury to his leg remains untreated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앞으로 모은 수갑 찬 두 손바닥을 바짝 밀착한 자세의 이현우가 눈물 맺힌 눈으로 윤성찬을 애절하게 올려다보는 시선.\n\nLOCATION (lock): At the central table inside a spacious detention-center visiting room, in ordinary daytime interior light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Visitation table (Positioned between 이현우 and 윤성찬) — A narrow oblique strip of tabletop appears beneath the joined hands; used as Maintains the shared dialogue axis and the physical distance underlying the plea; Handcuffs (Still fastening 이현우's wrists before their removal) — Visible around the wrists beneath his pressed palms; used as A small lower-frame detail makes the plea's constrained physical posture legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate daytime ambient illumination preserves the bruising and moisture in 이현우's eyes with restrained, humane contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A table stands in the middle of the spacious visiting room. 이현우: He is seated at the central table with his hands cuffed. His face is bruised, and the dog-bite injury to his leg remains untreated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S23sh3__bgfirst_bg.png",
     "asset_id": "198fa0af-9de1-4845-afac-576ce5d2bbb4",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S23sh3.png",
     "asset_id": "fef5f11b-4b93-49e9-b87e-547e83090ee5",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L30B01.png",
     "asset_id": "63113696-c93a-4d23-a087-83bee2a865c6",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선이 화면 우측 전경에 있는 윤성찬의 어깨와 얼굴 쪽을 향하고 있습니다.",
    "built_space": "회색 하단 벽면과 문이 있는 면회실 내부입니다. 프롬프트대로 이현우와 우측 인물 사이에 테이블이 위치해 있습니다.",
    "entities": "이현우는 레퍼런스의 얼굴, 헤어스타일, 의상과 일치하며 멍든 얼굴과 눈물이 구현되었습니다. 수갑을 차고 두 손바닥을 밀착하고 있습니다.",
    "hard_violations": [],
    "physics": "의자에 앉아 테이블 위로 팔을 올려 두 손을 지탱하고 있는 자연스러운 자세입니다."
   },
   {
    "label": "B",
    "direction": "이현우의 시선이 화면 우측 전경에 있는 윤성찬의 얼굴 부위를 향하고 있습니다.",
    "built_space": "면회실 내부이나, 프롬프트의 지시와 달리 면회 테이블이 이현우의 등 뒤쪽 배경에 멀리 떨어져 배치되어 있습니다.",
    "entities": "이현우의 외모, 의상, 멍든 얼굴과 눈물, 수갑 찬 두 손 등은 레퍼런스와 지시사항에 부합합니다.",
    "hard_violations": [
     "[gemini-pro] 인물 위치 오류: 프롬프트는 두 사람 사이 중앙 테이블에 앉아있다고 명시했으나, 테이블이 인물의 등 뒤 배경에 배치되어 지정된 공간 연출을 위반함.",
     "[gpt-high] 이현우가 탁자 뒤편에 앉아 윤성찬과 탁자를 사이에 두어야 하는 배치와 달리, 탁자가 이현우 뒤에 놓여 두 사람이 탁자 같은 쪽에서 마주한다."
    ],
    "physics": "테이블의 지지 없이 공중에 팔을 들어 올려 손을 모으고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 테이블 앞 좌석 배치와 수갑 찬 손, 애절한 표정 및 시선 처리를 지시사항에 맞게 잘 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "클로즈업 프레이밍은 좋으나, 인물이 테이블에 앉아있지 않고 테이블이 등 뒤에 배치되어 공간 설정을 완전히 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 화면 우측 전경에 있는 윤성찬의 어깨와 얼굴 쪽을 향하고 있습니다.",
        "built_space": "회색 하단 벽면과 문이 있는 면회실 내부입니다. 프롬프트대로 이현우와 우측 인물 사이에 테이블이 위치해 있습니다.",
        "entities": "이현우는 레퍼런스의 얼굴, 헤어스타일, 의상과 일치하며 멍든 얼굴과 눈물이 구현되었습니다. 수갑을 차고 두 손바닥을 밀착하고 있습니다.",
        "hard_violations": [],
        "physics": "의자에 앉아 테이블 위로 팔을 올려 두 손을 지탱하고 있는 자연스러운 자세입니다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 화면 우측 전경에 있는 윤성찬의 얼굴 부위를 향하고 있습니다.",
        "built_space": "면회실 내부이나, 프롬프트의 지시와 달리 면회 테이블이 이현우의 등 뒤쪽 배경에 멀리 떨어져 배치되어 있습니다.",
        "entities": "이현우의 외모, 의상, 멍든 얼굴과 눈물, 수갑 찬 두 손 등은 레퍼런스와 지시사항에 부합합니다.",
        "hard_violations": [
         "인물 위치 오류: 프롬프트는 두 사람 사이 중앙 테이블에 앉아있다고 명시했으나, 테이블이 인물의 등 뒤 배경에 배치되어 지정된 공간 연출을 위반함."
        ],
        "physics": "테이블의 지지 없이 공중에 팔을 들어 올려 손을 모으고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 테이블 앞 좌석 배치와 수갑 찬 손, 애절한 표정 및 시선 처리를 지시사항에 맞게 잘 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "클로즈업 프레이밍은 좋으나, 인물이 테이블에 앉아있지 않고 테이블이 등 뒤에 배치되어 공간 설정을 완전히 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 화면 우측 전경에 있는 윤성찬의 어깨와 얼굴 쪽을 향하고 있습니다.",
        "built_space": "회색 하단 벽면과 문이 있는 면회실 내부입니다. 프롬프트대로 이현우와 우측 인물 사이에 테이블이 위치해 있습니다.",
        "entities": "이현우는 레퍼런스의 얼굴, 헤어스타일, 의상과 일치하며 멍든 얼굴과 눈물이 구현되었습니다. 수갑을 차고 두 손바닥을 밀착하고 있습니다.",
        "hard_violations": [],
        "physics": "의자에 앉아 테이블 위로 팔을 올려 두 손을 지탱하고 있는 자연스러운 자세입니다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 화면 우측 전경에 있는 윤성찬의 얼굴 부위를 향하고 있습니다.",
        "built_space": "면회실 내부이나, 프롬프트의 지시와 달리 면회 테이블이 이현우의 등 뒤쪽 배경에 멀리 떨어져 배치되어 있습니다.",
        "entities": "이현우의 외모, 의상, 멍든 얼굴과 눈물, 수갑 찬 두 손 등은 레퍼런스와 지시사항에 부합합니다.",
        "hard_violations": [
         "인물 위치 오류: 프롬프트는 두 사람 사이 중앙 테이블에 앉아있다고 명시했으나, 테이블이 인물의 등 뒤 배경에 배치되어 지정된 공간 연출을 위반함."
        ],
        "physics": "테이블의 지지 없이 공중에 팔을 들어 올려 손을 모으고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "클로즈업과 눈물 어린 올려다보는 시선은 충실하지만, 탁자가 두 사람 사이가 아니라 이현우 뒤에 놓여 필수 공간 배치를 위반한다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "탁자를 사이에 둔 착석과 간절한 시선·합장·수갑은 맞지만, 넓은 어깨너머 구도로 촬영해 요구한 클로즈업과 좁은 탁자 띠를 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 얼굴과 눈동자를 화면 오른쪽 위로 향해 전경 남성의 얼굴을 올려다본다. 윤성찬을 향한 간청으로 읽히며, 상대 남성도 이현우 쪽으로 고개를 숙인다. 두 손은 얼굴 앞에서 손끝이 위를 향하고 손바닥끼리 밀착되어 있다.",
        "built_space": "목재 탁자 한 개가 이현우의 등 뒤로 넓게 펼쳐지고, 왼쪽 뒤에 검은 의자 등받이 한 개가 보인다. 회색과 흰색의 이중 도장 벽, 어두운 걸레받이는 장소 참조와 부합한다. 그러나 두 사람은 탁자를 사이에 두지 않고 같은 쪽에서 마주하는 배치다. 요구한 손 아래의 좁고 비스듬한 탁자 띠 대신 넓은 상판이 인물 뒤를 차지한다. 문과 창은 잘려 있어 확인할 수 없다.",
        "entities": "헝클어진 검은 머리의 한국계 청년 남성 한 명이 중심에 있고, 오른쪽에는 윤성찬으로 읽히는 남성의 머리와 어깨 일부가 있다. 이현우의 얼굴은 참조보다 다소 성숙하고 길게 보이지만, 낡은 짙은 회색 후드와 체격은 대체로 맞는다. 얼굴의 멍과 찰과상, 젖은 눈과 눈물 자국, 양쪽 손목을 채운 금속 수갑 및 연결 사슬이 보인다. 다리 부상은 적절히 프레임 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이현우가 탁자 뒤편에 앉아 윤성찬과 탁자를 사이에 두어야 하는 배치와 달리, 탁자가 이현우 뒤에 놓여 두 사람이 탁자 같은 쪽에서 마주한다."
        ],
        "physics": "들어 올린 두 손은 전완과 팔꿈치로 이어져 몸이 지지하며, 손바닥을 맞댄 자세는 가능하다. 수갑은 양 손목을 둘러싸고 짧은 사슬은 아래로 처진다. 좌석과 하체는 보이지 않아 착석 지지점을 직접 확인할 수 없지만, 몸이 공중에 떠 있다는 증거는 없다. 탁자는 보이는 검은 다리로 지지된다."
       },
       {
        "label": "B",
        "direction": "이현우의 눈과 들린 턱은 화면 오른쪽 위의 전경 남성 얼굴을 향한다. 윤성찬을 애절하게 올려다보는 방향이 성립하며, 윤성찬도 이현우를 향해 몸을 돌리고 있다. 두 손바닥은 가슴 앞에서 밀착되고 손끝은 위를 향한다.",
        "built_space": "이현우는 검은 등받이와 금속 테두리가 있는 의자 한 개에 앉아 있고, 목재 탁자 한 개가 그와 전경 남성 사이에 놓인다. 왼쪽의 좁은 유리창 달린 회색 문 한 개와 도어클로저, 벽 스위치 한 개, 뒤쪽 가로 게시판 한 개, 오른쪽 격자창 일부가 보이며 참조 장소와 잘 맞는다. 다만 방과 몸통을 많이 드러낸 어깨너머 구도여서 클로즈업이 아니고, 하단 상판도 좁은 띠보다 넓다.",
        "entities": "이현우는 참조와 유사한 젊은 한국계 남성 얼굴, 헝클어진 검은 머리, 마른 체격과 해진 회색 후드를 갖췄다. 얼굴에 멍과 붉어진 젖은 눈이 보이지만 눈물의 가시성은 A보다 약하다. 양 손목의 금속 수갑과 연결 사슬이 손 아래에 보인다. 오른쪽 전경에는 윤성찬으로 읽히는 검은 옷의 남성 뒷머리와 상체가 있다. 다리는 프레임 밖이므로 부상 여부는 평가할 수 없다. 게시판에 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우의 몸은 뒤에 보이는 의자와 자연스러운 착석 관계를 이루고, 양 전완과 팔꿈치는 탁자 가장자리에 기대어 모은 손을 지지한다. 수갑은 손목에 걸려 있으며 사슬은 중력 방향으로 처진다. 전경 남성의 하체는 잘려 있으나 부유하거나 불가능한 자세로 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "클로즈업과 눈물 어린 올려다보는 시선은 충실하지만, 탁자가 두 사람 사이가 아니라 이현우 뒤에 놓여 필수 공간 배치를 위반한다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "탁자를 사이에 둔 착석과 간절한 시선·합장·수갑은 맞지만, 넓은 어깨너머 구도로 촬영해 요구한 클로즈업과 좁은 탁자 띠를 놓쳤다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 얼굴과 눈동자를 화면 오른쪽 위로 향해 전경 남성의 얼굴을 올려다본다. 윤성찬을 향한 간청으로 읽히며, 상대 남성도 이현우 쪽으로 고개를 숙인다. 두 손은 얼굴 앞에서 손끝이 위를 향하고 손바닥끼리 밀착되어 있다.",
        "built_space": "목재 탁자 한 개가 이현우의 등 뒤로 넓게 펼쳐지고, 왼쪽 뒤에 검은 의자 등받이 한 개가 보인다. 회색과 흰색의 이중 도장 벽, 어두운 걸레받이는 장소 참조와 부합한다. 그러나 두 사람은 탁자를 사이에 두지 않고 같은 쪽에서 마주하는 배치다. 요구한 손 아래의 좁고 비스듬한 탁자 띠 대신 넓은 상판이 인물 뒤를 차지한다. 문과 창은 잘려 있어 확인할 수 없다.",
        "entities": "헝클어진 검은 머리의 한국계 청년 남성 한 명이 중심에 있고, 오른쪽에는 윤성찬으로 읽히는 남성의 머리와 어깨 일부가 있다. 이현우의 얼굴은 참조보다 다소 성숙하고 길게 보이지만, 낡은 짙은 회색 후드와 체격은 대체로 맞는다. 얼굴의 멍과 찰과상, 젖은 눈과 눈물 자국, 양쪽 손목을 채운 금속 수갑 및 연결 사슬이 보인다. 다리 부상은 적절히 프레임 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이현우가 탁자 뒤편에 앉아 윤성찬과 탁자를 사이에 두어야 하는 배치와 달리, 탁자가 이현우 뒤에 놓여 두 사람이 탁자 같은 쪽에서 마주한다."
        ],
        "physics": "들어 올린 두 손은 전완과 팔꿈치로 이어져 몸이 지지하며, 손바닥을 맞댄 자세는 가능하다. 수갑은 양 손목을 둘러싸고 짧은 사슬은 아래로 처진다. 좌석과 하체는 보이지 않아 착석 지지점을 직접 확인할 수 없지만, 몸이 공중에 떠 있다는 증거는 없다. 탁자는 보이는 검은 다리로 지지된다."
       },
       {
        "label": "A",
        "direction": "이현우의 눈과 들린 턱은 화면 오른쪽 위의 전경 남성 얼굴을 향한다. 윤성찬을 애절하게 올려다보는 방향이 성립하며, 윤성찬도 이현우를 향해 몸을 돌리고 있다. 두 손바닥은 가슴 앞에서 밀착되고 손끝은 위를 향한다.",
        "built_space": "이현우는 검은 등받이와 금속 테두리가 있는 의자 한 개에 앉아 있고, 목재 탁자 한 개가 그와 전경 남성 사이에 놓인다. 왼쪽의 좁은 유리창 달린 회색 문 한 개와 도어클로저, 벽 스위치 한 개, 뒤쪽 가로 게시판 한 개, 오른쪽 격자창 일부가 보이며 참조 장소와 잘 맞는다. 다만 방과 몸통을 많이 드러낸 어깨너머 구도여서 클로즈업이 아니고, 하단 상판도 좁은 띠보다 넓다.",
        "entities": "이현우는 참조와 유사한 젊은 한국계 남성 얼굴, 헝클어진 검은 머리, 마른 체격과 해진 회색 후드를 갖췄다. 얼굴에 멍과 붉어진 젖은 눈이 보이지만 눈물의 가시성은 A보다 약하다. 양 손목의 금속 수갑과 연결 사슬이 손 아래에 보인다. 오른쪽 전경에는 윤성찬으로 읽히는 검은 옷의 남성 뒷머리와 상체가 있다. 다리는 프레임 밖이므로 부상 여부는 평가할 수 없다. 게시판에 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우의 몸은 뒤에 보이는 의자와 자연스러운 착석 관계를 이루고, 양 전완과 팔꿈치는 탁자 가장자리에 기대어 모은 손을 지지한다. 수갑은 손목에 걸려 있으며 사슬은 중력 방향으로 처진다. 전경 남성의 하체는 잘려 있으나 부유하거나 불가능한 자세로 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.929
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.679
   },
   "violations": {
    "B": [
     "[gemini-pro] 인물 위치 오류: 프롬프트는 두 사람 사이 중앙 테이블에 앉아있다고 명시했으나, 테이블이 인물의 등 뒤 배경에 배치되어 지정된 공간 연출을 위반함.",
     "[gpt-high] 이현우가 탁자 뒤편에 앉아 윤성찬과 탁자를 사이에 두어야 하는 배치와 달리, 탁자가 이현우 뒤에 놓여 두 사람이 탁자 같은 쪽에서 마주한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 679
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 테이블 앞 좌석 배치와 수갑 찬 손, 애절한 표정 및 시선 처리를 지시사항에 맞게 잘 구현했습니다."
   },
   {
    "label": "B",
    "score": 679,
    "verdict_ko": "클로즈업 프레이밍은 좋으나, 인물이 테이블에 앉아있지 않고 테이블이 등 뒤에 배치되어 공간 설정을 완전히 위반했습니다.  ★위반: [gemini-pro] 인물 위치 오류: 프롬프트는 두 사람 사이 중앙 테이블에 앉아있다고 명시했으나, 테이블이 인물의 등 뒤 배경에 배치되어 지정된 공간 연출을 위반함. / [gpt-high] 이현우가 탁자 뒤편에 앉아 윤성찬과 탁자를 사이에 두어야 하는 배치와 달리, 탁자가 이현우 뒤에 놓여 두 사람이 탁자 같은 쪽에서 마주한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L30B01.png",
    "asset_id": "63113696-c93a-4d23-a087-83bee2a865c6",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd9c-8088-70a5-b1e3-4d40c9f20994",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S23sh3__bgfirst_bg.png",
   "bg_asset_id": "198fa0af-9de1-4845-afac-576ce5d2bbb4",
   "bg_record_key": "S23sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S23sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:27:20.117153+00:00",
  "fingerprint": "404c6c3d640996b318dc6c6508bf318273852583cc7f08675b4ee5b1982a7177",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S23sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S23sh3_sel.png",
  "source_sha256": "ee457ce6a2d397da2b93b80848de8360d107c9a9bcf89d5765a19f5a6f913062",
  "file": "S23sh3_cine.png",
  "staged_sha256": "2b57f491f2c96696260ea6c82a36e5236beadf2edd5b00cbd6477669bbd51686",
  "latency_ms": 13345
 },
 "S23sh7::signage": {
  "fp": "3eef51ecf902d74c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S23sh7": {
  "input_fingerprint": "742f477305a1c96d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬이 탁자 위 이현우 앞으로 고릴라 로봇 찰리의 사진을 손가락으로 누른 채 밀고 있는 mid-action의 손끝.\n\nLOCATION (lock): Across the tabletop in the center of the spacious detention-center visiting room, under even interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리 wanted photograph (Being slid toward 이현우 beneath 윤성찬's fingertip) — The printed face is visible obliquely to the camera and shows the gorilla robot 찰리; used as The focal object links the controlled hand gesture to the subject of the negotiation without replacing the live scene with the photograph's viewpoint; Visitation table (Supporting the sliding photograph) — Its upper surface and near edge are seen diagonally; used as Provides a continuous movement plane between the two cropped human figures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the visitation room's ambient continuity with sufficient tonal separation to read the fingertip and printed image without introducing glare effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The visiting-room table remains central, and the displayed wanted photograph depicts Charlie. 윤성찬: He wears his suit and holds the wanted photograph on display.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 윤성찬 right now, so 윤성찬's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 윤성찬: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬이 탁자 위 이현우 앞으로 고릴라 로봇 찰리의 사진을 손가락으로 누른 채 밀고 있는 mid-action의 손끝.\n\nLOCATION (lock): Across the tabletop in the center of the spacious detention-center visiting room, under even interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리 wanted photograph (Being slid toward 이현우 beneath 윤성찬's fingertip) — The printed face is visible obliquely to the camera and shows the gorilla robot 찰리; used as The focal object links the controlled hand gesture to the subject of the negotiation without replacing the live scene with the photograph's viewpoint; Visitation table (Supporting the sliding photograph) — Its upper surface and near edge are seen diagonally; used as Provides a continuous movement plane between the two cropped human figures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the visitation room's ambient continuity with sufficient tonal separation to read the fingertip and printed image without introducing glare effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The visiting-room table remains central, and the displayed wanted photograph depicts Charlie. 윤성찬: He wears his suit and holds the wanted photograph on display.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 윤성찬 right now, so 윤성찬's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 윤성찬: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬이 탁자 위 이현우 앞으로 고릴라 로봇 찰리의 사진을 손가락으로 누른 채 밀고 있는 mid-action의 손끝.\n\nLOCATION (lock): Across the tabletop in the center of the spacious detention-center visiting room, under even interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리 wanted photograph (Being slid toward 이현우 beneath 윤성찬's fingertip) — The printed face is visible obliquely to the camera and shows the gorilla robot 찰리; used as The focal object links the controlled hand gesture to the subject of the negotiation without replacing the live scene with the photograph's viewpoint; Visitation table (Supporting the sliding photograph) — Its upper surface and near edge are seen diagonally; used as Provides a continuous movement plane between the two cropped human figures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the visitation room's ambient continuity with sufficient tonal separation to read the fingertip and printed image without introducing glare effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The visiting-room table remains central, and the displayed wanted photograph depicts Charlie. 윤성찬: He wears his suit and holds the wanted photograph on display.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 윤성찬 right now, so 윤성찬's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 윤성찬: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "윤성찬의 검지손가락이 탁자 위의 고릴라 로봇 사진을 누르며 앞으로 밀어내고 있습니다.",
    "built_space": "나무 탁자 위가 얕은 심도의 클로즈업으로 촬영되었으며, 배경은 면회실 내부로 보이나 심하게 흐려져 있습니다.",
    "entities": "주름진 노인의 손과 남색 양복 소매, 흰 셔츠(윤성찬의 레퍼런스와 일치), 찰리 고릴라 로봇이 인쇄된 사진이 모두 정확하게 묘사되었습니다.",
    "hard_violations": [],
    "physics": "손가락 끝이 탁자 면에 놓인 사진을 단단히 누르며 현실적인 마찰과 무게감을 보여줍니다."
   },
   {
    "label": "B",
    "direction": "윤성찬의 검지손가락이 탁자 위의 고릴라 로봇 사진을 누르고 있습니다.",
    "built_space": "나무 탁자와 함께 면회실의 문과 벽면 구조가 보이나, 배경에 좌석과 인물들이 배치되어 있습니다.",
    "entities": "주름진 노인의 손과 양복 소매(윤성찬 일치), 로봇 사진이 존재하지만, 배경에 지시문에 없는 3명의 신원 미상 인물들이 등장합니다.",
    "hard_violations": [
     "[gemini-pro] invented people (배경에 지시문에 없는 인물 3명 추가)",
     "[gpt-high] 허용되지 않은 배경 인물 세 명을 추가했다."
    ],
    "physics": "손가락이 탁자 위 사진에 닿아 있으며 자연스럽게 지탱되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시된 클로즈업 프레이밍을 정확히 따랐으며, 노인의 손과 고릴라 로봇 사진의 상호작용을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트에서 엄격하게 금지한 '지시문에 없는 인물 추가' 규칙을 위반하여 배경에 세 명의 인물을 임의로 생성했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬의 검지손가락이 탁자 위의 고릴라 로봇 사진을 누르며 앞으로 밀어내고 있습니다.",
        "built_space": "나무 탁자 위가 얕은 심도의 클로즈업으로 촬영되었으며, 배경은 면회실 내부로 보이나 심하게 흐려져 있습니다.",
        "entities": "주름진 노인의 손과 남색 양복 소매, 흰 셔츠(윤성찬의 레퍼런스와 일치), 찰리 고릴라 로봇이 인쇄된 사진이 모두 정확하게 묘사되었습니다.",
        "hard_violations": [],
        "physics": "손가락 끝이 탁자 면에 놓인 사진을 단단히 누르며 현실적인 마찰과 무게감을 보여줍니다."
       },
       {
        "label": "B",
        "direction": "윤성찬의 검지손가락이 탁자 위의 고릴라 로봇 사진을 누르고 있습니다.",
        "built_space": "나무 탁자와 함께 면회실의 문과 벽면 구조가 보이나, 배경에 좌석과 인물들이 배치되어 있습니다.",
        "entities": "주름진 노인의 손과 양복 소매(윤성찬 일치), 로봇 사진이 존재하지만, 배경에 지시문에 없는 3명의 신원 미상 인물들이 등장합니다.",
        "hard_violations": [
         "invented people (배경에 지시문에 없는 인물 3명 추가)"
        ],
        "physics": "손가락이 탁자 위 사진에 닿아 있으며 자연스럽게 지탱되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시된 클로즈업 프레이밍을 정확히 따랐으며, 노인의 손과 고릴라 로봇 사진의 상호작용을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트에서 엄격하게 금지한 '지시문에 없는 인물 추가' 규칙을 위반하여 배경에 세 명의 인물을 임의로 생성했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬의 검지손가락이 탁자 위의 고릴라 로봇 사진을 누르며 앞으로 밀어내고 있습니다.",
        "built_space": "나무 탁자 위가 얕은 심도의 클로즈업으로 촬영되었으며, 배경은 면회실 내부로 보이나 심하게 흐려져 있습니다.",
        "entities": "주름진 노인의 손과 남색 양복 소매, 흰 셔츠(윤성찬의 레퍼런스와 일치), 찰리 고릴라 로봇이 인쇄된 사진이 모두 정확하게 묘사되었습니다.",
        "hard_violations": [],
        "physics": "손가락 끝이 탁자 면에 놓인 사진을 단단히 누르며 현실적인 마찰과 무게감을 보여줍니다."
       },
       {
        "label": "B",
        "direction": "윤성찬의 검지손가락이 탁자 위의 고릴라 로봇 사진을 누르고 있습니다.",
        "built_space": "나무 탁자와 함께 면회실의 문과 벽면 구조가 보이나, 배경에 좌석과 인물들이 배치되어 있습니다.",
        "entities": "주름진 노인의 손과 양복 소매(윤성찬 일치), 로봇 사진이 존재하지만, 배경에 지시문에 없는 3명의 신원 미상 인물들이 등장합니다.",
        "hard_violations": [
         "invented people (배경에 지시문에 없는 인물 3명 추가)"
        ],
        "physics": "손가락이 탁자 위 사진에 닿아 있으며 자연스럽게 지탱되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "손끝으로 사진을 누르는 동작과 대각선 탁자 모서리는 맞지만, 배경에 허용되지 않은 인물 세 명을 추가해 실격이다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "고령 남성의 정장 소매와 사진을 누르며 내미는 손끝을 정확히 클로즈업했지만, 받는 쪽에 대한 사진의 상하 방향과 누락된 가까운 탁자 모서리는 아쉽다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽에서 들어온 손의 검지가 사진 중앙을 누르며 왼쪽으로 뻗는다. 사진을 탁자 건너편보다 화면 왼쪽으로 미는 동작에 가깝고, 그 방향에 이현우는 보이지 않는다. 맞은편 중앙에는 빈 의자가 있다. 배경 인물들의 시선은 흐려 확인하기 어렵다.",
        "built_space": "목재 탁자 하나의 상판과 가까운 모서리가 대각선으로 보인다. 맞은편 빈 의자 하나와 오른쪽 아래 의자 일부가 있다. 왼쪽 회색 문 하나, 중앙 벽판 하나, 오른쪽 창 일부, 투톤 벽은 장소 참고와 대체로 맞는다. 그러나 탁자 너머 벽 쪽에 세 사람을 새로 배치했다.",
        "entities": "주름과 혈관이 드러난 고령 남성의 손, 짙은 정장 소매와 흰 셔츠 커프스는 윤성찬의 설정에 맞는다. 얼굴은 프레임 밖이다. 흰 테두리 사진에는 금속 고릴라 로봇이 인쇄되어 있어 찰리 소품으로 식별된다. 배경에는 젊은 동아시아계 남성처럼 보이는 인물 하나와 외형을 정확히 판별하기 어려운 성인 두 명이 추가되어 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "허용되지 않은 배경 인물 세 명을 추가했다."
        ],
        "physics": "사진은 탁자에 놓여 있고 검지 끝이 인쇄면에 닿는다. 약간 들린 종이 가장자리는 종이의 휨으로 설명되며 중앙은 상판이 지지한다. 손목과 팔은 화면 오른쪽 정장 소매로 이어진다. 누른 채 미는 자세는 물리적으로 가능하고, 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "화면 위쪽 윤성찬의 몸에서 손이 아래쪽으로 뻗어 사진 중앙을 누른다. 카메라 쪽의 화면 밖 상대에게 사진을 내미는 방향으로 읽힌다. 다만 인쇄된 로봇의 머리가 화면 아래쪽에 있어, 그쪽이 받는 이현우의 자리라면 사진은 그에게 거꾸로 놓인다. 사람의 눈과 시선은 보이지 않는다.",
        "built_space": "목재 탁자 하나가 화면 대부분을 채우고 뒤에는 회색 벽과 정장 몸통 일부만 보인다. 손끝과 사진에 집중한 클로즈업은 적절하다. 상판의 목재 재질은 참고와 이어지지만 긁힘이 더 두드러지고, 지시된 가까운 탁자 모서리는 프레임에 들어오지 않았다. 문·창·의자는 크롭 밖이므로 개수나 배치를 판단할 수 없다.",
        "entities": "고령 남성다운 손등의 주름, 혈관과 피부 질감, 짙은 정장 및 흰 커프스가 윤성찬의 나이와 복장에 부합한다. 얼굴과 머리는 보이지 않아 동일 인물 여부를 더 확인할 수 없다. 흰 테두리의 실물 사진에는 참고와 유사한 은회색 갑옷과 푸른 부품의 고릴라 로봇 찰리가 보인다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "사진은 상판에 평평하게 지지되고 검지 끝은 사진의 로봇 가슴 부분을 누른다. 나머지 손가락은 자연스럽게 굽혀져 있으며 손목은 소매와 팔로 연결된다. 팔을 앞으로 뻗어 종이를 미는 동작으로 가능한 자세다. 사진이나 신체가 근거 없이 떠 있지 않다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "손끝으로 사진을 누르는 동작과 대각선 탁자 모서리는 맞지만, 배경에 허용되지 않은 인물 세 명을 추가해 실격이다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "고령 남성의 정장 소매와 사진을 누르며 내미는 손끝을 정확히 클로즈업했지만, 받는 쪽에 대한 사진의 상하 방향과 누락된 가까운 탁자 모서리는 아쉽다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽에서 들어온 손의 검지가 사진 중앙을 누르며 왼쪽으로 뻗는다. 사진을 탁자 건너편보다 화면 왼쪽으로 미는 동작에 가깝고, 그 방향에 이현우는 보이지 않는다. 맞은편 중앙에는 빈 의자가 있다. 배경 인물들의 시선은 흐려 확인하기 어렵다.",
        "built_space": "목재 탁자 하나의 상판과 가까운 모서리가 대각선으로 보인다. 맞은편 빈 의자 하나와 오른쪽 아래 의자 일부가 있다. 왼쪽 회색 문 하나, 중앙 벽판 하나, 오른쪽 창 일부, 투톤 벽은 장소 참고와 대체로 맞는다. 그러나 탁자 너머 벽 쪽에 세 사람을 새로 배치했다.",
        "entities": "주름과 혈관이 드러난 고령 남성의 손, 짙은 정장 소매와 흰 셔츠 커프스는 윤성찬의 설정에 맞는다. 얼굴은 프레임 밖이다. 흰 테두리 사진에는 금속 고릴라 로봇이 인쇄되어 있어 찰리 소품으로 식별된다. 배경에는 젊은 동아시아계 남성처럼 보이는 인물 하나와 외형을 정확히 판별하기 어려운 성인 두 명이 추가되어 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "허용되지 않은 배경 인물 세 명을 추가했다."
        ],
        "physics": "사진은 탁자에 놓여 있고 검지 끝이 인쇄면에 닿는다. 약간 들린 종이 가장자리는 종이의 휨으로 설명되며 중앙은 상판이 지지한다. 손목과 팔은 화면 오른쪽 정장 소매로 이어진다. 누른 채 미는 자세는 물리적으로 가능하고, 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "화면 위쪽 윤성찬의 몸에서 손이 아래쪽으로 뻗어 사진 중앙을 누른다. 카메라 쪽의 화면 밖 상대에게 사진을 내미는 방향으로 읽힌다. 다만 인쇄된 로봇의 머리가 화면 아래쪽에 있어, 그쪽이 받는 이현우의 자리라면 사진은 그에게 거꾸로 놓인다. 사람의 눈과 시선은 보이지 않는다.",
        "built_space": "목재 탁자 하나가 화면 대부분을 채우고 뒤에는 회색 벽과 정장 몸통 일부만 보인다. 손끝과 사진에 집중한 클로즈업은 적절하다. 상판의 목재 재질은 참고와 이어지지만 긁힘이 더 두드러지고, 지시된 가까운 탁자 모서리는 프레임에 들어오지 않았다. 문·창·의자는 크롭 밖이므로 개수나 배치를 판단할 수 없다.",
        "entities": "고령 남성다운 손등의 주름, 혈관과 피부 질감, 짙은 정장 및 흰 커프스가 윤성찬의 나이와 복장에 부합한다. 얼굴과 머리는 보이지 않아 동일 인물 여부를 더 확인할 수 없다. 흰 테두리의 실물 사진에는 참고와 유사한 은회색 갑옷과 푸른 부품의 고릴라 로봇 찰리가 보인다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "사진은 상판에 평평하게 지지되고 검지 끝은 사진의 로봇 가슴 부분을 누른다. 나머지 손가락은 자연스럽게 굽혀져 있으며 손목은 소매와 팔로 연결된다. 팔을 앞으로 뻗어 종이를 미는 동작으로 가능한 자세다. 사진이나 신체가 근거 없이 떠 있지 않다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.472
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.222
   },
   "violations": {
    "B": [
     "[gemini-pro] invented people (배경에 지시문에 없는 인물 3명 추가)",
     "[gpt-high] 허용되지 않은 배경 인물 세 명을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 222
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 클로즈업 프레이밍을 정확히 따랐으며, 노인의 손과 고릴라 로봇 사진의 상호작용을 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 222,
    "verdict_ko": "프롬프트에서 엄격하게 금지한 '지시문에 없는 인물 추가' 규칙을 위반하여 배경에 세 명의 인물을 임의로 생성했습니다.  ★위반: [gemini-pro] invented people (배경에 지시문에 없는 인물 3명 추가) / [gpt-high] 허용되지 않은 배경 인물 세 명을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S23sh3_sel.png",
    "asset_id": "72d33a24-767e-49ca-a219-877c93579838",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:781789>",
    "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 찰리의 사진: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1187837>",
    "asset_id": "02e210a7-8018-48af-b07d-26cb94d64100",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd9d-5c7d-7e2e-95fa-c6ed1d91fffa",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S23sh3"
  }
 },
 "S23sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:27:33.791949+00:00",
  "fingerprint": "3594d677119435f84c89d1da3c1962528efa6fa29d74f38217b97e76d7cb0e9b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S23sh7_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S23sh7_sel.png",
  "source_sha256": "406228472fa20d2e2ef9e511177c73bd1c33ffd0f527673c03addf4343392edd",
  "file": "S23sh7_cine.png",
  "staged_sha256": "89a36e80b41d8dea210a935e0ea81bb3ab9fb16fad5177da671b2da97360ed23",
  "latency_ms": 11458
 },
 "S23sh11::signage": {
  "fp": "9e1891824d3189d7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S23sh11": {
  "input_fingerprint": "48347c17c0223482",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 상체를 탁자 쪽으로 쑥 내민 채 두 눈을 반짝이고 절박한 표정으로 앞을 바라보는 자세.\n\nLOCATION (lock): At the visitor-facing side of the central table inside the spacious detention-center visiting room, in daytime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Visitation table edge (Remaining between the negotiating men) — A thin diagonal edge is visible below 이현우's advancing torso; used as Measures his forward intrusion into the space without competing with the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established ambient illumination steady, retaining readable bruising and eye detail as hope sharpens his expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The photograph of Charlie remains available at the central visiting-room table; it has not been put away. 이현우: He remains at the table without handcuffs, with his facial bruises and dog-bitten leg unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 상체를 탁자 쪽으로 쑥 내민 채 두 눈을 반짝이고 절박한 표정으로 앞을 바라보는 자세.\n\nLOCATION (lock): At the visitor-facing side of the central table inside the spacious detention-center visiting room, in daytime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Visitation table edge (Remaining between the negotiating men) — A thin diagonal edge is visible below 이현우's advancing torso; used as Measures his forward intrusion into the space without competing with the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established ambient illumination steady, retaining readable bruising and eye detail as hope sharpens his expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The photograph of Charlie remains available at the central visiting-room table; it has not been put away. 이현우: He remains at the table without handcuffs, with his facial bruises and dog-bitten leg unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 상체를 탁자 쪽으로 쑥 내민 채 두 눈을 반짝이고 절박한 표정으로 앞을 바라보는 자세.\n\nLOCATION (lock): At the visitor-facing side of the central table inside the spacious detention-center visiting room, in daytime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Visitation table edge (Remaining between the negotiating men) — A thin diagonal edge is visible below 이현우's advancing torso; used as Measures his forward intrusion into the space without competing with the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established ambient illumination steady, retaining readable bruising and eye detail as hope sharpens his expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The photograph of Charlie remains available at the central visiting-room table; it has not been put away. 이현우: He remains at the table without handcuffs, with his facial bruises and dog-bitten leg unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선은 카메라 앞쪽 약간 오른쪽을 향하고 있음.",
    "built_space": "면회실 내부로 배경에 문과 창살이 있는 창문이 위치하며, 인물 앞에 가로 방향의 탁자 모서리가 있음.",
    "entities": "이현우의 얼굴, 상처, 낡은 후드티가 레퍼런스와 정확히 일치함. 탁자 위 사진은 프레임 밖으로 제외됨.",
    "hard_violations": [],
    "physics": "양손을 탁자 위에 짚어 앞으로 내민 상체의 무게를 자연스럽게 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "이현우의 시선은 화면 정면을 향하고 있음.",
    "built_space": "면회실 내부의 문과 창문이 배경에 있으며, 앞쪽 탁자 위에 잘린 인물 사진이 놓여 있음.",
    "entities": "이현우의 얼굴과 의상이 레퍼런스와 일치하며, 찰리로 추정되는 사진 일부가 탁자 위에 있음.",
    "hard_violations": [
     "[gemini-pro] 지지하는 손이나 물체 없이 허공에 비스듬히 떠 있는 흰 종이 객체",
     "[gpt-high] 탁자 위에 참조나 지시에서 요구하지 않은 길고 비스듬한 판 형태의 물체가 추가되어, 단일 탁자의 얇은 가장자리로 설명되지 않는 구조가 생겼다."
    ],
    "physics": "양손으로 탁자를 짚고 있으나, 화면 앞쪽의 긴 흰색 종이가 물리적으로 불가능하게 허공에 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 절박한 표정과 상체를 내민 자세를 잘 구현했으며 물리적 오류가 없어 가장 우수함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "화면 앞쪽에 아무런 지지 없이 허공에 떠 있는 종이가 존재하여 물리적 규칙을 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 카메라 앞쪽 약간 오른쪽을 향하고 있음.",
        "built_space": "면회실 내부로 배경에 문과 창살이 있는 창문이 위치하며, 인물 앞에 가로 방향의 탁자 모서리가 있음.",
        "entities": "이현우의 얼굴, 상처, 낡은 후드티가 레퍼런스와 정확히 일치함. 탁자 위 사진은 프레임 밖으로 제외됨.",
        "hard_violations": [],
        "physics": "양손을 탁자 위에 짚어 앞으로 내민 상체의 무게를 자연스럽게 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 화면 정면을 향하고 있음.",
        "built_space": "면회실 내부의 문과 창문이 배경에 있으며, 앞쪽 탁자 위에 잘린 인물 사진이 놓여 있음.",
        "entities": "이현우의 얼굴과 의상이 레퍼런스와 일치하며, 찰리로 추정되는 사진 일부가 탁자 위에 있음.",
        "hard_violations": [
         "지지하는 손이나 물체 없이 허공에 비스듬히 떠 있는 흰 종이 객체"
        ],
        "physics": "양손으로 탁자를 짚고 있으나, 화면 앞쪽의 긴 흰색 종이가 물리적으로 불가능하게 허공에 떠 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 절박한 표정과 상체를 내민 자세를 잘 구현했으며 물리적 오류가 없어 가장 우수함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "화면 앞쪽에 아무런 지지 없이 허공에 떠 있는 종이가 존재하여 물리적 규칙을 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 카메라 앞쪽 약간 오른쪽을 향하고 있음.",
        "built_space": "면회실 내부로 배경에 문과 창살이 있는 창문이 위치하며, 인물 앞에 가로 방향의 탁자 모서리가 있음.",
        "entities": "이현우의 얼굴, 상처, 낡은 후드티가 레퍼런스와 정확히 일치함. 탁자 위 사진은 프레임 밖으로 제외됨.",
        "hard_violations": [],
        "physics": "양손을 탁자 위에 짚어 앞으로 내민 상체의 무게를 자연스럽게 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 화면 정면을 향하고 있음.",
        "built_space": "면회실 내부의 문과 창문이 배경에 있으며, 앞쪽 탁자 위에 잘린 인물 사진이 놓여 있음.",
        "entities": "이현우의 얼굴과 의상이 레퍼런스와 일치하며, 찰리로 추정되는 사진 일부가 탁자 위에 있음.",
        "hard_violations": [
         "지지하는 손이나 물체 없이 허공에 비스듬히 떠 있는 흰 종이 객체"
        ],
        "physics": "양손으로 탁자를 짚고 있으나, 화면 앞쪽의 긴 흰색 종이가 물리적으로 불가능하게 허공에 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "앞으로 내민 상체와 절박한 눈빛은 맞지만, 지나치게 넓은 구도와 탁자 위에 추가된 정체불명의 비스듬한 판이 지정된 클로즈업과 공간의 연속성을 해친다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "탁자에 몸을 깊이 내밀고 앞의 상대를 바라보는 행동, 얼굴의 멍과 의상, 대각선 탁자 가장자리가 더 충실하지만 손과 탁자 면적이 커서 클로즈업의 밀도는 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈은 카메라 바로 옆의 정면 상대 위치를 향하고, 상체도 탁자 너머 그 방향으로 전진한다. 벌어진 입과 크게 뜬 정상적인 눈에 긴박함이 보인다. 손가락은 탁자 위에 펼쳐져 있으며 겨누는 도구는 없다.",
        "built_space": "뒤쪽에 세로 유리창이 있는 문 하나, 오른쪽에 격자창 하나, 왼쪽에 실내 유리 개구부 하나, 천장 조명 두 개와 의자 등받이 일부가 보인다. 회색 하단 벽과 밝은 상단 벽, 목재 탁자는 참조와 유사하지만 참조에서 확인되지 않은 공간까지 넓게 드러낸다. 탁자는 화면 아래를 넓게 차지하고, 요구된 얇은 가장자리 대신 별도의 긴 비스듬한 판이 탁자 위를 가른다.",
        "entities": "인물은 젊은 한국계 남성으로 보이는 이현우 한 명뿐이며, 헝클어진 검은 머리와 해지고 더러운 회색 후드가 참조와 대체로 맞는다. 볼과 코 주변의 멍도 보이나 참조와 상처 분포가 완전히 같지는 않다. 양 손목에 수갑은 없다. 아래쪽 사진 일부에는 얼굴이 잘려 보여 찰리의 사진인지 확인할 수 없다. 다리 부상은 프레임 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "탁자 위에 참조나 지시에서 요구하지 않은 길고 비스듬한 판 형태의 물체가 추가되어, 단일 탁자의 얇은 가장자리로 설명되지 않는 구조가 생겼다."
        ],
        "physics": "양손이 탁자에 밀착되고 팔꿈치가 굽혀져 있어 상체를 앞으로 미는 자세는 지지 가능하다. 하체는 가려져 있지만 인물이 공중에 떠 있다고 볼 근거는 없다. 추가된 판은 아래쪽 일부가 탁자 면에 닿아 보이지만, 위로 솟은 끝을 고정하거나 받치는 구조는 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "얼굴과 두 눈은 화면 오른쪽 앞의 프레임 밖 상대 위치를 향한다. 고개를 낮추고 상체를 탁자 쪽으로 깊이 내민 방향도 그 시선과 일치한다. 눈에는 자연스러운 반사광이 있고 입과 눈썹의 긴장으로 절박함을 표현하지만, 희망이 밝아지는 느낌은 비교적 절제되어 있다.",
        "built_space": "오른쪽 격자창 하나, 뒤쪽 벽판 일부 하나, 왼쪽 벽 스위치 하나와 의자 등받이 일부가 보인다. 두 색으로 나뉜 벽과 목재 탁자, 낮의 실내광이 참조 장소와 잘 이어진다. 이현우는 의자 앞에서 탁자에 몸을 기울이고 있으며, 몸 아래로 검은 테두리의 탁자 가장자리가 대각선으로 지나간다. 다만 넓은 탁자 면과 양손이 강조되어 지정된 얼굴 중심 클로즈업보다 여유가 많다. 불가능한 반사나 중복된 설비는 보이지 않는다.",
        "entities": "이현우 한 명만 보이며, 젊은 한국계 남성의 얼굴, 검은 헝클어진 머리, 마른 체형과 낡은 회색 후드가 참조에 가깝다. 볼과 콧등의 멍이 읽히며 눈은 정상적인 홍채와 동공을 유지한다. 드러난 손목에 수갑은 없다. 찰리의 사진은 보이지 않지만 탁자 전체가 나온 것은 아니므로 치워졌다고 단정할 수 없다. 다리 부상은 프레임 밖이고, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양 손바닥과 손가락이 탁자 표면에 닿고 팔이 몸의 전진을 받친다. 뒤에 의자 등받이가 보이며 하체는 가려져 있다. 낮게 숙인 몸은 팔과 화면 밖 하체로 지지할 수 있는 자세이고, 떠 있는 물체나 지지 없는 신체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "앞으로 내민 상체와 절박한 눈빛은 맞지만, 지나치게 넓은 구도와 탁자 위에 추가된 정체불명의 비스듬한 판이 지정된 클로즈업과 공간의 연속성을 해친다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "탁자에 몸을 깊이 내밀고 앞의 상대를 바라보는 행동, 얼굴의 멍과 의상, 대각선 탁자 가장자리가 더 충실하지만 손과 탁자 면적이 커서 클로즈업의 밀도는 부족하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 눈은 카메라 바로 옆의 정면 상대 위치를 향하고, 상체도 탁자 너머 그 방향으로 전진한다. 벌어진 입과 크게 뜬 정상적인 눈에 긴박함이 보인다. 손가락은 탁자 위에 펼쳐져 있으며 겨누는 도구는 없다.",
        "built_space": "뒤쪽에 세로 유리창이 있는 문 하나, 오른쪽에 격자창 하나, 왼쪽에 실내 유리 개구부 하나, 천장 조명 두 개와 의자 등받이 일부가 보인다. 회색 하단 벽과 밝은 상단 벽, 목재 탁자는 참조와 유사하지만 참조에서 확인되지 않은 공간까지 넓게 드러낸다. 탁자는 화면 아래를 넓게 차지하고, 요구된 얇은 가장자리 대신 별도의 긴 비스듬한 판이 탁자 위를 가른다.",
        "entities": "인물은 젊은 한국계 남성으로 보이는 이현우 한 명뿐이며, 헝클어진 검은 머리와 해지고 더러운 회색 후드가 참조와 대체로 맞는다. 볼과 코 주변의 멍도 보이나 참조와 상처 분포가 완전히 같지는 않다. 양 손목에 수갑은 없다. 아래쪽 사진 일부에는 얼굴이 잘려 보여 찰리의 사진인지 확인할 수 없다. 다리 부상은 프레임 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "탁자 위에 참조나 지시에서 요구하지 않은 길고 비스듬한 판 형태의 물체가 추가되어, 단일 탁자의 얇은 가장자리로 설명되지 않는 구조가 생겼다."
        ],
        "physics": "양손이 탁자에 밀착되고 팔꿈치가 굽혀져 있어 상체를 앞으로 미는 자세는 지지 가능하다. 하체는 가려져 있지만 인물이 공중에 떠 있다고 볼 근거는 없다. 추가된 판은 아래쪽 일부가 탁자 면에 닿아 보이지만, 위로 솟은 끝을 고정하거나 받치는 구조는 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "얼굴과 두 눈은 화면 오른쪽 앞의 프레임 밖 상대 위치를 향한다. 고개를 낮추고 상체를 탁자 쪽으로 깊이 내민 방향도 그 시선과 일치한다. 눈에는 자연스러운 반사광이 있고 입과 눈썹의 긴장으로 절박함을 표현하지만, 희망이 밝아지는 느낌은 비교적 절제되어 있다.",
        "built_space": "오른쪽 격자창 하나, 뒤쪽 벽판 일부 하나, 왼쪽 벽 스위치 하나와 의자 등받이 일부가 보인다. 두 색으로 나뉜 벽과 목재 탁자, 낮의 실내광이 참조 장소와 잘 이어진다. 이현우는 의자 앞에서 탁자에 몸을 기울이고 있으며, 몸 아래로 검은 테두리의 탁자 가장자리가 대각선으로 지나간다. 다만 넓은 탁자 면과 양손이 강조되어 지정된 얼굴 중심 클로즈업보다 여유가 많다. 불가능한 반사나 중복된 설비는 보이지 않는다.",
        "entities": "이현우 한 명만 보이며, 젊은 한국계 남성의 얼굴, 검은 헝클어진 머리, 마른 체형과 낡은 회색 후드가 참조에 가깝다. 볼과 콧등의 멍이 읽히며 눈은 정상적인 홍채와 동공을 유지한다. 드러난 손목에 수갑은 없다. 찰리의 사진은 보이지 않지만 탁자 전체가 나온 것은 아니므로 치워졌다고 단정할 수 없다. 다리 부상은 프레임 밖이고, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양 손바닥과 손가락이 탁자 표면에 닿고 팔이 몸의 전진을 받친다. 뒤에 의자 등받이가 보이며 하체는 가려져 있다. 낮게 숙인 몸은 팔과 화면 밖 하체로 지지할 수 있는 자세이고, 떠 있는 물체나 지지 없는 신체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.929
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.679
   },
   "violations": {
    "B": [
     "[gemini-pro] 지지하는 손이나 물체 없이 허공에 비스듬히 떠 있는 흰 종이 객체",
     "[gpt-high] 탁자 위에 참조나 지시에서 요구하지 않은 길고 비스듬한 판 형태의 물체가 추가되어, 단일 탁자의 얇은 가장자리로 설명되지 않는 구조가 생겼다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 679
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 절박한 표정과 상체를 내민 자세를 잘 구현했으며 물리적 오류가 없어 가장 우수함."
   },
   {
    "label": "B",
    "score": 679,
    "verdict_ko": "화면 앞쪽에 아무런 지지 없이 허공에 떠 있는 종이가 존재하여 물리적 규칙을 위반함.  ★위반: [gemini-pro] 지지하는 손이나 물체 없이 허공에 비스듬히 떠 있는 흰 종이 객체 / [gpt-high] 탁자 위에 참조나 지시에서 요구하지 않은 길고 비스듬한 판 형태의 물체가 추가되어, 단일 탁자의 얇은 가장자리로 설명되지 않는 구조가 생겼다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S23sh3_sel.png",
    "asset_id": "72d33a24-767e-49ca-a219-877c93579838",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd9e-194f-71d3-a912-5a8368cac272",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S23sh3"
  }
 },
 "S23sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:27:45.590496+00:00",
  "fingerprint": "64365e3d2fce2352c5c71151272400ab16658ae30a04c2f1ded28407e7c1db47",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S23sh11_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S23sh11_sel.png",
  "source_sha256": "5efc028f6cb4f950e937ec75b31cb1f1e61f21659ea00d30857b523c5af06c6e",
  "file": "S23sh11_cine.png",
  "staged_sha256": "c743c91fb01fba0a7e2de84323b77bfaffd0f19064c8a204909612a6d17381cc",
  "latency_ms": 11914
 },
 "S24sh3::signage": {
  "fp": "b0a2ca4352bcab72",
  "inscriptions": [
   {
    "text_native": "7-31",
    "source": "scene_text_quoted",
    "reason_ko": "컨테이너 현관문에 표기된 호수 번호이다.",
    "source_quote": "7-31"
   }
  ],
  "cues": [],
  "dropped": []
 },
 "era_assess::7c2f2de7a35b1ca8": {
  "subjects": [],
  "subject_text": "이현우의 컨테이너 7-31 앞\n컨테이너 주택들이 이어진 도로변 끝의 주거용 컨테이너. 금속 외벽과 손잡이 달린 출입문에 ‘7-31’ 번호가 표시되어 있다.",
  "identity": "canonical",
  "scope_id": "L21",
  "scope_role": "location_exterior",
  "scope_sha": "19810ec4ea4025c8"
 },
 "S24sh3::bgfirst_bg": {
  "input_fingerprint": "2f52bf4d457a2667",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 7-31이 적힌 컨테이너 현관문을 거세게 밀어내며 한 발을 안으로 뻗어 몸을 막 들이민 급박한 뒷모습.\n\nLOCATION (lock): At the open entrance threshold of a family container home in the refugee settlement, with daylight crossing into the common room.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Container interior entrance in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Container entrance door marked 7-31 (Being pushed inward by 이현우) — The exterior face bearing 7-31 remains visible at an oblique angle on screen right; used as Identifies the entrance while its moving edge reveals the route inside; the door occupies less than one third of the frame; Entrance threshold and opening (Being crossed by 이현우) — The opening lies ahead of his body, with the threshold below his advancing foot; used as Provides a clear screen-space destination for the following camera without inventing interior furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use setting-appropriate daytime ambient illumination with restrained contrast that keeps the threshold action and door number readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 7-31이 적힌 컨테이너 현관문을 거세게 밀어내며 한 발을 안으로 뻗어 몸을 막 들이민 급박한 뒷모습.\n\nLOCATION (lock): At the open entrance threshold of a family container home in the refugee settlement, with daylight crossing into the common room.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Container interior entrance in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Container entrance door marked 7-31 (Being pushed inward by 이현우) — The exterior face bearing 7-31 remains visible at an oblique angle on screen right; used as Identifies the entrance while its moving edge reveals the route inside; the door occupies less than one third of the frame; Entrance threshold and opening (Being crossed by 이현우) — The opening lies ahead of his body, with the threshold below his advancing foot; used as Provides a clear screen-space destination for the following camera without inventing interior furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use setting-appropriate daytime ambient illumination with restrained contrast that keeps the threshold action and door number readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S24sh3__bgfirst_bg.png",
  "asset_id": "a4ec703c-d1d1-4a87-b91f-2707849e359c",
  "input_asset_ids": [
   "75bf1138-e4d0-42ea-92a4-e52f8c38b7c8",
   "c70c70d0-fcb5-4bb6-a911-8841f7df9f3e",
   "b86d0aae-cb8e-4c9d-bb09-59a96b48673d"
  ]
 },
 "S24sh3": {
  "input_fingerprint": "91e25e3082625b44",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 7-31이 적힌 컨테이너 현관문을 거세게 밀어내며 한 발을 안으로 뻗어 몸을 막 들이민 급박한 뒷모습.\n\nLOCATION (lock): At the open entrance threshold of a family container home in the refugee settlement, with daylight crossing into the common room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Container interior entrance in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Container entrance door marked 7-31 (Being pushed inward by 이현우) — The exterior face bearing 7-31 remains visible at an oblique angle on screen right; used as Identifies the entrance while its moving edge reveals the route inside; the door occupies less than one third of the frame; Entrance threshold and opening (Being crossed by 이현우) — The opening lies ahead of his body, with the threshold below his advancing foot; used as Provides a clear screen-space destination for the following camera without inventing interior furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use setting-appropriate daytime ambient illumination with restrained contrast that keeps the threshold action and door number readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The home is still numbered 7-31, and its interior retains the disorder and damage from the militia search; no cleanup is described. 이현우: He is entering the container without handcuffs, retaining his facial bruises and dog-bitten leg. He now has the contact card obtained during the detention-room encounter, later revealed to be concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"7-31\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 7-31이 적힌 컨테이너 현관문을 거세게 밀어내며 한 발을 안으로 뻗어 몸을 막 들이민 급박한 뒷모습.\n\nLOCATION (lock): At the open entrance threshold of a family container home in the refugee settlement, with daylight crossing into the common room. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Container interior entrance in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Container entrance door marked 7-31 (Being pushed inward by 이현우) — The exterior face bearing 7-31 remains visible at an oblique angle on screen right; used as Identifies the entrance while its moving edge reveals the route inside; the door occupies less than one third of the frame; Entrance threshold and opening (Being crossed by 이현우) — The opening lies ahead of his body, with the threshold below his advancing foot; used as Provides a clear screen-space destination for the following camera without inventing interior furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use setting-appropriate daytime ambient illumination with restrained contrast that keeps the threshold action and door number readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The home is still numbered 7-31, and its interior retains the disorder and damage from the militia search; no cleanup is described. 이현우: He is entering the container without handcuffs, retaining his facial bruises and dog-bitten leg. He now has the contact card obtained during the detention-room encounter, later revealed to be concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"7-31\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 7-31이 적힌 컨테이너 현관문을 거세게 밀어내며 한 발을 안으로 뻗어 몸을 막 들이민 급박한 뒷모습.\n\nLOCATION (lock): At the open entrance threshold of a family container home in the refugee settlement, with daylight crossing into the common room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Container interior entrance in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Container entrance door marked 7-31 (Being pushed inward by 이현우) — The exterior face bearing 7-31 remains visible at an oblique angle on screen right; used as Identifies the entrance while its moving edge reveals the route inside; the door occupies less than one third of the frame; Entrance threshold and opening (Being crossed by 이현우) — The opening lies ahead of his body, with the threshold below his advancing foot; used as Provides a clear screen-space destination for the following camera without inventing interior furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use setting-appropriate daytime ambient illumination with restrained contrast that keeps the threshold action and door number readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The home is still numbered 7-31, and its interior retains the disorder and damage from the militia search; no cleanup is described. 이현우: He is entering the container without handcuffs, retaining his facial bruises and dog-bitten leg. He now has the contact card obtained during the detention-room encounter, later revealed to be concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"7-31\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S24sh3__bgfirst_bg.png",
     "asset_id": "a4ec703c-d1d1-4a87-b91f-2707849e359c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S24sh3.png",
     "asset_id": "75bf1138-e4d0-42ea-92a4-e52f8c38b7c8",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L21B01.png",
     "asset_id": "c70c70d0-fcb5-4bb6-a911-8841f7df9f3e",
     "role": "location_plate"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_primary_container_home_sel.png",
     "asset_id": "b86d0aae-cb8e-4c9d-bb09-59a96b48673d",
     "role": "structure_seed_look"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 몸과 시선은 컨테이너 내부를 향해 진입하고 있으며, 오른손은 바깥으로 열린 문짝의 겉면을 짚고 있음.",
    "built_space": "참조 이미지의 구조물 지침에 따라 파란색 컨테이너 외벽을 정확히 구현했고, 우측의 화분과 쓰레기통, 좌측의 바닥 덮개 등 공간의 세부 요소가 알맞게 배치됨. 문은 우측으로 바깥을 향해 열려 있음.",
    "entities": "이현우는 지시된 지저분한 후드티와 청바지를 입고 있으며 얼굴의 멍과 다리의 상처가 확인됨. 문짝에는 '7-31'이 정확히 표기되었고, 추가된 엑스트라는 없음.",
    "hard_violations": [],
    "physics": "왼발로 바닥을 단단히 딛고 오른발을 문지방 위로 뻗어 진입하는 체중 이동이 자연스러우며, 오른손은 문짝에 안정적으로 얹혀 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "이현우의 몸은 내부를 향하고 있고, 왼손은 문손잡이를 잡고 있음.",
    "built_space": "파란색 컨테이너와 우측 화분 등 주변 환경은 구현되었으나, '7-31' 텍스트가 문이 아닌 우측 벽면 명판에 적혀 있음. 문은 좌측으로 바깥을 향해 열려 있음.",
    "entities": "이현우의 착장과 외모는 참조와 일치함. 그러나 화면 좌측 배경 원경에 프롬프트에서 지시하지 않은 정체불명의 인물 두 명이 등장함.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 인물(화면 좌측 배경의 두 사람)이 임의로 생성됨.",
     "[gpt-high] 이현우 외에는 사람을 추가하지 말라는 조건과 달리 왼쪽 배경에 두 사람이 보인다."
    ],
    "physics": "왼발로 지면을 딛고 있으나 오른발이 공중에 과도하게 들려 있어 자연스러운 보행이나 밀고 들어가는 동작이라기보다 부자연스러운 도약 포즈에 가까움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 화면 구도와 '7-31'이 적힌 문의 위치를 매우 정확히 구현했으며, 지정되지 않은 인물을 완벽히 배제하고 캐릭터의 상처 디테일까지 훌륭하게 살려냄."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "컨테이너의 질감과 위치적 특성은 살렸으나 프롬프트에 없는 인물들이 배경에 등장하는 치명적인 오류를 범함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 몸과 시선은 컨테이너 내부를 향해 진입하고 있으며, 오른손은 바깥으로 열린 문짝의 겉면을 짚고 있음.",
        "built_space": "참조 이미지의 구조물 지침에 따라 파란색 컨테이너 외벽을 정확히 구현했고, 우측의 화분과 쓰레기통, 좌측의 바닥 덮개 등 공간의 세부 요소가 알맞게 배치됨. 문은 우측으로 바깥을 향해 열려 있음.",
        "entities": "이현우는 지시된 지저분한 후드티와 청바지를 입고 있으며 얼굴의 멍과 다리의 상처가 확인됨. 문짝에는 '7-31'이 정확히 표기되었고, 추가된 엑스트라는 없음.",
        "hard_violations": [],
        "physics": "왼발로 바닥을 단단히 딛고 오른발을 문지방 위로 뻗어 진입하는 체중 이동이 자연스러우며, 오른손은 문짝에 안정적으로 얹혀 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 몸은 내부를 향하고 있고, 왼손은 문손잡이를 잡고 있음.",
        "built_space": "파란색 컨테이너와 우측 화분 등 주변 환경은 구현되었으나, '7-31' 텍스트가 문이 아닌 우측 벽면 명판에 적혀 있음. 문은 좌측으로 바깥을 향해 열려 있음.",
        "entities": "이현우의 착장과 외모는 참조와 일치함. 그러나 화면 좌측 배경 원경에 프롬프트에서 지시하지 않은 정체불명의 인물 두 명이 등장함.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 인물(화면 좌측 배경의 두 사람)이 임의로 생성됨."
        ],
        "physics": "왼발로 지면을 딛고 있으나 오른발이 공중에 과도하게 들려 있어 자연스러운 보행이나 밀고 들어가는 동작이라기보다 부자연스러운 도약 포즈에 가까움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 화면 구도와 '7-31'이 적힌 문의 위치를 매우 정확히 구현했으며, 지정되지 않은 인물을 완벽히 배제하고 캐릭터의 상처 디테일까지 훌륭하게 살려냄."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "컨테이너의 질감과 위치적 특성은 살렸으나 프롬프트에 없는 인물들이 배경에 등장하는 치명적인 오류를 범함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 몸과 시선은 컨테이너 내부를 향해 진입하고 있으며, 오른손은 바깥으로 열린 문짝의 겉면을 짚고 있음.",
        "built_space": "참조 이미지의 구조물 지침에 따라 파란색 컨테이너 외벽을 정확히 구현했고, 우측의 화분과 쓰레기통, 좌측의 바닥 덮개 등 공간의 세부 요소가 알맞게 배치됨. 문은 우측으로 바깥을 향해 열려 있음.",
        "entities": "이현우는 지시된 지저분한 후드티와 청바지를 입고 있으며 얼굴의 멍과 다리의 상처가 확인됨. 문짝에는 '7-31'이 정확히 표기되었고, 추가된 엑스트라는 없음.",
        "hard_violations": [],
        "physics": "왼발로 바닥을 단단히 딛고 오른발을 문지방 위로 뻗어 진입하는 체중 이동이 자연스러우며, 오른손은 문짝에 안정적으로 얹혀 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 몸은 내부를 향하고 있고, 왼손은 문손잡이를 잡고 있음.",
        "built_space": "파란색 컨테이너와 우측 화분 등 주변 환경은 구현되었으나, '7-31' 텍스트가 문이 아닌 우측 벽면 명판에 적혀 있음. 문은 좌측으로 바깥을 향해 열려 있음.",
        "entities": "이현우의 착장과 외모는 참조와 일치함. 그러나 화면 좌측 배경 원경에 프롬프트에서 지시하지 않은 정체불명의 인물 두 명이 등장함.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 인물(화면 좌측 배경의 두 사람)이 임의로 생성됨."
        ],
        "physics": "왼발로 지면을 딛고 있으나 오른발이 공중에 과도하게 들려 있어 자연스러운 보행이나 밀고 들어가는 동작이라기보다 부자연스러운 도약 포즈에 가까움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "금지된 배경 인물 두 명이 추가됐고, 문이 화면 왼쪽의 바깥으로 열리며 번호도 문이 아닌 벽에 있어 핵심 연출을 어긴다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "한 사람의 급박한 뒷모습과 오른손으로 미는 번호 달린 문은 더 가깝지만, 문이 바깥으로 열리고 발도 아직 문턱을 넘지 않았으며 구조 참조를 따르지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리와 몸통은 열린 실내를 향하고 얼굴과 시선은 보이지 않는다. 왼손은 화면 왼쪽 문의 손잡이를 잡고 있다. 문짝은 실내가 아니라 외부 바닥 쪽으로 돌출되어 있어 안으로 거세게 미는 동작과 반대다. 오른발은 문턱 가까이 들려 있으나 다리를 안쪽으로 길게 뻗은 순간보다는 무릎을 굽혀 들어 올린 모습이다.",
        "built_space": "주 출입구 하나, 왼쪽으로 열린 문짝 하나, 오른쪽 벽의 번호판 하나와 호출 장치 하나, 상부 벽등 하나, 오른쪽 창 하나가 보인다. 외부에는 배수구 하나, 화분 하나와 일부 잘린 수거함이 있다. 장소 사진의 회색 외벽과 오른쪽 창 배치를 재현했지만, 우선권을 가진 구조 사진의 파란 문·외벽과 출입구 왼쪽 방범창 배치는 따르지 않았다. 입구는 중앙 상단에 있으나 문과 인물이 전경에서 크게 보인다. 번호는 요구된 문 외측이 아니라 고정 벽에 붙어 있다.",
        "entities": "중앙에는 헝클어진 검은 머리의 젊고 마른 남성 한 명이 있으며, 낡은 짙은 후드와 청바지, 운동화는 인물 참조에 가깝다. 뒷모습이라 한국계 얼굴의 동일성과 얼굴 멍은 확인할 수 없다. 개에게 물린 상처는 드러나지 않으며 숨긴 카드도 확인할 수 없다. 왼손에 수갑은 보이지 않는다. 실내 바닥에는 흩어진 물건이 보인다. '7-31'은 읽히지만 부착 위치가 다르다. 왼쪽 골목에 별도의 사람 두 명이 추가되어 있다.",
        "hard_violations": [
         "이현우 외에는 사람을 추가하지 말라는 조건과 달리 왼쪽 배경에 두 사람이 보인다."
        ],
        "physics": "왼발이 외부 콘크리트 바닥을 딛고 몸을 지지하고, 오른발은 무릎을 굽혀 들려 있어 한 발로 진입하는 자세 자체는 가능하다. 왼손은 손잡이와 접촉한다. 문은 문틀 쪽 경첩으로 지지되는 회전문으로 읽히며 공중에 뜬 물체는 없다. 다만 손잡이를 잡고 바깥으로 열린 문을 다루는 자세여서 요구된 안쪽 밀기와 맞지 않는다."
       },
       {
        "label": "B",
        "direction": "몸통과 앞다리는 출입구를 향하지만 얼굴은 오른손이 닿은 문 쪽으로 돌아가 있다. 오른손 손바닥은 번호 아래 문 표면을 민다. 번호가 적힌 면은 화면 오른쪽에서 비스듬히 보이지만, 문짝의 자유단이 외부 바닥 쪽으로 나와 있어 안으로 열리는 방향은 아니다. 앞선 오른발도 문턱 바깥에 남아 있어 몸을 막 안으로 들이민 순간에는 미치지 못한다.",
        "built_space": "출입구 하나와 오른쪽으로 열린 문짝 하나, 상부 차양 하나, 부분적으로 보이는 벽등 하나, 오른쪽 창 하나가 보인다. 배수구와 화분, 일부 잘린 수거함은 각각 하나다. 실내에는 신발 선반과 흩어진 신발·옷감, 천장등이 보인다. 번호는 별도의 벽 번호판 없이 문 표면에 있다. 문은 화면의 삼분의 일 미만을 차지하지만, 입구와 인물이 상당히 크게 잡혔다. 장소 사진의 회색 구조를 따라 구조 참조의 파란 외벽·문과 왼쪽 방범창 배치를 놓쳤다.",
        "entities": "사람은 한 명뿐이다. 검은 헝클어진 머리, 젊은 동아시아계 남성의 옆얼굴, 마른 체형과 낡은 짙은 후드·청바지·운동화는 인물 참조에 대체로 부합한다. 옆얼굴에는 붉은 멍 자국이 있고 오른쪽 종아리 부근에는 찢어진 천과 붉은 상처 흔적이 보이지만 개에게 물린 상처인지는 확정할 수 없다. 양손에 수갑은 없다. 신발 속 카드는 보이지 않으며 은닉 설정과 충돌하지 않는다. 문에는 정확히 '7-31'이 있고, 실내의 어질러진 상태도 보인다.",
        "hard_violations": [],
        "physics": "앞선 오른발은 문턱 앞 바닥에 닿아 있고, 뒤쪽 왼발은 뒤꿈치를 들고 앞부분으로 지면을 밀어내는 자세다. 체중 이동과 전진 동작은 물리적으로 가능하다. 오른손 손바닥과 손목은 문에 자연스럽게 이어져 접촉하며 문도 문틀 쪽에서 지지된다. 떠 있는 사람이나 물체는 없다. 그러나 실제로 보이는 문짝의 회전 위치는 바깥쪽 열림이어서 안으로 밀어 진입한다는 행동과 맞지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "금지된 배경 인물 두 명이 추가됐고, 문이 화면 왼쪽의 바깥으로 열리며 번호도 문이 아닌 벽에 있어 핵심 연출을 어긴다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "한 사람의 급박한 뒷모습과 오른손으로 미는 번호 달린 문은 더 가깝지만, 문이 바깥으로 열리고 발도 아직 문턱을 넘지 않았으며 구조 참조를 따르지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "머리와 몸통은 열린 실내를 향하고 얼굴과 시선은 보이지 않는다. 왼손은 화면 왼쪽 문의 손잡이를 잡고 있다. 문짝은 실내가 아니라 외부 바닥 쪽으로 돌출되어 있어 안으로 거세게 미는 동작과 반대다. 오른발은 문턱 가까이 들려 있으나 다리를 안쪽으로 길게 뻗은 순간보다는 무릎을 굽혀 들어 올린 모습이다.",
        "built_space": "주 출입구 하나, 왼쪽으로 열린 문짝 하나, 오른쪽 벽의 번호판 하나와 호출 장치 하나, 상부 벽등 하나, 오른쪽 창 하나가 보인다. 외부에는 배수구 하나, 화분 하나와 일부 잘린 수거함이 있다. 장소 사진의 회색 외벽과 오른쪽 창 배치를 재현했지만, 우선권을 가진 구조 사진의 파란 문·외벽과 출입구 왼쪽 방범창 배치는 따르지 않았다. 입구는 중앙 상단에 있으나 문과 인물이 전경에서 크게 보인다. 번호는 요구된 문 외측이 아니라 고정 벽에 붙어 있다.",
        "entities": "중앙에는 헝클어진 검은 머리의 젊고 마른 남성 한 명이 있으며, 낡은 짙은 후드와 청바지, 운동화는 인물 참조에 가깝다. 뒷모습이라 한국계 얼굴의 동일성과 얼굴 멍은 확인할 수 없다. 개에게 물린 상처는 드러나지 않으며 숨긴 카드도 확인할 수 없다. 왼손에 수갑은 보이지 않는다. 실내 바닥에는 흩어진 물건이 보인다. '7-31'은 읽히지만 부착 위치가 다르다. 왼쪽 골목에 별도의 사람 두 명이 추가되어 있다.",
        "hard_violations": [
         "이현우 외에는 사람을 추가하지 말라는 조건과 달리 왼쪽 배경에 두 사람이 보인다."
        ],
        "physics": "왼발이 외부 콘크리트 바닥을 딛고 몸을 지지하고, 오른발은 무릎을 굽혀 들려 있어 한 발로 진입하는 자세 자체는 가능하다. 왼손은 손잡이와 접촉한다. 문은 문틀 쪽 경첩으로 지지되는 회전문으로 읽히며 공중에 뜬 물체는 없다. 다만 손잡이를 잡고 바깥으로 열린 문을 다루는 자세여서 요구된 안쪽 밀기와 맞지 않는다."
       },
       {
        "label": "A",
        "direction": "몸통과 앞다리는 출입구를 향하지만 얼굴은 오른손이 닿은 문 쪽으로 돌아가 있다. 오른손 손바닥은 번호 아래 문 표면을 민다. 번호가 적힌 면은 화면 오른쪽에서 비스듬히 보이지만, 문짝의 자유단이 외부 바닥 쪽으로 나와 있어 안으로 열리는 방향은 아니다. 앞선 오른발도 문턱 바깥에 남아 있어 몸을 막 안으로 들이민 순간에는 미치지 못한다.",
        "built_space": "출입구 하나와 오른쪽으로 열린 문짝 하나, 상부 차양 하나, 부분적으로 보이는 벽등 하나, 오른쪽 창 하나가 보인다. 배수구와 화분, 일부 잘린 수거함은 각각 하나다. 실내에는 신발 선반과 흩어진 신발·옷감, 천장등이 보인다. 번호는 별도의 벽 번호판 없이 문 표면에 있다. 문은 화면의 삼분의 일 미만을 차지하지만, 입구와 인물이 상당히 크게 잡혔다. 장소 사진의 회색 구조를 따라 구조 참조의 파란 외벽·문과 왼쪽 방범창 배치를 놓쳤다.",
        "entities": "사람은 한 명뿐이다. 검은 헝클어진 머리, 젊은 동아시아계 남성의 옆얼굴, 마른 체형과 낡은 짙은 후드·청바지·운동화는 인물 참조에 대체로 부합한다. 옆얼굴에는 붉은 멍 자국이 있고 오른쪽 종아리 부근에는 찢어진 천과 붉은 상처 흔적이 보이지만 개에게 물린 상처인지는 확정할 수 없다. 양손에 수갑은 없다. 신발 속 카드는 보이지 않으며 은닉 설정과 충돌하지 않는다. 문에는 정확히 '7-31'이 있고, 실내의 어질러진 상태도 보인다.",
        "hard_violations": [],
        "physics": "앞선 오른발은 문턱 앞 바닥에 닿아 있고, 뒤쪽 왼발은 뒤꿈치를 들고 앞부분으로 지면을 밀어내는 자세다. 체중 이동과 전진 동작은 물리적으로 가능하다. 오른손 손바닥과 손목은 문에 자연스럽게 이어져 접촉하며 문도 문틀 쪽에서 지지된다. 떠 있는 사람이나 물체는 없다. 그러나 실제로 보이는 문짝의 회전 위치는 바깥쪽 열림이어서 안으로 밀어 진입한다는 행동과 맞지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.833
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.583
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트에 명시되지 않은 인물(화면 좌측 배경의 두 사람)이 임의로 생성됨.",
     "[gpt-high] 이현우 외에는 사람을 추가하지 말라는 조건과 달리 왼쪽 배경에 두 사람이 보인다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 583
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 화면 구도와 '7-31'이 적힌 문의 위치를 매우 정확히 구현했으며, 지정되지 않은 인물을 완벽히 배제하고 캐릭터의 상처 디테일까지 훌륭하게 살려냄."
   },
   {
    "label": "B",
    "score": 583,
    "verdict_ko": "컨테이너의 질감과 위치적 특성은 살렸으나 프롬프트에 없는 인물들이 배경에 등장하는 치명적인 오류를 범함.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 인물(화면 좌측 배경의 두 사람)이 임의로 생성됨. / [gpt-high] 이현우 외에는 사람을 추가하지 말라는 조건과 달리 왼쪽 배경에 두 사람이 보인다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L21B01.png",
    "asset_id": "c70c70d0-fcb5-4bb6-a911-8841f7df9f3e",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_primary_container_home_sel.png",
    "asset_id": "b86d0aae-cb8e-4c9d-bb09-59a96b48673d",
    "role": "structure_seed_look"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacd9e-dd56-77fd-b032-e80e2b898359",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S24sh3__bgfirst_bg.png",
   "bg_asset_id": "a4ec703c-d1d1-4a87-b91f-2707849e359c",
   "bg_record_key": "S24sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S24sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:27:57.929061+00:00",
  "fingerprint": "f2f1a86ae71cbfaaf4c8a0b51ee07e9648d9f7fad102cbc33f08eec65815a0da",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S24sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S24sh3_sel.png",
  "source_sha256": "91c3a5a50967f70eda9052de0282d4c5aae2da10040973918db528da78c85d8e",
  "file": "S24sh3_cine.png",
  "staged_sha256": "4a2e4f11a3341f9343c1d3719b3cb369bd71efb9fee2a22b3fb5ac0d2bb47912",
  "latency_ms": 18119
 },
 "S24sh7::signage": {
  "fp": "48f8595feb4036e3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S24sh7": {
  "input_fingerprint": "6eb5509822f31e2e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 페드로의 멱살을 두 손으로 꽉 틀어쥐고 눈에 핏대를 세운 채 무섭게 노려보는 구도.\n\nLOCATION (lock): Inside the family container home's common room just beyond the entrance, with daylight from the doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container interior (The confrontation has moved inside) — Only a limited interior background remains around the two figures; used as Retains spatial enclosure without adding unsupported furniture or fixtures; 페드로's collar (Held tightly in both of 이현우's hands) — Its gripped front section is visible below 페드로's foreground shoulder; used as Provides the physical contact point that makes the confrontation readable in the shoulder-overlap composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain setting-appropriate daytime ambient light with firm but restrained facial contrast, allowing the grip and strained expression to carry the threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed and damaged from the search, with no restoration established. 이현우: He is inside the container, uncuffed, with facial bruises and the persistent dog-bitten leg. He retains the contact card concealed in his shoe. 페드로: He remains inside the container after his hurried arrival.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 페드로의 멱살을 두 손으로 꽉 틀어쥐고 눈에 핏대를 세운 채 무섭게 노려보는 구도.\n\nLOCATION (lock): Inside the family container home's common room just beyond the entrance, with daylight from the doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container interior (The confrontation has moved inside) — Only a limited interior background remains around the two figures; used as Retains spatial enclosure without adding unsupported furniture or fixtures; 페드로's collar (Held tightly in both of 이현우's hands) — Its gripped front section is visible below 페드로's foreground shoulder; used as Provides the physical contact point that makes the confrontation readable in the shoulder-overlap composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain setting-appropriate daytime ambient light with firm but restrained facial contrast, allowing the grip and strained expression to carry the threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed and damaged from the search, with no restoration established. 이현우: He is inside the container, uncuffed, with facial bruises and the persistent dog-bitten leg. He retains the contact card concealed in his shoe. 페드로: He remains inside the container after his hurried arrival.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 페드로의 멱살을 두 손으로 꽉 틀어쥐고 눈에 핏대를 세운 채 무섭게 노려보는 구도.\n\nLOCATION (lock): Inside the family container home's common room just beyond the entrance, with daylight from the doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container interior (The confrontation has moved inside) — Only a limited interior background remains around the two figures; used as Retains spatial enclosure without adding unsupported furniture or fixtures; 페드로's collar (Held tightly in both of 이현우's hands) — Its gripped front section is visible below 페드로's foreground shoulder; used as Provides the physical contact point that makes the confrontation readable in the shoulder-overlap composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain setting-appropriate daytime ambient light with firm but restrained facial contrast, allowing the grip and strained expression to carry the threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed and damaged from the search, with no restoration established. 이현우: He is inside the container, uncuffed, with facial bruises and the persistent dog-bitten leg. He retains the contact card concealed in his shoe. 페드로: He remains inside the container after his hurried arrival.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선이 페드로의 얼굴을 향해 똑바로 고정됨.",
    "built_space": "화면 왼쪽에 파란색 골판지 형태의 컨테이너 철벽이 크게 자리 잡고 있으며, 이는 레퍼런스의 하얀 벽 및 창문 구조와 완전히 불일치함.",
    "entities": "이현우와 페드로의 복장 및 인물 형태가 레퍼런스와 일치함.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 없는 컨테이너 철벽이 등장하고 실내 구조가 완전히 다르게 변경됨"
    ],
    "physics": "이현우의 양손이 페드로의 가죽 재킷 앞부분을 단단히 쥐고 지탱함."
   },
   {
    "label": "B",
    "direction": "이현우가 핏발이 선 눈으로 페드로의 눈을 강렬하게 노려봄.",
    "built_space": "왼쪽의 창문과 선반, 출입문, 벽면의 사진들, 그리고 바닥에 흩어진 깨진 도자기 파편까지 레퍼런스의 공간 구조와 디테일이 정확히 유지됨.",
    "entities": "이현우의 상처 난 얼굴과 핏대 선 눈, 페드로의 비니와 가죽 재킷 등 명시된 외형이 모두 정확함.",
    "hard_violations": [],
    "physics": "이현우의 양손이 페드로의 멱살을 강하게 꽉 틀어쥐고 있으며 접촉면의 주름이 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스의 실내 구조와 바닥의 파편까지 완벽하게 유지했으며, 두 손으로 멱살을 쥔 채 핏발 선 눈으로 노려보는 묘사가 프롬프트와 정확히 일치함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물들의 행동은 구현되었으나, 레퍼런스의 하얀 벽을 무시하고 컨테이너 철벽을 생성하여 공간 연속성을 크게 훼손함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 페드로의 얼굴을 향해 똑바로 고정됨.",
        "built_space": "화면 왼쪽에 파란색 골판지 형태의 컨테이너 철벽이 크게 자리 잡고 있으며, 이는 레퍼런스의 하얀 벽 및 창문 구조와 완전히 불일치함.",
        "entities": "이현우와 페드로의 복장 및 인물 형태가 레퍼런스와 일치함.",
        "hard_violations": [
         "레퍼런스에 없는 컨테이너 철벽이 등장하고 실내 구조가 완전히 다르게 변경됨"
        ],
        "physics": "이현우의 양손이 페드로의 가죽 재킷 앞부분을 단단히 쥐고 지탱함."
       },
       {
        "label": "B",
        "direction": "이현우가 핏발이 선 눈으로 페드로의 눈을 강렬하게 노려봄.",
        "built_space": "왼쪽의 창문과 선반, 출입문, 벽면의 사진들, 그리고 바닥에 흩어진 깨진 도자기 파편까지 레퍼런스의 공간 구조와 디테일이 정확히 유지됨.",
        "entities": "이현우의 상처 난 얼굴과 핏대 선 눈, 페드로의 비니와 가죽 재킷 등 명시된 외형이 모두 정확함.",
        "hard_violations": [],
        "physics": "이현우의 양손이 페드로의 멱살을 강하게 꽉 틀어쥐고 있으며 접촉면의 주름이 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스의 실내 구조와 바닥의 파편까지 완벽하게 유지했으며, 두 손으로 멱살을 쥔 채 핏발 선 눈으로 노려보는 묘사가 프롬프트와 정확히 일치함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물들의 행동은 구현되었으나, 레퍼런스의 하얀 벽을 무시하고 컨테이너 철벽을 생성하여 공간 연속성을 크게 훼손함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 페드로의 얼굴을 향해 똑바로 고정됨.",
        "built_space": "화면 왼쪽에 파란색 골판지 형태의 컨테이너 철벽이 크게 자리 잡고 있으며, 이는 레퍼런스의 하얀 벽 및 창문 구조와 완전히 불일치함.",
        "entities": "이현우와 페드로의 복장 및 인물 형태가 레퍼런스와 일치함.",
        "hard_violations": [
         "레퍼런스에 없는 컨테이너 철벽이 등장하고 실내 구조가 완전히 다르게 변경됨"
        ],
        "physics": "이현우의 양손이 페드로의 가죽 재킷 앞부분을 단단히 쥐고 지탱함."
       },
       {
        "label": "B",
        "direction": "이현우가 핏발이 선 눈으로 페드로의 눈을 강렬하게 노려봄.",
        "built_space": "왼쪽의 창문과 선반, 출입문, 벽면의 사진들, 그리고 바닥에 흩어진 깨진 도자기 파편까지 레퍼런스의 공간 구조와 디테일이 정확히 유지됨.",
        "entities": "이현우의 상처 난 얼굴과 핏대 선 눈, 페드로의 비니와 가죽 재킷 등 명시된 외형이 모두 정확함.",
        "hard_violations": [],
        "physics": "이현우의 양손이 페드로의 멱살을 강하게 꽉 틀어쥐고 있으며 접촉면의 주름이 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "페드로의 전경 어깨 아래 드러나는 양손 멱살잡이와 이현우의 정면에 가까운 위협적 눈빛이 지정된 미디엄 숏 구도를 더 충실하게 구현한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "양손의 접촉과 상호 시선, 실내 연속성은 좋지만 두 사람을 옆모습으로 나란히 배치해 전경 어깨 중첩과 이현우의 눈빛 강조가 A보다 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 이현우의 눈은 왼쪽 페드로의 얼굴을 직접 겨냥하며, 페드로도 이현우 쪽을 바라본다. 이현우의 두 팔과 주먹은 페드로의 앞깃으로 모인다. 카메라를 바라보는 사람은 없다.",
        "built_space": "왼쪽 창 하나, 그 뒤쪽 목재 문틀의 개구부 하나, 왼쪽 낮은 수납장 하나, 중앙 뒤편 책장 하나, 오른쪽 뒤편 낮은 의자 하나가 보인다. 베이지색 벽과 목재 바닥, 벽의 사진, 바닥의 깨진 그릇이 이전 장소의 재료와 훼손 상태를 이어간다. 두 사람은 실내 중앙에 서 있고 페드로의 어깨가 왼쪽 전경을 차지한다. 오른쪽 뒤의 큰 검은 면은 참고 사진에서 명확히 확인되지 않아 배경 연속성에는 약간의 불확실성이 있다. 반사면이나 명백히 중복된 고정 설비는 없다.",
        "entities": "등장인물은 두 명뿐이다. 이현우는 젊은 한국계 남성으로 묘사되며 헝클어진 검은 머리, 얼굴 상처, 해진 짙은 회색 후드가 참고와 잘 맞는다. 페드로는 젊은 남성의 옆얼굴과 검은 비니, 낡은 검은 가죽 재킷으로 참고 인물에 부합한다. 눈은 붉게 충혈되어 있지만 정상적인 홍채와 동공을 유지한다. 멱살을 잡은 손과 소매는 이현우에게 연결된다. 다리의 부상과 신발 속 카드는 프레임 밖이므로 판단할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우의 두 손이 페드로의 재킷 앞깃을 각각 움켜쥐고 있으며, 손 주변의 가죽이 당겨지고 접힌다. 손목과 팔꿈치의 연결 및 앞으로 뻗은 상체가 잡아당기는 동작으로 성립한다. 두 사람의 발은 잘렸지만 몸통은 서 있는 자세로 이어지며 공중에 뜬 징후는 없다. 깨진 그릇은 바닥에, 배경 물건들은 수납장과 창턱에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "왼쪽 이현우는 오른쪽 페드로의 눈과 얼굴을 노려보고, 페드로도 이현우를 마주 본다. 두 손은 페드로의 목 아래 양쪽 앞깃에 향하며 실제로 접촉한다. 시선과 행동의 대상은 정확하다.",
        "built_space": "중앙 뒤 목재 문틀의 밝은 개구부 하나, 왼쪽 금속 골조가 드러난 입구 영역 하나, 왼쪽 아래 낮은 수납장 하나가 보인다. 수납장 위 병 두 개와 흰 컵, 파란 컵, 그릇 및 벽의 사진은 이전 사진의 생활 공간과 잘 연결된다. 인물들은 수납장 앞 실내에 서 있고 가구와 충돌하지 않는다. 허리 부근까지 담은 미디엄 숏이지만 양측 옆모습을 보여주는 배치여서 페드로의 전경 어깨를 통한 중첩은 약하다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "이현우와 페드로 두 명만 보인다. 이현우의 젊은 한국계 남성 얼굴, 검은 흐트러진 머리, 얼굴 상처와 낡은 후드는 참고와 부합한다. 페드로의 젊은 옆얼굴, 검은 비니와 가죽 재킷도 참고에 부합하며, 옆모습만으로 세부 얼굴 일치 여부를 완전히 확인하기는 어렵다. 이현우의 눈은 정상적인 형태를 유지하고, 양손과 해진 소매의 소유 관계도 명확하다. 하체 부상과 숨긴 카드는 잘려 있어 평가 대상이 아니다. 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우가 상체를 앞으로 기울이고 팔꿈치를 굽혀 페드로의 양쪽 깃을 잡는다. 두 주먹과 깃의 접촉, 당겨진 옷감의 주름이 분명하다. 페드로의 상체는 약간 뒤로 물러난 상태로 읽히며 들려 있지는 않다. 발은 프레임 밖이지만 부유나 불가능한 관절 연결은 보이지 않는다. 컵과 병, 그릇은 수납장 표면과 선반에 지지되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "페드로의 전경 어깨 아래 드러나는 양손 멱살잡이와 이현우의 정면에 가까운 위협적 눈빛이 지정된 미디엄 숏 구도를 더 충실하게 구현한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "양손의 접촉과 상호 시선, 실내 연속성은 좋지만 두 사람을 옆모습으로 나란히 배치해 전경 어깨 중첩과 이현우의 눈빛 강조가 A보다 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 이현우의 눈은 왼쪽 페드로의 얼굴을 직접 겨냥하며, 페드로도 이현우 쪽을 바라본다. 이현우의 두 팔과 주먹은 페드로의 앞깃으로 모인다. 카메라를 바라보는 사람은 없다.",
        "built_space": "왼쪽 창 하나, 그 뒤쪽 목재 문틀의 개구부 하나, 왼쪽 낮은 수납장 하나, 중앙 뒤편 책장 하나, 오른쪽 뒤편 낮은 의자 하나가 보인다. 베이지색 벽과 목재 바닥, 벽의 사진, 바닥의 깨진 그릇이 이전 장소의 재료와 훼손 상태를 이어간다. 두 사람은 실내 중앙에 서 있고 페드로의 어깨가 왼쪽 전경을 차지한다. 오른쪽 뒤의 큰 검은 면은 참고 사진에서 명확히 확인되지 않아 배경 연속성에는 약간의 불확실성이 있다. 반사면이나 명백히 중복된 고정 설비는 없다.",
        "entities": "등장인물은 두 명뿐이다. 이현우는 젊은 한국계 남성으로 묘사되며 헝클어진 검은 머리, 얼굴 상처, 해진 짙은 회색 후드가 참고와 잘 맞는다. 페드로는 젊은 남성의 옆얼굴과 검은 비니, 낡은 검은 가죽 재킷으로 참고 인물에 부합한다. 눈은 붉게 충혈되어 있지만 정상적인 홍채와 동공을 유지한다. 멱살을 잡은 손과 소매는 이현우에게 연결된다. 다리의 부상과 신발 속 카드는 프레임 밖이므로 판단할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우의 두 손이 페드로의 재킷 앞깃을 각각 움켜쥐고 있으며, 손 주변의 가죽이 당겨지고 접힌다. 손목과 팔꿈치의 연결 및 앞으로 뻗은 상체가 잡아당기는 동작으로 성립한다. 두 사람의 발은 잘렸지만 몸통은 서 있는 자세로 이어지며 공중에 뜬 징후는 없다. 깨진 그릇은 바닥에, 배경 물건들은 수납장과 창턱에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "왼쪽 이현우는 오른쪽 페드로의 눈과 얼굴을 노려보고, 페드로도 이현우를 마주 본다. 두 손은 페드로의 목 아래 양쪽 앞깃에 향하며 실제로 접촉한다. 시선과 행동의 대상은 정확하다.",
        "built_space": "중앙 뒤 목재 문틀의 밝은 개구부 하나, 왼쪽 금속 골조가 드러난 입구 영역 하나, 왼쪽 아래 낮은 수납장 하나가 보인다. 수납장 위 병 두 개와 흰 컵, 파란 컵, 그릇 및 벽의 사진은 이전 사진의 생활 공간과 잘 연결된다. 인물들은 수납장 앞 실내에 서 있고 가구와 충돌하지 않는다. 허리 부근까지 담은 미디엄 숏이지만 양측 옆모습을 보여주는 배치여서 페드로의 전경 어깨를 통한 중첩은 약하다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "이현우와 페드로 두 명만 보인다. 이현우의 젊은 한국계 남성 얼굴, 검은 흐트러진 머리, 얼굴 상처와 낡은 후드는 참고와 부합한다. 페드로의 젊은 옆얼굴, 검은 비니와 가죽 재킷도 참고에 부합하며, 옆모습만으로 세부 얼굴 일치 여부를 완전히 확인하기는 어렵다. 이현우의 눈은 정상적인 형태를 유지하고, 양손과 해진 소매의 소유 관계도 명확하다. 하체 부상과 숨긴 카드는 잘려 있어 평가 대상이 아니다. 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우가 상체를 앞으로 기울이고 팔꿈치를 굽혀 페드로의 양쪽 깃을 잡는다. 두 주먹과 깃의 접촉, 당겨진 옷감의 주름이 분명하다. 페드로의 상체는 약간 뒤로 물러난 상태로 읽히며 들려 있지는 않다. 발은 프레임 밖이지만 부유나 불가능한 관절 연결은 보이지 않는다. 컵과 병, 그릇은 수납장 표면과 선반에 지지되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.317,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.067,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스에 없는 컨테이너 철벽이 등장하고 실내 구조가 완전히 다르게 변경됨"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1067
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "레퍼런스의 실내 구조와 바닥의 파편까지 완벽하게 유지했으며, 두 손으로 멱살을 쥔 채 핏발 선 눈으로 노려보는 묘사가 프롬프트와 정확히 일치함."
   },
   {
    "label": "A",
    "score": 1067,
    "verdict_ko": "인물들의 행동은 구현되었으나, 레퍼런스의 하얀 벽을 무시하고 컨테이너 철벽을 생성하여 공간 연속성을 크게 훼손함.  ★위반: [gemini-pro] 레퍼런스에 없는 컨테이너 철벽이 등장하고 실내 구조가 완전히 다르게 변경됨"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S14sh8_sel.png",
    "asset_id": "ba2bf8c2-7bb5-4458-999d-a5b37587dff0",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:826925>",
    "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda0-05ee-73a3-bf3e-3bedf82674fe",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S14sh8"
  }
 },
 "S24sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:28:16.380378+00:00",
  "fingerprint": "89fd5054d42958ecdd3ed843d1b6907e3979e7878b86999ce7649d9be736ce4d",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S24sh7_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S24sh7_sel.png",
  "source_sha256": "602c8be65efd9d61a303a7ed5f5e5b38b02e37d989966e7ff73be4445e13aa54",
  "file": "S24sh7_cine.png",
  "staged_sha256": "22001bc91434714691581359fe46db84c2a1ae353827d110c8da77e0480a5224",
  "latency_ms": 11069
 },
 "S24sh9::signage": {
  "fp": "19a1afde0419ebe6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S24sh9": {
  "input_fingerprint": "1921b902375142de",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 멱살을 쥔 손을 느슨하게 놓은 채, 두 눈동자가 크게 확장된 충격 받은 표정 클로즈업.\n\nLOCATION (lock): In the common room inside the family container home, near the daylight-facing entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container-home interior (The confrontation remains inside the home); used as A narrow, softly resolved background margin preserves the interior setting without competing with the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination keeps the widening eyes and slackening facial tension legible without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 retains the disruption from the militia's search; no restoration has occurred. 이현우: He is inside the container with facial bruising and an unhealed dog-bite wound on his leg; his handcuffs have been removed. The contact card remains concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 멱살을 쥔 손을 느슨하게 놓은 채, 두 눈동자가 크게 확장된 충격 받은 표정 클로즈업.\n\nLOCATION (lock): In the common room inside the family container home, near the daylight-facing entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container-home interior (The confrontation remains inside the home); used as A narrow, softly resolved background margin preserves the interior setting without competing with the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination keeps the widening eyes and slackening facial tension legible without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 retains the disruption from the militia's search; no restoration has occurred. 이현우: He is inside the container with facial bruising and an unhealed dog-bite wound on his leg; his handcuffs have been removed. The contact card remains concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 멱살을 쥔 손을 느슨하게 놓은 채, 두 눈동자가 크게 확장된 충격 받은 표정 클로즈업.\n\nLOCATION (lock): In the common room inside the family container home, near the daylight-facing entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container-home interior (The confrontation remains inside the home); used as A narrow, softly resolved background margin preserves the interior setting without competing with the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination keeps the widening eyes and slackening facial tension legible without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 retains the disruption from the militia's search; no restoration has occurred. 이현우: He is inside the container with facial bruising and an unhealed dog-bite wound on his leg; his handcuffs have been removed. The contact card remains concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 프레임 왼쪽 밖을 향해 고정되어 있음.",
    "built_space": "레퍼런스와 일치하는 컨테이너 주택 내부. 창문, 목재 책상, 화분, 벽에 붙은 포스터 등의 고정 요소가 정확한 위치에 있으며 인물은 그 앞 전경에 위치함.",
    "entities": "이현우(젊은 얼굴, 헝클어진 검은 머리, 상처 난 얼굴, 낡은 후드티)가 정확히 구현됨. 지시대로 다른 인물은 나타나지 않음.",
    "hard_violations": [],
    "physics": "가슴 부근에서 빈 옷가지(또는 천)를 한 손으로 쥐고 있으며, 팔의 자세에 의해 물리적으로 자연스럽게 지탱됨."
   },
   {
    "label": "B",
    "direction": "시선은 프레임 왼쪽의 비니를 쓴 인물의 얼굴을 향하고 있음.",
    "built_space": "실내이나 레퍼런스의 주요 배경 요소(책상 등)가 보이지 않으며, 뒤편에 임의의 벽 조명이 추가되는 등 공간 구조가 다름.",
    "entities": "이현우는 요구된 외모와 일치하나, 프롬프트에서 금지한 다른 인물(비니를 쓰고 가죽 재킷을 입은 남성)이 프레임 왼쪽에 등장함.",
    "hard_violations": [
     "[gemini-pro] 프롬프트가 명시적으로 배제할 것을 지시한 레퍼런스의 다른 인물을 프레임에 포함시킴 (invented/forbidden people).",
     "[gpt-high] 이 장면에 재등장시키지 말라고 명시한 이전 상대 남성의 얼굴·몸·의복을 화면 왼쪽에 그대로 포함했다."
    ],
    "physics": "오른손은 상대방의 옷을 잡고 있고 왼손은 허공에 느슨하게 풀려 있으며, 신체 접촉은 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 대로 레퍼런스의 다른 인물을 정확히 배제했으며, 일치하는 배경 속에서 이현우의 확장된 눈동자와 충격받은 표정을 성공적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트가 명시적으로 금지한 레퍼런스 상의 다른 인물을 그대로 프레임에 포함시켰으며, 배경 구조도 레퍼런스와 다르게 왜곡되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 프레임 왼쪽 밖을 향해 고정되어 있음.",
        "built_space": "레퍼런스와 일치하는 컨테이너 주택 내부. 창문, 목재 책상, 화분, 벽에 붙은 포스터 등의 고정 요소가 정확한 위치에 있으며 인물은 그 앞 전경에 위치함.",
        "entities": "이현우(젊은 얼굴, 헝클어진 검은 머리, 상처 난 얼굴, 낡은 후드티)가 정확히 구현됨. 지시대로 다른 인물은 나타나지 않음.",
        "hard_violations": [],
        "physics": "가슴 부근에서 빈 옷가지(또는 천)를 한 손으로 쥐고 있으며, 팔의 자세에 의해 물리적으로 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "시선은 프레임 왼쪽의 비니를 쓴 인물의 얼굴을 향하고 있음.",
        "built_space": "실내이나 레퍼런스의 주요 배경 요소(책상 등)가 보이지 않으며, 뒤편에 임의의 벽 조명이 추가되는 등 공간 구조가 다름.",
        "entities": "이현우는 요구된 외모와 일치하나, 프롬프트에서 금지한 다른 인물(비니를 쓰고 가죽 재킷을 입은 남성)이 프레임 왼쪽에 등장함.",
        "hard_violations": [
         "프롬프트가 명시적으로 배제할 것을 지시한 레퍼런스의 다른 인물을 프레임에 포함시킴 (invented/forbidden people)."
        ],
        "physics": "오른손은 상대방의 옷을 잡고 있고 왼손은 허공에 느슨하게 풀려 있으며, 신체 접촉은 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 대로 레퍼런스의 다른 인물을 정확히 배제했으며, 일치하는 배경 속에서 이현우의 확장된 눈동자와 충격받은 표정을 성공적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트가 명시적으로 금지한 레퍼런스 상의 다른 인물을 그대로 프레임에 포함시켰으며, 배경 구조도 레퍼런스와 다르게 왜곡되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 프레임 왼쪽 밖을 향해 고정되어 있음.",
        "built_space": "레퍼런스와 일치하는 컨테이너 주택 내부. 창문, 목재 책상, 화분, 벽에 붙은 포스터 등의 고정 요소가 정확한 위치에 있으며 인물은 그 앞 전경에 위치함.",
        "entities": "이현우(젊은 얼굴, 헝클어진 검은 머리, 상처 난 얼굴, 낡은 후드티)가 정확히 구현됨. 지시대로 다른 인물은 나타나지 않음.",
        "hard_violations": [],
        "physics": "가슴 부근에서 빈 옷가지(또는 천)를 한 손으로 쥐고 있으며, 팔의 자세에 의해 물리적으로 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "시선은 프레임 왼쪽의 비니를 쓴 인물의 얼굴을 향하고 있음.",
        "built_space": "실내이나 레퍼런스의 주요 배경 요소(책상 등)가 보이지 않으며, 뒤편에 임의의 벽 조명이 추가되는 등 공간 구조가 다름.",
        "entities": "이현우는 요구된 외모와 일치하나, 프롬프트에서 금지한 다른 인물(비니를 쓰고 가죽 재킷을 입은 남성)이 프레임 왼쪽에 등장함.",
        "hard_violations": [
         "프롬프트가 명시적으로 배제할 것을 지시한 레퍼런스의 다른 인물을 프레임에 포함시킴 (invented/forbidden people)."
        ],
        "physics": "오른손은 상대방의 옷을 잡고 있고 왼손은 허공에 느슨하게 풀려 있으며, 신체 접촉은 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "충격 받은 얼굴과 손을 놓기 시작하는 동작은 더 가깝지만, 등장 금지된 이전 상대의 얼굴·몸·옷을 크게 재등장시킨 중대 위반이 결정적이다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "이현우만 등장하고 기존 실내를 유지해 우세하지만, 얼굴 클로즈업 대신 넓은 상반신 구도이며 멱살을 느슨하게 놓기보다 옷감을 자기 가슴 앞에 움켜쥐고 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 크게 뜬 두 눈은 화면 왼쪽에 보이는 상대 남성의 얼굴을 향한다. 상대 역시 이현우 쪽을 바라본다. 아래쪽 손은 상대의 옷깃을 향해 붙어 있고, 위쪽 손은 옷깃에서 떨어지며 손가락을 펴기 시작한다.",
        "built_space": "얼굴 뒤로 밝은 벽, 오른쪽 목재 문틀 하나, 왼쪽의 밝은 개구부 일부, 천장 조명 하나와 흐릿한 벽 부착물들이 보인다. 주간 실내라는 조건은 맞지만, 배경 여백이 좁지만은 않으며 이전 공간의 정확한 연결 관계는 흐림과 가림 때문에 확인하기 어렵다. 불가능한 반사나 명백한 설비 중복은 보이지 않는다.",
        "entities": "이현우는 검은 헝클어진 머리의 젊은 한국계 남성으로 표현되며, 얼굴 상처와 낡고 찢어진 짙은 회색 후드가 이전 사진과 대체로 맞는다. 자연스러운 홍채와 동공을 유지한 채 눈을 크게 뜨고 입을 벌렸다. 그러나 왼쪽에는 이전 사진의 검은 비니를 쓴 상대 남성의 얼굴과 상체, 겉옷까지 보인다. 다리 상처와 신발 속 카드는 구도 밖이므로 판단 대상이 아니다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "이 장면에 재등장시키지 말라고 명시한 이전 상대 남성의 얼굴·몸·의복을 화면 왼쪽에 그대로 포함했다."
        ],
        "physics": "아래 손은 상대의 옷깃에 접촉해 옷감을 잡고 있고, 위 손은 후드 소매로 이어지는 팔에 의해 지지된다. 두 몸은 화면 밖 하체로 이어지는 정상적인 상체 자세이며 공중 부유는 없다. 한 손을 풀기 시작하는 동작은 가능하지만, 다른 손은 여전히 옷깃을 움켜쥐고 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 카메라보다 왼쪽의 화면 밖 지점을 크게 뜬 눈으로 바라본다. 시선 대상 자체는 보이지 않는다. 보이는 손은 화면 밖 상대에게 뻗어 있지 않고, 자기 가슴 앞의 어두운 옷감을 향해 당겨져 있다.",
        "built_space": "왼쪽 창 하나와 그 옆 밝은 출입구 하나, 창가 화분 두 개, 목재 책상 하나, 책상 아래 의자 하나, 뒤쪽 어두운 통로와 오른쪽 목재 문틀이 보인다. 밝은 벽과 목재 바닥, 낮의 채광은 이전 실내와 잘 연결된다. 다만 책상과 여러 화분까지 넓고 비교적 선명하게 드러나, 얼굴 주변에 좁고 부드러운 배경 여백만 남기라는 구도와 다르다. 설비 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "이현우 한 명만 보이며, 젊은 한국계 남성의 얼굴, 검은 헝클어진 머리, 얼굴의 멍과 상처, 낡은 회색 후드가 참조와 대체로 일치한다. 눈은 정상적인 인간 눈이며 놀란 표정이 분명하다. 손에는 어두운 겉옷의 깃처럼 보이는 천이 있지만, 누군가 입은 옷의 멱살이라는 연결은 보이지 않는다. 손목에 수갑은 없으며 다리 상처와 숨긴 카드는 구도 밖이다. 배경 인쇄물의 글자는 판독되지 않는다.",
        "hard_violations": [],
        "physics": "옷감은 이현우의 손가락에 확실히 잡혀 아래로 늘어지므로 지지가 있다. 손목과 팔은 후드 소매로 자연스럽게 이어지고 상체 자세도 물리적으로 가능하다. 다만 손가락이 옷감을 단단히 감싸고 있어, 멱살을 느슨하게 놓는 순간보다는 옷감을 가슴 쪽에 붙들고 있는 정지 동작으로 읽힌다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "충격 받은 얼굴과 손을 놓기 시작하는 동작은 더 가깝지만, 등장 금지된 이전 상대의 얼굴·몸·옷을 크게 재등장시킨 중대 위반이 결정적이다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "이현우만 등장하고 기존 실내를 유지해 우세하지만, 얼굴 클로즈업 대신 넓은 상반신 구도이며 멱살을 느슨하게 놓기보다 옷감을 자기 가슴 앞에 움켜쥐고 있다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 크게 뜬 두 눈은 화면 왼쪽에 보이는 상대 남성의 얼굴을 향한다. 상대 역시 이현우 쪽을 바라본다. 아래쪽 손은 상대의 옷깃을 향해 붙어 있고, 위쪽 손은 옷깃에서 떨어지며 손가락을 펴기 시작한다.",
        "built_space": "얼굴 뒤로 밝은 벽, 오른쪽 목재 문틀 하나, 왼쪽의 밝은 개구부 일부, 천장 조명 하나와 흐릿한 벽 부착물들이 보인다. 주간 실내라는 조건은 맞지만, 배경 여백이 좁지만은 않으며 이전 공간의 정확한 연결 관계는 흐림과 가림 때문에 확인하기 어렵다. 불가능한 반사나 명백한 설비 중복은 보이지 않는다.",
        "entities": "이현우는 검은 헝클어진 머리의 젊은 한국계 남성으로 표현되며, 얼굴 상처와 낡고 찢어진 짙은 회색 후드가 이전 사진과 대체로 맞는다. 자연스러운 홍채와 동공을 유지한 채 눈을 크게 뜨고 입을 벌렸다. 그러나 왼쪽에는 이전 사진의 검은 비니를 쓴 상대 남성의 얼굴과 상체, 겉옷까지 보인다. 다리 상처와 신발 속 카드는 구도 밖이므로 판단 대상이 아니다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "이 장면에 재등장시키지 말라고 명시한 이전 상대 남성의 얼굴·몸·의복을 화면 왼쪽에 그대로 포함했다."
        ],
        "physics": "아래 손은 상대의 옷깃에 접촉해 옷감을 잡고 있고, 위 손은 후드 소매로 이어지는 팔에 의해 지지된다. 두 몸은 화면 밖 하체로 이어지는 정상적인 상체 자세이며 공중 부유는 없다. 한 손을 풀기 시작하는 동작은 가능하지만, 다른 손은 여전히 옷깃을 움켜쥐고 있다."
       },
       {
        "label": "A",
        "direction": "이현우는 카메라보다 왼쪽의 화면 밖 지점을 크게 뜬 눈으로 바라본다. 시선 대상 자체는 보이지 않는다. 보이는 손은 화면 밖 상대에게 뻗어 있지 않고, 자기 가슴 앞의 어두운 옷감을 향해 당겨져 있다.",
        "built_space": "왼쪽 창 하나와 그 옆 밝은 출입구 하나, 창가 화분 두 개, 목재 책상 하나, 책상 아래 의자 하나, 뒤쪽 어두운 통로와 오른쪽 목재 문틀이 보인다. 밝은 벽과 목재 바닥, 낮의 채광은 이전 실내와 잘 연결된다. 다만 책상과 여러 화분까지 넓고 비교적 선명하게 드러나, 얼굴 주변에 좁고 부드러운 배경 여백만 남기라는 구도와 다르다. 설비 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "이현우 한 명만 보이며, 젊은 한국계 남성의 얼굴, 검은 헝클어진 머리, 얼굴의 멍과 상처, 낡은 회색 후드가 참조와 대체로 일치한다. 눈은 정상적인 인간 눈이며 놀란 표정이 분명하다. 손에는 어두운 겉옷의 깃처럼 보이는 천이 있지만, 누군가 입은 옷의 멱살이라는 연결은 보이지 않는다. 손목에 수갑은 없으며 다리 상처와 숨긴 카드는 구도 밖이다. 배경 인쇄물의 글자는 판독되지 않는다.",
        "hard_violations": [],
        "physics": "옷감은 이현우의 손가락에 확실히 잡혀 아래로 늘어지므로 지지가 있다. 손목과 팔은 후드 소매로 자연스럽게 이어지고 상체 자세도 물리적으로 가능하다. 다만 손가락이 옷감을 단단히 감싸고 있어, 멱살을 느슨하게 놓는 순간보다는 옷감을 가슴 쪽에 붙들고 있는 정지 동작으로 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.975
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.725
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트가 명시적으로 배제할 것을 지시한 레퍼런스의 다른 인물을 프레임에 포함시킴 (invented/forbidden people).",
     "[gpt-high] 이 장면에 재등장시키지 말라고 명시한 이전 상대 남성의 얼굴·몸·의복을 화면 왼쪽에 그대로 포함했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 725
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 대로 레퍼런스의 다른 인물을 정확히 배제했으며, 일치하는 배경 속에서 이현우의 확장된 눈동자와 충격받은 표정을 성공적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 725,
    "verdict_ko": "프롬프트가 명시적으로 금지한 레퍼런스 상의 다른 인물을 그대로 프레임에 포함시켰으며, 배경 구조도 레퍼런스와 다르게 왜곡되었습니다.  ★위반: [gemini-pro] 프롬프트가 명시적으로 배제할 것을 지시한 레퍼런스의 다른 인물을 프레임에 포함시킴 (invented/forbidden people). / [gpt-high] 이 장면에 재등장시키지 말라고 명시한 이전 상대 남성의 얼굴·몸·의복을 화면 왼쪽에 그대로 포함했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S24sh7_sel.png",
    "asset_id": "35e8fe3e-bff5-49aa-987c-56d9101d24c5",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda0-bc84-7ad7-8fb0-72982f1675a1",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S24sh7"
  }
 },
 "S24sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:28:27.792308+00:00",
  "fingerprint": "b1b728cd5785a27a5778588c7e26c95df4da57bca44a8db062025edc64c337c2",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S24sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S24sh9_sel.png",
  "source_sha256": "6c6738bd412329b9db5be4a241556c55b407a26856d1c4a4b08beaa25dfaf767",
  "file": "S24sh9_cine.png",
  "staged_sha256": "495afe6adfad5a05e6d92e335b54cbd1ae1c7c13849fd830165c2c29199eaccd",
  "latency_ms": 11513
 },
 "S25sh15::signage": {
  "fp": "18cc6df0d3eb2126",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S25sh15": {
  "input_fingerprint": "918a202996c89698",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 찰리의 낡은 금속 가슴 장갑판 틈새에서 둥근 링 모양의 푸른빛이 눈부시게 뿜어져 나오는 찰나.\n\nLOCATION (lock): In the dancing area of an open refugee-settlement reception lot beneath makeshift canopies, immediately after the surrounding lights go out. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reception tent (Erected over the open-lot reception area) — Only a peripheral section of the tent enclosure is visible beyond the torso; used as Provides a restrained environmental margin around the chest-focused composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dazzling blue ring breaks through the surrounding blackout, revealing the edges of the worn chest plates before the wider reception lighting returns.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception occupies a tented vacant lot, with the jukebox stopped and the streetlights extinguished by the premature blackout. Charlie retains his old coat and hat, worn metal body and faded chest logo; a ring-shaped light is now emerging from his chest.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 찰리의 낡은 금속 가슴 장갑판 틈새에서 둥근 링 모양의 푸른빛이 눈부시게 뿜어져 나오는 찰나.\n\nLOCATION (lock): In the dancing area of an open refugee-settlement reception lot beneath makeshift canopies, immediately after the surrounding lights go out. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reception tent (Erected over the open-lot reception area) — Only a peripheral section of the tent enclosure is visible beyond the torso; used as Provides a restrained environmental margin around the chest-focused composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dazzling blue ring breaks through the surrounding blackout, revealing the edges of the worn chest plates before the wider reception lighting returns.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception occupies a tented vacant lot, with the jukebox stopped and the streetlights extinguished by the premature blackout. Charlie retains his old coat and hat, worn metal body and faded chest logo; a ring-shaped light is now emerging from his chest.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 찰리의 낡은 금속 가슴 장갑판 틈새에서 둥근 링 모양의 푸른빛이 눈부시게 뿜어져 나오는 찰나.\n\nLOCATION (lock): In the dancing area of an open refugee-settlement reception lot beneath makeshift canopies, immediately after the surrounding lights go out. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reception tent (Erected over the open-lot reception area) — Only a peripheral section of the tent enclosure is visible beyond the torso; used as Provides a restrained environmental margin around the chest-focused composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dazzling blue ring breaks through the surrounding blackout, revealing the edges of the worn chest plates before the wider reception lighting returns.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception occupies a tented vacant lot, with the jukebox stopped and the streetlights extinguished by the premature blackout. Charlie retains his old coat and hat, worn metal body and faded chest logo; a ring-shaped light is now emerging from his chest.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "캐릭터가 정면 카메라 방향을 응시하고 있으며, 가슴 중앙의 갈라진 틈에서 푸른빛이 정면으로 뿜어져 나옴.",
    "built_space": "캐릭터 뒤로 난민 수용소의 텐트 천막 가장자리가 배경 양옆을 채우고 있어 요구된 제한적인 환경 여백을 충족함.",
    "entities": "찰리(고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈, 낡은 코트와 페도라 모자, 희미한 로고가 있는 금속 가슴 장갑판), 틈새에서 새어 나오는 푸른색 링 모양의 빛 모두 프롬프트와 일치함.",
    "hard_violations": [],
    "physics": "자연스럽게 서 있는 자세이며, 모자와 코트가 몸체에 안정적으로 착용되어 있음."
   },
   {
    "label": "B",
    "direction": "캐릭터가 시선을 살짝 아래로 향하고 있으며, 가슴 중앙의 빛이 정면을 향해 빛남.",
    "built_space": "배경에 천막과 파이프 지지대가 제한적으로 보여 클로즈업 프레이밍에 맞는 구조를 갖춤.",
    "entities": "모자와 코트, 푸른색 링 모양의 빛은 존재하나, 얼굴 형태가 프롬프트에 명시된 '고릴라형 기계 머리'가 아닌 일반적인 각진 로봇 얼굴임.",
    "hard_violations": [],
    "physics": "서 있는 자세와 의상의 처짐 등 물리적인 오류는 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "고릴라형 기계 머리의 특징과 가슴 장갑판 틈새에서 뿜어져 나오는 푸른 링 모양의 빛, 클로즈업 프레이밍을 모두 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "가슴의 푸른빛과 의상은 표현되었으나, 명시된 '고릴라형 기계 머리'가 아닌 일반적인 휴머노이드 형태의 얼굴이 묘사되어 인물 정체성 조건에 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캐릭터가 정면 카메라 방향을 응시하고 있으며, 가슴 중앙의 갈라진 틈에서 푸른빛이 정면으로 뿜어져 나옴.",
        "built_space": "캐릭터 뒤로 난민 수용소의 텐트 천막 가장자리가 배경 양옆을 채우고 있어 요구된 제한적인 환경 여백을 충족함.",
        "entities": "찰리(고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈, 낡은 코트와 페도라 모자, 희미한 로고가 있는 금속 가슴 장갑판), 틈새에서 새어 나오는 푸른색 링 모양의 빛 모두 프롬프트와 일치함.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 자세이며, 모자와 코트가 몸체에 안정적으로 착용되어 있음."
       },
       {
        "label": "B",
        "direction": "캐릭터가 시선을 살짝 아래로 향하고 있으며, 가슴 중앙의 빛이 정면을 향해 빛남.",
        "built_space": "배경에 천막과 파이프 지지대가 제한적으로 보여 클로즈업 프레이밍에 맞는 구조를 갖춤.",
        "entities": "모자와 코트, 푸른색 링 모양의 빛은 존재하나, 얼굴 형태가 프롬프트에 명시된 '고릴라형 기계 머리'가 아닌 일반적인 각진 로봇 얼굴임.",
        "hard_violations": [],
        "physics": "서 있는 자세와 의상의 처짐 등 물리적인 오류는 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "고릴라형 기계 머리의 특징과 가슴 장갑판 틈새에서 뿜어져 나오는 푸른 링 모양의 빛, 클로즈업 프레이밍을 모두 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "가슴의 푸른빛과 의상은 표현되었으나, 명시된 '고릴라형 기계 머리'가 아닌 일반적인 휴머노이드 형태의 얼굴이 묘사되어 인물 정체성 조건에 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "캐릭터가 정면 카메라 방향을 응시하고 있으며, 가슴 중앙의 갈라진 틈에서 푸른빛이 정면으로 뿜어져 나옴.",
        "built_space": "캐릭터 뒤로 난민 수용소의 텐트 천막 가장자리가 배경 양옆을 채우고 있어 요구된 제한적인 환경 여백을 충족함.",
        "entities": "찰리(고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈, 낡은 코트와 페도라 모자, 희미한 로고가 있는 금속 가슴 장갑판), 틈새에서 새어 나오는 푸른색 링 모양의 빛 모두 프롬프트와 일치함.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 자세이며, 모자와 코트가 몸체에 안정적으로 착용되어 있음."
       },
       {
        "label": "B",
        "direction": "캐릭터가 시선을 살짝 아래로 향하고 있으며, 가슴 중앙의 빛이 정면을 향해 빛남.",
        "built_space": "배경에 천막과 파이프 지지대가 제한적으로 보여 클로즈업 프레이밍에 맞는 구조를 갖춤.",
        "entities": "모자와 코트, 푸른색 링 모양의 빛은 존재하나, 얼굴 형태가 프롬프트에 명시된 '고릴라형 기계 머리'가 아닌 일반적인 각진 로봇 얼굴임.",
        "hard_violations": [],
        "physics": "서 있는 자세와 의상의 처짐 등 물리적인 오류는 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "얼굴 일부를 잘라 가슴과 눈부신 푸른 링을 크게 잡은 구도가 핵심 순간에 더 충실하지만, 고릴라형 머리의 정체성은 부족하고 천막도 다소 많이 보입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "고릴라형 기계 머리와 낡은 복장은 잘 맞지만, 얼굴·상체·천막 입구까지 넓게 담아 가슴 중심 클로즈업보다 인물 초상에 가까워졌습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리와 눈은 아래쪽 가슴 방향으로 기울어 있습니다. 가슴 중앙의 링은 몸통 전면을 향하며 카메라에 거의 원형으로 보입니다. 푸른빛이 주변 장갑판 가장자리와 코트 안쪽을 밝힙니다. 무기나 이동 동작은 없습니다.",
        "built_space": "뒤에 천막 한 동의 어두운 측면 천, 왼쪽 수직 지지대, 상단 가로 프레임과 고정 밧줄이 보입니다. 찰리는 천막 앞에 있으며 다른 사람이나 추가 설비는 없습니다. 다만 왼쪽 천막과 하늘이 상당한 면적을 차지해 배경을 주변부로만 제한하라는 지시에는 덜 정확합니다. 반사상은 없습니다.",
        "entities": "기계 인물 한 명이며 낡은 모자와 코트, 마모된 금속 몸통, 흰색 각진 얼굴판과 선 모양 입이 보입니다. 눈은 작은 검은 구멍에 가깝고 얼굴은 고릴라형보다 길쭉한 인간형입니다. 중앙 장갑판 틈의 청색 고리는 크고 눈부시며 내부가 구별되어 원반으로 보이지 않습니다. 가슴에는 희미한 도형 자국이 있으나 읽을 수 있는 글자는 없습니다. 주크박스와 가로등은 화면 밖이므로 확인할 수 없습니다.",
        "hard_violations": [],
        "physics": "발은 프레임 밖이지만 몸통은 곧게 서 있고 부유를 암시하는 자세가 아닙니다. 머리는 목 구조에, 어깨와 가슴 장갑은 몸통에 연결되어 있습니다. 코트는 어깨에 걸려 아래로 처지고 링은 가슴 장갑 내부에 고정된 발광 부품처럼 보입니다. 금속과 천에 청색 반사광이 생겨 광원이 공간 안에 존재하는 것으로 읽힙니다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴과 눈은 카메라 쪽 정면을 향하며 가슴을 내려다보지는 않습니다. 가슴의 발광 링도 정면을 향하고 주변 판 틈으로 빛이 번집니다. 지시된 시선 목표나 이동 동작은 없습니다.",
        "built_space": "천막 한 동의 입구가 뒤에 크게 보이고, 양쪽으로 걷어 묶은 천 두 폭과 상단 덮개가 드러납니다. 찰리는 입구 바로 앞 중앙에 서 있습니다. 구성 자체는 가능한 천막 구조지만 입구 전체가 배경의 상당 부분을 차지하여 제한적인 환경 여백이라는 요구와 어긋납니다. 추가 인물이나 불가능한 반사는 없습니다.",
        "entities": "기계 인물 한 명으로 고릴라형 코와 주둥이, 흰 얼굴판, 각진 머리 외장, 선 모양 입이 보입니다. 눈은 단순한 점보다 눈꺼풀과 렌즈가 구체적으로 표현되어 있습니다. 낡은 모자·코트·금속 몸통과 희미한 가슴 도형이 있으며 읽을 수 있는 문자는 없습니다. 푸른 고리는 가슴 중앙 판의 벌어진 틈에 있지만 위아래가 끊겨 두 호처럼 보입니다. 주변 조명은 꺼져 있고 위쪽에 희미한 하늘이 보입니다.",
        "hard_violations": [],
        "physics": "몸통은 수직으로 서 있으며 프레임 밖의 하체 지지를 의심할 부유 징후는 없습니다. 모자는 머리 위에 놓이고 코트는 어깨에서 자연스럽게 내려옵니다. 갈라진 가슴판은 몸통에 부착되어 있고 발광부도 그 내부에 자리합니다. 푸른빛이 판 가장자리와 턱 아래에 반사되며, 지지 없이 떠 있는 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "얼굴 일부를 잘라 가슴과 눈부신 푸른 링을 크게 잡은 구도가 핵심 순간에 더 충실하지만, 고릴라형 머리의 정체성은 부족하고 천막도 다소 많이 보입니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "고릴라형 기계 머리와 낡은 복장은 잘 맞지만, 얼굴·상체·천막 입구까지 넓게 담아 가슴 중심 클로즈업보다 인물 초상에 가까워졌습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "머리와 눈은 아래쪽 가슴 방향으로 기울어 있습니다. 가슴 중앙의 링은 몸통 전면을 향하며 카메라에 거의 원형으로 보입니다. 푸른빛이 주변 장갑판 가장자리와 코트 안쪽을 밝힙니다. 무기나 이동 동작은 없습니다.",
        "built_space": "뒤에 천막 한 동의 어두운 측면 천, 왼쪽 수직 지지대, 상단 가로 프레임과 고정 밧줄이 보입니다. 찰리는 천막 앞에 있으며 다른 사람이나 추가 설비는 없습니다. 다만 왼쪽 천막과 하늘이 상당한 면적을 차지해 배경을 주변부로만 제한하라는 지시에는 덜 정확합니다. 반사상은 없습니다.",
        "entities": "기계 인물 한 명이며 낡은 모자와 코트, 마모된 금속 몸통, 흰색 각진 얼굴판과 선 모양 입이 보입니다. 눈은 작은 검은 구멍에 가깝고 얼굴은 고릴라형보다 길쭉한 인간형입니다. 중앙 장갑판 틈의 청색 고리는 크고 눈부시며 내부가 구별되어 원반으로 보이지 않습니다. 가슴에는 희미한 도형 자국이 있으나 읽을 수 있는 글자는 없습니다. 주크박스와 가로등은 화면 밖이므로 확인할 수 없습니다.",
        "hard_violations": [],
        "physics": "발은 프레임 밖이지만 몸통은 곧게 서 있고 부유를 암시하는 자세가 아닙니다. 머리는 목 구조에, 어깨와 가슴 장갑은 몸통에 연결되어 있습니다. 코트는 어깨에 걸려 아래로 처지고 링은 가슴 장갑 내부에 고정된 발광 부품처럼 보입니다. 금속과 천에 청색 반사광이 생겨 광원이 공간 안에 존재하는 것으로 읽힙니다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴과 눈은 카메라 쪽 정면을 향하며 가슴을 내려다보지는 않습니다. 가슴의 발광 링도 정면을 향하고 주변 판 틈으로 빛이 번집니다. 지시된 시선 목표나 이동 동작은 없습니다.",
        "built_space": "천막 한 동의 입구가 뒤에 크게 보이고, 양쪽으로 걷어 묶은 천 두 폭과 상단 덮개가 드러납니다. 찰리는 입구 바로 앞 중앙에 서 있습니다. 구성 자체는 가능한 천막 구조지만 입구 전체가 배경의 상당 부분을 차지하여 제한적인 환경 여백이라는 요구와 어긋납니다. 추가 인물이나 불가능한 반사는 없습니다.",
        "entities": "기계 인물 한 명으로 고릴라형 코와 주둥이, 흰 얼굴판, 각진 머리 외장, 선 모양 입이 보입니다. 눈은 단순한 점보다 눈꺼풀과 렌즈가 구체적으로 표현되어 있습니다. 낡은 모자·코트·금속 몸통과 희미한 가슴 도형이 있으며 읽을 수 있는 문자는 없습니다. 푸른 고리는 가슴 중앙 판의 벌어진 틈에 있지만 위아래가 끊겨 두 호처럼 보입니다. 주변 조명은 꺼져 있고 위쪽에 희미한 하늘이 보입니다.",
        "hard_violations": [],
        "physics": "몸통은 수직으로 서 있으며 프레임 밖의 하체 지지를 의심할 부유 징후는 없습니다. 모자는 머리 위에 놓이고 코트는 어깨에서 자연스럽게 내려옵니다. 갈라진 가슴판은 몸통에 부착되어 있고 발광부도 그 내부에 자리합니다. 푸른빛이 판 가장자리와 턱 아래에 반사되며, 지지 없이 떠 있는 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.5
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1750,
   "B": 1500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "고릴라형 기계 머리의 특징과 가슴 장갑판 틈새에서 뿜어져 나오는 푸른 링 모양의 빛, 클로즈업 프레이밍을 모두 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "가슴의 푸른빛과 의상은 표현되었으나, 명시된 '고릴라형 기계 머리'가 아닌 일반적인 휴머노이드 형태의 얼굴이 묘사되어 인물 정체성 조건에 어긋납니다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda1-79ea-7561-8ef2-7a4b132fc532",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S25sh15::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:28:39.627274+00:00",
  "fingerprint": "9c2c3fb82c1e262fb924ebf4e289c2ae5e781453e43dc1c547b833d5db1b1223",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S25sh15_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S25sh15_sel.png",
  "source_sha256": "379b3c2480626c1f4ad413abf316d5d3dd43bd7c0627e1d21c7649851e8b2c42",
  "file": "S25sh15_cine.png",
  "staged_sha256": "cda93a0ec9cf1f0e575f79a97ac9c620ad360904328ddccf65852e7cc0bb3296",
  "latency_ms": 19538
 },
 "S25sh24::signage": {
  "fp": "55fe2d92ae320a89",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S25sh24": {
  "input_fingerprint": "611d2540c9342a53",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 트럭 조수석에서 내린 박철진이 매서운 눈빛으로 난민들을 훑어보는 위압적인 전신.\n\nLOCATION (lock): Beside a newly arrived militia truck at the edge of the illuminated outdoor reception lot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Militia truck (Stopped, with the passenger-side exit open following 박철진's descent) — The passenger side is seen obliquely behind 박철진; used as Anchors his arrival in the left third without enlarging the vehicle beyond the human action; Open-lot reception area (Guests are scattering after the militia's arrival); used as Leaves visible escape space to the right and separates the arriving threat from the fleeing guests.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The reception remains brightly illuminated after its restoration, with controlled contrast giving 박철진's arrival severity without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception and surrounding settlement are brightly illuminated again, although some over-bright bulbs have burst, and the jukebox is playing again. The militia truck has arrived beside the tented lot; Charlie still has his old coat and hat. 박철진: He has disembarked at the reception and is directing the search.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 트럭 조수석에서 내린 박철진이 매서운 눈빛으로 난민들을 훑어보는 위압적인 전신.\n\nLOCATION (lock): Beside a newly arrived militia truck at the edge of the illuminated outdoor reception lot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Militia truck (Stopped, with the passenger-side exit open following 박철진's descent) — The passenger side is seen obliquely behind 박철진; used as Anchors his arrival in the left third without enlarging the vehicle beyond the human action; Open-lot reception area (Guests are scattering after the militia's arrival); used as Leaves visible escape space to the right and separates the arriving threat from the fleeing guests.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The reception remains brightly illuminated after its restoration, with controlled contrast giving 박철진's arrival severity without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception and surrounding settlement are brightly illuminated again, although some over-bright bulbs have burst, and the jukebox is playing again. The militia truck has arrived beside the tented lot; Charlie still has his old coat and hat. 박철진: He has disembarked at the reception and is directing the search.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 트럭 조수석에서 내린 박철진이 매서운 눈빛으로 난민들을 훑어보는 위압적인 전신.\n\nLOCATION (lock): Beside a newly arrived militia truck at the edge of the illuminated outdoor reception lot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Militia truck (Stopped, with the passenger-side exit open following 박철진's descent) — The passenger side is seen obliquely behind 박철진; used as Anchors his arrival in the left third without enlarging the vehicle beyond the human action; Open-lot reception area (Guests are scattering after the militia's arrival); used as Leaves visible escape space to the right and separates the arriving threat from the fleeing guests.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The reception remains brightly illuminated after its restoration, with controlled contrast giving 박철진's arrival severity without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception and surrounding settlement are brightly illuminated again, although some over-bright bulbs have burst, and the jukebox is playing again. The militia truck has arrived beside the tented lot; Charlie still has his old coat and hat. 박철진: He has disembarked at the reception and is directing the search.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "박철진의 시선은 프레임 우측에서 흩어지는 난민들을 향해 정확히 고정되어 있음.",
    "built_space": "좌측에 문이 열린 군용 트럭 1대, 우측 배경에 밝게 조명이 켜진 비닐하우스형 천막 1동과 주크박스 1대가 배치됨.",
    "entities": "박철진은 참조 이미지의 모자와 군복을 완벽히 착용함. 지시된 다국적 난민들과 켜진 주크박스가 정확히 식별됨.",
    "hard_violations": [],
    "physics": "박철진은 지면에 안정적으로 서 있으며, 달리는 난민들 역시 발이 땅에 닿아 있거나 주행의 역학을 자연스럽게 보여줌."
   },
   {
    "label": "B",
    "direction": "박철진은 프레임 우측으로 달아나는 난민들을 바라보고 있음.",
    "built_space": "좌측에 군용 트럭 1대, 배경에 어두운 색상의 군용 천막들이 있으나 명시된 밝은 접수처 공간은 보이지 않음.",
    "entities": "박철진의 의상과 얼굴은 일치하나 참조 사진의 모자가 누락됨. 주크박스 객체가 화면에 존재하지 않음.",
    "hard_violations": [],
    "physics": "박철진은 바닥을 딛고 서 있으며, 달리는 인물들의 체중 이동과 발의 접지가 지면에 의해 정상적으로 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트에 명시된 밝게 빛나는 접수처와 주크박스를 훌륭하게 구현했으며, 박철진의 외형(모자 포함)과 흩어지는 난민들의 묘사가 지시사항과 잘 일치합니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "박철진이 레퍼런스의 모자를 착용하지 않았으며, 필수 요소인 밝게 켜진 접수처와 주크박스 배경이 완전히 누락되어 현장감이 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 시선은 프레임 우측에서 흩어지는 난민들을 향해 정확히 고정되어 있음.",
        "built_space": "좌측에 문이 열린 군용 트럭 1대, 우측 배경에 밝게 조명이 켜진 비닐하우스형 천막 1동과 주크박스 1대가 배치됨.",
        "entities": "박철진은 참조 이미지의 모자와 군복을 완벽히 착용함. 지시된 다국적 난민들과 켜진 주크박스가 정확히 식별됨.",
        "hard_violations": [],
        "physics": "박철진은 지면에 안정적으로 서 있으며, 달리는 난민들 역시 발이 땅에 닿아 있거나 주행의 역학을 자연스럽게 보여줌."
       },
       {
        "label": "B",
        "direction": "박철진은 프레임 우측으로 달아나는 난민들을 바라보고 있음.",
        "built_space": "좌측에 군용 트럭 1대, 배경에 어두운 색상의 군용 천막들이 있으나 명시된 밝은 접수처 공간은 보이지 않음.",
        "entities": "박철진의 의상과 얼굴은 일치하나 참조 사진의 모자가 누락됨. 주크박스 객체가 화면에 존재하지 않음.",
        "hard_violations": [],
        "physics": "박철진은 바닥을 딛고 서 있으며, 달리는 인물들의 체중 이동과 발의 접지가 지면에 의해 정상적으로 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트에 명시된 밝게 빛나는 접수처와 주크박스를 훌륭하게 구현했으며, 박철진의 외형(모자 포함)과 흩어지는 난민들의 묘사가 지시사항과 잘 일치합니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "박철진이 레퍼런스의 모자를 착용하지 않았으며, 필수 요소인 밝게 켜진 접수처와 주크박스 배경이 완전히 누락되어 현장감이 떨어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 시선은 프레임 우측에서 흩어지는 난민들을 향해 정확히 고정되어 있음.",
        "built_space": "좌측에 문이 열린 군용 트럭 1대, 우측 배경에 밝게 조명이 켜진 비닐하우스형 천막 1동과 주크박스 1대가 배치됨.",
        "entities": "박철진은 참조 이미지의 모자와 군복을 완벽히 착용함. 지시된 다국적 난민들과 켜진 주크박스가 정확히 식별됨.",
        "hard_violations": [],
        "physics": "박철진은 지면에 안정적으로 서 있으며, 달리는 난민들 역시 발이 땅에 닿아 있거나 주행의 역학을 자연스럽게 보여줌."
       },
       {
        "label": "B",
        "direction": "박철진은 프레임 우측으로 달아나는 난민들을 바라보고 있음.",
        "built_space": "좌측에 군용 트럭 1대, 배경에 어두운 색상의 군용 천막들이 있으나 명시된 밝은 접수처 공간은 보이지 않음.",
        "entities": "박철진의 의상과 얼굴은 일치하나 참조 사진의 모자가 누락됨. 주크박스 객체가 화면에 존재하지 않음.",
        "hard_violations": [],
        "physics": "박철진은 바닥을 딛고 서 있으며, 달리는 인물들의 체중 이동과 발의 접지가 지면에 의해 정상적으로 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "전신과 열린 조수석 문, 낡은 천막은 잘 담았지만 박철진이 난민들의 반대쪽을 바라보고, 복구된 접수장의 밝은 조명과 참조의 군모가 빠졌습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "황혼의 밝은 접수장과 열린 조수석 문, 군모를 쓴 박철진의 전신 및 오른쪽으로 흩어지는 난민들을 더 충실하게 구현했으나, 시선의 표적과 기존 천막의 연속성은 다소 약합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 얼굴과 눈은 화면 왼쪽을 향하지만, 난민 네 명은 모두 그의 오른쪽에 있습니다. 따라서 보이는 난민들을 매섭게 훑어보는 시선으로 연결되지 않습니다. 오른쪽 세 명은 오른쪽으로 달리고, 중앙 뒤쪽 남성은 앞쪽으로 나오며 박철진 쪽을 보는 듯합니다. 무기나 겨누는 물체는 없습니다.",
        "built_space": "왼쪽에 군용 트럭 한 대와 열린 출입문 하나가 있고 박철진은 문 앞 지면에 서 있습니다. 오른쪽에는 묶어 젖힌 입구의 짙은 천막 하나, 중앙에는 베이지색 천막 하나와 뒤쪽 청회색 천막 일부가 보입니다. 짙은 천막의 재질과 입구는 장소 참조와 유사합니다. 오른쪽으로 빠져나갈 공간은 확보했지만, 접수장이 밝게 다시 켜졌다는 설정을 뒷받침할 조명은 보이지 않습니다. 불가능한 반사나 중복 출입문은 없습니다.",
        "entities": "박철진은 회색 섞인 짧은 머리와 수염을 가진 중년 남성으로, 참조의 얼굴 인상과 체격에 대체로 부합합니다. 올리브색 군복, 검은 완장, 검은 전투화는 있으나 참조의 군모는 없습니다. 트럭과 흩어지는 난민 네 명이 보이며, 난민들은 배경 행동 지시에 해당합니다. 이전 장면의 인물은 재등장하지 않습니다. 주크박스는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "박철진은 두 장화로 땅을 딛고 있으며 팔과 손이 자연스럽게 내려와 있습니다. 달리는 난민들은 지면을 딛는 발이나 지면 가까이 내려오는 발, 굽힌 무릎과 기울어진 몸통을 보여 달리기로 가능한 자세입니다. 트럭은 타이어로 지지되고 열린 문은 차체에 연결되어 있습니다. 상자와 잡동사니도 바닥이나 다른 물체에 놓여 있어 근거 없이 떠 있는 대상은 없습니다."
       },
       {
        "label": "B",
        "direction": "박철진의 몸은 거의 정면이고 얼굴과 시선은 화면 오른쪽 전방으로 조금 향합니다. 오른쪽으로 달아나는 난민들의 영역과는 연결되지만, 특정 난민을 노려보는 표적은 선명하지 않습니다. 배경과 오른쪽 전경의 난민들은 대체로 오른쪽 출구 방향으로 달립니다. 무기나 겨누는 물체는 없습니다.",
        "built_space": "왼쪽에 군용 트럭 한 대, 열린 조수석 출입문 하나, 출입용 발판이 보이며 박철진은 그 아래 지면에 서 있습니다. 전신 주변에 여백이 있는 와이드 구도이고 오른쪽으로 난민들의 이동 공간이 이어집니다. 중앙의 밝은 천막 한 동과 왼쪽의 낮은 천막, 주크박스 한 대, 탁자와 의자, 여러 고정 조명이 접수장 기능을 드러냅니다. 다만 중앙 천막은 참조의 짙고 낡은 천막보다 밝고 반투명해 장소 연속성이 약합니다. 트럭도 화면 왼쪽에서 다소 큰 비중을 차지합니다. 불가능한 반사나 중복 출입문은 보이지 않습니다.",
        "entities": "박철진은 중년 남성이며 회색 수염, 군모, 올리브색 야전복과 바지, 검은 완장과 전투화가 참조와 대체로 일치합니다. 얼굴은 참조보다 어둡고 작게 보이지만 인상과 체격은 유사합니다. 여러 피부색의 난민들이 오른쪽으로 흩어지고 군용 트럭과 켜진 주크박스가 보입니다. 난민들은 배경 행동 지시에 해당하며 이전 장면의 인물은 복제되지 않았습니다. 황혼의 하늘과 밝게 복구된 접수장 조명이 함께 표현되어 있습니다.",
        "hard_violations": [],
        "physics": "박철진의 양발은 자갈 바닥에 닿아 있고 체중을 지탱합니다. 난민들은 지지발, 뒤로 들린 발, 굽힌 무릎과 앞으로 기운 상체를 보여 실제 달리기로 가능한 자세입니다. 오른쪽 전경 남성도 한쪽 발을 지면에 내리며 다른 다리를 뒤로 접고 있습니다. 트럭은 바퀴로, 출입문은 경첩으로 지지되며 천막과 조명에는 기둥이나 프레임이 있습니다. 주크박스와 가구는 바닥에 놓여 있고 지지 없이 떠 있는 대상은 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전신과 열린 조수석 문, 낡은 천막은 잘 담았지만 박철진이 난민들의 반대쪽을 바라보고, 복구된 접수장의 밝은 조명과 참조의 군모가 빠졌습니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "황혼의 밝은 접수장과 열린 조수석 문, 군모를 쓴 박철진의 전신 및 오른쪽으로 흩어지는 난민들을 더 충실하게 구현했으나, 시선의 표적과 기존 천막의 연속성은 다소 약합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "박철진의 얼굴과 눈은 화면 왼쪽을 향하지만, 난민 네 명은 모두 그의 오른쪽에 있습니다. 따라서 보이는 난민들을 매섭게 훑어보는 시선으로 연결되지 않습니다. 오른쪽 세 명은 오른쪽으로 달리고, 중앙 뒤쪽 남성은 앞쪽으로 나오며 박철진 쪽을 보는 듯합니다. 무기나 겨누는 물체는 없습니다.",
        "built_space": "왼쪽에 군용 트럭 한 대와 열린 출입문 하나가 있고 박철진은 문 앞 지면에 서 있습니다. 오른쪽에는 묶어 젖힌 입구의 짙은 천막 하나, 중앙에는 베이지색 천막 하나와 뒤쪽 청회색 천막 일부가 보입니다. 짙은 천막의 재질과 입구는 장소 참조와 유사합니다. 오른쪽으로 빠져나갈 공간은 확보했지만, 접수장이 밝게 다시 켜졌다는 설정을 뒷받침할 조명은 보이지 않습니다. 불가능한 반사나 중복 출입문은 없습니다.",
        "entities": "박철진은 회색 섞인 짧은 머리와 수염을 가진 중년 남성으로, 참조의 얼굴 인상과 체격에 대체로 부합합니다. 올리브색 군복, 검은 완장, 검은 전투화는 있으나 참조의 군모는 없습니다. 트럭과 흩어지는 난민 네 명이 보이며, 난민들은 배경 행동 지시에 해당합니다. 이전 장면의 인물은 재등장하지 않습니다. 주크박스는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "박철진은 두 장화로 땅을 딛고 있으며 팔과 손이 자연스럽게 내려와 있습니다. 달리는 난민들은 지면을 딛는 발이나 지면 가까이 내려오는 발, 굽힌 무릎과 기울어진 몸통을 보여 달리기로 가능한 자세입니다. 트럭은 타이어로 지지되고 열린 문은 차체에 연결되어 있습니다. 상자와 잡동사니도 바닥이나 다른 물체에 놓여 있어 근거 없이 떠 있는 대상은 없습니다."
       },
       {
        "label": "A",
        "direction": "박철진의 몸은 거의 정면이고 얼굴과 시선은 화면 오른쪽 전방으로 조금 향합니다. 오른쪽으로 달아나는 난민들의 영역과는 연결되지만, 특정 난민을 노려보는 표적은 선명하지 않습니다. 배경과 오른쪽 전경의 난민들은 대체로 오른쪽 출구 방향으로 달립니다. 무기나 겨누는 물체는 없습니다.",
        "built_space": "왼쪽에 군용 트럭 한 대, 열린 조수석 출입문 하나, 출입용 발판이 보이며 박철진은 그 아래 지면에 서 있습니다. 전신 주변에 여백이 있는 와이드 구도이고 오른쪽으로 난민들의 이동 공간이 이어집니다. 중앙의 밝은 천막 한 동과 왼쪽의 낮은 천막, 주크박스 한 대, 탁자와 의자, 여러 고정 조명이 접수장 기능을 드러냅니다. 다만 중앙 천막은 참조의 짙고 낡은 천막보다 밝고 반투명해 장소 연속성이 약합니다. 트럭도 화면 왼쪽에서 다소 큰 비중을 차지합니다. 불가능한 반사나 중복 출입문은 보이지 않습니다.",
        "entities": "박철진은 중년 남성이며 회색 수염, 군모, 올리브색 야전복과 바지, 검은 완장과 전투화가 참조와 대체로 일치합니다. 얼굴은 참조보다 어둡고 작게 보이지만 인상과 체격은 유사합니다. 여러 피부색의 난민들이 오른쪽으로 흩어지고 군용 트럭과 켜진 주크박스가 보입니다. 난민들은 배경 행동 지시에 해당하며 이전 장면의 인물은 복제되지 않았습니다. 황혼의 하늘과 밝게 복구된 접수장 조명이 함께 표현되어 있습니다.",
        "hard_violations": [],
        "physics": "박철진의 양발은 자갈 바닥에 닿아 있고 체중을 지탱합니다. 난민들은 지지발, 뒤로 들린 발, 굽힌 무릎과 앞으로 기운 상체를 보여 실제 달리기로 가능한 자세입니다. 오른쪽 전경 남성도 한쪽 발을 지면에 내리며 다른 다리를 뒤로 접고 있습니다. 트럭은 바퀴로, 출입문은 경첩으로 지지되며 천막과 조명에는 기둥이나 프레임이 있습니다. 주크박스와 가구는 바닥에 놓여 있고 지지 없이 떠 있는 대상은 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.339
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.339
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1339
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트에 명시된 밝게 빛나는 접수처와 주크박스를 훌륭하게 구현했으며, 박철진의 외형(모자 포함)과 흩어지는 난민들의 묘사가 지시사항과 잘 일치합니다."
   },
   {
    "label": "B",
    "score": 1339,
    "verdict_ko": "박철진이 레퍼런스의 모자를 착용하지 않았으며, 필수 요소인 밝게 켜진 접수처와 주크박스 배경이 완전히 누락되어 현장감이 떨어집니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S25sh15_sel.png",
    "asset_id": "5203ebe6-e1f8-4f57-af34-931e6889185e",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda2-b7a7-732d-a8a8-33c49103ad40",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh15"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S25sh24::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:28:59.487939+00:00",
  "fingerprint": "118dd6bfef66a7167f16b91a9bcd21109c1dd9b9b233169594c250737c8c8a12",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S25sh24_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S25sh24_sel.png",
  "source_sha256": "5f44dcfb268da9b7cb6facb925fd6de196dc8643a5b29714c9d62a8464ea89f7",
  "file": "S25sh24_cine.png",
  "staged_sha256": "5c2ff70064059d3df721e1dabc976e6f2921e13e017b4273695a3c7585691502",
  "latency_ms": 14331
 },
 "S25sh29::signage": {
  "fp": "bd242b7f54d9f808",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::1bc0004a5c3ba72b": {
  "subjects": [],
  "subject_text": "인천 난민촌 피로연장 천막과 공터\n컨테이너 주거지 사이 빈 공터에 천막을 친 간이 행사장. 중앙 바닥이 넓게 비어 있고 주변에 전등과 가로등이 있다.",
  "identity": "canonical",
  "scope_id": "L31",
  "scope_role": "location_exterior",
  "scope_sha": "4df90112573cf5be"
 },
 "S25sh29::bgfirst_bg": {
  "input_fingerprint": "f94b9f2a5dd012ce",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 골목 모퉁이 벽면에 이현우, 앰버, 라울, 페드로, 찰리가 등을 바짝 기댄 채 숨죽인 긴장된 자세.\n\nLOCATION (lock): Behind a wall corner in a secluded refugee-settlement alley away from the reception lot, outdoors at night.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Outer corner of the sheltering alley wall in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Alley corner wall (The group is concealed behind the corner with their backs against its adjoining face) — The sheltering wall recedes obliquely across the image, with its outer corner at the left edge; used as Establishes the shared hiding surface and blocks a direct view of the militia's position; Concealed alley floor (Occupied by the five companions after their hurried stop); used as Keeps their differing foot placements and compressed spacing visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime visibility and restrained contrast preserve the concealed alley's tension without specifying an unsupported light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 골목 모퉁이 벽면에 이현우, 앰버, 라울, 페드로, 찰리가 등을 바짝 기댄 채 숨죽인 긴장된 자세.\n\nLOCATION (lock): Behind a wall corner in a secluded refugee-settlement alley away from the reception lot, outdoors at night.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Outer corner of the sheltering alley wall in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Alley corner wall (The group is concealed behind the corner with their backs against its adjoining face) — The sheltering wall recedes obliquely across the image, with its outer corner at the left edge; used as Establishes the shared hiding surface and blocks a direct view of the militia's position; Concealed alley floor (Occupied by the five companions after their hurried stop); used as Keeps their differing foot placements and compressed spacing visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime visibility and restrained contrast preserve the concealed alley's tension without specifying an unsupported light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S25sh29__bgfirst_bg.png",
  "asset_id": "a7e553a8-5eb4-4796-bad9-a8df7e6955f1",
  "input_asset_ids": [
   "2e756d97-d5c0-4dc9-88d5-254ef25684d3",
   "24131eeb-c517-44d7-890e-15c4f0bbe013"
  ]
 },
 "S25sh29": {
  "input_fingerprint": "ef0a7bad3a46a3d4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 골목 모퉁이 벽면에 이현우, 앰버, 라울, 페드로, 찰리가 등을 바짝 기댄 채 숨죽인 긴장된 자세.\n\nLOCATION (lock): Behind a wall corner in a secluded refugee-settlement alley away from the reception lot, outdoors at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Outer corner of the sheltering alley wall in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Alley corner wall (The group is concealed behind the corner with their backs against its adjoining face) — The sheltering wall recedes obliquely across the image, with its outer corner at the left edge; used as Establishes the shared hiding surface and blocks a direct view of the militia's position; Concealed alley floor (Occupied by the five companions after their hurried stop); used as Keeps their differing foot placements and compressed spacing visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime visibility and restrained contrast preserve the concealed alley's tension without specifying an unsupported light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement remains illuminated after the blackout reversal, with some bulbs burst. Charlie is concealed behind an alley wall in his old coat and hat; the manhole cover in the rubbish dump has not yet been moved. 이현우: He is hiding behind the alley wall after running, retaining his facial bruises, leg wound and outer garment. His contact card remains concealed in his shoe. 앰버: She is hiding behind the alley wall after the escape run. 라울: He is hiding behind the alley wall, breathless from running. 페드로: He is hiding behind the alley wall, breathless from running.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 골목 모퉁이 벽면에 이현우, 앰버, 라울, 페드로, 찰리가 등을 바짝 기댄 채 숨죽인 긴장된 자세.\n\nLOCATION (lock): Behind a wall corner in a secluded refugee-settlement alley away from the reception lot, outdoors at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Outer corner of the sheltering alley wall in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Alley corner wall (The group is concealed behind the corner with their backs against its adjoining face) — The sheltering wall recedes obliquely across the image, with its outer corner at the left edge; used as Establishes the shared hiding surface and blocks a direct view of the militia's position; Concealed alley floor (Occupied by the five companions after their hurried stop); used as Keeps their differing foot placements and compressed spacing visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime visibility and restrained contrast preserve the concealed alley's tension without specifying an unsupported light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement remains illuminated after the blackout reversal, with some bulbs burst. Charlie is concealed behind an alley wall in his old coat and hat; the manhole cover in the rubbish dump has not yet been moved. 이현우: He is hiding behind the alley wall after running, retaining his facial bruises, leg wound and outer garment. His contact card remains concealed in his shoe. 앰버: She is hiding behind the alley wall after the escape run. 라울: He is hiding behind the alley wall, breathless from running. 페드로: He is hiding behind the alley wall, breathless from running.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 골목 모퉁이 벽면에 이현우, 앰버, 라울, 페드로, 찰리가 등을 바짝 기댄 채 숨죽인 긴장된 자세.\n\nLOCATION (lock): Behind a wall corner in a secluded refugee-settlement alley away from the reception lot, outdoors at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Outer corner of the sheltering alley wall in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Alley corner wall (The group is concealed behind the corner with their backs against its adjoining face) — The sheltering wall recedes obliquely across the image, with its outer corner at the left edge; used as Establishes the shared hiding surface and blocks a direct view of the militia's position; Concealed alley floor (Occupied by the five companions after their hurried stop); used as Keeps their differing foot placements and compressed spacing visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime visibility and restrained contrast preserve the concealed alley's tension without specifying an unsupported light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement remains illuminated after the blackout reversal, with some bulbs burst. Charlie is concealed behind an alley wall in his old coat and hat; the manhole cover in the rubbish dump has not yet been moved. 이현우: He is hiding behind the alley wall after running, retaining his facial bruises, leg wound and outer garment. His contact card remains concealed in his shoe. 앰버: She is hiding behind the alley wall after the escape run. 라울: He is hiding behind the alley wall, breathless from running. 페드로: He is hiding behind the alley wall, breathless from running.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S25sh29__bgfirst_bg.png",
     "asset_id": "a7e553a8-5eb4-4796-bad9-a8df7e6955f1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S25sh29.png",
     "asset_id": "2e756d97-d5c0-4dc9-88d5-254ef25684d3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:826925>",
     "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1597480>",
     "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_outdoor_reception_sel.png",
     "asset_id": "24131eeb-c517-44d7-890e-15c4f0bbe013",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:826925>",
     "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1597480>",
     "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "캐릭터들은 골목 밖을 향해 시선을 던지고 있으며, 몸통은 앞으로 기울어져 있음.",
    "built_space": "화면 중앙 좌측의 벽을 기준으로 인물들이 서 있으며, 왼쪽으로 깊게 파인 골목 안쪽 배경에 민병대로 보이는 두 명의 추가 인물이 배치되어 시야 차단 지시를 위반함.",
    "entities": "이현우, 앰버, 라울, 찰리의 복장 및 외형은 대체로 부합하나, 페드로가 레퍼런스에 명시된 비니를 착용하지 않은 상태로 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 명시된 액션('등을 바짝 기댄 채') 위반: 모든 인물이 벽에 등을 기대지 않고 앞으로 체중을 실어 손이나 어깨로 벽을 짚고 있음.",
     "[gemini-pro] 캐릭터 외형 위반: 페드로의 고정된 의상인 비니가 누락됨.",
     "[gemini-pro] 프레이밍 지시 위반: '민병대의 위치를 직접적으로 볼 수 없게 가린다'는 지시를 무시하고 왼쪽 골목 배경에 직접적으로 2명의 인물을 노출함.",
     "[gpt-high] 숏에 허용된 다섯 인물 외에 왼쪽 골목 끝에 두 사람을 추가했다."
    ],
    "physics": "발은 땅에 닿아 있으나, 등을 기대어 숨죽인 자세가 아니라 앞으로 나아가려는 듯 체중이 앞쪽(벽면을 짚은 손과 무릎)으로 쏠려 있어 프롬프트의 상황에 맞지 않음."
   },
   {
    "label": "B",
    "direction": "캐릭터들의 시선과 얼굴 방향이 프레임 왼쪽의 골목 모퉁이 너머를 향해 집중되어 있음.",
    "built_space": "화면 중앙 좌측에 골목 모퉁이가 위치하고 오른쪽으로 은신처 벽이 이어짐. 배경 우측에는 레퍼런스 이미지에서 가져온 특정 형태의 가로등과 아파트 단지가 자연스럽게 배치됨.",
    "entities": "이현우(상처 입은 얼굴, 후드티), 앰버(금발, 원피스), 페드로(비니 착용), 라울(어두운 피부, 초록색 티셔츠), 찰리(고릴라 머리, 코트와 모자, 가슴 발광부)의 특징이 모두 프롬프트 및 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "이현우와 좁은 공간의 페드로는 벽에 등을 대고 체중을 지탱하고 있으며, 다른 인물들도 좁은 골목 바닥에 발을 단단히 디딘 채 자연스럽게 서서 밀착된 포즈를 유지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "인물들이 벽에 '등을 바짝 기댄' 긴장된 자세를 지시문대로 완벽히 구현했으며, 페드로의 비니를 포함한 캐릭터 외형과 레퍼런스의 배경 요소(가로등, 아파트)를 훌륭하게 조화시킨 결과물입니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "명시된 '등을 바짝 기댄' 자세 대신 벽에 손을 짚고 몸을 앞으로 숙인 형태를 취했으며, 페드로의 비니가 누락되고 시야가 차단되어야 할 배경에 다른 인물들이 직접 노출되는 위반 사항이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "캐릭터들의 시선과 얼굴 방향이 프레임 왼쪽의 골목 모퉁이 너머를 향해 집중되어 있음.",
        "built_space": "화면 중앙 좌측에 골목 모퉁이가 위치하고 오른쪽으로 은신처 벽이 이어짐. 배경 우측에는 레퍼런스 이미지에서 가져온 특정 형태의 가로등과 아파트 단지가 자연스럽게 배치됨.",
        "entities": "이현우(상처 입은 얼굴, 후드티), 앰버(금발, 원피스), 페드로(비니 착용), 라울(어두운 피부, 초록색 티셔츠), 찰리(고릴라 머리, 코트와 모자, 가슴 발광부)의 특징이 모두 프롬프트 및 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "이현우와 좁은 공간의 페드로는 벽에 등을 대고 체중을 지탱하고 있으며, 다른 인물들도 좁은 골목 바닥에 발을 단단히 디딘 채 자연스럽게 서서 밀착된 포즈를 유지함."
       },
       {
        "label": "A",
        "direction": "캐릭터들은 골목 밖을 향해 시선을 던지고 있으며, 몸통은 앞으로 기울어져 있음.",
        "built_space": "화면 중앙 좌측의 벽을 기준으로 인물들이 서 있으며, 왼쪽으로 깊게 파인 골목 안쪽 배경에 민병대로 보이는 두 명의 추가 인물이 배치되어 시야 차단 지시를 위반함.",
        "entities": "이현우, 앰버, 라울, 찰리의 복장 및 외형은 대체로 부합하나, 페드로가 레퍼런스에 명시된 비니를 착용하지 않은 상태로 묘사됨.",
        "hard_violations": [
         "명시된 액션('등을 바짝 기댄 채') 위반: 모든 인물이 벽에 등을 기대지 않고 앞으로 체중을 실어 손이나 어깨로 벽을 짚고 있음.",
         "캐릭터 외형 위반: 페드로의 고정된 의상인 비니가 누락됨.",
         "프레이밍 지시 위반: '민병대의 위치를 직접적으로 볼 수 없게 가린다'는 지시를 무시하고 왼쪽 골목 배경에 직접적으로 2명의 인물을 노출함."
        ],
        "physics": "발은 땅에 닿아 있으나, 등을 기대어 숨죽인 자세가 아니라 앞으로 나아가려는 듯 체중이 앞쪽(벽면을 짚은 손과 무릎)으로 쏠려 있어 프롬프트의 상황에 맞지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "인물들이 벽에 '등을 바짝 기댄' 긴장된 자세를 지시문대로 완벽히 구현했으며, 페드로의 비니를 포함한 캐릭터 외형과 레퍼런스의 배경 요소(가로등, 아파트)를 훌륭하게 조화시킨 결과물입니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "명시된 '등을 바짝 기댄' 자세 대신 벽에 손을 짚고 몸을 앞으로 숙인 형태를 취했으며, 페드로의 비니가 누락되고 시야가 차단되어야 할 배경에 다른 인물들이 직접 노출되는 위반 사항이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "캐릭터들의 시선과 얼굴 방향이 프레임 왼쪽의 골목 모퉁이 너머를 향해 집중되어 있음.",
        "built_space": "화면 중앙 좌측에 골목 모퉁이가 위치하고 오른쪽으로 은신처 벽이 이어짐. 배경 우측에는 레퍼런스 이미지에서 가져온 특정 형태의 가로등과 아파트 단지가 자연스럽게 배치됨.",
        "entities": "이현우(상처 입은 얼굴, 후드티), 앰버(금발, 원피스), 페드로(비니 착용), 라울(어두운 피부, 초록색 티셔츠), 찰리(고릴라 머리, 코트와 모자, 가슴 발광부)의 특징이 모두 프롬프트 및 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "이현우와 좁은 공간의 페드로는 벽에 등을 대고 체중을 지탱하고 있으며, 다른 인물들도 좁은 골목 바닥에 발을 단단히 디딘 채 자연스럽게 서서 밀착된 포즈를 유지함."
       },
       {
        "label": "A",
        "direction": "캐릭터들은 골목 밖을 향해 시선을 던지고 있으며, 몸통은 앞으로 기울어져 있음.",
        "built_space": "화면 중앙 좌측의 벽을 기준으로 인물들이 서 있으며, 왼쪽으로 깊게 파인 골목 안쪽 배경에 민병대로 보이는 두 명의 추가 인물이 배치되어 시야 차단 지시를 위반함.",
        "entities": "이현우, 앰버, 라울, 찰리의 복장 및 외형은 대체로 부합하나, 페드로가 레퍼런스에 명시된 비니를 착용하지 않은 상태로 묘사됨.",
        "hard_violations": [
         "명시된 액션('등을 바짝 기댄 채') 위반: 모든 인물이 벽에 등을 기대지 않고 앞으로 체중을 실어 손이나 어깨로 벽을 짚고 있음.",
         "캐릭터 외형 위반: 페드로의 고정된 의상인 비니가 누락됨.",
         "프레이밍 지시 위반: '민병대의 위치를 직접적으로 볼 수 없게 가린다'는 지시를 무시하고 왼쪽 골목 배경에 직접적으로 2명의 인물을 노출함."
        ],
        "physics": "발은 땅에 닿아 있으나, 등을 기대어 숨죽인 자세가 아니라 앞으로 나아가려는 듯 체중이 앞쪽(벽면을 짚은 손과 무릎)으로 쏠려 있어 프롬프트의 상황에 맞지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "다섯 인물만 등장하고 왼쪽 전경 모퉁이와 은폐된 와이드 구도는 맞지만, 전원이 등을 벽에 바짝 붙인 배치와 장소 참조의 고정 구조는 재현하지 못했다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "황혼과 긴장된 정지 동작은 표현했지만, 골목에 두 인물을 추가하고 위협 방향을 직접 노출하여 인원 제한과 은폐 구도를 위반했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 몸을 조금 내밀어 왼쪽 모퉁이 너머를 보고, 앰버·라울·페드로도 대체로 같은 왼쪽 방향을 경계한다. 찰리의 얼굴도 왼쪽 전방을 향한다. 시선의 대상은 화면 밖에 있으며 민병대는 보이지 않는다. 무기나 겨누는 소품은 없다.",
        "built_space": "큰 건물 모퉁이가 화면 중간 왼쪽 전경에 있고, 인접 벽은 오른쪽 뒤로 이어진다. 외벽 계량기 형태의 상자 세 개, 창문과 출입구, 뒤쪽 가로등 한 개, 바닥의 닫힌 맨홀 뚜껑 한 개가 보인다. 다섯 인물은 모퉁이 오른쪽에 압축되어 있지만, 이현우 뒤로 겹쳐 선 배치라 모두의 등이 같은 벽면에 밀착했다고 보기는 어렵다. 참조의 돌담·기와 마감과 열린 흙바닥 대신 높은 미장 건물과 포장 골목을 보여 주어 동일 장소의 구조를 확인할 수 없다.",
        "entities": "이현우·앰버·라울·페드로·찰리로 구분되는 다섯 인물만 보인다. 이현우의 젊은 한국계 남성 외모, 헝클어진 검은 머리, 어두운 후드와 찢어진 청바지는 부합하며 얼굴에 상처가 보이지만 다리 상처는 분명하지 않다. 앰버는 금발 여자아이이고 회색 원피스와 연녹색 겉옷은 맞지만 참조의 마스크가 없다. 라울은 짙은 피부의 어린아이로 올리브색 상의와 짧은 청바지를 입었으며 뒤로 묶은 머리는 잘 확인되지 않는다. 페드로는 젊은 남성으로 검은 비니를 썼지만 몸이 가려져 의복 비교가 제한된다. 찰리는 낡은 코트와 모자를 착용한 기계 인물이지만, 지정된 흰 얼굴판·점 눈·선 입 대신 어두운 고릴라형 얼굴이다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "이현우·앰버·라울·찰리의 보이는 신발은 바닥에 닿아 있고 몸의 하중을 지탱한다. 페드로의 하체는 다른 인물에게 가려져 접지를 직접 확인하기 어렵지만 공중에 떠 있는 정황은 없다. 앰버의 손은 이현우의 팔에, 찰리의 손은 라울의 어깨 부근에 놓여 있다. 코트와 의복은 몸에서 자연스럽게 늘어지며 맨홀 뚜껑과 쓰레기는 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽 골목을 살피며 모퉁이에 손을 댄다. 앰버·라울·페드로도 이현우 또는 그 너머 왼쪽을 보고, 찰리는 고개를 왼쪽 아래로 기울인다. 왼쪽 골목 끝에는 두 사람이 보이므로 경계 대상이 화면 안에 직접 드러난다. 이는 벽이 민병대 위치의 직접적인 시야를 차단한다는 구도와 맞지 않는다.",
        "built_space": "모퉁이는 화면 중간 왼쪽에 있고 넓은 미장 벽이 오른쪽으로 뻗는다. 오른쪽에 수직 배관 두 개와 쓰레기통 한 개, 왼쪽 통로에 최소 세 개의 켜진 벽등과 닫힌 맨홀 뚜껑 한 개가 보인다. 다섯 주인공은 벽을 따라 나란히 배치되어 발 위치가 잘 보이지만, 이현우는 벽에 등을 대기보다 옆으로 돌아 손을 짚고 있고 나머지도 상체를 앞으로 숙였다. 왼쪽 통로를 멀리까지 열어 보여 은폐 효과가 약하다. 참조의 돌담·기와·흙바닥과 달라 고정된 장소 구조는 일치하지 않는다.",
        "entities": "전경의 다섯 주인공 외에 왼쪽 골목 끝에 별도의 인물 두 명이 있다. 이현우는 검은 머리의 젊은 한국계 남성으로 후드와 청바지, 얼굴의 멍이 부합한다. 앰버는 금발 여자아이로 원피스와 겉옷은 맞지만 마스크가 없다. 라울은 짙은 피부의 어린아이이며 뒤로 묶은 머리, 올리브색 티셔츠와 짧은 청바지가 부합한다. 페드로는 젊은 남성이고 가죽 재킷과 찢어진 검은 바지는 맞지만 참조의 비니가 없다. 찰리는 코트와 모자를 쓴 기계 인물이지만 지정된 흰 얼굴판과 단순한 눈·입이 아니라 어두운 얼굴과 밝은 파란 눈으로 표현되었다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "숏에 허용된 다섯 인물 외에 왼쪽 골목 끝에 두 사람을 추가했다."
        ],
        "physics": "다섯 주인공 모두 신발이 바닥에 닿아 있다. 이현우는 벌린 다리와 벽을 짚은 손으로 몸을 지탱하고, 라울과 페드로는 무릎을 굽혀 달리다 멈춘 자세를 취한다. 찰리도 양발로 서 있으며 손은 몸 앞에 내려져 있다. 추가된 원경 인물들도 지면에 서 있다. 지지 없는 신체나 공중에 뜬 물건은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "다섯 인물만 등장하고 왼쪽 전경 모퉁이와 은폐된 와이드 구도는 맞지만, 전원이 등을 벽에 바짝 붙인 배치와 장소 참조의 고정 구조는 재현하지 못했다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "황혼과 긴장된 정지 동작은 표현했지만, 골목에 두 인물을 추가하고 위협 방향을 직접 노출하여 인원 제한과 은폐 구도를 위반했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 몸을 조금 내밀어 왼쪽 모퉁이 너머를 보고, 앰버·라울·페드로도 대체로 같은 왼쪽 방향을 경계한다. 찰리의 얼굴도 왼쪽 전방을 향한다. 시선의 대상은 화면 밖에 있으며 민병대는 보이지 않는다. 무기나 겨누는 소품은 없다.",
        "built_space": "큰 건물 모퉁이가 화면 중간 왼쪽 전경에 있고, 인접 벽은 오른쪽 뒤로 이어진다. 외벽 계량기 형태의 상자 세 개, 창문과 출입구, 뒤쪽 가로등 한 개, 바닥의 닫힌 맨홀 뚜껑 한 개가 보인다. 다섯 인물은 모퉁이 오른쪽에 압축되어 있지만, 이현우 뒤로 겹쳐 선 배치라 모두의 등이 같은 벽면에 밀착했다고 보기는 어렵다. 참조의 돌담·기와 마감과 열린 흙바닥 대신 높은 미장 건물과 포장 골목을 보여 주어 동일 장소의 구조를 확인할 수 없다.",
        "entities": "이현우·앰버·라울·페드로·찰리로 구분되는 다섯 인물만 보인다. 이현우의 젊은 한국계 남성 외모, 헝클어진 검은 머리, 어두운 후드와 찢어진 청바지는 부합하며 얼굴에 상처가 보이지만 다리 상처는 분명하지 않다. 앰버는 금발 여자아이이고 회색 원피스와 연녹색 겉옷은 맞지만 참조의 마스크가 없다. 라울은 짙은 피부의 어린아이로 올리브색 상의와 짧은 청바지를 입었으며 뒤로 묶은 머리는 잘 확인되지 않는다. 페드로는 젊은 남성으로 검은 비니를 썼지만 몸이 가려져 의복 비교가 제한된다. 찰리는 낡은 코트와 모자를 착용한 기계 인물이지만, 지정된 흰 얼굴판·점 눈·선 입 대신 어두운 고릴라형 얼굴이다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "이현우·앰버·라울·찰리의 보이는 신발은 바닥에 닿아 있고 몸의 하중을 지탱한다. 페드로의 하체는 다른 인물에게 가려져 접지를 직접 확인하기 어렵지만 공중에 떠 있는 정황은 없다. 앰버의 손은 이현우의 팔에, 찰리의 손은 라울의 어깨 부근에 놓여 있다. 코트와 의복은 몸에서 자연스럽게 늘어지며 맨홀 뚜껑과 쓰레기는 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "이현우는 왼쪽 골목을 살피며 모퉁이에 손을 댄다. 앰버·라울·페드로도 이현우 또는 그 너머 왼쪽을 보고, 찰리는 고개를 왼쪽 아래로 기울인다. 왼쪽 골목 끝에는 두 사람이 보이므로 경계 대상이 화면 안에 직접 드러난다. 이는 벽이 민병대 위치의 직접적인 시야를 차단한다는 구도와 맞지 않는다.",
        "built_space": "모퉁이는 화면 중간 왼쪽에 있고 넓은 미장 벽이 오른쪽으로 뻗는다. 오른쪽에 수직 배관 두 개와 쓰레기통 한 개, 왼쪽 통로에 최소 세 개의 켜진 벽등과 닫힌 맨홀 뚜껑 한 개가 보인다. 다섯 주인공은 벽을 따라 나란히 배치되어 발 위치가 잘 보이지만, 이현우는 벽에 등을 대기보다 옆으로 돌아 손을 짚고 있고 나머지도 상체를 앞으로 숙였다. 왼쪽 통로를 멀리까지 열어 보여 은폐 효과가 약하다. 참조의 돌담·기와·흙바닥과 달라 고정된 장소 구조는 일치하지 않는다.",
        "entities": "전경의 다섯 주인공 외에 왼쪽 골목 끝에 별도의 인물 두 명이 있다. 이현우는 검은 머리의 젊은 한국계 남성으로 후드와 청바지, 얼굴의 멍이 부합한다. 앰버는 금발 여자아이로 원피스와 겉옷은 맞지만 마스크가 없다. 라울은 짙은 피부의 어린아이이며 뒤로 묶은 머리, 올리브색 티셔츠와 짧은 청바지가 부합한다. 페드로는 젊은 남성이고 가죽 재킷과 찢어진 검은 바지는 맞지만 참조의 비니가 없다. 찰리는 코트와 모자를 쓴 기계 인물이지만 지정된 흰 얼굴판과 단순한 눈·입이 아니라 어두운 얼굴과 밝은 파란 눈으로 표현되었다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "숏에 허용된 다섯 인물 외에 왼쪽 골목 끝에 두 사람을 추가했다."
        ],
        "physics": "다섯 주인공 모두 신발이 바닥에 닿아 있다. 이현우는 벌린 다리와 벽을 짚은 손으로 몸을 지탱하고, 라울과 페드로는 무릎을 굽혀 달리다 멈춘 자세를 취한다. 찰리도 양발로 서 있으며 손은 몸 앞에 내려져 있다. 추가된 원경 인물들도 지면에 서 있다. 지지 없는 신체나 공중에 뜬 물건은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.9,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.65,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 명시된 액션('등을 바짝 기댄 채') 위반: 모든 인물이 벽에 등을 기대지 않고 앞으로 체중을 실어 손이나 어깨로 벽을 짚고 있음.",
     "[gemini-pro] 캐릭터 외형 위반: 페드로의 고정된 의상인 비니가 누락됨.",
     "[gemini-pro] 프레이밍 지시 위반: '민병대의 위치를 직접적으로 볼 수 없게 가린다'는 지시를 무시하고 왼쪽 골목 배경에 직접적으로 2명의 인물을 노출함.",
     "[gpt-high] 숏에 허용된 다섯 인물 외에 왼쪽 골목 끝에 두 사람을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 650
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "인물들이 벽에 '등을 바짝 기댄' 긴장된 자세를 지시문대로 완벽히 구현했으며, 페드로의 비니를 포함한 캐릭터 외형과 레퍼런스의 배경 요소(가로등, 아파트)를 훌륭하게 조화시킨 결과물입니다."
   },
   {
    "label": "A",
    "score": 650,
    "verdict_ko": "명시된 '등을 바짝 기댄' 자세 대신 벽에 손을 짚고 몸을 앞으로 숙인 형태를 취했으며, 페드로의 비니가 누락되고 시야가 차단되어야 할 배경에 다른 인물들이 직접 노출되는 위반 사항이 발생했습니다.  ★위반: [gemini-pro] 명시된 액션('등을 바짝 기댄 채') 위반: 모든 인물이 벽에 등을 기대지 않고 앞으로 체중을 실어 손이나 어깨로 벽을 짚고 있음. / [gemini-pro] 캐릭터 외형 위반: 페드로의 고정된 의상인 비니가 누락됨. / [gemini-pro] 프레이밍 지시 위반: '민병대의 위치를 직접적으로 볼 수 없게 가린다'는 지시를 무시하고 왼쪽 골목 배경에 직접적으로 2명의 인물을 노출함. / [gpt-high] 숏에 허용된 다섯 인물 외에 왼쪽 골목 끝에 두 사람을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_outdoor_reception_sel.png",
    "asset_id": "24131eeb-c517-44d7-890e-15c4f0bbe013",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:943979>",
    "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:826925>",
    "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1597480>",
    "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda3-a21d-7217-9a84-4018652200f2",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S25sh29__bgfirst_bg.png",
   "bg_asset_id": "a7e553a8-5eb4-4796-bad9-a8df7e6955f1",
   "bg_record_key": "S25sh29::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S25sh29::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:29:14.230519+00:00",
  "fingerprint": "5dd6853710f11668f95b88c3c4e416bcd3db70c5e3ac10a7ab84328bf7e427b9",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S25sh29_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S25sh29_sel.png",
  "source_sha256": "d61f195822c4d42a54dfec2cdccfb70a235c7188a8f1bedfa365146ffac1b7fc",
  "file": "S25sh29_cine.png",
  "staged_sha256": "24e3c9bbefbf9e8241413257062946042178ebb21aad8c0d253cdf4f8ef874af",
  "latency_ms": 13689
 },
 "S26sh5::signage": {
  "fp": "c0a868bcf02ca104",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S26sh5": {
  "input_fingerprint": "610a9ee7adb5efae",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진의 커다란 손아귀가 아시아계 취객 난민의 턱을 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): In the center of the outdoor reception lot, where militia guards hold the gathered wedding guests at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reception ground (The refugee has been forced to kneel before 박철진); used as A softly resolved lower background establishes the captive's lowered position without introducing additional action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate nighttime illumination keeps the gripping fingers and compressed facial contours readable with taut, restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception retains its tent, jukebox and restored lighting, including the bulbs already burst by the power surge. The rubbish-dump manhole cover has been pushed aside, and Charlie is moving through the sewer in his old coat and hat. 박철진: He is conducting the interrogation in front of the gathered reception guests, with one hand extended in a gripping posture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진의 커다란 손아귀가 아시아계 취객 난민의 턱을 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): In the center of the outdoor reception lot, where militia guards hold the gathered wedding guests at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reception ground (The refugee has been forced to kneel before 박철진); used as A softly resolved lower background establishes the captive's lowered position without introducing additional action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate nighttime illumination keeps the gripping fingers and compressed facial contours readable with taut, restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception retains its tent, jukebox and restored lighting, including the bulbs already burst by the power surge. The rubbish-dump manhole cover has been pushed aside, and Charlie is moving through the sewer in his old coat and hat. 박철진: He is conducting the interrogation in front of the gathered reception guests, with one hand extended in a gripping posture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진의 커다란 손아귀가 아시아계 취객 난민의 턱을 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): In the center of the outdoor reception lot, where militia guards hold the gathered wedding guests at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reception ground (The refugee has been forced to kneel before 박철진); used as A softly resolved lower background establishes the captive's lowered position without introducing additional action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate nighttime illumination keeps the gripping fingers and compressed facial contours readable with taut, restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception retains its tent, jukebox and restored lighting, including the bulbs already burst by the power surge. The rubbish-dump manhole cover has been pushed aside, and Charlie is moving through the sewer in his old coat and hat. 박철진: He is conducting the interrogation in front of the gathered reception guests, with one hand extended in a gripping posture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "군복을 입은 팔이 프레임 밖에서 뻗어 나와 꿇어앉은 난민의 턱을 빈틈없이 꽉 쥐고 있으며, 난민의 시선은 멍하게 허공을 향합니다.",
    "built_space": "야외 수용소 환경으로, 이전 샷과 일치하는 트럭 바퀴, 불이 켜진 주크박스, 천막이 흐릿한 배경에 정확히 배치되어 있습니다.",
    "entities": "레퍼런스와 일치하는 군복 및 완장을 착용한 박철진(팔과 몸통 일부만 노출)과 아시아계 난민 남성이 등장하며, 다른 인물은 없습니다.",
    "hard_violations": [
     "[gpt-high] 읽을 수 있는 글자를 금지했는데 완장의 흰 한글 및 로마자 일부가 판독 가능하게 보인다."
    ],
    "physics": "무릎을 꿇은 난민의 체중이 지면에 잘 지탱되어 있고, 턱을 거머쥔 손의 밀착감과 힘이 물리적으로 자연스럽게 작용합니다."
   },
   {
    "label": "B",
    "direction": "박철진이 꿇어앉은 난민의 턱을 잡고 내려다보고 있으며, 난민은 얼굴을 찡그린 채 위를 올려다봅니다.",
    "built_space": "트럭, 주크박스, 천막 등 이전 샷의 요소들이 배경에 보이지만, 불필요하게 넓은 앵글로 인해 환경이 넓게 노출되었습니다.",
    "entities": "박철진(얼굴과 복장 묘사 일치)과 아시아계 난민이 중앙에 위치하며, 배경에 샷 텍스트에 언급되지 않은 무장 경비병 2명이 서 있습니다.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 인물(배경의 무장 경비병 2명) 임의 추가",
     "[gpt-high] 샷 텍스트에 없는 무장 경비 두 명을 배경에 추가했다.",
     "[gpt-high] 읽을 수 있는 글자를 금지했는데 완장의 흰 한글 일부가 판독 가능하게 보인다."
    ],
    "physics": "두 인물 모두 지면에 닿아 무게 중심이 정상적으로 지지되며, 손으로 턱을 쥔 접촉 부위도 물리적으로 타당합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "요청된 클로즈업 프레이밍을 정확히 구현하였으며, 샷 텍스트에 명시된 인물 외에 다른 사람을 추가하지 않아 지시사항을 훌륭히 준수했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지정된 클로즈업이 아닌 미디엄 샷으로 앵글을 넓혔으며, 샷 텍스트에 없는 배경 인물(경비병 2명)을 임의로 추가하여 치명적인 위반을 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "군복을 입은 팔이 프레임 밖에서 뻗어 나와 꿇어앉은 난민의 턱을 빈틈없이 꽉 쥐고 있으며, 난민의 시선은 멍하게 허공을 향합니다.",
        "built_space": "야외 수용소 환경으로, 이전 샷과 일치하는 트럭 바퀴, 불이 켜진 주크박스, 천막이 흐릿한 배경에 정확히 배치되어 있습니다.",
        "entities": "레퍼런스와 일치하는 군복 및 완장을 착용한 박철진(팔과 몸통 일부만 노출)과 아시아계 난민 남성이 등장하며, 다른 인물은 없습니다.",
        "hard_violations": [],
        "physics": "무릎을 꿇은 난민의 체중이 지면에 잘 지탱되어 있고, 턱을 거머쥔 손의 밀착감과 힘이 물리적으로 자연스럽게 작용합니다."
       },
       {
        "label": "B",
        "direction": "박철진이 꿇어앉은 난민의 턱을 잡고 내려다보고 있으며, 난민은 얼굴을 찡그린 채 위를 올려다봅니다.",
        "built_space": "트럭, 주크박스, 천막 등 이전 샷의 요소들이 배경에 보이지만, 불필요하게 넓은 앵글로 인해 환경이 넓게 노출되었습니다.",
        "entities": "박철진(얼굴과 복장 묘사 일치)과 아시아계 난민이 중앙에 위치하며, 배경에 샷 텍스트에 언급되지 않은 무장 경비병 2명이 서 있습니다.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 인물(배경의 무장 경비병 2명) 임의 추가"
        ],
        "physics": "두 인물 모두 지면에 닿아 무게 중심이 정상적으로 지지되며, 손으로 턱을 쥔 접촉 부위도 물리적으로 타당합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "요청된 클로즈업 프레이밍을 정확히 구현하였으며, 샷 텍스트에 명시된 인물 외에 다른 사람을 추가하지 않아 지시사항을 훌륭히 준수했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지정된 클로즈업이 아닌 미디엄 샷으로 앵글을 넓혔으며, 샷 텍스트에 없는 배경 인물(경비병 2명)을 임의로 추가하여 치명적인 위반을 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "군복을 입은 팔이 프레임 밖에서 뻗어 나와 꿇어앉은 난민의 턱을 빈틈없이 꽉 쥐고 있으며, 난민의 시선은 멍하게 허공을 향합니다.",
        "built_space": "야외 수용소 환경으로, 이전 샷과 일치하는 트럭 바퀴, 불이 켜진 주크박스, 천막이 흐릿한 배경에 정확히 배치되어 있습니다.",
        "entities": "레퍼런스와 일치하는 군복 및 완장을 착용한 박철진(팔과 몸통 일부만 노출)과 아시아계 난민 남성이 등장하며, 다른 인물은 없습니다.",
        "hard_violations": [],
        "physics": "무릎을 꿇은 난민의 체중이 지면에 잘 지탱되어 있고, 턱을 거머쥔 손의 밀착감과 힘이 물리적으로 자연스럽게 작용합니다."
       },
       {
        "label": "B",
        "direction": "박철진이 꿇어앉은 난민의 턱을 잡고 내려다보고 있으며, 난민은 얼굴을 찡그린 채 위를 올려다봅니다.",
        "built_space": "트럭, 주크박스, 천막 등 이전 샷의 요소들이 배경에 보이지만, 불필요하게 넓은 앵글로 인해 환경이 넓게 노출되었습니다.",
        "entities": "박철진(얼굴과 복장 묘사 일치)과 아시아계 난민이 중앙에 위치하며, 배경에 샷 텍스트에 언급되지 않은 무장 경비병 2명이 서 있습니다.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 인물(배경의 무장 경비병 2명) 임의 추가"
        ],
        "physics": "두 인물 모두 지면에 닿아 무게 중심이 정상적으로 지지되며, 손으로 턱을 쥔 접촉 부위도 물리적으로 타당합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "턱을 감싸는 손과 박철진의 외형은 맞지만, 클로즈업보다 넓은 구도에 불필요한 무장 인물 두 명과 읽히는 완장 글자가 들어갔다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "손과 난민의 얼굴에 집중한 클로즈업과 낮아진 위치는 더 정확하지만, 꽉 틀어쥐기보다 턱을 받치는 동작이며 완장 글자 노출도 위반이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진은 아래의 난민 얼굴을 내려다보고, 난민은 화면 왼쪽 위 박철진 쪽으로 시선을 올린다. 손은 난민의 턱과 양쪽 볼을 감싼다. 뒤쪽 오른쪽 경비의 총구는 화면 오른쪽 아래 지면 쪽을 향하며, 다른 경비의 무기 방향은 흐려 확정하기 어렵다.",
        "built_space": "왼쪽에 군용 트럭 한 대와 열린 문 하나, 오른쪽에 밝은 천막 한 동과 주크박스 한 대가 보인다. 중앙 뒤에는 두 등으로 된 조명과 따뜻한 보조등이 있고, 자갈 바닥이 전경까지 이어져 이전 장소의 주요 재료와 설비를 유지한다. 그러나 두 인물의 상반신과 뒤쪽 경비 두 명까지 보여 손과 턱 중심의 클로즈업보다 넓다.",
        "entities": "박철진의 중년 남성 얼굴, 희끗한 수염, 군모와 올리브색 군복은 참조에 가깝다. 난민은 검은 머리의 아시아계 성인 남성으로 보이며 흐트러진 옷과 벌어진 입이 취기나 무력함을 암시한다. 샷 텍스트가 등장시키지 않은 무장 경비 두 명이 추가되었다. 완장에 흰 한글 일부가 읽히게 노출된다. 프레임 밖의 하객·하수도·찰리는 확인 대상이 아니다.",
        "hard_violations": [
         "샷 텍스트에 없는 무장 경비 두 명을 배경에 추가했다.",
         "읽을 수 있는 글자를 금지했는데 완장의 흰 한글 일부가 판독 가능하게 보인다."
        ],
        "physics": "잡는 손은 소매에서 이어진 손목과 팔에 연결되어 있고, 손바닥과 손가락이 턱 및 볼에 접촉한다. 난민의 머리는 목에 정상적으로 연결되어 있으며 박철진의 손이 얼굴 방향을 제어한다. 무릎과 발은 프레임 밖이라 실제 무릎 접촉은 확인할 수 없지만 공중에 떠 있는 모습은 아니다. 배경 경비들은 발을 지면에 딛고 무기를 손으로 들고 있다. 얼굴 압박은 보이나 빈틈없이 강하게 조인 변형은 제한적이다."
       },
       {
        "label": "B",
        "direction": "난민은 화면 왼쪽 위, 프레임 밖 박철진의 얼굴이 있을 방향을 본다. 박철진의 팔은 왼쪽 위에서 내려와 난민의 턱 밑에 닿는다. 손가락은 턱 아래를 위로 받치며, 턱 양옆을 강하게 조이는 방향성은 약하다. 무기나 별도의 지향 물체는 보이지 않는다.",
        "built_space": "왼쪽 배경에는 군용 트럭 한 대의 차체와 바퀴 하나, 오른쪽에는 천막 한 동과 주크박스 한 대가 보인다. 따뜻한 등 하나와 천막의 밝은 조명이 밤의 자갈 마당을 비춘다. 난민의 얼굴과 손이 화면 대부분을 차지하고 아래쪽의 흐린 자갈 바닥이 낮아진 위치를 설명해 지정된 클로즈업에 더 가깝다. 추가 사람이나 중복 설비는 보이지 않는다.",
        "entities": "난민은 검은 머리와 낡은 어두운 겉옷을 가진 아시아계 성인 남성으로 보인다. 눈은 정상적인 홍채와 동공을 유지하지만 표정에서 취기가 강하게 드러나지는 않는다. 박철진은 몸통 일부와 팔·손만 보이고, 올리브색 군복과 검은 완장이 참조의 복장에 부합한다. 얼굴과 모자는 올바른 크롭 밖이므로 동일 인물 여부를 얼굴로 확인할 수 없다. 완장의 흰 한글 및 로마자 일부는 선명하게 노출된다.",
        "hard_violations": [
         "읽을 수 있는 글자를 금지했는데 완장의 흰 한글 및 로마자 일부가 판독 가능하게 보인다."
        ],
        "physics": "손목은 군복 소매에서 자연스럽게 이어지고, 펼친 손바닥과 굽힌 손가락이 턱 아래에 접촉해 머리를 들어 올리는 힘을 준다. 머리와 목의 연결 및 손의 지지는 물리적으로 가능하며 떠 있는 신체나 물체는 없다. 하체는 프레임 밖이라 무릎의 지면 접촉은 직접 확인되지 않는다. 다만 볼과 턱의 압축이 거의 없어 강압적으로 꽉 틀어쥔 순간보다는 턱을 받쳐 올리는 순간으로 읽힌다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "턱을 감싸는 손과 박철진의 외형은 맞지만, 클로즈업보다 넓은 구도에 불필요한 무장 인물 두 명과 읽히는 완장 글자가 들어갔다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "손과 난민의 얼굴에 집중한 클로즈업과 낮아진 위치는 더 정확하지만, 꽉 틀어쥐기보다 턱을 받치는 동작이며 완장 글자 노출도 위반이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박철진은 아래의 난민 얼굴을 내려다보고, 난민은 화면 왼쪽 위 박철진 쪽으로 시선을 올린다. 손은 난민의 턱과 양쪽 볼을 감싼다. 뒤쪽 오른쪽 경비의 총구는 화면 오른쪽 아래 지면 쪽을 향하며, 다른 경비의 무기 방향은 흐려 확정하기 어렵다.",
        "built_space": "왼쪽에 군용 트럭 한 대와 열린 문 하나, 오른쪽에 밝은 천막 한 동과 주크박스 한 대가 보인다. 중앙 뒤에는 두 등으로 된 조명과 따뜻한 보조등이 있고, 자갈 바닥이 전경까지 이어져 이전 장소의 주요 재료와 설비를 유지한다. 그러나 두 인물의 상반신과 뒤쪽 경비 두 명까지 보여 손과 턱 중심의 클로즈업보다 넓다.",
        "entities": "박철진의 중년 남성 얼굴, 희끗한 수염, 군모와 올리브색 군복은 참조에 가깝다. 난민은 검은 머리의 아시아계 성인 남성으로 보이며 흐트러진 옷과 벌어진 입이 취기나 무력함을 암시한다. 샷 텍스트가 등장시키지 않은 무장 경비 두 명이 추가되었다. 완장에 흰 한글 일부가 읽히게 노출된다. 프레임 밖의 하객·하수도·찰리는 확인 대상이 아니다.",
        "hard_violations": [
         "샷 텍스트에 없는 무장 경비 두 명을 배경에 추가했다.",
         "읽을 수 있는 글자를 금지했는데 완장의 흰 한글 일부가 판독 가능하게 보인다."
        ],
        "physics": "잡는 손은 소매에서 이어진 손목과 팔에 연결되어 있고, 손바닥과 손가락이 턱 및 볼에 접촉한다. 난민의 머리는 목에 정상적으로 연결되어 있으며 박철진의 손이 얼굴 방향을 제어한다. 무릎과 발은 프레임 밖이라 실제 무릎 접촉은 확인할 수 없지만 공중에 떠 있는 모습은 아니다. 배경 경비들은 발을 지면에 딛고 무기를 손으로 들고 있다. 얼굴 압박은 보이나 빈틈없이 강하게 조인 변형은 제한적이다."
       },
       {
        "label": "A",
        "direction": "난민은 화면 왼쪽 위, 프레임 밖 박철진의 얼굴이 있을 방향을 본다. 박철진의 팔은 왼쪽 위에서 내려와 난민의 턱 밑에 닿는다. 손가락은 턱 아래를 위로 받치며, 턱 양옆을 강하게 조이는 방향성은 약하다. 무기나 별도의 지향 물체는 보이지 않는다.",
        "built_space": "왼쪽 배경에는 군용 트럭 한 대의 차체와 바퀴 하나, 오른쪽에는 천막 한 동과 주크박스 한 대가 보인다. 따뜻한 등 하나와 천막의 밝은 조명이 밤의 자갈 마당을 비춘다. 난민의 얼굴과 손이 화면 대부분을 차지하고 아래쪽의 흐린 자갈 바닥이 낮아진 위치를 설명해 지정된 클로즈업에 더 가깝다. 추가 사람이나 중복 설비는 보이지 않는다.",
        "entities": "난민은 검은 머리와 낡은 어두운 겉옷을 가진 아시아계 성인 남성으로 보인다. 눈은 정상적인 홍채와 동공을 유지하지만 표정에서 취기가 강하게 드러나지는 않는다. 박철진은 몸통 일부와 팔·손만 보이고, 올리브색 군복과 검은 완장이 참조의 복장에 부합한다. 얼굴과 모자는 올바른 크롭 밖이므로 동일 인물 여부를 얼굴로 확인할 수 없다. 완장의 흰 한글 및 로마자 일부는 선명하게 노출된다.",
        "hard_violations": [
         "읽을 수 있는 글자를 금지했는데 완장의 흰 한글 및 로마자 일부가 판독 가능하게 보인다."
        ],
        "physics": "손목은 군복 소매에서 자연스럽게 이어지고, 펼친 손바닥과 굽힌 손가락이 턱 아래에 접촉해 머리를 들어 올리는 힘을 준다. 머리와 목의 연결 및 손의 지지는 물리적으로 가능하며 떠 있는 신체나 물체는 없다. 하체는 프레임 밖이라 무릎의 지면 접촉은 직접 확인되지 않는다. 다만 볼과 턱의 압축이 거의 없어 강압적으로 꽉 틀어쥔 순간보다는 턱을 받쳐 올리는 순간으로 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.775
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.525
   },
   "violations": {
    "B": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 인물(배경의 무장 경비병 2명) 임의 추가",
     "[gpt-high] 샷 텍스트에 없는 무장 경비 두 명을 배경에 추가했다.",
     "[gpt-high] 읽을 수 있는 글자를 금지했는데 완장의 흰 한글 일부가 판독 가능하게 보인다."
    ],
    "A": [
     "[gpt-high] 읽을 수 있는 글자를 금지했는데 완장의 흰 한글 및 로마자 일부가 판독 가능하게 보인다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 525
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "요청된 클로즈업 프레이밍을 정확히 구현하였으며, 샷 텍스트에 명시된 인물 외에 다른 사람을 추가하지 않아 지시사항을 훌륭히 준수했습니다.  ★위반: [gpt-high] 읽을 수 있는 글자를 금지했는데 완장의 흰 한글 및 로마자 일부가 판독 가능하게 보인다."
   },
   {
    "label": "B",
    "score": 525,
    "verdict_ko": "지정된 클로즈업이 아닌 미디엄 샷으로 앵글을 넓혔으며, 샷 텍스트에 없는 배경 인물(경비병 2명)을 임의로 추가하여 치명적인 위반을 범했습니다.  ★위반: [gemini-pro] 샷 텍스트에 명시되지 않은 인물(배경의 무장 경비병 2명) 임의 추가 / [gpt-high] 샷 텍스트에 없는 무장 경비 두 명을 배경에 추가했다. / [gpt-high] 읽을 수 있는 글자를 금지했는데 완장의 흰 한글 일부가 판독 가능하게 보인다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S25sh24_sel.png",
    "asset_id": "aa3108e8-a4e6-4351-a482-cf8206917efc",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda4-83b9-7756-bc2a-d8367eb7c0c4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh24"
  }
 },
 "S26sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:29:28.240765+00:00",
  "fingerprint": "596e4c7d1c7759ccbaa526a4c0d20198553a8f813c3ece34deed4e26d8b265fd",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S26sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S26sh5_sel.png",
  "source_sha256": "7efd7bc68d214225f87f88c631894056586e8c52d8c860e9c0401141a7e3c707",
  "file": "S26sh5_cine.png",
  "staged_sha256": "e507ff7e1f190346fd1a404366f1222f4568309614e5dd4c2f17f522e8aa7cb0",
  "latency_ms": 12335
 },
 "S26sh8::signage": {
  "fp": "cd7d524a9ae2b746",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S26sh8": {
  "input_fingerprint": "e5ef81779920766c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 고개를 비스듬히 꺾은 채 미연을 향해 비릿한 미소를 머금고 매섭게 쏘아보는 시선.\n\nLOCATION (lock): Within the guarded crowd at the outdoor refugee-settlement reception lot at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the interrogation's nighttime illumination and restrained contrast, allowing the revealed eyes and slight smile to carry the menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception retains its restored lighting and previously burst bulbs, while the rubbish-dump manhole remains uncovered. Charlie continues through the sewer in his old coat and hat. 박철진: He remains at the interrogation, his attention now directed toward the front row.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 고개를 비스듬히 꺾은 채 미연을 향해 비릿한 미소를 머금고 매섭게 쏘아보는 시선.\n\nLOCATION (lock): Within the guarded crowd at the outdoor refugee-settlement reception lot at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the interrogation's nighttime illumination and restrained contrast, allowing the revealed eyes and slight smile to carry the menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception retains its restored lighting and previously burst bulbs, while the rubbish-dump manhole remains uncovered. Charlie continues through the sewer in his old coat and hat. 박철진: He remains at the interrogation, his attention now directed toward the front row.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 고개를 비스듬히 꺾은 채 미연을 향해 비릿한 미소를 머금고 매섭게 쏘아보는 시선.\n\nLOCATION (lock): Within the guarded crowd at the outdoor refugee-settlement reception lot at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the interrogation's nighttime illumination and restrained contrast, allowing the revealed eyes and slight smile to carry the menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception retains its restored lighting and previously burst bulbs, while the rubbish-dump manhole remains uncovered. Charlie continues through the sewer in his old coat and hat. 박철진: He remains at the interrogation, his attention now directed toward the front row.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "박철진이 고개를 약간 기울인 채 전경에 등진 인물(미연으로 추정)을 향해 시선을 던지고 있음.",
    "built_space": "야외 난민 수용소 공터. 화면 좌측에 군용 트럭 일부가, 우측 배경에 조명이 켜진 천막이 배치됨.",
    "entities": "박철진의 얼굴은 이전 샷과 일치하나, 헌병의 녹색 재킷을 입고 있음. 전경 인물의 검은 소매에 흰색 글자가 보임.",
    "hard_violations": [
     "[gemini-pro] 다른 인물의 의상을 박철진에게 적용(복장 전이 금지 지침 위반)",
     "[gemini-pro] 전경 인물 소매에 읽을 수 있는 텍스트 노출(텍스트 누출)",
     "[gpt-high] 등장이 허용되지 않은 남성의 머리와 상체를 오른쪽 전경에 추가했으며, 이전 장면에서 제외하라고 한 인물의 외양도 이어졌다.",
     "[gpt-high] 글자를 읽을 수 없어야 한다는 지시와 달리 검은 완장에 판독 가능한 흰색 한글이 노출됐다."
    ],
    "physics": "모든 인물과 물체가 지면 및 신체에 자연스럽게 지지되어 있으며, 떠 있는 요소 없음."
   },
   {
    "label": "B",
    "direction": "박철진이 고개를 비스듬히 꺾은 채 화면 밖의 대상(미연)을 향해 날카롭게 시선을 고정하고 있음.",
    "built_space": "야외 공터. 이전 샷에서 확립된 주크박스 형태의 조명 구조물이 좌측 배경에 정확히 유지되어 있음.",
    "entities": "박철진의 얼굴 및 회색 셔츠가 이전 샷과 완벽히 일치함. 배경에 초점이 흐려진 군중(다수의 엑스트라)이 보임.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 배경 인물 다수 추가(extra bodies/invented people)",
     "[gpt-high] 이전 장면에서 재등장시키지 말라고 명시한 심문받던 남성의 얼굴과 복장을 중심 인물로 재현했다.",
     "[gpt-high] 박철진만 등장할 수 있는 화면에 여러 배경 인물을 추가했다."
    ],
    "physics": "모든 인물이 지면에 올바르게 서 있거나 자세를 유지하고 있으며, 물리적으로 어색하게 떠 있는 물체 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이전 장면의 의상(회색 셔츠) 및 배경 구조물 연속성을 완벽히 유지하고 요구된 비릿한 미소를 훌륭히 연출했으나, 텍스트에 없는 배경 군중이 포함됨."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "표정 연출이 덜 극적이며, 이전 장면 속 헌병의 의상을 대상에게 잘못 입혔고 전경 인물 소매에 텍스트 누출이 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진이 고개를 약간 기울인 채 전경에 등진 인물(미연으로 추정)을 향해 시선을 던지고 있음.",
        "built_space": "야외 난민 수용소 공터. 화면 좌측에 군용 트럭 일부가, 우측 배경에 조명이 켜진 천막이 배치됨.",
        "entities": "박철진의 얼굴은 이전 샷과 일치하나, 헌병의 녹색 재킷을 입고 있음. 전경 인물의 검은 소매에 흰색 글자가 보임.",
        "hard_violations": [
         "다른 인물의 의상을 박철진에게 적용(복장 전이 금지 지침 위반)",
         "전경 인물 소매에 읽을 수 있는 텍스트 노출(텍스트 누출)"
        ],
        "physics": "모든 인물과 물체가 지면 및 신체에 자연스럽게 지지되어 있으며, 떠 있는 요소 없음."
       },
       {
        "label": "B",
        "direction": "박철진이 고개를 비스듬히 꺾은 채 화면 밖의 대상(미연)을 향해 날카롭게 시선을 고정하고 있음.",
        "built_space": "야외 공터. 이전 샷에서 확립된 주크박스 형태의 조명 구조물이 좌측 배경에 정확히 유지되어 있음.",
        "entities": "박철진의 얼굴 및 회색 셔츠가 이전 샷과 완벽히 일치함. 배경에 초점이 흐려진 군중(다수의 엑스트라)이 보임.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 배경 인물 다수 추가(extra bodies/invented people)"
        ],
        "physics": "모든 인물이 지면에 올바르게 서 있거나 자세를 유지하고 있으며, 물리적으로 어색하게 떠 있는 물체 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이전 장면의 의상(회색 셔츠) 및 배경 구조물 연속성을 완벽히 유지하고 요구된 비릿한 미소를 훌륭히 연출했으나, 텍스트에 없는 배경 군중이 포함됨."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "표정 연출이 덜 극적이며, 이전 장면 속 헌병의 의상을 대상에게 잘못 입혔고 전경 인물 소매에 텍스트 누출이 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박철진이 고개를 약간 기울인 채 전경에 등진 인물(미연으로 추정)을 향해 시선을 던지고 있음.",
        "built_space": "야외 난민 수용소 공터. 화면 좌측에 군용 트럭 일부가, 우측 배경에 조명이 켜진 천막이 배치됨.",
        "entities": "박철진의 얼굴은 이전 샷과 일치하나, 헌병의 녹색 재킷을 입고 있음. 전경 인물의 검은 소매에 흰색 글자가 보임.",
        "hard_violations": [
         "다른 인물의 의상을 박철진에게 적용(복장 전이 금지 지침 위반)",
         "전경 인물 소매에 읽을 수 있는 텍스트 노출(텍스트 누출)"
        ],
        "physics": "모든 인물과 물체가 지면 및 신체에 자연스럽게 지지되어 있으며, 떠 있는 요소 없음."
       },
       {
        "label": "B",
        "direction": "박철진이 고개를 비스듬히 꺾은 채 화면 밖의 대상(미연)을 향해 날카롭게 시선을 고정하고 있음.",
        "built_space": "야외 공터. 이전 샷에서 확립된 주크박스 형태의 조명 구조물이 좌측 배경에 정확히 유지되어 있음.",
        "entities": "박철진의 얼굴 및 회색 셔츠가 이전 샷과 완벽히 일치함. 배경에 초점이 흐려진 군중(다수의 엑스트라)이 보임.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 배경 인물 다수 추가(extra bodies/invented people)"
        ],
        "physics": "모든 인물이 지면에 올바르게 서 있거나 자세를 유지하고 있으며, 물리적으로 어색하게 떠 있는 물체 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "고개를 기울인 얼굴 클로즈업과 비릿한 미소는 더 가깝지만, 박철진 대신 이전 장면의 심문받던 남성을 재현하고 금지된 배경 인물까지 추가했다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "군복은 일부 맞지만 미연이 아닌 전경 남성을 바라보는 어깨너머 구도이며, 제외된 인물과 판독 가능한 완장 글자까지 남아 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "주인공은 고개를 화면 오른쪽으로 기울이고 눈동자를 카메라 오른쪽 가까운 화면 밖으로 돌린다. 미연은 보이지 않아 실제 시선 도착점은 확인되지 않지만, 화면 밖 미연을 바라보는 배치는 가능하다. 입꼬리는 올라가 있으나 눈매의 매서움은 비교적 약하다. 배경 사람들의 시선은 흐려 판별하기 어렵다. 무기나 겨냥하는 소품은 없다.",
        "built_space": "왼쪽에 대형 차량 한 대의 일부와 바퀴 하나, 뒤쪽에 천막·임시 건물, 중앙 왼쪽에 노란 아치형 발광 설비 하나가 보인다. 오른쪽 위에는 한 기둥에 투광등 두 개가 달려 있다. 원본의 차량·자갈 바닥·야간 임시 시설이라는 재료감은 이어지지만, 원본에서 보인 단일 조명과 배치가 정확히 같다고 확인되지는 않는다. 얼굴 중심 클로즈업이며 배경 설비는 원거리 크기로 표현됐다.",
        "entities": "중앙에는 검은 머리와 짙은 회색 셔츠를 가진 중년 동아시아계 외양의 남성이 있다. 이는 모자, 회색 수염, 올리브색 군복을 갖춘 박철진 인물 참조가 아니라 이전 장면에서 턱을 붙잡힌 남성의 얼굴과 복장에 가깝다. 배경에는 왼쪽 한 명과 오른쪽 여러 명의 추가 인물이 보인다. 미연은 화면에 없으며 찰리·맨홀·하수도 등도 클로즈업 밖이므로 그 부재 자체는 결함이 아니다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "이전 장면에서 재등장시키지 말라고 명시한 심문받던 남성의 얼굴과 복장을 중심 인물로 재현했다.",
         "박철진만 등장할 수 있는 화면에 여러 배경 인물을 추가했다."
        ],
        "physics": "머리는 목에 정상적으로 연결되고 기울어진 어깨와 상체가 자세를 지탱한다. 하체는 클로즈업 밖이므로 발의 접지는 확인할 수 없으나 공중에 떠 있다는 징후는 없다. 배경 인물들은 지면에 서 있는 배치이며 차량 바퀴는 자갈 바닥에 놓여 있다. 손에 든 물건이나 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "군복 남성은 고개를 화면 왼쪽으로 크게 기울인 채 오른쪽 전경 남성의 얼굴을 직접 올려다본다. 시선의 도착점은 화면 밖 미연이 아니라 바로 앞의 짧은 검은 머리 남성이다. 전경 남성도 군복 남성을 마주하는 방향이며, 요청된 앞줄 미연을 향한 시선보다 두 남성 사이의 심문 대치로 읽힌다. 무기나 겨냥하는 물건은 없다.",
        "built_space": "왼쪽 배경에 차량 한 대의 차체와 큰 바퀴 하나, 바닥에 자갈, 뒤쪽에 따뜻한 투광등 하나와 임시 건물의 밝은 개구부가 보인다. 차량과 조명의 질감은 이전 장면과 가깝다. 군복 남성은 차량 앞에서 몸을 기울이고, 다른 남성의 머리와 어깨가 오른쪽 전경을 크게 차지한다. 얼굴 단독 클로즈업보다 두 사람의 어깨너머 근접 대화 구도로 구성됐다. 불가능한 반사나 명백하게 중복된 고정 설비는 없다.",
        "entities": "중심 남성은 중년 동아시아계 외양으로 올리브색 군복과 검은 완장을 착용한다. 군복은 박철진의 복장 방향에 맞지만, 참조의 얼굴·회색 수염·모자와는 다르고 노출된 검은 머리도 일치하지 않는다. 오른쪽 전경에는 이전 장면의 심문받던 남성을 연상시키는 검은 머리와 회색 옷의 남성이 추가됐다. 미연으로 확인되는 인물은 없다. 완장에는 흰색 한글 글자 일부가 판독 가능한 상태로 남아 있다.",
        "hard_violations": [
         "등장이 허용되지 않은 남성의 머리와 상체를 오른쪽 전경에 추가했으며, 이전 장면에서 제외하라고 한 인물의 외양도 이어졌다.",
         "글자를 읽을 수 없어야 한다는 지시와 달리 검은 완장에 판독 가능한 흰색 한글이 노출됐다."
        ],
        "physics": "군복 남성의 기울어진 머리는 목과 기울어진 몸통으로 이어져 실제로 가능한 전방 숙임 자세다. 두 인물의 하체는 화면 밖이므로 서 있는지 앉았는지는 확정할 수 없지만, 지지 없이 떠 있는 모습은 아니다. 차량 바퀴는 자갈 지면에 닿아 있다. 손에 든 소품이나 공중에 매달린 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "고개를 기울인 얼굴 클로즈업과 비릿한 미소는 더 가깝지만, 박철진 대신 이전 장면의 심문받던 남성을 재현하고 금지된 배경 인물까지 추가했다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "군복은 일부 맞지만 미연이 아닌 전경 남성을 바라보는 어깨너머 구도이며, 제외된 인물과 판독 가능한 완장 글자까지 남아 있다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "주인공은 고개를 화면 오른쪽으로 기울이고 눈동자를 카메라 오른쪽 가까운 화면 밖으로 돌린다. 미연은 보이지 않아 실제 시선 도착점은 확인되지 않지만, 화면 밖 미연을 바라보는 배치는 가능하다. 입꼬리는 올라가 있으나 눈매의 매서움은 비교적 약하다. 배경 사람들의 시선은 흐려 판별하기 어렵다. 무기나 겨냥하는 소품은 없다.",
        "built_space": "왼쪽에 대형 차량 한 대의 일부와 바퀴 하나, 뒤쪽에 천막·임시 건물, 중앙 왼쪽에 노란 아치형 발광 설비 하나가 보인다. 오른쪽 위에는 한 기둥에 투광등 두 개가 달려 있다. 원본의 차량·자갈 바닥·야간 임시 시설이라는 재료감은 이어지지만, 원본에서 보인 단일 조명과 배치가 정확히 같다고 확인되지는 않는다. 얼굴 중심 클로즈업이며 배경 설비는 원거리 크기로 표현됐다.",
        "entities": "중앙에는 검은 머리와 짙은 회색 셔츠를 가진 중년 동아시아계 외양의 남성이 있다. 이는 모자, 회색 수염, 올리브색 군복을 갖춘 박철진 인물 참조가 아니라 이전 장면에서 턱을 붙잡힌 남성의 얼굴과 복장에 가깝다. 배경에는 왼쪽 한 명과 오른쪽 여러 명의 추가 인물이 보인다. 미연은 화면에 없으며 찰리·맨홀·하수도 등도 클로즈업 밖이므로 그 부재 자체는 결함이 아니다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "이전 장면에서 재등장시키지 말라고 명시한 심문받던 남성의 얼굴과 복장을 중심 인물로 재현했다.",
         "박철진만 등장할 수 있는 화면에 여러 배경 인물을 추가했다."
        ],
        "physics": "머리는 목에 정상적으로 연결되고 기울어진 어깨와 상체가 자세를 지탱한다. 하체는 클로즈업 밖이므로 발의 접지는 확인할 수 없으나 공중에 떠 있다는 징후는 없다. 배경 인물들은 지면에 서 있는 배치이며 차량 바퀴는 자갈 바닥에 놓여 있다. 손에 든 물건이나 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "군복 남성은 고개를 화면 왼쪽으로 크게 기울인 채 오른쪽 전경 남성의 얼굴을 직접 올려다본다. 시선의 도착점은 화면 밖 미연이 아니라 바로 앞의 짧은 검은 머리 남성이다. 전경 남성도 군복 남성을 마주하는 방향이며, 요청된 앞줄 미연을 향한 시선보다 두 남성 사이의 심문 대치로 읽힌다. 무기나 겨냥하는 물건은 없다.",
        "built_space": "왼쪽 배경에 차량 한 대의 차체와 큰 바퀴 하나, 바닥에 자갈, 뒤쪽에 따뜻한 투광등 하나와 임시 건물의 밝은 개구부가 보인다. 차량과 조명의 질감은 이전 장면과 가깝다. 군복 남성은 차량 앞에서 몸을 기울이고, 다른 남성의 머리와 어깨가 오른쪽 전경을 크게 차지한다. 얼굴 단독 클로즈업보다 두 사람의 어깨너머 근접 대화 구도로 구성됐다. 불가능한 반사나 명백하게 중복된 고정 설비는 없다.",
        "entities": "중심 남성은 중년 동아시아계 외양으로 올리브색 군복과 검은 완장을 착용한다. 군복은 박철진의 복장 방향에 맞지만, 참조의 얼굴·회색 수염·모자와는 다르고 노출된 검은 머리도 일치하지 않는다. 오른쪽 전경에는 이전 장면의 심문받던 남성을 연상시키는 검은 머리와 회색 옷의 남성이 추가됐다. 미연으로 확인되는 인물은 없다. 완장에는 흰색 한글 글자 일부가 판독 가능한 상태로 남아 있다.",
        "hard_violations": [
         "등장이 허용되지 않은 남성의 머리와 상체를 오른쪽 전경에 추가했으며, 이전 장면에서 제외하라고 한 인물의 외양도 이어졌다.",
         "글자를 읽을 수 없어야 한다는 지시와 달리 검은 완장에 판독 가능한 흰색 한글이 노출됐다."
        ],
        "physics": "군복 남성의 기울어진 머리는 목과 기울어진 몸통으로 이어져 실제로 가능한 전방 숙임 자세다. 두 인물의 하체는 화면 밖이므로 서 있는지 앉았는지는 확정할 수 없지만, 지지 없이 떠 있는 모습은 아니다. 차량 바퀴는 자갈 지면에 닿아 있다. 손에 든 소품이나 공중에 매달린 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.071,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.821,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 다른 인물의 의상을 박철진에게 적용(복장 전이 금지 지침 위반)",
     "[gemini-pro] 전경 인물 소매에 읽을 수 있는 텍스트 노출(텍스트 누출)",
     "[gpt-high] 등장이 허용되지 않은 남성의 머리와 상체를 오른쪽 전경에 추가했으며, 이전 장면에서 제외하라고 한 인물의 외양도 이어졌다.",
     "[gpt-high] 글자를 읽을 수 없어야 한다는 지시와 달리 검은 완장에 판독 가능한 흰색 한글이 노출됐다."
    ],
    "B": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 배경 인물 다수 추가(extra bodies/invented people)",
     "[gpt-high] 이전 장면에서 재등장시키지 말라고 명시한 심문받던 남성의 얼굴과 복장을 중심 인물로 재현했다.",
     "[gpt-high] 박철진만 등장할 수 있는 화면에 여러 배경 인물을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 821
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "이전 장면의 의상(회색 셔츠) 및 배경 구조물 연속성을 완벽히 유지하고 요구된 비릿한 미소를 훌륭히 연출했으나, 텍스트에 없는 배경 군중이 포함됨.  ★위반: [gemini-pro] 샷 텍스트에 명시되지 않은 배경 인물 다수 추가(extra bodies/invented people) / [gpt-high] 이전 장면에서 재등장시키지 말라고 명시한 심문받던 남성의 얼굴과 복장을 중심 인물로 재현했다. / [gpt-high] 박철진만 등장할 수 있는 화면에 여러 배경 인물을 추가했다."
   },
   {
    "label": "A",
    "score": 821,
    "verdict_ko": "표정 연출이 덜 극적이며, 이전 장면 속 헌병의 의상을 대상에게 잘못 입혔고 전경 인물 소매에 텍스트 누출이 발생함.  ★위반: [gemini-pro] 다른 인물의 의상을 박철진에게 적용(복장 전이 금지 지침 위반) / [gemini-pro] 전경 인물 소매에 읽을 수 있는 텍스트 노출(텍스트 누출) / [gpt-high] 등장이 허용되지 않은 남성의 머리와 상체를 오른쪽 전경에 추가했으며, 이전 장면에서 제외하라고 한 인물의 외양도 이어졌다. / [gpt-high] 글자를 읽을 수 없어야 한다는 지시와 달리 검은 완장에 판독 가능한 흰색 한글이 노출됐다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S26sh5_sel.png",
    "asset_id": "903d84a8-4ed4-4faa-8d4f-ff9e3e5933d0",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda5-4e7d-775e-994a-04f24e30eb10",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S26sh5"
  }
 },
 "S26sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:29:40.914506+00:00",
  "fingerprint": "cbd91eee83e063951c0c44bdee81f83fc42f10e0ce14cea0d4be119fdaea8bc2",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S26sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S26sh8_sel.png",
  "source_sha256": "d71f1337ba1299a3ab0e957112f692da654a9e80175aac59a1991a89bba00640",
  "file": "S26sh8_cine.png",
  "staged_sha256": "bf2d147d11ce8ec0bba83c61cafefb592422947496ff724acd795f7343676066",
  "latency_ms": 12743
 },
 "S26sh12::signage": {
  "fp": "db9dcce025413c7a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::758bc4bb1aab5450": {
  "subjects": [],
  "subject_text": "인천 난민촌 지하 하수도\n맨홀 아래 사다리로 연결되는 어두운 지하 하수 통로. 길게 뻗은 통로 중간에 두 방향으로 갈라지는 분기점이 있다.",
  "identity": "canonical",
  "scope_id": "L32",
  "scope_role": "location_interior",
  "scope_sha": "aec19388934a494f"
 },
 "S26sh12::bgfirst_bg": {
  "input_fingerprint": "9eedb7359fc5df6a",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 기침하는 앰버를 자신의 등에 단단히 업은 채 좁은 하수도를 향해 한쪽 발로 바닥을 강하게 차내며 질주하는 미드액션 자세.\n\nLOCATION (lock): Inside a dark underground sewer passage beyond a fork beneath the refugee settlement, with only faint ambient visibility.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Sewer branch continuing in the escape direction in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Chosen sewer branch (The pair is entering this branch after separating from 페드로 and 라울) — The passage recedes from the foreground toward the upper-right background; used as Gives the push-off a readable destination and preserves the rear-side tracking route; Sewer floor (이현우's driving foot is pushing against it); used as Keeps the point of propulsion visible beneath the carried pair.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination separates the intertwined bodies from the sewer passage without adding an unsupported source, color, or atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 기침하는 앰버를 자신의 등에 단단히 업은 채 좁은 하수도를 향해 한쪽 발로 바닥을 강하게 차내며 질주하는 미드액션 자세.\n\nLOCATION (lock): Inside a dark underground sewer passage beyond a fork beneath the refugee settlement, with only faint ambient visibility.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Sewer branch continuing in the escape direction in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Chosen sewer branch (The pair is entering this branch after separating from 페드로 and 라울) — The passage recedes from the foreground toward the upper-right background; used as Gives the push-off a readable destination and preserves the rear-side tracking route; Sewer floor (이현우's driving foot is pushing against it); used as Keeps the point of propulsion visible beneath the carried pair.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination separates the intertwined bodies from the sewer passage without adding an unsupported source, color, or atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S26sh12__bgfirst_bg.png",
  "asset_id": "cbdab119-6360-4bbb-b01c-a8f50193bfec",
  "input_asset_ids": [
   "370c5e80-0275-4785-b540-af69ec16bd57",
   "ad0f9f4b-0a55-45e6-9c39-7261e4894a80"
  ]
 },
 "S26sh12": {
  "input_fingerprint": "8a613e49ffc72f8f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 기침하는 앰버를 자신의 등에 단단히 업은 채 좁은 하수도를 향해 한쪽 발로 바닥을 강하게 차내며 질주하는 미드액션 자세.\n\nLOCATION (lock): Inside a dark underground sewer passage beyond a fork beneath the refugee settlement, with only faint ambient visibility. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Sewer branch continuing in the escape direction in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Chosen sewer branch (The pair is entering this branch after separating from 페드로 and 라울) — The passage recedes from the foreground toward the upper-right background; used as Gives the push-off a readable destination and preserves the rear-side tracking route; Sewer floor (이현우's driving foot is pushing against it); used as Keeps the point of propulsion visible beneath the carried pair.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination separates the intertwined bodies from the sewer passage without adding an unsupported source, color, or atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer route has reached a fork, and the rubbish-dump entrance cover remains displaced. Charlie follows along the chosen branch in his old coat and hat. 이현우: He runs in a piggyback-carrying posture without his outer garment, retaining facial bruising and his injured leg. His contact card remains concealed in his shoe. 앰버: She is being carried piggyback, coughing more severely, with the outer garment still covering her mouth.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 기침하는 앰버를 자신의 등에 단단히 업은 채 좁은 하수도를 향해 한쪽 발로 바닥을 강하게 차내며 질주하는 미드액션 자세.\n\nLOCATION (lock): Inside a dark underground sewer passage beyond a fork beneath the refugee settlement, with only faint ambient visibility. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Sewer branch continuing in the escape direction in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Chosen sewer branch (The pair is entering this branch after separating from 페드로 and 라울) — The passage recedes from the foreground toward the upper-right background; used as Gives the push-off a readable destination and preserves the rear-side tracking route; Sewer floor (이현우's driving foot is pushing against it); used as Keeps the point of propulsion visible beneath the carried pair.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination separates the intertwined bodies from the sewer passage without adding an unsupported source, color, or atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer route has reached a fork, and the rubbish-dump entrance cover remains displaced. Charlie follows along the chosen branch in his old coat and hat. 이현우: He runs in a piggyback-carrying posture without his outer garment, retaining facial bruising and his injured leg. His contact card remains concealed in his shoe. 앰버: She is being carried piggyback, coughing more severely, with the outer garment still covering her mouth.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 기침하는 앰버를 자신의 등에 단단히 업은 채 좁은 하수도를 향해 한쪽 발로 바닥을 강하게 차내며 질주하는 미드액션 자세.\n\nLOCATION (lock): Inside a dark underground sewer passage beyond a fork beneath the refugee settlement, with only faint ambient visibility. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Sewer branch continuing in the escape direction in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Chosen sewer branch (The pair is entering this branch after separating from 페드로 and 라울) — The passage recedes from the foreground toward the upper-right background; used as Gives the push-off a readable destination and preserves the rear-side tracking route; Sewer floor (이현우's driving foot is pushing against it); used as Keeps the point of propulsion visible beneath the carried pair.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination separates the intertwined bodies from the sewer passage without adding an unsupported source, color, or atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer route has reached a fork, and the rubbish-dump entrance cover remains displaced. Charlie follows along the chosen branch in his old coat and hat. 이현우: He runs in a piggyback-carrying posture without his outer garment, retaining facial bruising and his injured leg. His contact card remains concealed in his shoe. 앰버: She is being carried piggyback, coughing more severely, with the outer garment still covering her mouth.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S26sh12__bgfirst_bg.png",
     "asset_id": "cbdab119-6360-4bbb-b01c-a8f50193bfec",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S26sh12.png",
     "asset_id": "370c5e80-0275-4785-b540-af69ec16bd57",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L32B03.png",
     "asset_id": "ad0f9f4b-0a55-45e6-9c39-7261e4894a80",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물들이 화면 우측 상단으로 뻗은 하수도 통로를 향해 이동하고 있음.",
    "built_space": "레퍼런스와 동일한 갈림길 구조의 지하 하수도. 인물들이 우측 분기점으로 진입하는 위치에 자연스럽게 배치됨.",
    "entities": "이현우는 지시대로 겉옷을 벗은 상태(반팔 티셔츠)이며 얼굴에 멍이 보임. 앰버는 등에 업힌 채 겉옷 소매로 입을 가리고 기침하고 있음.",
    "hard_violations": [],
    "physics": "이현우의 왼발이 바닥의 물을 강하게 차며 밀어내고 있고, 두 팔로 앰버의 다리를 안정적으로 받치고 있는 역동적인 질주 자세가 잘 표현됨."
   },
   {
    "label": "B",
    "direction": "인물들이 우측 상단 분기점이 아닌 카메라 정면(앞쪽)을 향해 달려오고 있음.",
    "built_space": "레퍼런스의 하수도 공간이나, 카메라를 향해 다가오는 구도로 인해 지시된 탈출 방향(우상단)과 공간적 맥락이 어긋남.",
    "entities": "이현우가 벗어야 할 겉옷(어두운 후드티)을 그대로 착용하고 있음. 앰버는 마스크를 쓰고 기침하는 모습.",
    "hard_violations": [
     "[gemini-pro] 이현우가 겉옷을 벗은 상태여야 한다는 복장 지시 위반 (후드티 착용)",
     "[gemini-pro] 우상단 통로 방향으로 질주해야 하는 프레임 레이아웃 지시 위반 (카메라 정면으로 이동)"
    ],
    "physics": "이현우의 오른발이 바닥의 물웅덩이를 딛고 있으며 앰버를 업은 무게중심과 자세는 성립함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "우상단 통로로 질주하는 지정된 구도와 겉옷을 벗은 이현우의 복장 등 프롬프트의 요구사항을 정확히 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물이 카메라 정면을 향해 달려 프레이밍 지시를 어겼으며, 이현우가 겉옷을 그대로 입고 있어 복장 계약을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들이 화면 우측 상단으로 뻗은 하수도 통로를 향해 이동하고 있음.",
        "built_space": "레퍼런스와 동일한 갈림길 구조의 지하 하수도. 인물들이 우측 분기점으로 진입하는 위치에 자연스럽게 배치됨.",
        "entities": "이현우는 지시대로 겉옷을 벗은 상태(반팔 티셔츠)이며 얼굴에 멍이 보임. 앰버는 등에 업힌 채 겉옷 소매로 입을 가리고 기침하고 있음.",
        "hard_violations": [],
        "physics": "이현우의 왼발이 바닥의 물을 강하게 차며 밀어내고 있고, 두 팔로 앰버의 다리를 안정적으로 받치고 있는 역동적인 질주 자세가 잘 표현됨."
       },
       {
        "label": "B",
        "direction": "인물들이 우측 상단 분기점이 아닌 카메라 정면(앞쪽)을 향해 달려오고 있음.",
        "built_space": "레퍼런스의 하수도 공간이나, 카메라를 향해 다가오는 구도로 인해 지시된 탈출 방향(우상단)과 공간적 맥락이 어긋남.",
        "entities": "이현우가 벗어야 할 겉옷(어두운 후드티)을 그대로 착용하고 있음. 앰버는 마스크를 쓰고 기침하는 모습.",
        "hard_violations": [
         "이현우가 겉옷을 벗은 상태여야 한다는 복장 지시 위반 (후드티 착용)",
         "우상단 통로 방향으로 질주해야 하는 프레임 레이아웃 지시 위반 (카메라 정면으로 이동)"
        ],
        "physics": "이현우의 오른발이 바닥의 물웅덩이를 딛고 있으며 앰버를 업은 무게중심과 자세는 성립함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "우상단 통로로 질주하는 지정된 구도와 겉옷을 벗은 이현우의 복장 등 프롬프트의 요구사항을 정확히 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물이 카메라 정면을 향해 달려 프레이밍 지시를 어겼으며, 이현우가 겉옷을 그대로 입고 있어 복장 계약을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들이 화면 우측 상단으로 뻗은 하수도 통로를 향해 이동하고 있음.",
        "built_space": "레퍼런스와 동일한 갈림길 구조의 지하 하수도. 인물들이 우측 분기점으로 진입하는 위치에 자연스럽게 배치됨.",
        "entities": "이현우는 지시대로 겉옷을 벗은 상태(반팔 티셔츠)이며 얼굴에 멍이 보임. 앰버는 등에 업힌 채 겉옷 소매로 입을 가리고 기침하고 있음.",
        "hard_violations": [],
        "physics": "이현우의 왼발이 바닥의 물을 강하게 차며 밀어내고 있고, 두 팔로 앰버의 다리를 안정적으로 받치고 있는 역동적인 질주 자세가 잘 표현됨."
       },
       {
        "label": "B",
        "direction": "인물들이 우측 상단 분기점이 아닌 카메라 정면(앞쪽)을 향해 달려오고 있음.",
        "built_space": "레퍼런스의 하수도 공간이나, 카메라를 향해 다가오는 구도로 인해 지시된 탈출 방향(우상단)과 공간적 맥락이 어긋남.",
        "entities": "이현우가 벗어야 할 겉옷(어두운 후드티)을 그대로 착용하고 있음. 앰버는 마스크를 쓰고 기침하는 모습.",
        "hard_violations": [
         "이현우가 겉옷을 벗은 상태여야 한다는 복장 지시 위반 (후드티 착용)",
         "우상단 통로 방향으로 질주해야 하는 프레임 레이아웃 지시 위반 (카메라 정면으로 이동)"
        ],
        "physics": "이현우의 오른발이 바닥의 물웅덩이를 딛고 있으며 앰버를 업은 무게중심과 자세는 성립함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "하수도 구조와 업기 동작은 맞지만 오른쪽 위 탈출로를 등지고 카메라 쪽으로 달리며, 현우가 겉옷을 입고 있어 핵심 진행 방향과 의상 연속성을 어깁니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽 위 통로로 진입하는 후측면 와이드 숏과 강한 발차기 추진, 겉옷을 벗은 업기 자세가 정확하지만 앰버의 하의와 입을 덮는 옷의 연속성은 덜 정확합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우의 얼굴과 몸통, 지면을 딛는 발은 화면 왼쪽 앞과 카메라 쪽을 향합니다. 목적지인 오른쪽 위 하수도 통로는 두 사람 뒤에 있어 그 통로로 진입하는 방향이 아닙니다. 앰버는 눈을 감고 고개를 숙여 입을 가립니다.",
        "built_space": "중앙 콘크리트 분기벽 하나를 사이에 두고 왼쪽과 오른쪽 통로 두 개가 보입니다. 왼쪽 조명 하나, 오른쪽 천장의 연속 조명, 오른쪽 벽의 배관 한 줄과 젖은 바닥은 장소 참조와 잘 맞습니다. 두 사람의 전신과 발밑을 포함한 와이드 구도이며 오른쪽 통로도 배경에 있지만, 인물은 그곳에서 전경으로 나오는 배치입니다. 바닥 반사는 천장 조명과 양립합니다.",
        "entities": "등장인물은 청소년 남성과 금발 여자아이 두 명뿐입니다. 현우의 검은 머리, 젊은 한국계 얼굴, 얼굴 멍, 해진 청바지와 운동화는 부합하지만 벗어야 할 검은 후드 겉옷을 여전히 입었습니다. 앰버는 참조와 유사한 금발 어린아이로 녹색 재킷과 갈색 부츠를 착용하며, 입에는 겉옷 대신 흰 마스크가 보입니다. 기침하는 표정은 읽히나 다리 부상은 분명하지 않습니다. 신발 속 카드는 보이지 않으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "현우의 화면 왼쪽 발은 젖은 바닥에 닿아 물을 튀기고, 반대쪽 다리는 뒤로 굽혀져 있어 달리기의 지지와 회수가 가능합니다. 앰버는 현우의 등에 몸을 기대고 팔을 어깨에 걸쳤으며 현우의 손이 허벅지 아래를 받칩니다. 지지 없이 떠 있는 사람이나 물체는 없습니다. 다만 자세는 목표 통로로 차고 들어가기보다 카메라 쪽으로 달려오는 순간입니다."
       },
       {
        "label": "B",
        "direction": "현우의 등은 카메라에 보이고 머리와 상체는 오른쪽 위 배경의 통로 입구를 향합니다. 뒤로 길게 뻗은 다리와 앞으로 기울인 몸이 같은 탈출 방향을 만듭니다. 앰버는 고개를 숙이고 옷에 입을 묻어 기침하며 카메라를 보지 않습니다.",
        "built_space": "중앙 분기벽 하나와 그 양옆의 통로 두 개, 왼쪽 조명 하나, 오른쪽 천장의 연속 조명과 벽 배관이 보이며 참조의 낮은 콘크리트 하수도 구조를 유지합니다. 두 사람은 분기벽 오른쪽 입구에서 오른쪽 위로 멀어지는 통로에 진입합니다. 전신과 추진 지점을 담은 후측면 와이드 숏이고 젖은 바닥의 반사도 조명 위치에 맞습니다. 다만 인물과 공간은 희미한 주변광이라는 요구보다 다소 밝습니다.",
        "entities": "현우와 앰버 두 명만 보입니다. 현우는 검은 헝클어진 머리의 젊은 남성이며 겉옷 없이 오염된 반팔 상의, 청바지, 운동화를 착용합니다. 얼굴은 후측면으로 가려져 정확한 얼굴 일치와 얼굴 멍을 확인하기 어렵습니다. 앰버는 금발 어린 여자아이로 녹색 재킷과 갈색 부츠는 참조와 유사하지만, 참조의 드레스 아래 맨다리 대신 바지 형태의 하의가 보입니다. 자기 녹색 옷을 입에 대고 기침하여 입 가림은 구현했으나 현우가 벗어 준 겉옷의 유지 여부는 불명확합니다. 다리 부상도 확실하지 않으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "현우는 상체를 앞으로 숙이고 한쪽 다리를 뒤로 길게 펴며, 뒤쪽 신발 바로 아래의 물보라가 바닥을 차고 떠나는 추진 지점을 보여 줍니다. 다른 발은 몸 아래에서 다음 착지를 준비하는 달리기 보폭으로 읽힙니다. 앰버의 몸은 현우의 등에 실리고 허벅지는 현우의 팔과 손으로 받쳐져 있어 업기의 하중 경로가 있습니다. 발이 들린 순간도 직전 지면 접촉과 진행할 바닥이 보여 무근거한 공중 부양은 아닙니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "하수도 구조와 업기 동작은 맞지만 오른쪽 위 탈출로를 등지고 카메라 쪽으로 달리며, 현우가 겉옷을 입고 있어 핵심 진행 방향과 의상 연속성을 어깁니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽 위 통로로 진입하는 후측면 와이드 숏과 강한 발차기 추진, 겉옷을 벗은 업기 자세가 정확하지만 앰버의 하의와 입을 덮는 옷의 연속성은 덜 정확합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "현우의 얼굴과 몸통, 지면을 딛는 발은 화면 왼쪽 앞과 카메라 쪽을 향합니다. 목적지인 오른쪽 위 하수도 통로는 두 사람 뒤에 있어 그 통로로 진입하는 방향이 아닙니다. 앰버는 눈을 감고 고개를 숙여 입을 가립니다.",
        "built_space": "중앙 콘크리트 분기벽 하나를 사이에 두고 왼쪽과 오른쪽 통로 두 개가 보입니다. 왼쪽 조명 하나, 오른쪽 천장의 연속 조명, 오른쪽 벽의 배관 한 줄과 젖은 바닥은 장소 참조와 잘 맞습니다. 두 사람의 전신과 발밑을 포함한 와이드 구도이며 오른쪽 통로도 배경에 있지만, 인물은 그곳에서 전경으로 나오는 배치입니다. 바닥 반사는 천장 조명과 양립합니다.",
        "entities": "등장인물은 청소년 남성과 금발 여자아이 두 명뿐입니다. 현우의 검은 머리, 젊은 한국계 얼굴, 얼굴 멍, 해진 청바지와 운동화는 부합하지만 벗어야 할 검은 후드 겉옷을 여전히 입었습니다. 앰버는 참조와 유사한 금발 어린아이로 녹색 재킷과 갈색 부츠를 착용하며, 입에는 겉옷 대신 흰 마스크가 보입니다. 기침하는 표정은 읽히나 다리 부상은 분명하지 않습니다. 신발 속 카드는 보이지 않으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "현우의 화면 왼쪽 발은 젖은 바닥에 닿아 물을 튀기고, 반대쪽 다리는 뒤로 굽혀져 있어 달리기의 지지와 회수가 가능합니다. 앰버는 현우의 등에 몸을 기대고 팔을 어깨에 걸쳤으며 현우의 손이 허벅지 아래를 받칩니다. 지지 없이 떠 있는 사람이나 물체는 없습니다. 다만 자세는 목표 통로로 차고 들어가기보다 카메라 쪽으로 달려오는 순간입니다."
       },
       {
        "label": "A",
        "direction": "현우의 등은 카메라에 보이고 머리와 상체는 오른쪽 위 배경의 통로 입구를 향합니다. 뒤로 길게 뻗은 다리와 앞으로 기울인 몸이 같은 탈출 방향을 만듭니다. 앰버는 고개를 숙이고 옷에 입을 묻어 기침하며 카메라를 보지 않습니다.",
        "built_space": "중앙 분기벽 하나와 그 양옆의 통로 두 개, 왼쪽 조명 하나, 오른쪽 천장의 연속 조명과 벽 배관이 보이며 참조의 낮은 콘크리트 하수도 구조를 유지합니다. 두 사람은 분기벽 오른쪽 입구에서 오른쪽 위로 멀어지는 통로에 진입합니다. 전신과 추진 지점을 담은 후측면 와이드 숏이고 젖은 바닥의 반사도 조명 위치에 맞습니다. 다만 인물과 공간은 희미한 주변광이라는 요구보다 다소 밝습니다.",
        "entities": "현우와 앰버 두 명만 보입니다. 현우는 검은 헝클어진 머리의 젊은 남성이며 겉옷 없이 오염된 반팔 상의, 청바지, 운동화를 착용합니다. 얼굴은 후측면으로 가려져 정확한 얼굴 일치와 얼굴 멍을 확인하기 어렵습니다. 앰버는 금발 어린 여자아이로 녹색 재킷과 갈색 부츠는 참조와 유사하지만, 참조의 드레스 아래 맨다리 대신 바지 형태의 하의가 보입니다. 자기 녹색 옷을 입에 대고 기침하여 입 가림은 구현했으나 현우가 벗어 준 겉옷의 유지 여부는 불명확합니다. 다리 부상도 확실하지 않으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "현우는 상체를 앞으로 숙이고 한쪽 다리를 뒤로 길게 펴며, 뒤쪽 신발 바로 아래의 물보라가 바닥을 차고 떠나는 추진 지점을 보여 줍니다. 다른 발은 몸 아래에서 다음 착지를 준비하는 달리기 보폭으로 읽힙니다. 앰버의 몸은 현우의 등에 실리고 허벅지는 현우의 팔과 손으로 받쳐져 있어 업기의 하중 경로가 있습니다. 발이 들린 순간도 직전 지면 접촉과 진행할 바닥이 보여 무근거한 공중 부양은 아닙니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.929
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.679
   },
   "violations": {
    "B": [
     "[gemini-pro] 이현우가 겉옷을 벗은 상태여야 한다는 복장 지시 위반 (후드티 착용)",
     "[gemini-pro] 우상단 통로 방향으로 질주해야 하는 프레임 레이아웃 지시 위반 (카메라 정면으로 이동)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 679
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "우상단 통로로 질주하는 지정된 구도와 겉옷을 벗은 이현우의 복장 등 프롬프트의 요구사항을 정확히 구현했습니다."
   },
   {
    "label": "B",
    "score": 679,
    "verdict_ko": "인물이 카메라 정면을 향해 달려 프레이밍 지시를 어겼으며, 이현우가 겉옷을 그대로 입고 있어 복장 계약을 위반했습니다.  ★위반: [gemini-pro] 이현우가 겉옷을 벗은 상태여야 한다는 복장 지시 위반 (후드티 착용) / [gemini-pro] 우상단 통로 방향으로 질주해야 하는 프레임 레이아웃 지시 위반 (카메라 정면으로 이동)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L32B03.png",
    "asset_id": "ad0f9f4b-0a55-45e6-9c39-7261e4894a80",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda6-1fe8-737f-993e-7d2dcd7ff5eb",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S26sh12__bgfirst_bg.png",
   "bg_asset_id": "cbdab119-6360-4bbb-b01c-a8f50193bfec",
   "bg_record_key": "S26sh12::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S26sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:29:54.082924+00:00",
  "fingerprint": "3dfb4b045f9f65281656a782ef0a29ed93965cc8f913890fb20bb8d3e3df92ea",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S26sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S26sh12_sel.png",
  "source_sha256": "9ca37b38cfdbbe632746b6940a6c927d27344d2ecb1b9e6e4988bbc90d954a8c",
  "file": "S26sh12_cine.png",
  "staged_sha256": "793bab390128089405a44c165b49bcf0ba5e072acfd7dcd89a1f1d43fba7d3bb",
  "latency_ms": 12458
 },
 "S27sh14::signage": {
  "fp": "4db1192ca8d569e5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S27sh14": {
  "input_fingerprint": "5de83f15bb311362",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거대한 자동차가 도로 한가운데 선 찰리의 코앞에서 거칠게 정지하며 타이어 흙먼지를 일으키는 찰나.\n\nLOCATION (lock): In the middle of a refugee-settlement roadway at night, directly in front of a van braking with its headlights on. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Large vehicle (Stopping abruptly immediately in front of 찰리) — Its front and a limited section of its near side are visible at an oblique angle; used as Occupies the left third, defining the near-collision gap without placing the camera on the approach axis; Roadway (찰리 and the braking vehicle occupy the road); used as Provides scale and a visible separation between the lateral camera position and the vehicle's route; Tire-raised dust (Briefly thrown up at the braking instant); used as A localized low accent around the tires conveys the abrupt stop without obscuring 찰리.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The vehicle's headlights create the immediate brightness contrast against the nighttime street, with no additional lighting source introduced.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The street's manhole cover has been opened, and streetlights continue cycling on and off along Charlie's route. The van is stopped with its headlights on and a crude halogen cross on its roof; Charlie retains his coat and hat, and the small bird has flown away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거대한 자동차가 도로 한가운데 선 찰리의 코앞에서 거칠게 정지하며 타이어 흙먼지를 일으키는 찰나.\n\nLOCATION (lock): In the middle of a refugee-settlement roadway at night, directly in front of a van braking with its headlights on. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Large vehicle (Stopping abruptly immediately in front of 찰리) — Its front and a limited section of its near side are visible at an oblique angle; used as Occupies the left third, defining the near-collision gap without placing the camera on the approach axis; Roadway (찰리 and the braking vehicle occupy the road); used as Provides scale and a visible separation between the lateral camera position and the vehicle's route; Tire-raised dust (Briefly thrown up at the braking instant); used as A localized low accent around the tires conveys the abrupt stop without obscuring 찰리.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The vehicle's headlights create the immediate brightness contrast against the nighttime street, with no additional lighting source introduced.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The street's manhole cover has been opened, and streetlights continue cycling on and off along Charlie's route. The van is stopped with its headlights on and a crude halogen cross on its roof; Charlie retains his coat and hat, and the small bird has flown away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거대한 자동차가 도로 한가운데 선 찰리의 코앞에서 거칠게 정지하며 타이어 흙먼지를 일으키는 찰나.\n\nLOCATION (lock): In the middle of a refugee-settlement roadway at night, directly in front of a van braking with its headlights on. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Large vehicle (Stopping abruptly immediately in front of 찰리) — Its front and a limited section of its near side are visible at an oblique angle; used as Occupies the left third, defining the near-collision gap without placing the camera on the approach axis; Roadway (찰리 and the braking vehicle occupy the road); used as Provides scale and a visible separation between the lateral camera position and the vehicle's route; Tire-raised dust (Briefly thrown up at the braking instant); used as A localized low accent around the tires conveys the abrupt stop without obscuring 찰리.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The vehicle's headlights create the immediate brightness contrast against the nighttime street, with no additional lighting source introduced.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The street's manhole cover has been opened, and streetlights continue cycling on and off along Charlie's route. The van is stopped with its headlights on and a crude halogen cross on its roof; Charlie retains his coat and hat, and the small bird has flown away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "밴이 찰리를 향해 정지하며 헤드라이트가 인물을 비춤.",
    "built_space": "레퍼런스 구조물은 존재하나, 차량과 인물 간의 크기 비례가 심하게 왜곡됨.",
    "entities": "코트와 모자를 착용했으나, 기계 얼굴이 아닌 실제 털이 있는 고릴라 얼굴로 묘사됨. 밴 지붕의 십자가와 타이어 먼지 존재함.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 스테이징 (인물이 열린 맨홀 구멍 안에 서 있으며 주변 환경 및 차량 대비 비정상적으로 소인처럼 묘사됨)"
    ],
    "physics": "차량은 도로 위에 있고 인물은 맨홀 내부에 서 있음."
   },
   {
    "label": "B",
    "direction": "밴이 찰리를 향해 정지하며 헤드라이트가 정면의 찰리를 비춤.",
    "built_space": "레퍼런스의 컨테이너와 골목이 정확히 구현됨. 맨홀 뚜껑은 바닥에 닫힌 상태임.",
    "entities": "찰리는 각진 기계 머리, 흰 얼굴판, 점 형태의 눈과 선 형태의 입을 정확히 갖췄으나 모자가 없음. 밴 지붕의 십자가와 타이어 먼지가 확인됨.",
    "hard_violations": [],
    "physics": "차량과 인물 모두 도로 표면에 닿아 있으며 타이어 주변 흙먼지가 일어남."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "모자와 열린 맨홀 상태가 누락되었으나, 찰리의 기계 얼굴 세부 묘사가 매우 정확하며 전체적인 스케일과 구도가 자연스러움."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 얼굴이 명시된 기계 형태를 전혀 따르지 않았으며, 맨홀 안 인물의 크기가 비정상적으로 축소되어 스테이징에 실패함."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "밴이 찰리를 향해 정지하며 헤드라이트가 정면의 찰리를 비춤.",
        "built_space": "레퍼런스의 컨테이너와 골목이 정확히 구현됨. 맨홀 뚜껑은 바닥에 닫힌 상태임.",
        "entities": "찰리는 각진 기계 머리, 흰 얼굴판, 점 형태의 눈과 선 형태의 입을 정확히 갖췄으나 모자가 없음. 밴 지붕의 십자가와 타이어 먼지가 확인됨.",
        "hard_violations": [],
        "physics": "차량과 인물 모두 도로 표면에 닿아 있으며 타이어 주변 흙먼지가 일어남."
       },
       {
        "label": "A",
        "direction": "밴이 찰리를 향해 정지하며 헤드라이트가 인물을 비춤.",
        "built_space": "레퍼런스 구조물은 존재하나, 차량과 인물 간의 크기 비례가 심하게 왜곡됨.",
        "entities": "코트와 모자를 착용했으나, 기계 얼굴이 아닌 실제 털이 있는 고릴라 얼굴로 묘사됨. 밴 지붕의 십자가와 타이어 먼지 존재함.",
        "hard_violations": [
         "물리적으로 불가능한 스테이징 (인물이 열린 맨홀 구멍 안에 서 있으며 주변 환경 및 차량 대비 비정상적으로 소인처럼 묘사됨)"
        ],
        "physics": "차량은 도로 위에 있고 인물은 맨홀 내부에 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "모자와 열린 맨홀 상태가 누락되었으나, 찰리의 기계 얼굴 세부 묘사가 매우 정확하며 전체적인 스케일과 구도가 자연스러움."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 얼굴이 명시된 기계 형태를 전혀 따르지 않았으며, 맨홀 안 인물의 크기가 비정상적으로 축소되어 스테이징에 실패함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "밴이 찰리를 향해 정지하며 헤드라이트가 정면의 찰리를 비춤.",
        "built_space": "레퍼런스의 컨테이너와 골목이 정확히 구현됨. 맨홀 뚜껑은 바닥에 닫힌 상태임.",
        "entities": "찰리는 각진 기계 머리, 흰 얼굴판, 점 형태의 눈과 선 형태의 입을 정확히 갖췄으나 모자가 없음. 밴 지붕의 십자가와 타이어 먼지가 확인됨.",
        "hard_violations": [],
        "physics": "차량과 인물 모두 도로 표면에 닿아 있으며 타이어 주변 흙먼지가 일어남."
       },
       {
        "label": "A",
        "direction": "밴이 찰리를 향해 정지하며 헤드라이트가 인물을 비춤.",
        "built_space": "레퍼런스 구조물은 존재하나, 차량과 인물 간의 크기 비례가 심하게 왜곡됨.",
        "entities": "코트와 모자를 착용했으나, 기계 얼굴이 아닌 실제 털이 있는 고릴라 얼굴로 묘사됨. 밴 지붕의 십자가와 타이어 먼지 존재함.",
        "hard_violations": [
         "물리적으로 불가능한 스테이징 (인물이 열린 맨홀 구멍 안에 서 있으며 주변 환경 및 차량 대비 비정상적으로 소인처럼 묘사됨)"
        ],
        "physics": "차량은 도로 위에 있고 인물은 맨홀 내부에 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "코트·모자와 장소 세부의 충실도는 떨어지지만, 측면 와이드 구도에서 차량과 찰리의 간격이 더 좁아 최우선인 ‘코앞 급정지’ 연출에 더 가깝다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "장소와 고릴라형 머리·코트·모자·열린 맨홀은 더 충실하지만, 범퍼와 찰리 사이의 넓은 빈 공간이 ‘코앞에서 거칠게 정지’하는 핵심 순간을 약화한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "밴은 화면 왼쪽에서 오른쪽의 찰리를 향하고, 찰리의 얼굴과 몸도 왼쪽 차량 전면을 향한다. 전조등은 두 대상 사이의 먼지와 찰리 쪽 도로를 밝힌다. 카메라는 차량 접근축을 벗어나 전면과 가까운 측면을 함께 본다.",
        "built_space": "중앙 컨테이너 전면에 창 하나와 검은 문 하나, 상단의 작은 조명 하나가 있으며 오른쪽에 좁은 골목이 이어진다. 창과 문의 좌우 관계는 참조와 맞지만, 위층은 창이 여러 개 달린 높은 컨테이너로 바뀌어 참조의 낮은 적층 구조와 다르다. 확성기 달린 전신주 하나와 골목의 켜진 가로등 두 개가 보인다. 찰리와 차량은 모두 연석 아래 차도에 있다. 찰리 오른쪽의 맨홀 덮개 하나는 닫혀 보인다.",
        "entities": "보이는 인물은 찰리 한 명이다. 각진 기계 머리, 흰 얼굴판, 점 같은 눈과 짧은 선 형태 입은 있지만 고릴라형보다는 단순한 로봇 얼굴에 가깝다. 짧은 재킷을 입고 모자는 쓰지 않아 유지해야 할 코트·모자와 다르다. 큰 밴 한 대, 켜진 전조등 두 개, 지붕의 발광 십자가 하나와 타이어 주변 먼지가 있다. 새는 없다. 명확하게 읽히는 문구는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 두 발을 노면에 대고 무릎을 굽히며 상체를 약간 뒤로 물려 제동에 반응하는 자세다. 밴은 타이어로 도로에 지지되고 십자가는 지붕의 금속 지지대에 고정되어 있다. 먼지는 바퀴 주변에서 발생하지만 일부가 범퍼 위까지 올라와 요구된 낮고 국소적인 효과보다 퍼져 있다. 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "밴 전면은 오른쪽의 찰리를 향하고 찰리는 왼쪽 밴을 바라본다. 카메라는 접근축 옆에서 전면과 측면을 비스듬히 본다. 다만 범퍼와 찰리 사이에 상당한 도로 폭이 남아 있어 차량이 코앞까지 다가왔다는 관계는 약하다.",
        "built_space": "컨테이너 전면의 창 하나, 검은 문 하나, 상단 조명 하나가 참조와 같은 순서로 배치되어 있다. 뒤쪽 녹슨 컨테이너와 실외기 하나, 왼쪽의 확성기·가로등이 붙은 전신주, 오른쪽 골목과 담장도 참조에 더 가깝다. 골목에는 켜진 가로등 두 개가 보인다. 찰리 뒤쪽 차도에는 열린 맨홀 하나와 옆으로 밀린 덮개 하나가 보이며 찰리는 구멍 바깥 노면에 서 있다. 밴은 지정된 왼쪽 3분의 1보다 넓게 화면을 차지한다.",
        "entities": "인물은 찰리 한 명뿐이다. 흰 얼굴판이 있는 고릴라형 기계 머리, 어두운 모자와 긴 코트가 보인다. 얼굴은 요구된 단순한 점 눈·선 입보다 굴곡과 기계 세부가 많다. 낡은 대형 밴 한 대, 켜진 전조등 두 개, 지붕의 조잡한 발광 십자가 하나, 타이어 주변 먼지가 있다. 새는 없으며 읽을 수 있는 글자는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 두 발로 차도에 서 있고 팔은 아래로 자연스럽게 늘어져 있다. 차량은 바퀴로 지지되며 지붕 십자가에는 고정 프레임이 있다. 낮게 번지는 먼지와 차량 전면의 흐림은 급제동 순간으로 해석할 수 있고 찰리를 가리지 않는다. 맨홀 덮개는 노면에 놓여 있으며 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "코트·모자와 장소 세부의 충실도는 떨어지지만, 측면 와이드 구도에서 차량과 찰리의 간격이 더 좁아 최우선인 ‘코앞 급정지’ 연출에 더 가깝다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "장소와 고릴라형 머리·코트·모자·열린 맨홀은 더 충실하지만, 범퍼와 찰리 사이의 넓은 빈 공간이 ‘코앞에서 거칠게 정지’하는 핵심 순간을 약화한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "밴은 화면 왼쪽에서 오른쪽의 찰리를 향하고, 찰리의 얼굴과 몸도 왼쪽 차량 전면을 향한다. 전조등은 두 대상 사이의 먼지와 찰리 쪽 도로를 밝힌다. 카메라는 차량 접근축을 벗어나 전면과 가까운 측면을 함께 본다.",
        "built_space": "중앙 컨테이너 전면에 창 하나와 검은 문 하나, 상단의 작은 조명 하나가 있으며 오른쪽에 좁은 골목이 이어진다. 창과 문의 좌우 관계는 참조와 맞지만, 위층은 창이 여러 개 달린 높은 컨테이너로 바뀌어 참조의 낮은 적층 구조와 다르다. 확성기 달린 전신주 하나와 골목의 켜진 가로등 두 개가 보인다. 찰리와 차량은 모두 연석 아래 차도에 있다. 찰리 오른쪽의 맨홀 덮개 하나는 닫혀 보인다.",
        "entities": "보이는 인물은 찰리 한 명이다. 각진 기계 머리, 흰 얼굴판, 점 같은 눈과 짧은 선 형태 입은 있지만 고릴라형보다는 단순한 로봇 얼굴에 가깝다. 짧은 재킷을 입고 모자는 쓰지 않아 유지해야 할 코트·모자와 다르다. 큰 밴 한 대, 켜진 전조등 두 개, 지붕의 발광 십자가 하나와 타이어 주변 먼지가 있다. 새는 없다. 명확하게 읽히는 문구는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 두 발을 노면에 대고 무릎을 굽히며 상체를 약간 뒤로 물려 제동에 반응하는 자세다. 밴은 타이어로 도로에 지지되고 십자가는 지붕의 금속 지지대에 고정되어 있다. 먼지는 바퀴 주변에서 발생하지만 일부가 범퍼 위까지 올라와 요구된 낮고 국소적인 효과보다 퍼져 있다. 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "밴 전면은 오른쪽의 찰리를 향하고 찰리는 왼쪽 밴을 바라본다. 카메라는 접근축 옆에서 전면과 측면을 비스듬히 본다. 다만 범퍼와 찰리 사이에 상당한 도로 폭이 남아 있어 차량이 코앞까지 다가왔다는 관계는 약하다.",
        "built_space": "컨테이너 전면의 창 하나, 검은 문 하나, 상단 조명 하나가 참조와 같은 순서로 배치되어 있다. 뒤쪽 녹슨 컨테이너와 실외기 하나, 왼쪽의 확성기·가로등이 붙은 전신주, 오른쪽 골목과 담장도 참조에 더 가깝다. 골목에는 켜진 가로등 두 개가 보인다. 찰리 뒤쪽 차도에는 열린 맨홀 하나와 옆으로 밀린 덮개 하나가 보이며 찰리는 구멍 바깥 노면에 서 있다. 밴은 지정된 왼쪽 3분의 1보다 넓게 화면을 차지한다.",
        "entities": "인물은 찰리 한 명뿐이다. 흰 얼굴판이 있는 고릴라형 기계 머리, 어두운 모자와 긴 코트가 보인다. 얼굴은 요구된 단순한 점 눈·선 입보다 굴곡과 기계 세부가 많다. 낡은 대형 밴 한 대, 켜진 전조등 두 개, 지붕의 조잡한 발광 십자가 하나, 타이어 주변 먼지가 있다. 새는 없으며 읽을 수 있는 글자는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 두 발로 차도에 서 있고 팔은 아래로 자연스럽게 늘어져 있다. 차량은 바퀴로 지지되며 지붕 십자가에는 고정 프레임이 있다. 낮게 번지는 먼지와 차량 전면의 흐림은 급제동 순간으로 해석할 수 있고 찰리를 가리지 않는다. 맨홀 덮개는 노면에 놓여 있으며 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.286,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.036,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 스테이징 (인물이 열린 맨홀 구멍 안에 서 있으며 주변 환경 및 차량 대비 비정상적으로 소인처럼 묘사됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1036
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "모자와 열린 맨홀 상태가 누락되었으나, 찰리의 기계 얼굴 세부 묘사가 매우 정확하며 전체적인 스케일과 구도가 자연스러움."
   },
   {
    "label": "A",
    "score": 1036,
    "verdict_ko": "찰리의 얼굴이 명시된 기계 형태를 전혀 따르지 않았으며, 맨홀 안 인물의 크기가 비정상적으로 축소되어 스테이징에 실패함.  ★위반: [gemini-pro] 물리적으로 불가능한 스테이징 (인물이 열린 맨홀 구멍 안에 서 있으며 주변 환경 및 차량 대비 비정상적으로 소인처럼 묘사됨)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_refugee_residential_lanes_sel.png",
    "asset_id": "fb1c15d7-9723-4177-bf7c-6320da46869c",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda6-ee01-796d-a0a9-14b119052df8",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S27sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:30:06.885659+00:00",
  "fingerprint": "03e6fb13638282491293e607d8f65c49cad0398e61064f5b508a7a6e1d693e49",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S27sh14_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S27sh14_sel.png",
  "source_sha256": "3f278d856a7acea3c2b3497928c44cb0dfcfc3bdf8961a4b989312260580d565",
  "file": "S27sh14_cine.png",
  "staged_sha256": "9f165e643ff1e8ca76f404858f931fc3845963bbbedcfb69568ea40beaf192e0",
  "latency_ms": 14467
 },
 "S27sh20::signage": {
  "fp": "7609470eecbdbd5f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S27sh20": {
  "input_fingerprint": "13551216393dfcda",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 헤드라이트 불빛을 벗어난 신부가 눈을 동그랗게 뜨고 찰리의 거대한 덩치를 올려다보는 놀란 얼굴 클로즈업.\n\nLOCATION (lock): On the roadway beside the stopped van, just outside the glare of its headlights. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Roadway beside the stopped vehicle (The priest has approached 찰리); used as A narrow background margin grounds the elevated close-up in the street.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: As the priest clears the headlight glare, his face becomes readable against the nighttime surroundings while the established headlight contrast remains.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the stopped vehicle, its headlights, and the surrounding nighttime road surfaces. EXCLUDE the airborne tire dust from the braking instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van remains stopped with its headlights on and its crude halogen roof cross in place; the street's manhole remains open. Charlie retains his old coat and hat, the small bird is gone, and the streetlights continue their established on-and-off behavior. 신부: He has stepped out of the van and approached along the road, still wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 헤드라이트 불빛을 벗어난 신부가 눈을 동그랗게 뜨고 찰리의 거대한 덩치를 올려다보는 놀란 얼굴 클로즈업.\n\nLOCATION (lock): On the roadway beside the stopped van, just outside the glare of its headlights. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Roadway beside the stopped vehicle (The priest has approached 찰리); used as A narrow background margin grounds the elevated close-up in the street.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: As the priest clears the headlight glare, his face becomes readable against the nighttime surroundings while the established headlight contrast remains.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the stopped vehicle, its headlights, and the surrounding nighttime road surfaces. EXCLUDE the airborne tire dust from the braking instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van remains stopped with its headlights on and its crude halogen roof cross in place; the street's manhole remains open. Charlie retains his old coat and hat, the small bird is gone, and the streetlights continue their established on-and-off behavior. 신부: He has stepped out of the van and approached along the road, still wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 헤드라이트 불빛을 벗어난 신부가 눈을 동그랗게 뜨고 찰리의 거대한 덩치를 올려다보는 놀란 얼굴 클로즈업.\n\nLOCATION (lock): On the roadway beside the stopped van, just outside the glare of its headlights. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Roadway beside the stopped vehicle (The priest has approached 찰리); used as A narrow background margin grounds the elevated close-up in the street.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: As the priest clears the headlight glare, his face becomes readable against the nighttime surroundings while the established headlight contrast remains.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the stopped vehicle, its headlights, and the surrounding nighttime road surfaces. EXCLUDE the airborne tire dust from the braking instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van remains stopped with its headlights on and its crude halogen roof cross in place; the street's manhole remains open. Charlie retains his old coat and hat, the small bird is gone, and the streetlights continue their established on-and-off behavior. 신부: He has stepped out of the van and approached along the road, still wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "신부가 위쪽을 향해 놀란 눈으로 시선을 던지고 있음.",
    "built_space": "밴의 전면부와 지붕 십자가, 뒤편의 컨테이너 건물들이 배경에 넓게 배치됨.",
    "entities": "신부(레퍼런스 일치), 십자가가 달린 밴, 지시와 달리 포함된 타이어 먼지.",
    "hard_violations": [
     "[gpt-high] 한 대여야 할 정차 밴에 별도의 앞유리와 지붕·차체가 뒤쪽으로 중복되어, 추가 차량이 겹쳐진 공간 구조를 만든다."
    ],
    "physics": "지면에 서서 고개를 들고 있는 자세가 물리적으로 문제없이 표현됨."
   },
   {
    "label": "B",
    "direction": "신부의 시선이 화면 상단 밖의 거대한 대상을 향해 고정되어 있음.",
    "built_space": "밴의 측면과 도로 바닥, 맨홀 뚜껑이 카메라 아래쪽으로 깔리며 배경을 구성함.",
    "entities": "신부(레퍼런스와 일치하는 외모 및 복장), 정차된 밴의 일부.",
    "hard_violations": [],
    "physics": "인물이 지면에 안정적으로 서 있으며, 시선 처리에 맞는 목의 각도가 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "명시적으로 제외를 지시한 타이어 먼지를 정확히 제거했으며, 하이앵글(elevated) 클로즈업을 통해 도로 바닥을 배경으로 삼은 프레이밍 지시를 충실히 따랐습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프롬프트에서 제외하라고 명시한 브레이크 타이어 먼지가 배경에 짙게 남아있으며, 지시된 부감(elevated) 구도가 아닌 로우 앵글에 가까워 배경 지시를 벗어났습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "신부의 시선이 화면 상단 밖의 거대한 대상을 향해 고정되어 있음.",
        "built_space": "밴의 측면과 도로 바닥, 맨홀 뚜껑이 카메라 아래쪽으로 깔리며 배경을 구성함.",
        "entities": "신부(레퍼런스와 일치하는 외모 및 복장), 정차된 밴의 일부.",
        "hard_violations": [],
        "physics": "인물이 지면에 안정적으로 서 있으며, 시선 처리에 맞는 목의 각도가 자연스러움."
       },
       {
        "label": "A",
        "direction": "신부가 위쪽을 향해 놀란 눈으로 시선을 던지고 있음.",
        "built_space": "밴의 전면부와 지붕 십자가, 뒤편의 컨테이너 건물들이 배경에 넓게 배치됨.",
        "entities": "신부(레퍼런스 일치), 십자가가 달린 밴, 지시와 달리 포함된 타이어 먼지.",
        "hard_violations": [],
        "physics": "지면에 서서 고개를 들고 있는 자세가 물리적으로 문제없이 표현됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "명시적으로 제외를 지시한 타이어 먼지를 정확히 제거했으며, 하이앵글(elevated) 클로즈업을 통해 도로 바닥을 배경으로 삼은 프레이밍 지시를 충실히 따랐습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프롬프트에서 제외하라고 명시한 브레이크 타이어 먼지가 배경에 짙게 남아있으며, 지시된 부감(elevated) 구도가 아닌 로우 앵글에 가까워 배경 지시를 벗어났습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "신부의 시선이 화면 상단 밖의 거대한 대상을 향해 고정되어 있음.",
        "built_space": "밴의 측면과 도로 바닥, 맨홀 뚜껑이 카메라 아래쪽으로 깔리며 배경을 구성함.",
        "entities": "신부(레퍼런스와 일치하는 외모 및 복장), 정차된 밴의 일부.",
        "hard_violations": [],
        "physics": "인물이 지면에 안정적으로 서 있으며, 시선 처리에 맞는 목의 각도가 자연스러움."
       },
       {
        "label": "A",
        "direction": "신부가 위쪽을 향해 놀란 눈으로 시선을 던지고 있음.",
        "built_space": "밴의 전면부와 지붕 십자가, 뒤편의 컨테이너 건물들이 배경에 넓게 배치됨.",
        "entities": "신부(레퍼런스 일치), 십자가가 달린 밴, 지시와 달리 포함된 타이어 먼지.",
        "hard_violations": [],
        "physics": "지면에 서서 고개를 들고 있는 자세가 물리적으로 문제없이 표현됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "높은 위치에서 내려다본 얼굴 클로즈업과 동그랗게 뜬 눈이 핵심 순간에 더 충실하지만, 맨홀이 닫혀 있고 제외해야 할 먼지성 연무가 남아 있다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "신부의 올려다보는 시선과 열린 맨홀은 맞지만, 밴의 차체·앞유리가 중복되고 배경을 크게 드러내 좁은 도로 여백 중심의 클로즈업에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "신부는 턱을 들고 두 눈을 크게 뜬 채 높은 카메라 부근을 바라본다. 화면 밖 위쪽에 선 찰리를 올려다보는 관계로 자연스럽게 읽힌다. 찰리 자체는 보이지 않아 정확한 눈맞춤 지점까지 확인할 수는 없다. 왼쪽 밴의 전조등은 도로를 비추고 신부의 얼굴은 그 직접적인 빛줄기 바깥에서 읽힌다.",
        "built_space": "왼쪽 위에 밴 한 대의 전면 일부와 켜진 전조등 두 개가 잘려 보이고, 얼굴 주위로 아스팔트와 오른쪽 맨홀 덮개 하나가 보인다. 높은 카메라에서 내려다본 도로 배경은 요구된 구도에 잘 맞는다. 다만 맨홀은 열린 구멍 없이 덮개로 막혀 있어 명시된 지속 상태와 다르다. 지붕 십자가와 가로등은 화면 밖이라 평가할 수 없다.",
        "entities": "보이는 사람은 신부 한 명뿐이다. 인물 참조와 유사한 밝은 피부의 60대 남성으로, 뒤로 넘긴 짧은 회색 머리, 노년의 주름, 짧은 회색 수염, 검은 성직복과 흰 칼라가 확인된다. 눈은 정상적인 사람 눈이며 놀람은 눈꺼풀과 입의 연기로 표현된다. 밴과 도로는 참조의 낡은 재질을 따른다. 찰리와 새는 보이지 않으며 읽을 수 있는 글자도 없다. 도로에는 제외하라고 한 제동 먼지를 연상시키는 연무가 남아 있다.",
        "hard_violations": [],
        "physics": "머리와 목, 어깨가 자연스럽게 연결되어 있고 목을 젖힌 자세도 가능하다. 몸통 아래와 발은 클로즈업 밖이므로 지면 접촉을 직접 확인할 수 없지만 부유하는 모습은 아니다. 밴과 맨홀 덮개는 도로에 놓여 있으며, 지지 없이 떠 있는 물체나 비정상적인 신체 구조는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "신부의 턱과 눈동자는 화면 위쪽을 향한다. 화면 밖의 키 큰 찰리를 바라보는 방향과는 맞지만 A보다 눈을 동그랗게 뜬 놀람이 약하다. 왼쪽 전조등은 신부 뒤쪽 도로와 연무를 강하게 비춘다.",
        "built_space": "왼쪽에 큰 밴의 전면이 있고 그 뒤로 별도의 앞유리와 지붕 윤곽이 다시 나타나 밴이 두 대 겹친 듯한 구조다. 완전한 지붕 십자가 하나와 왼쪽 끝의 잘린 발광물도 보인다. 오른쪽 아래에는 열린 맨홀 하나와 옆으로 옮겨진 덮개 하나가 있다. 뒤쪽의 적층 컨테이너, 창문, 확성기와 오른쪽 골목의 가로등 두 개는 참조 장소를 따른다. 그러나 건물과 차량을 넓게 보여 주어 도로가 좁은 배경 여백으로만 남는 구도는 아니다.",
        "entities": "신부 한 명의 회색 머리, 노년의 얼굴, 밝은 피부, 짧은 회색 수염과 흰 성직자 칼라는 인물 참조에 대체로 맞는다. 찰리나 이전 장면의 인물은 보이지 않는다. 열린 맨홀과 켜진 전조등, 지붕 십자가는 확인되지만 차량 형상이 중복된다. 전조등 앞에는 제외 대상인 제동 먼지처럼 보이는 짙은 연무가 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "한 대여야 할 정차 밴에 별도의 앞유리와 지붕·차체가 뒤쪽으로 중복되어, 추가 차량이 겹쳐진 공간 구조를 만든다."
        ],
        "physics": "신부의 머리와 어깨는 자연스럽게 이어지고 위를 보는 목의 자세도 물리적으로 가능하다. 발은 화면 밖이므로 접지 상태는 확인할 수 없다. 열린 맨홀 옆 덮개는 도로에 놓여 있고 전조등과 십자가는 차량에 부착되어 있다. 다만 중복된 차량 윤곽의 접합은 하나의 실제 차체 구조로 성립하기 어렵다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "높은 위치에서 내려다본 얼굴 클로즈업과 동그랗게 뜬 눈이 핵심 순간에 더 충실하지만, 맨홀이 닫혀 있고 제외해야 할 먼지성 연무가 남아 있다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "신부의 올려다보는 시선과 열린 맨홀은 맞지만, 밴의 차체·앞유리가 중복되고 배경을 크게 드러내 좁은 도로 여백 중심의 클로즈업에서 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "신부는 턱을 들고 두 눈을 크게 뜬 채 높은 카메라 부근을 바라본다. 화면 밖 위쪽에 선 찰리를 올려다보는 관계로 자연스럽게 읽힌다. 찰리 자체는 보이지 않아 정확한 눈맞춤 지점까지 확인할 수는 없다. 왼쪽 밴의 전조등은 도로를 비추고 신부의 얼굴은 그 직접적인 빛줄기 바깥에서 읽힌다.",
        "built_space": "왼쪽 위에 밴 한 대의 전면 일부와 켜진 전조등 두 개가 잘려 보이고, 얼굴 주위로 아스팔트와 오른쪽 맨홀 덮개 하나가 보인다. 높은 카메라에서 내려다본 도로 배경은 요구된 구도에 잘 맞는다. 다만 맨홀은 열린 구멍 없이 덮개로 막혀 있어 명시된 지속 상태와 다르다. 지붕 십자가와 가로등은 화면 밖이라 평가할 수 없다.",
        "entities": "보이는 사람은 신부 한 명뿐이다. 인물 참조와 유사한 밝은 피부의 60대 남성으로, 뒤로 넘긴 짧은 회색 머리, 노년의 주름, 짧은 회색 수염, 검은 성직복과 흰 칼라가 확인된다. 눈은 정상적인 사람 눈이며 놀람은 눈꺼풀과 입의 연기로 표현된다. 밴과 도로는 참조의 낡은 재질을 따른다. 찰리와 새는 보이지 않으며 읽을 수 있는 글자도 없다. 도로에는 제외하라고 한 제동 먼지를 연상시키는 연무가 남아 있다.",
        "hard_violations": [],
        "physics": "머리와 목, 어깨가 자연스럽게 연결되어 있고 목을 젖힌 자세도 가능하다. 몸통 아래와 발은 클로즈업 밖이므로 지면 접촉을 직접 확인할 수 없지만 부유하는 모습은 아니다. 밴과 맨홀 덮개는 도로에 놓여 있으며, 지지 없이 떠 있는 물체나 비정상적인 신체 구조는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "신부의 턱과 눈동자는 화면 위쪽을 향한다. 화면 밖의 키 큰 찰리를 바라보는 방향과는 맞지만 A보다 눈을 동그랗게 뜬 놀람이 약하다. 왼쪽 전조등은 신부 뒤쪽 도로와 연무를 강하게 비춘다.",
        "built_space": "왼쪽에 큰 밴의 전면이 있고 그 뒤로 별도의 앞유리와 지붕 윤곽이 다시 나타나 밴이 두 대 겹친 듯한 구조다. 완전한 지붕 십자가 하나와 왼쪽 끝의 잘린 발광물도 보인다. 오른쪽 아래에는 열린 맨홀 하나와 옆으로 옮겨진 덮개 하나가 있다. 뒤쪽의 적층 컨테이너, 창문, 확성기와 오른쪽 골목의 가로등 두 개는 참조 장소를 따른다. 그러나 건물과 차량을 넓게 보여 주어 도로가 좁은 배경 여백으로만 남는 구도는 아니다.",
        "entities": "신부 한 명의 회색 머리, 노년의 얼굴, 밝은 피부, 짧은 회색 수염과 흰 성직자 칼라는 인물 참조에 대체로 맞는다. 찰리나 이전 장면의 인물은 보이지 않는다. 열린 맨홀과 켜진 전조등, 지붕 십자가는 확인되지만 차량 형상이 중복된다. 전조등 앞에는 제외 대상인 제동 먼지처럼 보이는 짙은 연무가 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "한 대여야 할 정차 밴에 별도의 앞유리와 지붕·차체가 뒤쪽으로 중복되어, 추가 차량이 겹쳐진 공간 구조를 만든다."
        ],
        "physics": "신부의 머리와 어깨는 자연스럽게 이어지고 위를 보는 목의 자세도 물리적으로 가능하다. 발은 화면 밖이므로 접지 상태는 확인할 수 없다. 열린 맨홀 옆 덮개는 도로에 놓여 있고 전조등과 십자가는 차량에 부착되어 있다. 다만 중복된 차량 윤곽의 접합은 하나의 실제 차체 구조로 성립하기 어렵다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.905,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.655,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gpt-high] 한 대여야 할 정차 밴에 별도의 앞유리와 지붕·차체가 뒤쪽으로 중복되어, 추가 차량이 겹쳐진 공간 구조를 만든다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 655
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "명시적으로 제외를 지시한 타이어 먼지를 정확히 제거했으며, 하이앵글(elevated) 클로즈업을 통해 도로 바닥을 배경으로 삼은 프레이밍 지시를 충실히 따랐습니다."
   },
   {
    "label": "A",
    "score": 655,
    "verdict_ko": "프롬프트에서 제외하라고 명시한 브레이크 타이어 먼지가 배경에 짙게 남아있으며, 지시된 부감(elevated) 구도가 아닌 로우 앵글에 가까워 배경 지시를 벗어났습니다.  ★위반: [gpt-high] 한 대여야 할 정차 밴에 별도의 앞유리와 지붕·차체가 뒤쪽으로 중복되어, 추가 차량이 겹쳐진 공간 구조를 만든다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S27sh14_sel.png",
    "asset_id": "f9e3a09f-302a-41c0-bce9-c3994d61b9f7",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:672084>",
    "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda7-dad3-7155-aff3-4b148bb14b77",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S27sh14"
  }
 },
 "S27sh20::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:30:21.686156+00:00",
  "fingerprint": "97ca9adbd75bb3fb4e30d308aa48af595f42f9d31e92d5df4d5cf025fdf1f7e3",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S27sh20_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S27sh20_sel.png",
  "source_sha256": "5bd6edc8bcc121652e2e04118d6e942aa4dbbe5e7f60840d1a03e5e28614de34",
  "file": "S27sh20_cine.png",
  "staged_sha256": "4f926ac336ee95b6e84bd55cb9b9c027f89f64a75d7ccbe129d24c38734da7e7",
  "latency_ms": 10870
 },
 "S27sh22::signage": {
  "fp": "75ecf59fe7b933c1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S27sh22": {
  "input_fingerprint": "0b2f1996e790b86c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 골목 구석에서 구도환이, 자동차 뒷문 안으로 상체를 숙여 반쯤 들어간 찰리의 뒷모습을 매섭게 주시하는 시선.\n\nLOCATION (lock): In a dark alley recess overlooking the stopped van's open rear doorway on the refugee-settlement street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open rear doorway receiving 찰리 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Alley corner (구도환 watches from its dark, concealed side) — The inner side of the corner borders the camera's view toward the street; used as Provides a narrow foreground occlusion beside the observer without masking his line of sight; Vehicle rear doorway (Open, with 찰리's upper body halfway inside) — Seen from an oblique rear angle, exposing the entry opening and 찰리's back; used as Anchors the observed action in the center-right background and establishes that 찰리 is entering rather than leaving.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep 구도환 within the alley's established darkness while preserving enough nighttime visibility to read 찰리's boarding posture beyond him.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The stopped van retains its halogen roof cross and the boxes spaced through its rear seating area. Charlie is boarding in his old coat and hat; the street manhole remains open and the small bird is no longer perched on him. 구도환: He remains nearby, closely observing the boarding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 골목 구석에서 구도환이, 자동차 뒷문 안으로 상체를 숙여 반쯤 들어간 찰리의 뒷모습을 매섭게 주시하는 시선.\n\nLOCATION (lock): In a dark alley recess overlooking the stopped van's open rear doorway on the refugee-settlement street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open rear doorway receiving 찰리 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Alley corner (구도환 watches from its dark, concealed side) — The inner side of the corner borders the camera's view toward the street; used as Provides a narrow foreground occlusion beside the observer without masking his line of sight; Vehicle rear doorway (Open, with 찰리's upper body halfway inside) — Seen from an oblique rear angle, exposing the entry opening and 찰리's back; used as Anchors the observed action in the center-right background and establishes that 찰리 is entering rather than leaving.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep 구도환 within the alley's established darkness while preserving enough nighttime visibility to read 찰리's boarding posture beyond him.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The stopped van retains its halogen roof cross and the boxes spaced through its rear seating area. Charlie is boarding in his old coat and hat; the street manhole remains open and the small bird is no longer perched on him. 구도환: He remains nearby, closely observing the boarding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 골목 구석에서 구도환이, 자동차 뒷문 안으로 상체를 숙여 반쯤 들어간 찰리의 뒷모습을 매섭게 주시하는 시선.\n\nLOCATION (lock): In a dark alley recess overlooking the stopped van's open rear doorway on the refugee-settlement street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open rear doorway receiving 찰리 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Alley corner (구도환 watches from its dark, concealed side) — The inner side of the corner borders the camera's view toward the street; used as Provides a narrow foreground occlusion beside the observer without masking his line of sight; Vehicle rear doorway (Open, with 찰리's upper body halfway inside) — Seen from an oblique rear angle, exposing the entry opening and 찰리's back; used as Anchors the observed action in the center-right background and establishes that 찰리 is entering rather than leaving.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep 구도환 within the alley's established darkness while preserving enough nighttime visibility to read 찰리's boarding posture beyond him.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The stopped van retains its halogen roof cross and the boxes spaced through its rear seating area. Charlie is boarding in his old coat and hat; the street manhole remains open and the small bird is no longer perched on him. 구도환: He remains nearby, closely observing the boarding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "구도환은 찰리가 아닌 카메라 렌즈를 정면으로 응시함. 찰리는 밴 내부를 향하고 있음.",
    "built_space": "왼쪽 전경에 골목 벽이 위치함. 밴의 뒷문이 열려 있고, 바닥의 맨홀은 닫힌 상태임.",
    "entities": "구도환은 레퍼런스와 일치하나, 찰리 역시 구도환과 동일한 중절모와 트렌치코트를 입고 있음. 밴 지붕에 밝게 빛나는 십자가 조명이 있음.",
    "hard_violations": [],
    "physics": "인물들이 땅을 제대로 딛고 서 있으며 물리적 지지 상태에 이상 없음."
   },
   {
    "label": "B",
    "direction": "구도환은 밴 쪽에 있는 찰리를 향해 시선을 고정함. 찰리는 밴 내부를 향해 상체를 기울이고 있음.",
    "built_space": "왼쪽 어두운 골목 벽이 전경을 가림. 밴 뒷문이 개방되어 내부에 상자들이 보이며, 바닥의 맨홀은 뚜껑이 열려 있음.",
    "entities": "구도환은 레퍼런스 의상과 일치함. 찰리는 낡은 코트와 비니를 착용함. 밴 지붕에 십자가 구조물이 있으나 불이 켜져 있지 않음.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥을 안정적으로 딛고 있으며 찰리가 상체를 숙인 자세가 자연스럽게 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도환이 찰리가 아닌 카메라를 응시하고 닫힌 맨홀을 묘사했으며 찰리의 복장이 중복되어 핵심 연출 지시를 위반함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "구도환의 정확한 시선 방향, 상체를 숙인 찰리의 자세, 열린 맨홀 등 프롬프트의 상황과 환경 디테일을 훌륭하게 구현함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "구도환은 찰리가 아닌 카메라 렌즈를 정면으로 응시함. 찰리는 밴 내부를 향하고 있음.",
        "built_space": "왼쪽 전경에 골목 벽이 위치함. 밴의 뒷문이 열려 있고, 바닥의 맨홀은 닫힌 상태임.",
        "entities": "구도환은 레퍼런스와 일치하나, 찰리 역시 구도환과 동일한 중절모와 트렌치코트를 입고 있음. 밴 지붕에 밝게 빛나는 십자가 조명이 있음.",
        "hard_violations": [],
        "physics": "인물들이 땅을 제대로 딛고 서 있으며 물리적 지지 상태에 이상 없음."
       },
       {
        "label": "B",
        "direction": "구도환은 밴 쪽에 있는 찰리를 향해 시선을 고정함. 찰리는 밴 내부를 향해 상체를 기울이고 있음.",
        "built_space": "왼쪽 어두운 골목 벽이 전경을 가림. 밴 뒷문이 개방되어 내부에 상자들이 보이며, 바닥의 맨홀은 뚜껑이 열려 있음.",
        "entities": "구도환은 레퍼런스 의상과 일치함. 찰리는 낡은 코트와 비니를 착용함. 밴 지붕에 십자가 구조물이 있으나 불이 켜져 있지 않음.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥을 안정적으로 딛고 있으며 찰리가 상체를 숙인 자세가 자연스럽게 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도환이 찰리가 아닌 카메라를 응시하고 닫힌 맨홀을 묘사했으며 찰리의 복장이 중복되어 핵심 연출 지시를 위반함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "구도환의 정확한 시선 방향, 상체를 숙인 찰리의 자세, 열린 맨홀 등 프롬프트의 상황과 환경 디테일을 훌륭하게 구현함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "구도환은 찰리가 아닌 카메라 렌즈를 정면으로 응시함. 찰리는 밴 내부를 향하고 있음.",
        "built_space": "왼쪽 전경에 골목 벽이 위치함. 밴의 뒷문이 열려 있고, 바닥의 맨홀은 닫힌 상태임.",
        "entities": "구도환은 레퍼런스와 일치하나, 찰리 역시 구도환과 동일한 중절모와 트렌치코트를 입고 있음. 밴 지붕에 밝게 빛나는 십자가 조명이 있음.",
        "hard_violations": [],
        "physics": "인물들이 땅을 제대로 딛고 서 있으며 물리적 지지 상태에 이상 없음."
       },
       {
        "label": "B",
        "direction": "구도환은 밴 쪽에 있는 찰리를 향해 시선을 고정함. 찰리는 밴 내부를 향해 상체를 기울이고 있음.",
        "built_space": "왼쪽 어두운 골목 벽이 전경을 가림. 밴 뒷문이 개방되어 내부에 상자들이 보이며, 바닥의 맨홀은 뚜껑이 열려 있음.",
        "entities": "구도환은 레퍼런스 의상과 일치함. 찰리는 낡은 코트와 비니를 착용함. 밴 지붕에 십자가 구조물이 있으나 불이 켜져 있지 않음.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥을 안정적으로 딛고 있으며 찰리가 상체를 숙인 자세가 자연스럽게 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도환이 오른쪽 차량 쪽을 엿보는 관계는 더 충실하지만, 찰리의 상체가 거의 서 있고 지붕 십자가가 보이지 않아 승차 순간과 소품 연속성이 부족하다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "찰리의 굽힌 승차 자세와 지붕 십자가는 맞지만, 구도환이 찰리가 아닌 카메라 쪽을 바라보아 장면의 핵심인 감시 시선이 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 구도환은 얼굴과 눈을 화면 오른쪽 거리 방향으로 돌려 차량 뒤의 찰리를 엿보는 것으로 읽힌다. 찰리는 등을 카메라에 보이고 차량 내부를 향하지만, 상체를 반쯤 집어넣기보다는 열린 문 앞에 거의 똑바로 서 있다.",
        "built_space": "왼쪽 전경의 골목 벽 뒤에 구도환이 있고, 중앙 오른쪽에 밴의 후면 출입구 하나와 열린 문짝 두 개가 보인다. 벽의 어두운 면이 화면 중앙까지 넓게 차지해 요구된 좁은 전경 가림보다 두껍다. 차량 내부에는 적어도 세 개의 상자가 있으며 오른쪽 노면에는 열린 맨홀 하나가 있다. 녹슨 적층 컨테이너, 창문, 확성기와 야간 가로등은 참고 장소의 재료와 시설을 대체로 이어간다.",
        "entities": "인물은 구도환과 뒷모습의 찰리 두 명뿐이다. 구도환은 중년 남성의 얼굴, 희끗한 수염, 짙은 중절모와 낡은 갈색 계열 코트로 참고 인물에 가깝다. 찰리는 낡은 짙은 코트와 니트 모자를 착용했으며 얼굴이 가려져 나이와 민족적 외모는 판별할 수 없다. 밴과 상자, 열린 맨홀은 보이지만, 노출된 지붕에는 유지되어야 할 발광 십자가가 보이지 않는다. 새나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 두 발은 노면에 닿아 있고 왼손은 후문 가장자리 부근에 놓여 있어 지지 없는 몸이나 물체는 없다. 다만 양발을 땅에 둔 거의 직립 자세라 상체를 깊이 숙여 들어가는 동작은 약하다. 구도환의 하체 접점은 가려져 있지만 서서 몸을 내민 자세에 물리적 모순은 없다. 밴은 바퀴로, 상자는 차량 내부 바닥으로 지지된다."
       },
       {
        "label": "B",
        "direction": "구도환의 얼굴과 눈은 거의 정면인 카메라 쪽을 향하며, 오른쪽 뒤에 있는 찰리에게 시선이 닿지 않는다. 찰리는 등을 보인 채 머리와 어깨를 후면 출입구 안쪽으로 숙이고 있어 차량에 들어가는 방향과 행동은 분명하다.",
        "built_space": "구도환은 왼쪽 어두운 골목 안에, 밴의 열린 후면 출입구는 중앙 오른쪽에 배치되어 있다. 출입구 하나와 바깥으로 열린 문짝 두 개, 지붕 십자가 하나가 보인다. 골목 기둥이 두 공간 사이를 구분한다. 배경의 적층 컨테이너, 확성기, 창문과 출입문은 참고 장소와 유사하다. 노출된 도로에는 열린 맨홀이 확인되지 않으며 차량 내부 상자도 어둠 속에서 식별되지 않는다.",
        "entities": "구도환과 찰리로 읽히는 두 사람만 등장한다. 구도환의 중년 남성 외모, 수염, 중절모와 낡은 코트는 참고와 대체로 맞는다. 찰리는 낡은 긴 코트와 챙 있는 모자를 쓰고 있으며 얼굴이 보이지 않아 세부 정체성은 확인할 수 없다. 오래된 밴과 발광 지붕 십자가는 유지되었다. 새나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 양발을 노면에 둔 채 허리와 무릎을 굽혀 상체를 차량 안으로 기울인다. 손의 정확한 접점은 내부 어둠에 가려져 있지만 지면에 닿은 다리로 체중을 지탱할 수 있는 자세다. 구도환 역시 자연스럽게 서 있으며 발은 프레임 밖이다. 문짝은 경첩에, 십자가는 지붕 장착대에 지지되어 있고 부유하거나 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도환이 오른쪽 차량 쪽을 엿보는 관계는 더 충실하지만, 찰리의 상체가 거의 서 있고 지붕 십자가가 보이지 않아 승차 순간과 소품 연속성이 부족하다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리의 굽힌 승차 자세와 지붕 십자가는 맞지만, 구도환이 찰리가 아닌 카메라 쪽을 바라보아 장면의 핵심인 감시 시선이 어긋난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 구도환은 얼굴과 눈을 화면 오른쪽 거리 방향으로 돌려 차량 뒤의 찰리를 엿보는 것으로 읽힌다. 찰리는 등을 카메라에 보이고 차량 내부를 향하지만, 상체를 반쯤 집어넣기보다는 열린 문 앞에 거의 똑바로 서 있다.",
        "built_space": "왼쪽 전경의 골목 벽 뒤에 구도환이 있고, 중앙 오른쪽에 밴의 후면 출입구 하나와 열린 문짝 두 개가 보인다. 벽의 어두운 면이 화면 중앙까지 넓게 차지해 요구된 좁은 전경 가림보다 두껍다. 차량 내부에는 적어도 세 개의 상자가 있으며 오른쪽 노면에는 열린 맨홀 하나가 있다. 녹슨 적층 컨테이너, 창문, 확성기와 야간 가로등은 참고 장소의 재료와 시설을 대체로 이어간다.",
        "entities": "인물은 구도환과 뒷모습의 찰리 두 명뿐이다. 구도환은 중년 남성의 얼굴, 희끗한 수염, 짙은 중절모와 낡은 갈색 계열 코트로 참고 인물에 가깝다. 찰리는 낡은 짙은 코트와 니트 모자를 착용했으며 얼굴이 가려져 나이와 민족적 외모는 판별할 수 없다. 밴과 상자, 열린 맨홀은 보이지만, 노출된 지붕에는 유지되어야 할 발광 십자가가 보이지 않는다. 새나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 두 발은 노면에 닿아 있고 왼손은 후문 가장자리 부근에 놓여 있어 지지 없는 몸이나 물체는 없다. 다만 양발을 땅에 둔 거의 직립 자세라 상체를 깊이 숙여 들어가는 동작은 약하다. 구도환의 하체 접점은 가려져 있지만 서서 몸을 내민 자세에 물리적 모순은 없다. 밴은 바퀴로, 상자는 차량 내부 바닥으로 지지된다."
       },
       {
        "label": "A",
        "direction": "구도환의 얼굴과 눈은 거의 정면인 카메라 쪽을 향하며, 오른쪽 뒤에 있는 찰리에게 시선이 닿지 않는다. 찰리는 등을 보인 채 머리와 어깨를 후면 출입구 안쪽으로 숙이고 있어 차량에 들어가는 방향과 행동은 분명하다.",
        "built_space": "구도환은 왼쪽 어두운 골목 안에, 밴의 열린 후면 출입구는 중앙 오른쪽에 배치되어 있다. 출입구 하나와 바깥으로 열린 문짝 두 개, 지붕 십자가 하나가 보인다. 골목 기둥이 두 공간 사이를 구분한다. 배경의 적층 컨테이너, 확성기, 창문과 출입문은 참고 장소와 유사하다. 노출된 도로에는 열린 맨홀이 확인되지 않으며 차량 내부 상자도 어둠 속에서 식별되지 않는다.",
        "entities": "구도환과 찰리로 읽히는 두 사람만 등장한다. 구도환의 중년 남성 외모, 수염, 중절모와 낡은 코트는 참고와 대체로 맞는다. 찰리는 낡은 긴 코트와 챙 있는 모자를 쓰고 있으며 얼굴이 보이지 않아 세부 정체성은 확인할 수 없다. 오래된 밴과 발광 지붕 십자가는 유지되었다. 새나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 양발을 노면에 둔 채 허리와 무릎을 굽혀 상체를 차량 안으로 기울인다. 손의 정확한 접점은 내부 어둠에 가려져 있지만 지면에 닿은 다리로 체중을 지탱할 수 있는 자세다. 구도환 역시 자연스럽게 서 있으며 발은 프레임 밖이다. 문짝은 경첩에, 십자가는 지붕 장착대에 지지되어 있고 부유하거나 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.238,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.238,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 1238,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1238,
    "verdict_ko": "구도환이 찰리가 아닌 카메라를 응시하고 닫힌 맨홀을 묘사했으며 찰리의 복장이 중복되어 핵심 연출 지시를 위반함."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "구도환의 정확한 시선 방향, 상체를 숙인 찰리의 자세, 열린 맨홀 등 프롬프트의 상황과 환경 디테일을 훌륭하게 구현함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S27sh14_sel.png",
    "asset_id": "f9e3a09f-302a-41c0-bce9-c3994d61b9f7",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929233>",
    "asset_id": "1297d7df-4792-4e39-bdf8-1c0f5b59585c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda8-8e2c-7311-b490-fa662317de39",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S27sh14"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S27sh22::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:30:32.894568+00:00",
  "fingerprint": "0668416a54d2053a0db610a04a83938fc89392b5218a172bab2c1bb7f97f2042",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S27sh22_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S27sh22_sel.png",
  "source_sha256": "61421a694b4ec1888647f4c89a7e30557ee8e7386773e2d9b02ede05e59867f3",
  "file": "S27sh22_cine.png",
  "staged_sha256": "963cfb830251d84141ba9eee120ab1eff94897b87fef113ad46b4ff4296d4471",
  "latency_ms": 10198
 },
 "S28sh6::signage": {
  "fp": "b37b2c6b3750b600",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::b40a4732f264f785": {
  "subjects": [],
  "subject_text": "신부의 승합차가 통과하는 민병대 검문 지점\n컨테이너 주거지 사이 도로에 마련된 야간 검문 구간. 어두운 노면을 따라 가로등이 이어지고 도로 가장자리에 건물들이 보인다.",
  "identity": "canonical",
  "scope_id": "L34",
  "scope_role": "location_exterior",
  "scope_sha": "98e744804b8f50f6"
 },
 "groupbg::supply_van_interior": {
  "input_fingerprint": "db6938a6e3659dae",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "supply_van_interior",
    "tags": [
     "S28sh6",
     "S28sh7"
    ]
   },
   "context_sig": "004c76b52b40af79"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: Inside the rear passenger and cargo area of a van, between scattered cardboard boxes with nighttime light entering through its windows.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n신부의 승합차가 통과하는 민병대 검문 지점: 어두운 도로변에 민병대원들이 바리케이드를 치고 손전등으로 통행 차량을 비추는 임시 검문소. (특징: 차창을 비추는 날카로운 손전등 불빛; 민병대 완장을 찬 무장 대원들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 마치 구급대 차량처럼, 뒷좌석에는 물건 박스가 듬성듬성.\n- 민병대. 창문을 내리라고 지시하자, 찡긋 미소를 짓는 신부.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: Inside the rear passenger and cargo area of a van, between scattered cardboard boxes with nighttime light entering through its windows.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n신부의 승합차가 통과하는 민병대 검문 지점: 어두운 도로변에 민병대원들이 바리케이드를 치고 손전등으로 통행 차량을 비추는 임시 검문소. (특징: 차창을 비추는 날카로운 손전등 불빛; 민병대 완장을 찬 무장 대원들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 마치 구급대 차량처럼, 뒷좌석에는 물건 박스가 듬성듬성.\n- 민병대. 창문을 내리라고 지시하자, 찡긋 미소를 짓는 신부.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_supply_van_interior_deaa25.png",
  "asset_id": "bea47285-24c3-4f7a-a723-52f9d4b2e93f",
  "input_asset_ids": [
   "63d83325-d1b8-4ba9-8d9b-7fc9b13331b5"
  ],
  "origin_tag": "S28sh6",
  "place_text": "Inside the rear passenger and cargo area of a van, between scattered cardboard boxes with nighttime light entering through its windows.",
  "origin_inputs": {
   "place_text": "Inside the rear passenger and cargo area of a van, between scattered cardboard boxes with nighttime light entering through its windows.",
   "time_of_day_en": "night",
   "conti_asset_id": "63d83325-d1b8-4ba9-8d9b-7fc9b13331b5"
  }
 },
 "S28sh6::bgfirst_bg": {
  "input_fingerprint": "e64fff8d5edba574",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 찰리의 거대한 금속 상체를 종이상자들 틈새 바닥으로 거칠게 짓누르는 힘이 실린 자세.\n\nLOCATION (lock): Inside the rear passenger and cargo area of a van, between scattered cardboard boxes with nighttime light entering through its windows.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Paper boxes (Sparsely arranged around the hiding gap) — Their upper edges and inward-facing sides border the bodies without masking the pressing hands; used as Fragmented edge framing that makes concealment and restricted space immediately legible; Rear-compartment floor (Exposed between the boxes); used as Visible destination of the downward pressure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime interior exposure preserves readable faces, hands, and metal contours without specifying an additional light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 찰리의 거대한 금속 상체를 종이상자들 틈새 바닥으로 거칠게 짓누르는 힘이 실린 자세.\n\nLOCATION (lock): Inside the rear passenger and cargo area of a van, between scattered cardboard boxes with nighttime light entering through its windows.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Paper boxes (Sparsely arranged around the hiding gap) — Their upper edges and inward-facing sides border the bodies without masking the pressing hands; used as Fragmented edge framing that makes concealment and restricted space immediately legible; Rear-compartment floor (Exposed between the boxes); used as Visible destination of the downward pressure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime interior exposure preserves readable faces, hands, and metal contours without specifying an additional light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S28sh6__bgfirst_bg.png",
  "asset_id": "e8037db3-b1e3-46d6-a566-dfa1c6b985d6",
  "input_asset_ids": [
   "63d83325-d1b8-4ba9-8d9b-7fc9b13331b5",
   "bea47285-24c3-4f7a-a723-52f9d4b2e93f"
  ]
 },
 "S28sh6": {
  "input_fingerprint": "4ad96af67c9b032f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 거대한 금속 상체를 종이상자들 틈새 바닥으로 거칠게 짓누르는 힘이 실린 자세.\n\nLOCATION (lock): Inside the rear passenger and cargo area of a van, between scattered cardboard boxes with nighttime light entering through its windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Paper boxes (Sparsely arranged around the hiding gap) — Their upper edges and inward-facing sides border the bodies without masking the pressing hands; used as Fragmented edge framing that makes concealment and restricted space immediately legible; Rear-compartment floor (Exposed between the boxes); used as Visible destination of the downward pressure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime interior exposure preserves readable faces, hands, and metal contours without specifying an additional light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boxes remain spaced through the van's rear seating area, providing hiding gaps, and the crude halogen cross remains on the roof. Charlie is being concealed low among the boxes, still wearing his old coat and hat. 이현우: He crouches low in the rear box-filled area, with facial bruising and the injured leg. His outer garment remains removed, and the contact card remains hidden in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 거대한 금속 상체를 종이상자들 틈새 바닥으로 거칠게 짓누르는 힘이 실린 자세.\n\nLOCATION (lock): Inside the rear passenger and cargo area of a van, between scattered cardboard boxes with nighttime light entering through its windows. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Paper boxes (Sparsely arranged around the hiding gap) — Their upper edges and inward-facing sides border the bodies without masking the pressing hands; used as Fragmented edge framing that makes concealment and restricted space immediately legible; Rear-compartment floor (Exposed between the boxes); used as Visible destination of the downward pressure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime interior exposure preserves readable faces, hands, and metal contours without specifying an additional light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boxes remain spaced through the van's rear seating area, providing hiding gaps, and the crude halogen cross remains on the roof. Charlie is being concealed low among the boxes, still wearing his old coat and hat. 이현우: He crouches low in the rear box-filled area, with facial bruising and the injured leg. His outer garment remains removed, and the contact card remains hidden in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 거대한 금속 상체를 종이상자들 틈새 바닥으로 거칠게 짓누르는 힘이 실린 자세.\n\nLOCATION (lock): Inside the rear passenger and cargo area of a van, between scattered cardboard boxes with nighttime light entering through its windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Paper boxes (Sparsely arranged around the hiding gap) — Their upper edges and inward-facing sides border the bodies without masking the pressing hands; used as Fragmented edge framing that makes concealment and restricted space immediately legible; Rear-compartment floor (Exposed between the boxes); used as Visible destination of the downward pressure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime interior exposure preserves readable faces, hands, and metal contours without specifying an additional light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boxes remain spaced through the van's rear seating area, providing hiding gaps, and the crude halogen cross remains on the roof. Charlie is being concealed low among the boxes, still wearing his old coat and hat. 이현우: He crouches low in the rear box-filled area, with facial bruising and the injured leg. His outer garment remains removed, and the contact card remains hidden in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S28sh6__bgfirst_bg.png",
     "asset_id": "e8037db3-b1e3-46d6-a566-dfa1c6b985d6",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S28sh6.png",
     "asset_id": "63d83325-d1b8-4ba9-8d9b-7fc9b13331b5",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1597480>",
     "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_supply_van_interior_deaa25.png",
     "asset_id": "bea47285-24c3-4f7a-a723-52f9d4b2e93f",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1597480>",
     "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선과 양손이 바닥에 엎드린 찰리의 등을 향해 아래로 똑바로 향하고 있음.",
    "built_space": "야간의 밴 내부. 종이상자들이 인물 주변으로 느슨하게 배치되어 있고 바닥면이 노출됨. 천장에는 일반적인 차량 실내등만 보이며 할로겐 십자가는 없음.",
    "entities": "이현우는 상처 난 얼굴의 한국인 남성이나 지시와 다르게 겉옷(후드티)을 입고 있음. 찰리는 흰 얼굴판, 점 형태의 눈, 선 형태의 입을 가졌으며 명시된 대로 코트와 모자를 착용함.",
    "hard_violations": [],
    "physics": "이현우가 무릎을 굽히고 체중을 실어 찰리를 누르고 있으며, 찰리의 묵직한 금속 상체는 밴 바닥면에 안정적으로 밀착되어 지탱됨."
   },
   {
    "label": "B",
    "direction": "이현우의 시선은 아래를 향하고, 양손은 등을 대고 누운 찰리의 가슴 부위를 누르고 있음.",
    "built_space": "밴 내부. 종이상자들이 주변에 쌓여 있음. 천장에 테이프로 만든 X자 형태가 붙어 있으나 할로겐 광원은 아님.",
    "entities": "이현우는 겉옷을 벗은 상태(긴팔 티셔츠)로 지시를 충족함. 찰리는 코트와 모자가 없고, 지시된 흰 얼굴판 대신 완전히 금속으로 된 맨 얼굴을 하고 있음.",
    "hard_violations": [],
    "physics": "이현우가 무릎을 꿇고 찰리 위에 웅크린 자세이며, 찰리는 바닥에 등을 대고 안정적으로 누워 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이현우가 겉옷(후드티)을 벗지 않은 점과 천장의 할로겐 십자가가 누락된 점은 아쉽지만, 찰리의 구체적인 기계 얼굴 형태(흰 얼굴판, 점 눈)와 의상(코트, 모자)을 정확히 구현했으며 거칠게 짓누르는 액션의 무게감이 훌륭합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이현우의 겉옷 탈의 조건은 충족했으나, 찰리의 핵심적인 흰 얼굴판 묘사가 전혀 없고 코트와 모자마저 완전히 누락되어 프롬프트의 캐릭터 지시를 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 양손이 바닥에 엎드린 찰리의 등을 향해 아래로 똑바로 향하고 있음.",
        "built_space": "야간의 밴 내부. 종이상자들이 인물 주변으로 느슨하게 배치되어 있고 바닥면이 노출됨. 천장에는 일반적인 차량 실내등만 보이며 할로겐 십자가는 없음.",
        "entities": "이현우는 상처 난 얼굴의 한국인 남성이나 지시와 다르게 겉옷(후드티)을 입고 있음. 찰리는 흰 얼굴판, 점 형태의 눈, 선 형태의 입을 가졌으며 명시된 대로 코트와 모자를 착용함.",
        "hard_violations": [],
        "physics": "이현우가 무릎을 굽히고 체중을 실어 찰리를 누르고 있으며, 찰리의 묵직한 금속 상체는 밴 바닥면에 안정적으로 밀착되어 지탱됨."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 아래를 향하고, 양손은 등을 대고 누운 찰리의 가슴 부위를 누르고 있음.",
        "built_space": "밴 내부. 종이상자들이 주변에 쌓여 있음. 천장에 테이프로 만든 X자 형태가 붙어 있으나 할로겐 광원은 아님.",
        "entities": "이현우는 겉옷을 벗은 상태(긴팔 티셔츠)로 지시를 충족함. 찰리는 코트와 모자가 없고, 지시된 흰 얼굴판 대신 완전히 금속으로 된 맨 얼굴을 하고 있음.",
        "hard_violations": [],
        "physics": "이현우가 무릎을 꿇고 찰리 위에 웅크린 자세이며, 찰리는 바닥에 등을 대고 안정적으로 누워 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이현우가 겉옷(후드티)을 벗지 않은 점과 천장의 할로겐 십자가가 누락된 점은 아쉽지만, 찰리의 구체적인 기계 얼굴 형태(흰 얼굴판, 점 눈)와 의상(코트, 모자)을 정확히 구현했으며 거칠게 짓누르는 액션의 무게감이 훌륭합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이현우의 겉옷 탈의 조건은 충족했으나, 찰리의 핵심적인 흰 얼굴판 묘사가 전혀 없고 코트와 모자마저 완전히 누락되어 프롬프트의 캐릭터 지시를 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 양손이 바닥에 엎드린 찰리의 등을 향해 아래로 똑바로 향하고 있음.",
        "built_space": "야간의 밴 내부. 종이상자들이 인물 주변으로 느슨하게 배치되어 있고 바닥면이 노출됨. 천장에는 일반적인 차량 실내등만 보이며 할로겐 십자가는 없음.",
        "entities": "이현우는 상처 난 얼굴의 한국인 남성이나 지시와 다르게 겉옷(후드티)을 입고 있음. 찰리는 흰 얼굴판, 점 형태의 눈, 선 형태의 입을 가졌으며 명시된 대로 코트와 모자를 착용함.",
        "hard_violations": [],
        "physics": "이현우가 무릎을 굽히고 체중을 실어 찰리를 누르고 있으며, 찰리의 묵직한 금속 상체는 밴 바닥면에 안정적으로 밀착되어 지탱됨."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 아래를 향하고, 양손은 등을 대고 누운 찰리의 가슴 부위를 누르고 있음.",
        "built_space": "밴 내부. 종이상자들이 주변에 쌓여 있음. 천장에 테이프로 만든 X자 형태가 붙어 있으나 할로겐 광원은 아님.",
        "entities": "이현우는 겉옷을 벗은 상태(긴팔 티셔츠)로 지시를 충족함. 찰리는 코트와 모자가 없고, 지시된 흰 얼굴판 대신 완전히 금속으로 된 맨 얼굴을 하고 있음.",
        "hard_violations": [],
        "physics": "이현우가 무릎을 꿇고 찰리 위에 웅크린 자세이며, 찰리는 바닥에 등을 대고 안정적으로 누워 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "힘을 실어 금속 가슴을 누르는 동작은 선명하지만, 압박의 도착점인 바닥이 잘 드러나지 않고 찰리의 흰 얼굴판·낡은 코트·모자가 빠졌다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "상자 틈의 노출된 바닥으로 찰리의 상체를 누르는 관계와 찰리의 얼굴·코트·모자를 더 충실히 구현했지만, 현우가 겉옷인 후드티를 그대로 입고 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 찰리의 가슴과 자기 손이 닿는 곳을 내려다본다. 양손을 금속 가슴판에 대고 아래로 누르므로 압박 대상과 방향은 맞는다. 바로 누운 찰리의 얼굴은 위쪽을 향하며 현우를 똑바로 바라보지는 않는다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "밴 뒤쪽에서 앞좌석 방향을 보는 실내다. 중앙 천장등 한 개, 좌우 창과 기둥, 왼쪽 안전벨트, 뒤편의 여러 좌석 등받이가 보이며 참조의 기본 배치와 재질을 유지한다. 상자는 왼쪽 전경·중경과 오른쪽 전경·중경에 있고 중앙 틈에 두 인물이 놓였다. 오른쪽에는 높이 쌓인 상자가 있어 틈이 참조보다 조밀하다. 손은 가리지 않지만 찰리의 몸이 틈 대부분을 채워 목적지인 바닥은 가장자리에서만 조금 보인다. 천장에는 참조에 없던 교차형 검은 부재가 보인다. 불가능한 반사나 명백한 고정 설비 중복은 없다.",
        "entities": "등장자는 현우와 찰리뿐이다. 현우는 헝클어진 검은 머리, 젊은 한국계 남성으로 보이는 얼굴, 볼의 멍을 갖췄으며 참조와 대체로 부합한다. 후드 겉옷을 벗고 얇은 상의를 입은 점도 지시와 맞는다. 찰리는 거대하고 마모된 금속 몸체지만, 얼굴은 흰 얼굴판과 점눈·선입 대신 짙은 금속의 고릴라형 얼굴이다. 코트와 모자는 화면에 드러난 상체와 머리에 없다. 골판지 상자와 야간 창밖은 적합하며 읽을 수 있는 글자는 없다. 지붕 바깥 십자가와 신발 속 카드는 이 구도에서 확인할 수 없다.",
        "hard_violations": [],
        "physics": "현우는 다리를 접고 낮게 웅크린 채 상체를 앞으로 기울이고, 양 손바닥으로 찰리의 가슴을 지지점 삼아 체중을 전달한다. 하체의 바닥 접점 일부는 가려져 있지만 공중에 떠 있는 모습은 아니다. 찰리는 등을 바닥 쪽에 둔 채 누워 있고 몸과 팔이 바닥 높이에 놓여 있어 지지가 자연스럽다. 상자는 바닥이나 아래 상자 위에 놓여 있다. 양손의 접촉과 힘을 주는 얼굴은 거칠게 누르는 행동으로 읽힌다."
       },
       {
        "label": "B",
        "direction": "현우는 찰리의 등·어깨 쪽 접촉 지점을 내려다보며 양손으로 상체를 아래로 누른다. 찰리는 엎드린 방향으로 낮아져 얼굴과 시선이 전경의 노출된 바닥 쪽을 향한다. 현우의 압박 방향이 상자 사이 바닥이라는 지정된 목표와 명확히 이어진다.",
        "built_space": "밴 후방에서 전방을 바라보는 실내로, 중앙 천장등 한 개, 좌우 창과 안전벨트, 뒤편 좌석 등받이가 참조와 유사하게 배치된다. 왼쪽에는 전경의 큰 상자와 뒤쪽의 세운 상자가 있고, 오른쪽에는 앞뒤로 떨어진 상자들이 중앙 통로를 둘러싼다. 현우는 통로 왼쪽에서 웅크리고 찰리는 그 앞의 틈에 낮게 엎드려 있다. 상자 가장자리와 안쪽 면이 몸을 둘러싸되 누르는 손은 가리지 않는다. 찰리의 얼굴과 손 앞에 바닥이 노출되어 압박의 목적지가 읽힌다. 불가능한 반사나 설비 중복은 보이지 않는다.",
        "entities": "현우와 찰리만 등장한다. 현우는 젊은 한국계 남성으로 보이는 얼굴, 검은 헝클어진 머리와 얼굴 멍을 갖췄다. 다만 참조의 어두운 후드티를 그대로 입어 겉옷을 벗은 상태라는 지시와 어긋난다. 찰리는 각진 기계 머리, 흰 얼굴판, 작은 점 형태 눈과 선 형태 입, 거대한 금속 상체를 갖췄으며 낡은 코트와 모자도 착용했다. 상자·금속·천의 질감과 야간 창밖이 구체적으로 보이고 읽을 수 있는 글자는 없다. 다리 부상은 보이는 부분만으로 뚜렷하게 확인되지 않으며, 신발 속 카드와 외부 지붕 십자가는 확인할 수 없다.",
        "hard_violations": [],
        "physics": "현우의 화면 왼쪽 신발은 바닥에 닿고 무릎은 굽혀져 있으며, 앞으로 기울인 몸의 힘이 찰리의 금속 상체에 닿은 양손으로 전달된다. 찰리의 두 기계 손과 전완은 바닥에 놓여 낮은 몸을 받친다. 아직 팔로 버티는 상체를 현우가 더 낮게 누르는 순간으로 물리적으로 성립한다. 코트는 몸 위에 처지고 모자는 머리에 얹혀 있으며, 상자는 바닥 또는 다른 상자에 지지된다. 지지 없이 떠 있는 몸이나 물건은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "힘을 실어 금속 가슴을 누르는 동작은 선명하지만, 압박의 도착점인 바닥이 잘 드러나지 않고 찰리의 흰 얼굴판·낡은 코트·모자가 빠졌다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "상자 틈의 노출된 바닥으로 찰리의 상체를 누르는 관계와 찰리의 얼굴·코트·모자를 더 충실히 구현했지만, 현우가 겉옷인 후드티를 그대로 입고 있다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 찰리의 가슴과 자기 손이 닿는 곳을 내려다본다. 양손을 금속 가슴판에 대고 아래로 누르므로 압박 대상과 방향은 맞는다. 바로 누운 찰리의 얼굴은 위쪽을 향하며 현우를 똑바로 바라보지는 않는다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "밴 뒤쪽에서 앞좌석 방향을 보는 실내다. 중앙 천장등 한 개, 좌우 창과 기둥, 왼쪽 안전벨트, 뒤편의 여러 좌석 등받이가 보이며 참조의 기본 배치와 재질을 유지한다. 상자는 왼쪽 전경·중경과 오른쪽 전경·중경에 있고 중앙 틈에 두 인물이 놓였다. 오른쪽에는 높이 쌓인 상자가 있어 틈이 참조보다 조밀하다. 손은 가리지 않지만 찰리의 몸이 틈 대부분을 채워 목적지인 바닥은 가장자리에서만 조금 보인다. 천장에는 참조에 없던 교차형 검은 부재가 보인다. 불가능한 반사나 명백한 고정 설비 중복은 없다.",
        "entities": "등장자는 현우와 찰리뿐이다. 현우는 헝클어진 검은 머리, 젊은 한국계 남성으로 보이는 얼굴, 볼의 멍을 갖췄으며 참조와 대체로 부합한다. 후드 겉옷을 벗고 얇은 상의를 입은 점도 지시와 맞는다. 찰리는 거대하고 마모된 금속 몸체지만, 얼굴은 흰 얼굴판과 점눈·선입 대신 짙은 금속의 고릴라형 얼굴이다. 코트와 모자는 화면에 드러난 상체와 머리에 없다. 골판지 상자와 야간 창밖은 적합하며 읽을 수 있는 글자는 없다. 지붕 바깥 십자가와 신발 속 카드는 이 구도에서 확인할 수 없다.",
        "hard_violations": [],
        "physics": "현우는 다리를 접고 낮게 웅크린 채 상체를 앞으로 기울이고, 양 손바닥으로 찰리의 가슴을 지지점 삼아 체중을 전달한다. 하체의 바닥 접점 일부는 가려져 있지만 공중에 떠 있는 모습은 아니다. 찰리는 등을 바닥 쪽에 둔 채 누워 있고 몸과 팔이 바닥 높이에 놓여 있어 지지가 자연스럽다. 상자는 바닥이나 아래 상자 위에 놓여 있다. 양손의 접촉과 힘을 주는 얼굴은 거칠게 누르는 행동으로 읽힌다."
       },
       {
        "label": "A",
        "direction": "현우는 찰리의 등·어깨 쪽 접촉 지점을 내려다보며 양손으로 상체를 아래로 누른다. 찰리는 엎드린 방향으로 낮아져 얼굴과 시선이 전경의 노출된 바닥 쪽을 향한다. 현우의 압박 방향이 상자 사이 바닥이라는 지정된 목표와 명확히 이어진다.",
        "built_space": "밴 후방에서 전방을 바라보는 실내로, 중앙 천장등 한 개, 좌우 창과 안전벨트, 뒤편 좌석 등받이가 참조와 유사하게 배치된다. 왼쪽에는 전경의 큰 상자와 뒤쪽의 세운 상자가 있고, 오른쪽에는 앞뒤로 떨어진 상자들이 중앙 통로를 둘러싼다. 현우는 통로 왼쪽에서 웅크리고 찰리는 그 앞의 틈에 낮게 엎드려 있다. 상자 가장자리와 안쪽 면이 몸을 둘러싸되 누르는 손은 가리지 않는다. 찰리의 얼굴과 손 앞에 바닥이 노출되어 압박의 목적지가 읽힌다. 불가능한 반사나 설비 중복은 보이지 않는다.",
        "entities": "현우와 찰리만 등장한다. 현우는 젊은 한국계 남성으로 보이는 얼굴, 검은 헝클어진 머리와 얼굴 멍을 갖췄다. 다만 참조의 어두운 후드티를 그대로 입어 겉옷을 벗은 상태라는 지시와 어긋난다. 찰리는 각진 기계 머리, 흰 얼굴판, 작은 점 형태 눈과 선 형태 입, 거대한 금속 상체를 갖췄으며 낡은 코트와 모자도 착용했다. 상자·금속·천의 질감과 야간 창밖이 구체적으로 보이고 읽을 수 있는 글자는 없다. 다리 부상은 보이는 부분만으로 뚜렷하게 확인되지 않으며, 신발 속 카드와 외부 지붕 십자가는 확인할 수 없다.",
        "hard_violations": [],
        "physics": "현우의 화면 왼쪽 신발은 바닥에 닿고 무릎은 굽혀져 있으며, 앞으로 기울인 몸의 힘이 찰리의 금속 상체에 닿은 양손으로 전달된다. 찰리의 두 기계 손과 전완은 바닥에 놓여 낮은 몸을 받친다. 아직 팔로 버티는 상체를 현우가 더 낮게 누르는 순간으로 물리적으로 성립한다. 코트는 몸 위에 처지고 모자는 머리에 얹혀 있으며, 상자는 바닥 또는 다른 상자에 지지된다. 지지 없이 떠 있는 몸이나 물건은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.179
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.179
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "이현우가 겉옷(후드티)을 벗지 않은 점과 천장의 할로겐 십자가가 누락된 점은 아쉽지만, 찰리의 구체적인 기계 얼굴 형태(흰 얼굴판, 점 눈)와 의상(코트, 모자)을 정확히 구현했으며 거칠게 짓누르는 액션의 무게감이 훌륭합니다."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "이현우의 겉옷 탈의 조건은 충족했으나, 찰리의 핵심적인 흰 얼굴판 묘사가 전혀 없고 코트와 모자마저 완전히 누락되어 프롬프트의 캐릭터 지시를 심각하게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_supply_van_interior_deaa25.png",
    "asset_id": "bea47285-24c3-4f7a-a723-52f9d4b2e93f",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1597480>",
    "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda9-3641-7797-a999-0eaff8b04818",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S28sh6__bgfirst_bg.png",
   "bg_asset_id": "e8037db3-b1e3-46d6-a566-dfa1c6b985d6",
   "bg_record_key": "S28sh6::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "supply_van_interior",
   "groupbg_asset_id": "bea47285-24c3-4f7a-a723-52f9d4b2e93f"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S28sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:30:43.554401+00:00",
  "fingerprint": "0eaa920e90ff75b2c6335032ada46186126b22eefa9976521e2dd8283a9c82ac",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S28sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S28sh6_sel.png",
  "source_sha256": "ff9469e615010c526063feade75bf1dd8d74e55ea4e779954700fd236e6b8618",
  "file": "S28sh6_cine.png",
  "staged_sha256": "86841b730931d4d0c8f726d217953ed291913f27ba243934ce51961605f43477",
  "latency_ms": 10570
 },
 "S28sh7::signage": {
  "fp": "02fbe31ec9a5e000",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S28sh7::bgfirst_bg": {
  "input_fingerprint": "43b2c57fe98ab5b7",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 운전석 옆 창문 밖에서 소총을 멘 아시아계 민병대원이 굳은 표정으로 서 있는 구도.\n\nLOCATION (lock): Outside the driver's side window of a van stopped at a militia checkpoint at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Driver's window surround (The window is the checkpoint's point of contact; its opening state is not emphasized) — Seen obliquely from inside the cabin, bordering the militia member outside; used as A practical frame separating the protected interior from the inspection; Slung rifle (Carried by the militia member) — Seen obliquely alongside his torso rather than aimed toward the lens; used as Secondary evidence of the checkpoint's threat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient exposure keeps the militia member's stern expression readable without anticipating the cross's illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 운전석 옆 창문 밖에서 소총을 멘 아시아계 민병대원이 굳은 표정으로 서 있는 구도.\n\nLOCATION (lock): Outside the driver's side window of a van stopped at a militia checkpoint at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Driver's window surround (The window is the checkpoint's point of contact; its opening state is not emphasized) — Seen obliquely from inside the cabin, bordering the militia member outside; used as A practical frame separating the protected interior from the inspection; Slung rifle (Carried by the militia member) — Seen obliquely alongside his torso rather than aimed toward the lens; used as Secondary evidence of the checkpoint's threat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient exposure keeps the militia member's stern expression readable without anticipating the cross's illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S28sh7__bgfirst_bg.png",
  "asset_id": "8c5bfd54-d57c-46a8-9763-a40b90a81062",
  "input_asset_ids": [
   "cf805284-ac6f-42f5-aa8b-e44c8f58c686",
   "b4c18cf3-5ae5-4429-8e1c-57108458b82e"
  ]
 },
 "S28sh7": {
  "input_fingerprint": "1e90a44dc1042b4b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 운전석 옆 창문 밖에서 소총을 멘 아시아계 민병대원이 굳은 표정으로 서 있는 구도.\n\nLOCATION (lock): Outside the driver's side window of a van stopped at a militia checkpoint at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Driver's window surround (The window is the checkpoint's point of contact; its opening state is not emphasized) — Seen obliquely from inside the cabin, bordering the militia member outside; used as A practical frame separating the protected interior from the inspection; Slung rifle (Carried by the militia member) — Seen obliquely alongside his torso rather than aimed toward the lens; used as Secondary evidence of the checkpoint's threat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient exposure keeps the militia member's stern expression readable without anticipating the cross's illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rear boxes remain arranged as cover, with Charlie concealed among them in his old coat and hat. The roof-mounted halogen cross is still present; the driver's window is being requested open, with no completed lowering yet specified.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 아시아계 민병대원 right now, so 아시아계 민병대원's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 아시아계 민병대원: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 운전석 옆 창문 밖에서 소총을 멘 아시아계 민병대원이 굳은 표정으로 서 있는 구도.\n\nLOCATION (lock): Outside the driver's side window of a van stopped at a militia checkpoint at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Driver's window surround (The window is the checkpoint's point of contact; its opening state is not emphasized) — Seen obliquely from inside the cabin, bordering the militia member outside; used as A practical frame separating the protected interior from the inspection; Slung rifle (Carried by the militia member) — Seen obliquely alongside his torso rather than aimed toward the lens; used as Secondary evidence of the checkpoint's threat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient exposure keeps the militia member's stern expression readable without anticipating the cross's illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rear boxes remain arranged as cover, with Charlie concealed among them in his old coat and hat. The roof-mounted halogen cross is still present; the driver's window is being requested open, with no completed lowering yet specified.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 아시아계 민병대원 right now, so 아시아계 민병대원's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 아시아계 민병대원: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 운전석 옆 창문 밖에서 소총을 멘 아시아계 민병대원이 굳은 표정으로 서 있는 구도.\n\nLOCATION (lock): Outside the driver's side window of a van stopped at a militia checkpoint at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Driver's window surround (The window is the checkpoint's point of contact; its opening state is not emphasized) — Seen obliquely from inside the cabin, bordering the militia member outside; used as A practical frame separating the protected interior from the inspection; Slung rifle (Carried by the militia member) — Seen obliquely alongside his torso rather than aimed toward the lens; used as Secondary evidence of the checkpoint's threat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient exposure keeps the militia member's stern expression readable without anticipating the cross's illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rear boxes remain arranged as cover, with Charlie concealed among them in his old coat and hat. The roof-mounted halogen cross is still present; the driver's window is being requested open, with no completed lowering yet specified.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 아시아계 민병대원 right now, so 아시아계 민병대원's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 아시아계 민병대원: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S28sh7__bgfirst_bg.png",
     "asset_id": "8c5bfd54-d57c-46a8-9763-a40b90a81062",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S28sh7.png",
     "asset_id": "cf805284-ac6f-42f5-aa8b-e44c8f58c686",
     "role": "conti_light"
    },
    {
     "label": "PROP REFERENCE — 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:667686>",
     "asset_id": "fcb46f42-5d47-4486-85bb-3542f9f839f4",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_militia_road_checkpoint_sel.png",
     "asset_id": "b4c18cf3-5ae5-4429-8e1c-57108458b82e",
     "role": "location_seed_bg"
    },
    {
     "label": "PROP REFERENCE — 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:667686>",
     "asset_id": "fcb46f42-5d47-4486-85bb-3542f9f839f4",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "민병대원이 프레임 왼쪽 전경에 있는 운전자를 바라보고 있으며, 소총은 사격 자세가 아닌 몸 쪽에 비스듬히 메어짐.",
    "built_space": "차량 내부에서 창밖을 바라보는 시점으로, 배경에 레퍼런스의 도로, 가로등, 펜스가 배치되어 있음.",
    "entities": "아시아계 남성 민병대원이 굳은 표정을 짓고 있으나, 지시문에 없는 인물이 전경에 추가됨. 소총 디자인이 레퍼런스와 다름.",
    "hard_violations": [
     "[gemini-pro] 지시문(Shot Text)에 언급되지 않은 인물(전경의 운전자)이 임의로 추가됨",
     "[gpt-high] 숏 텍스트에 없는 운전자를 왼쪽 전경에 추가했습니다.",
     "[gpt-high] 숏 텍스트에 없는 별도의 경계병을 배경에 추가했습니다."
    ],
    "physics": "민병대원의 손이 소총을 자연스럽게 받치고 있음."
   },
   {
    "label": "B",
    "direction": "민병대원이 차량 내부(카메라 방향)를 응시하고 있으며, 소총은 몸 앞에 멘 상태임.",
    "built_space": "차량 내부에서 운전석 창문틀을 시각적 프레임으로 활용해 밖을 바라보는 구도. 프레임 좌측에 엄폐용 박스들이 배치되어 있고, 배경에 레퍼런스의 횡단보도 선이 보임.",
    "entities": "아시아계 남성 민병대원, 명시된 표정, 레퍼런스와 일치하는 형태의 소총 개머리판, 그리고 뒷좌석의 박스들이 정확히 묘사됨.",
    "hard_violations": [],
    "physics": "민병대원의 오른손이 무기의 무게를 지탱하며 자연스럽게 쥐고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지시된 프레이밍과 위치를 정확히 구현했으며, 창문틀을 활용한 구도 및 소총의 레퍼런스 일치도가 높습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지시문에 없는 인물이 전경에 크게 등장하여 치명적인 오류를 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "민병대원이 프레임 왼쪽 전경에 있는 운전자를 바라보고 있으며, 소총은 사격 자세가 아닌 몸 쪽에 비스듬히 메어짐.",
        "built_space": "차량 내부에서 창밖을 바라보는 시점으로, 배경에 레퍼런스의 도로, 가로등, 펜스가 배치되어 있음.",
        "entities": "아시아계 남성 민병대원이 굳은 표정을 짓고 있으나, 지시문에 없는 인물이 전경에 추가됨. 소총 디자인이 레퍼런스와 다름.",
        "hard_violations": [
         "지시문(Shot Text)에 언급되지 않은 인물(전경의 운전자)이 임의로 추가됨"
        ],
        "physics": "민병대원의 손이 소총을 자연스럽게 받치고 있음."
       },
       {
        "label": "B",
        "direction": "민병대원이 차량 내부(카메라 방향)를 응시하고 있으며, 소총은 몸 앞에 멘 상태임.",
        "built_space": "차량 내부에서 운전석 창문틀을 시각적 프레임으로 활용해 밖을 바라보는 구도. 프레임 좌측에 엄폐용 박스들이 배치되어 있고, 배경에 레퍼런스의 횡단보도 선이 보임.",
        "entities": "아시아계 남성 민병대원, 명시된 표정, 레퍼런스와 일치하는 형태의 소총 개머리판, 그리고 뒷좌석의 박스들이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "민병대원의 오른손이 무기의 무게를 지탱하며 자연스럽게 쥐고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지시된 프레이밍과 위치를 정확히 구현했으며, 창문틀을 활용한 구도 및 소총의 레퍼런스 일치도가 높습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지시문에 없는 인물이 전경에 크게 등장하여 치명적인 오류를 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "민병대원이 프레임 왼쪽 전경에 있는 운전자를 바라보고 있으며, 소총은 사격 자세가 아닌 몸 쪽에 비스듬히 메어짐.",
        "built_space": "차량 내부에서 창밖을 바라보는 시점으로, 배경에 레퍼런스의 도로, 가로등, 펜스가 배치되어 있음.",
        "entities": "아시아계 남성 민병대원이 굳은 표정을 짓고 있으나, 지시문에 없는 인물이 전경에 추가됨. 소총 디자인이 레퍼런스와 다름.",
        "hard_violations": [
         "지시문(Shot Text)에 언급되지 않은 인물(전경의 운전자)이 임의로 추가됨"
        ],
        "physics": "민병대원의 손이 소총을 자연스럽게 받치고 있음."
       },
       {
        "label": "B",
        "direction": "민병대원이 차량 내부(카메라 방향)를 응시하고 있으며, 소총은 몸 앞에 멘 상태임.",
        "built_space": "차량 내부에서 운전석 창문틀을 시각적 프레임으로 활용해 밖을 바라보는 구도. 프레임 좌측에 엄폐용 박스들이 배치되어 있고, 배경에 레퍼런스의 횡단보도 선이 보임.",
        "entities": "아시아계 남성 민병대원, 명시된 표정, 레퍼런스와 일치하는 형태의 소총 개머리판, 그리고 뒷좌석의 박스들이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "민병대원의 오른손이 무기의 무게를 지탱하며 자연스럽게 쥐고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "불필요한 인물 없이 창밖 민병대원과 아래로 향한 소총을 구현했지만, 실내가 차지하는 비중이 크고 창문이 이미 일부 내려간 모습은 지시와 어긋납니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "민병대원의 크기와 굳은 표정은 적절하지만, 금지된 운전자와 배경 경계병을 추가했으며 소총도 참조와 다른 돌격소총 형태입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "민병대원은 운전석 창을 통해 실내 카메라 쪽을 바라봅니다. 소총은 개머리판이 어깨 가까이 있고 총열이 몸통을 따라 화면 오른쪽 아래로 향하며, 렌즈나 사람을 겨누지 않습니다.",
        "built_space": "실내에서 운전석 옆 창문 하나를 비스듬히 바라봅니다. 운전석 등받이 하나, 운전대 하나, 외부 사이드미러 하나와 앞유리 일부가 보이며, 민병대원은 문 바깥에 서 있습니다. 실내 후방에는 상자들이 쌓여 있습니다. 창에는 중간 높이의 유리 상단선이 있어 이미 일부 내려간 것으로 읽힙니다. 바깥의 차선·횡단보도·철조망 울타리는 참조 장소의 요소와 통하지만, 특징적인 가로등과 산업시설의 일치 여부는 어두운 배경에서 확인하기 어렵습니다.",
        "entities": "보이는 사람은 중년으로 보이는 아시아계 남성 민병대원 한 명입니다. 모자와 어두운 야전 재킷을 착용했고 표정은 굳어 있습니다. 소총에는 조준경과 구멍이 난 개머리판이 있어 참조의 검은 정밀소총에 비교적 가깝지만, 총열과 세부 구조가 가려져 정확한 일치는 확인할 수 없습니다. 그의 맨손과 소매가 소총에 닿아 있습니다. 후방 상자는 보이고 찰리는 드러나지 않습니다. 지붕 십자가는 화면 밖이므로 판단 대상이 아닙니다. 읽을 수 있는 문구는 없습니다.",
        "hard_violations": [],
        "physics": "하체는 문 아래로 가려지지만 상체는 창밖에 정상적으로 서 있는 자세이며 공중에 뜬 징후가 없습니다. 소총은 어깨 쪽 슬링과 총몸을 잡은 손으로 지지됩니다. 상자는 후방 적재 공간에 쌓여 있고, 보이는 좌석과 차량 부품도 정상적으로 고정되어 있습니다."
       },
       {
        "label": "B",
        "direction": "민병대원은 화면 왼쪽 전경의 운전자를 향해 시선을 둡니다. 운전자도 민병대원 쪽으로 고개를 돌리고 있습니다. 소총은 가슴 앞에서 아래쪽으로 향해 렌즈를 겨누지 않습니다. 먼 경계병의 정확한 시선은 판별하기 어렵습니다.",
        "built_space": "운전석 창문 하나, 운전대 하나, 사이드미러 하나가 보이며 창밖 민병대원은 허리 위 크기로 배치됩니다. 그러나 운전자 뒤에서 보는 구도로 바뀌면서 전경 인물이 화면 왼쪽을 크게 차지합니다. 바깥에는 참조와 유사한 철조망 울타리, 가로등 열, 붉은 보도와 산업시설이 보입니다. 참조에서 확인되지 않는 경비 부스 하나와 콘크리트 차단물도 추가되었습니다. 사이드미러의 도로와 가로등 반사는 이 시점에서 불가능하다고 단정할 근거가 없습니다.",
        "entities": "중앙에는 젊은 성인으로 보이는 아시아계 남성 민병대원이 굳은 표정으로 서 있고, 녹색 군복과 전술 장구를 착용했습니다. 이외에 성인 남성 운전자 한 명과 배경 경계병 한 명이 추가되어 총 세 사람이 보입니다. 소총은 각진 개머리판과 굽은 탄창을 지닌 돌격소총 형태로, 참조의 긴 총열과 대형 조준경을 가진 볼트액션 정밀소총과 다릅니다. 민병대원의 맨손과 손목은 소총 손잡이 부근에 놓여 있습니다. 찰리·후방 상자·지붕 십자가는 이 구도에서 보이지 않으며, 읽을 수 있는 문구는 없습니다.",
        "hard_violations": [
         "숏 텍스트에 없는 운전자를 왼쪽 전경에 추가했습니다.",
         "숏 텍스트에 없는 별도의 경계병을 배경에 추가했습니다."
        ],
        "physics": "민병대원의 소총은 어깨를 지나는 슬링과 손으로 지지되고, 검지는 총몸 옆을 따라 놓여 있습니다. 민병대원의 하체는 가려져 있으나 서 있는 상체 자세는 자연스럽습니다. 전경 운전자는 운전석에 앉은 배치이며, 배경 경계병은 발을 도로에 딛고 있습니다. 지지 없이 떠 있는 신체나 물체는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "불필요한 인물 없이 창밖 민병대원과 아래로 향한 소총을 구현했지만, 실내가 차지하는 비중이 크고 창문이 이미 일부 내려간 모습은 지시와 어긋납니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "민병대원의 크기와 굳은 표정은 적절하지만, 금지된 운전자와 배경 경계병을 추가했으며 소총도 참조와 다른 돌격소총 형태입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "민병대원은 운전석 창을 통해 실내 카메라 쪽을 바라봅니다. 소총은 개머리판이 어깨 가까이 있고 총열이 몸통을 따라 화면 오른쪽 아래로 향하며, 렌즈나 사람을 겨누지 않습니다.",
        "built_space": "실내에서 운전석 옆 창문 하나를 비스듬히 바라봅니다. 운전석 등받이 하나, 운전대 하나, 외부 사이드미러 하나와 앞유리 일부가 보이며, 민병대원은 문 바깥에 서 있습니다. 실내 후방에는 상자들이 쌓여 있습니다. 창에는 중간 높이의 유리 상단선이 있어 이미 일부 내려간 것으로 읽힙니다. 바깥의 차선·횡단보도·철조망 울타리는 참조 장소의 요소와 통하지만, 특징적인 가로등과 산업시설의 일치 여부는 어두운 배경에서 확인하기 어렵습니다.",
        "entities": "보이는 사람은 중년으로 보이는 아시아계 남성 민병대원 한 명입니다. 모자와 어두운 야전 재킷을 착용했고 표정은 굳어 있습니다. 소총에는 조준경과 구멍이 난 개머리판이 있어 참조의 검은 정밀소총에 비교적 가깝지만, 총열과 세부 구조가 가려져 정확한 일치는 확인할 수 없습니다. 그의 맨손과 소매가 소총에 닿아 있습니다. 후방 상자는 보이고 찰리는 드러나지 않습니다. 지붕 십자가는 화면 밖이므로 판단 대상이 아닙니다. 읽을 수 있는 문구는 없습니다.",
        "hard_violations": [],
        "physics": "하체는 문 아래로 가려지지만 상체는 창밖에 정상적으로 서 있는 자세이며 공중에 뜬 징후가 없습니다. 소총은 어깨 쪽 슬링과 총몸을 잡은 손으로 지지됩니다. 상자는 후방 적재 공간에 쌓여 있고, 보이는 좌석과 차량 부품도 정상적으로 고정되어 있습니다."
       },
       {
        "label": "A",
        "direction": "민병대원은 화면 왼쪽 전경의 운전자를 향해 시선을 둡니다. 운전자도 민병대원 쪽으로 고개를 돌리고 있습니다. 소총은 가슴 앞에서 아래쪽으로 향해 렌즈를 겨누지 않습니다. 먼 경계병의 정확한 시선은 판별하기 어렵습니다.",
        "built_space": "운전석 창문 하나, 운전대 하나, 사이드미러 하나가 보이며 창밖 민병대원은 허리 위 크기로 배치됩니다. 그러나 운전자 뒤에서 보는 구도로 바뀌면서 전경 인물이 화면 왼쪽을 크게 차지합니다. 바깥에는 참조와 유사한 철조망 울타리, 가로등 열, 붉은 보도와 산업시설이 보입니다. 참조에서 확인되지 않는 경비 부스 하나와 콘크리트 차단물도 추가되었습니다. 사이드미러의 도로와 가로등 반사는 이 시점에서 불가능하다고 단정할 근거가 없습니다.",
        "entities": "중앙에는 젊은 성인으로 보이는 아시아계 남성 민병대원이 굳은 표정으로 서 있고, 녹색 군복과 전술 장구를 착용했습니다. 이외에 성인 남성 운전자 한 명과 배경 경계병 한 명이 추가되어 총 세 사람이 보입니다. 소총은 각진 개머리판과 굽은 탄창을 지닌 돌격소총 형태로, 참조의 긴 총열과 대형 조준경을 가진 볼트액션 정밀소총과 다릅니다. 민병대원의 맨손과 손목은 소총 손잡이 부근에 놓여 있습니다. 찰리·후방 상자·지붕 십자가는 이 구도에서 보이지 않으며, 읽을 수 있는 문구는 없습니다.",
        "hard_violations": [
         "숏 텍스트에 없는 운전자를 왼쪽 전경에 추가했습니다.",
         "숏 텍스트에 없는 별도의 경계병을 배경에 추가했습니다."
        ],
        "physics": "민병대원의 소총은 어깨를 지나는 슬링과 손으로 지지되고, 검지는 총몸 옆을 따라 놓여 있습니다. 민병대원의 하체는 가려져 있으나 서 있는 상체 자세는 자연스럽습니다. 전경 운전자는 운전석에 앉은 배치이며, 배경 경계병은 발을 도로에 딛고 있습니다. 지지 없이 떠 있는 신체나 물체는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.661,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.411,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지시문(Shot Text)에 언급되지 않은 인물(전경의 운전자)이 임의로 추가됨",
     "[gpt-high] 숏 텍스트에 없는 운전자를 왼쪽 전경에 추가했습니다.",
     "[gpt-high] 숏 텍스트에 없는 별도의 경계병을 배경에 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 411
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 프레이밍과 위치를 정확히 구현했으며, 창문틀을 활용한 구도 및 소총의 레퍼런스 일치도가 높습니다."
   },
   {
    "label": "A",
    "score": 411,
    "verdict_ko": "지시문에 없는 인물이 전경에 크게 등장하여 치명적인 오류를 범했습니다.  ★위반: [gemini-pro] 지시문(Shot Text)에 언급되지 않은 인물(전경의 운전자)이 임의로 추가됨 / [gpt-high] 숏 텍스트에 없는 운전자를 왼쪽 전경에 추가했습니다. / [gpt-high] 숏 텍스트에 없는 별도의 경계병을 배경에 추가했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_militia_road_checkpoint_sel.png",
    "asset_id": "b4c18cf3-5ae5-4429-8e1c-57108458b82e",
    "role": "location_seed_bg"
   },
   {
    "label": "PROP REFERENCE — 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:667686>",
    "asset_id": "fcb46f42-5d47-4486-85bb-3542f9f839f4",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacda9-e769-7a3d-982a-29321d85a696",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S28sh7__bgfirst_bg.png",
   "bg_asset_id": "8c5bfd54-d57c-46a8-9763-a40b90a81062",
   "bg_record_key": "S28sh7::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S28sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:30:54.556621+00:00",
  "fingerprint": "98aa8d9cd4654ae05f3f1926f7707c841ec835b4fdf040679d4ec8a49802b4e8",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S28sh7_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S28sh7_sel.png",
  "source_sha256": "1b68939c4ec0a086722438f273eaf6d443290445167d5dd343de876b7f39aa3c",
  "file": "S28sh7_cine.png",
  "staged_sha256": "0d61f0127495bff59a8a06d32842bb201948cbeacd0f35c46275916bc42dc418",
  "latency_ms": 11158
 },
 "S28sh9::signage": {
  "fp": "2879ce894f56222c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S28sh9": {
  "input_fingerprint": "327c9af9b6fefbc6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 승합차 지붕 위의 낡은 십자가 장식에서 눈부신 불빛이 막 뿜어져 나오는 순간.\n\nLOCATION (lock): On the exterior roof of the checkpoint-stopped van, at its crudely made illuminated cross ornament. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Roof-mounted cross ornament in the upper-center of the frame, midground; Narrow van roof section beneath the cross in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Roof-mounted cross ornament (Worn, crudely fashioned, and fixed to the van roof) — Its cross-shaped face is seen obliquely from below; used as Upper-frame focal object and the visual explanation for safe passage; Van roof (Supporting the cross ornament) — A narrow oblique section remains beneath the cross; used as Spatial reference connecting the reveal to the vehicle.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The halogen cross switches on with dazzling brightness against the night, with controlled highlight bloom preserving its recognizable outline.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crude halogen cross on the van roof is now lit. The rear boxes remain in place, with Charlie hidden among them in his old coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 승합차 지붕 위의 낡은 십자가 장식에서 눈부신 불빛이 막 뿜어져 나오는 순간.\n\nLOCATION (lock): On the exterior roof of the checkpoint-stopped van, at its crudely made illuminated cross ornament. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Roof-mounted cross ornament in the upper-center of the frame, midground; Narrow van roof section beneath the cross in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Roof-mounted cross ornament (Worn, crudely fashioned, and fixed to the van roof) — Its cross-shaped face is seen obliquely from below; used as Upper-frame focal object and the visual explanation for safe passage; Van roof (Supporting the cross ornament) — A narrow oblique section remains beneath the cross; used as Spatial reference connecting the reveal to the vehicle.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The halogen cross switches on with dazzling brightness against the night, with controlled highlight bloom preserving its recognizable outline.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crude halogen cross on the van roof is now lit. The rear boxes remain in place, with Charlie hidden among them in his old coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 승합차 지붕 위의 낡은 십자가 장식에서 눈부신 불빛이 막 뿜어져 나오는 순간.\n\nLOCATION (lock): On the exterior roof of the checkpoint-stopped van, at its crudely made illuminated cross ornament. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Roof-mounted cross ornament in the upper-center of the frame, midground; Narrow van roof section beneath the cross in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Roof-mounted cross ornament (Worn, crudely fashioned, and fixed to the van roof) — Its cross-shaped face is seen obliquely from below; used as Upper-frame focal object and the visual explanation for safe passage; Van roof (Supporting the cross ornament) — A narrow oblique section remains beneath the cross; used as Spatial reference connecting the reveal to the vehicle.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The halogen cross switches on with dazzling brightness against the night, with controlled highlight bloom preserving its recognizable outline.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crude halogen cross on the van roof is now lit. The rear boxes remain in place, with Charlie hidden among them in his old coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 도로 정면을 향해 있으며, 십자가 장식에서 빛이 사방으로 뿜어져 나옵니다.",
    "built_space": "야간 도로. 레퍼런스 이미지의 왼쪽 공장 건물, 오른쪽 철조망, 가로등 배치가 정확히 재현되었습니다. 카메라는 차량 지붕 위에 위치합니다.",
    "entities": "빛을 발하는 투박한 형태의 십자가 장식이 승합차 지붕에 있습니다. 인물이나 지붕 위 상자는 보이지 않습니다.",
    "hard_violations": [
     "[gpt-high] 십자가가 달린 승합차와 별개인 차량 지붕 형태의 대형 구조물을 전경에 추가했습니다."
    ],
    "physics": "십자가는 승합차 지붕에 고정되어 무게를 지탱받고 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 승합차 지붕을 따라 울타리 쪽을 바라보며, 십자가에서 강한 빛이 발산됩니다.",
    "built_space": "야간. 레퍼런스와 달리 철조망이 왼쪽에 위치하여 장소의 구조적 배치가 어긋나 있습니다.",
    "entities": "빛나는 십자가와 그 뒤로 상자들이 지붕에 놓여 있습니다. 샷 텍스트에 없는 수많은 사람들이 차량 주변에 서 있습니다.",
    "hard_violations": [
     "[gemini-pro] 발명된 사람 (샷 텍스트에 명시되지 않은 다수의 군중이 프레임에 임의로 추가됨)",
     "[gpt-high] 인물이 등장하지 않는 십자가 공개 장면에 차량 양쪽의 군중을 추가했습니다.",
     "[gpt-high] 참조 장소에서 개방된 부지여야 하는 도로 왼쪽에 연속 철책을 추가하여 고정 공간 구조를 바꾸었습니다."
    ],
    "physics": "십자가와 상자들은 차량 지붕에 놓여 있고, 사람들은 바닥에 서서 지탱받고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 장소와 샷 텍스트의 조명 및 구도를 전반적으로 잘 구현하였으나, 짐칸의 상자가 보이지 않는 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "샷 텍스트에 전혀 언급되지 않은 다수의 군중을 임의로 화면에 추가하여 인물 지침을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 도로 정면을 향해 있으며, 십자가 장식에서 빛이 사방으로 뿜어져 나옵니다.",
        "built_space": "야간 도로. 레퍼런스 이미지의 왼쪽 공장 건물, 오른쪽 철조망, 가로등 배치가 정확히 재현되었습니다. 카메라는 차량 지붕 위에 위치합니다.",
        "entities": "빛을 발하는 투박한 형태의 십자가 장식이 승합차 지붕에 있습니다. 인물이나 지붕 위 상자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "십자가는 승합차 지붕에 고정되어 무게를 지탱받고 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 승합차 지붕을 따라 울타리 쪽을 바라보며, 십자가에서 강한 빛이 발산됩니다.",
        "built_space": "야간. 레퍼런스와 달리 철조망이 왼쪽에 위치하여 장소의 구조적 배치가 어긋나 있습니다.",
        "entities": "빛나는 십자가와 그 뒤로 상자들이 지붕에 놓여 있습니다. 샷 텍스트에 없는 수많은 사람들이 차량 주변에 서 있습니다.",
        "hard_violations": [
         "발명된 사람 (샷 텍스트에 명시되지 않은 다수의 군중이 프레임에 임의로 추가됨)"
        ],
        "physics": "십자가와 상자들은 차량 지붕에 놓여 있고, 사람들은 바닥에 서서 지탱받고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 장소와 샷 텍스트의 조명 및 구도를 전반적으로 잘 구현하였으나, 짐칸의 상자가 보이지 않는 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "샷 텍스트에 전혀 언급되지 않은 다수의 군중을 임의로 화면에 추가하여 인물 지침을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 도로 정면을 향해 있으며, 십자가 장식에서 빛이 사방으로 뿜어져 나옵니다.",
        "built_space": "야간 도로. 레퍼런스 이미지의 왼쪽 공장 건물, 오른쪽 철조망, 가로등 배치가 정확히 재현되었습니다. 카메라는 차량 지붕 위에 위치합니다.",
        "entities": "빛을 발하는 투박한 형태의 십자가 장식이 승합차 지붕에 있습니다. 인물이나 지붕 위 상자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "십자가는 승합차 지붕에 고정되어 무게를 지탱받고 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 승합차 지붕을 따라 울타리 쪽을 바라보며, 십자가에서 강한 빛이 발산됩니다.",
        "built_space": "야간. 레퍼런스와 달리 철조망이 왼쪽에 위치하여 장소의 구조적 배치가 어긋나 있습니다.",
        "entities": "빛나는 십자가와 그 뒤로 상자들이 지붕에 놓여 있습니다. 샷 텍스트에 없는 수많은 사람들이 차량 주변에 서 있습니다.",
        "hard_violations": [
         "발명된 사람 (샷 텍스트에 명시되지 않은 다수의 군중이 프레임에 임의로 추가됨)"
        ],
        "physics": "십자가와 상자들은 차량 지붕에 놓여 있고, 사람들은 바닥에 서서 지탱받고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "십자가의 점등과 고정 장치는 구현했지만, 지시되지 않은 군중과 왼쪽 철책을 추가했고 지붕을 넓게 내려다보아 지정 구도를 어겼습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "참조 장소와 낡은 십자가의 눈부신 점등은 더 충실하지만, 전경에 별도 차량 지붕 같은 구조물을 추가했고 좁은 지붕만 보여야 할 구도를 벗어났습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "십자가의 발광면은 거의 정면으로 카메라를 향하며, 아래에서 비스듬히 올려다본 면으로 읽히지 않습니다. 왼쪽 앞의 검은 머리 청년은 십자가 쪽을 올려다보지만, 주변 사람들의 시선은 차량과 전방 등으로 나뉩니다. 조준하는 무기나 이동 중인 물체는 보이지 않습니다.",
        "built_space": "십자가 한 개가 받침판과 긴 보강재를 통해 승합차 지붕 한 면에 고정되어 있습니다. 지붕은 하단 중앙의 좁은 조각이 아니라 화면 하단 상당 부분을 차지하고, 카메라는 그 윗면을 내려다봅니다. 뒤쪽에는 여러 상자, 오른쪽에는 열린 차량 문이 보입니다. 도로 양쪽에 철책이 있어, 왼쪽은 개방된 부지이고 오른쪽에 철책이 있는 참조 장소와 다릅니다.",
        "entities": "거칠고 낡은 금속 십자가, 내부의 밝은 전구, 마모된 금속 지붕과 뒤쪽 상자들은 보입니다. 그러나 차량 양쪽에 남녀 성인으로 보이는 인물이 열 명 이상 추가되어 있습니다. 앞쪽 청년은 검은 머리의 젊은 동아시아계 남성처럼 보이지만, 이 장면에는 그를 포함한 사람의 노출 자체가 지시되지 않았습니다. 찰리의 기계 머리는 보이지 않으며, 숨겨진 상태이므로 그 점은 결함이 아닙니다. 판독 가능한 문자는 확인되지 않습니다.",
        "hard_violations": [
         "인물이 등장하지 않는 십자가 공개 장면에 차량 양쪽의 군중을 추가했습니다.",
         "참조 장소에서 개방된 부지여야 하는 도로 왼쪽에 연속 철책을 추가하여 고정 공간 구조를 바꾸었습니다."
        ],
        "physics": "십자가 하단의 볼트와 받침판, 지붕 위 보강재가 하중을 지지합니다. 상자들은 차량 뒤쪽 적재 영역에 쌓여 있으며 공중에 떠 있지 않습니다. 사람들은 차 옆에 서 있는 자세이고 발은 대체로 프레임 밖이거나 가려져 있어, 무지지 부유로 판단할 근거는 없습니다. 전구의 빛이 금속 지붕에 반사되는 관계도 가능합니다."
       },
       {
        "label": "B",
        "direction": "십자가는 약간 비스듬하게 카메라를 향하고 빛은 중심에서 퍼집니다. 다만 지붕 윗면과 받침을 내려다보는 시점이어서, 십자가의 면을 아래에서 올려다보라는 요구에는 맞지 않습니다. 사람의 시선이나 무기 조준, 이동 방향은 없습니다.",
        "built_space": "십자가 한 개와 받침 한 개가 승합차 지붕에 놓여 있습니다. 왼쪽의 공장·빈 부지·가로수와 오른쪽의 철책, 양쪽 가로등은 참조 장소의 배치를 대체로 유지합니다. 그러나 승합차의 앞 유리와 측면까지 노출되며, 화면 맨 아래에는 그 승합차 지붕과 분리된 또 하나의 넓은 골진 차량 지붕 같은 면이 겹쳐 있습니다. 따라서 지정된 단일 승합차의 좁은 지붕 구성이 아닙니다.",
        "entities": "낡은 십자가형 부재, 강하게 켜진 전구, 노출 전선과 결속재, 마모된 승합차가 보이며 사람은 없습니다. 십자가 외곽은 남아 있지만 중심부는 강한 번짐으로 상당히 날아갑니다. 뒤쪽 상자와 숨은 찰리는 이 구도에서 확인되지 않으므로 누락으로 판단하지 않습니다. 차량의 종이와 표식은 흐려서 판독 가능한 글자는 확인되지 않습니다.",
        "hard_violations": [
         "십자가가 달린 승합차와 별개인 차량 지붕 형태의 대형 구조물을 전경에 추가했습니다."
        ],
        "physics": "십자가는 하단 받침과 결속재로 지지되고, 전선은 십자가에서 지붕 쪽으로 이어집니다. 승합차는 하단이 가려져 있지만 정상적으로 서 있는 차량 자세입니다. 추가 전경 구조물도 화면 아래로 이어져 있어 공중에 떠 있다고 판단할 근거는 없습니다. 조명이 지붕 금속에 만드는 반사와 그림자는 물리적으로 가능합니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "십자가의 점등과 고정 장치는 구현했지만, 지시되지 않은 군중과 왼쪽 철책을 추가했고 지붕을 넓게 내려다보아 지정 구도를 어겼습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "참조 장소와 낡은 십자가의 눈부신 점등은 더 충실하지만, 전경에 별도 차량 지붕 같은 구조물을 추가했고 좁은 지붕만 보여야 할 구도를 벗어났습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "십자가의 발광면은 거의 정면으로 카메라를 향하며, 아래에서 비스듬히 올려다본 면으로 읽히지 않습니다. 왼쪽 앞의 검은 머리 청년은 십자가 쪽을 올려다보지만, 주변 사람들의 시선은 차량과 전방 등으로 나뉩니다. 조준하는 무기나 이동 중인 물체는 보이지 않습니다.",
        "built_space": "십자가 한 개가 받침판과 긴 보강재를 통해 승합차 지붕 한 면에 고정되어 있습니다. 지붕은 하단 중앙의 좁은 조각이 아니라 화면 하단 상당 부분을 차지하고, 카메라는 그 윗면을 내려다봅니다. 뒤쪽에는 여러 상자, 오른쪽에는 열린 차량 문이 보입니다. 도로 양쪽에 철책이 있어, 왼쪽은 개방된 부지이고 오른쪽에 철책이 있는 참조 장소와 다릅니다.",
        "entities": "거칠고 낡은 금속 십자가, 내부의 밝은 전구, 마모된 금속 지붕과 뒤쪽 상자들은 보입니다. 그러나 차량 양쪽에 남녀 성인으로 보이는 인물이 열 명 이상 추가되어 있습니다. 앞쪽 청년은 검은 머리의 젊은 동아시아계 남성처럼 보이지만, 이 장면에는 그를 포함한 사람의 노출 자체가 지시되지 않았습니다. 찰리의 기계 머리는 보이지 않으며, 숨겨진 상태이므로 그 점은 결함이 아닙니다. 판독 가능한 문자는 확인되지 않습니다.",
        "hard_violations": [
         "인물이 등장하지 않는 십자가 공개 장면에 차량 양쪽의 군중을 추가했습니다.",
         "참조 장소에서 개방된 부지여야 하는 도로 왼쪽에 연속 철책을 추가하여 고정 공간 구조를 바꾸었습니다."
        ],
        "physics": "십자가 하단의 볼트와 받침판, 지붕 위 보강재가 하중을 지지합니다. 상자들은 차량 뒤쪽 적재 영역에 쌓여 있으며 공중에 떠 있지 않습니다. 사람들은 차 옆에 서 있는 자세이고 발은 대체로 프레임 밖이거나 가려져 있어, 무지지 부유로 판단할 근거는 없습니다. 전구의 빛이 금속 지붕에 반사되는 관계도 가능합니다."
       },
       {
        "label": "A",
        "direction": "십자가는 약간 비스듬하게 카메라를 향하고 빛은 중심에서 퍼집니다. 다만 지붕 윗면과 받침을 내려다보는 시점이어서, 십자가의 면을 아래에서 올려다보라는 요구에는 맞지 않습니다. 사람의 시선이나 무기 조준, 이동 방향은 없습니다.",
        "built_space": "십자가 한 개와 받침 한 개가 승합차 지붕에 놓여 있습니다. 왼쪽의 공장·빈 부지·가로수와 오른쪽의 철책, 양쪽 가로등은 참조 장소의 배치를 대체로 유지합니다. 그러나 승합차의 앞 유리와 측면까지 노출되며, 화면 맨 아래에는 그 승합차 지붕과 분리된 또 하나의 넓은 골진 차량 지붕 같은 면이 겹쳐 있습니다. 따라서 지정된 단일 승합차의 좁은 지붕 구성이 아닙니다.",
        "entities": "낡은 십자가형 부재, 강하게 켜진 전구, 노출 전선과 결속재, 마모된 승합차가 보이며 사람은 없습니다. 십자가 외곽은 남아 있지만 중심부는 강한 번짐으로 상당히 날아갑니다. 뒤쪽 상자와 숨은 찰리는 이 구도에서 확인되지 않으므로 누락으로 판단하지 않습니다. 차량의 종이와 표식은 흐려서 판독 가능한 글자는 확인되지 않습니다.",
        "hard_violations": [
         "십자가가 달린 승합차와 별개인 차량 지붕 형태의 대형 구조물을 전경에 추가했습니다."
        ],
        "physics": "십자가는 하단 받침과 결속재로 지지되고, 전선은 십자가에서 지붕 쪽으로 이어집니다. 승합차는 하단이 가려져 있지만 정상적으로 서 있는 차량 자세입니다. 추가 전경 구조물도 화면 아래로 이어져 있어 공중에 떠 있다고 판단할 근거는 없습니다. 조명이 지붕 금속에 만드는 반사와 그림자는 물리적으로 가능합니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.095
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.845
   },
   "violations": {
    "B": [
     "[gemini-pro] 발명된 사람 (샷 텍스트에 명시되지 않은 다수의 군중이 프레임에 임의로 추가됨)",
     "[gpt-high] 인물이 등장하지 않는 십자가 공개 장면에 차량 양쪽의 군중을 추가했습니다.",
     "[gpt-high] 참조 장소에서 개방된 부지여야 하는 도로 왼쪽에 연속 철책을 추가하여 고정 공간 구조를 바꾸었습니다."
    ],
    "A": [
     "[gpt-high] 십자가가 달린 승합차와 별개인 차량 지붕 형태의 대형 구조물을 전경에 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 845
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "레퍼런스의 장소와 샷 텍스트의 조명 및 구도를 전반적으로 잘 구현하였으나, 짐칸의 상자가 보이지 않는 점이 아쉽습니다.  ★위반: [gpt-high] 십자가가 달린 승합차와 별개인 차량 지붕 형태의 대형 구조물을 전경에 추가했습니다."
   },
   {
    "label": "B",
    "score": 845,
    "verdict_ko": "샷 텍스트에 전혀 언급되지 않은 다수의 군중을 임의로 화면에 추가하여 인물 지침을 심각하게 위반했습니다.  ★위반: [gemini-pro] 발명된 사람 (샷 텍스트에 명시되지 않은 다수의 군중이 프레임에 임의로 추가됨) / [gpt-high] 인물이 등장하지 않는 십자가 공개 장면에 차량 양쪽의 군중을 추가했습니다. / [gpt-high] 참조 장소에서 개방된 부지여야 하는 도로 왼쪽에 연속 철책을 추가하여 고정 공간 구조를 바꾸었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_militia_road_checkpoint_sel.png",
    "asset_id": "b4c18cf3-5ae5-4429-8e1c-57108458b82e",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdaa-a088-77eb-a63e-87deba1474b4",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S28sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:31:06.042623+00:00",
  "fingerprint": "68cef1869e5407ee492871aea418f30431bc62e88804f4585ec2563090e1f9c3",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S28sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S28sh9_sel.png",
  "source_sha256": "269c051f9941fd30d374991901cc7491796282148d79b95a72e4fdc734f6ab0b",
  "file": "S28sh9_cine.png",
  "staged_sha256": "0462708b39c936db5ce8ebe85a0e43160428ff7b6b0a6e82aa9179b1fc06600f",
  "latency_ms": 15249
 },
 "S29sh8::signage": {
  "fp": "13bbf8bc956d3c4d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::f98300cfa1b2b623": {
  "subjects": [],
  "subject_text": "인천 성당 지하실 계단\n성당 1층에서 지하 기도 공간으로 내려가는 실내 계단. 지하 방향으로 어둡게 이어지며 좁은 범위의 촛불이 계단을 비춘다.",
  "identity": "canonical",
  "scope_id": "L35",
  "scope_role": "location_interior",
  "scope_sha": "49812a3699032803"
 },
 "S29sh8::bgfirst_bg": {
  "input_fingerprint": "ac249f905df285cf",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 신부가 검지손가락을 자신의 입술에 바짝 가져다 댄 채 매서운 눈빛을 보내는 동작.\n\nLOCATION (lock): At the foot of the stairway inside a church's sparse basement prayer room, lit by the priest's candle.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Candlelight supports intimate facial modeling and readable eyes while the silence gesture carries the tension without optical distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 신부가 검지손가락을 자신의 입술에 바짝 가져다 댄 채 매서운 눈빛을 보내는 동작.\n\nLOCATION (lock): At the foot of the stairway inside a church's sparse basement prayer room, lit by the priest's candle.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Candlelight supports intimate facial modeling and readable eyes while the silence gesture carries the tension without optical distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S29sh8__bgfirst_bg.png",
  "asset_id": "4b90ee72-e082-487f-92a0-660e8b0ece7e",
  "input_asset_ids": [
   "5fb34c39-5546-49d4-be4e-ba102f8e791b",
   "d5f56d4b-9f05-4a32-becb-1f0727ee98b9"
  ]
 },
 "S29sh8": {
  "input_fingerprint": "341673da3b1e8d7a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 신부가 검지손가락을 자신의 입술에 바짝 가져다 댄 채 매서운 눈빛을 보내는 동작.\n\nLOCATION (lock): At the foot of the stairway inside a church's sparse basement prayer room, lit by the priest's candle. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Candlelight supports intimate facial modeling and readable eyes while the silence gesture carries the tension without optical distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement prayer room contains a cross, a small desk, chairs and a small harmonium, with the candle brought downstairs providing light. Charlie remains there in his old coat and hat; the church entrance is still shut before it is answered. 신부: He is preparing to ascend the stairs, wearing his clerical collar and carrying the candle brought down earlier.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 신부가 검지손가락을 자신의 입술에 바짝 가져다 댄 채 매서운 눈빛을 보내는 동작.\n\nLOCATION (lock): At the foot of the stairway inside a church's sparse basement prayer room, lit by the priest's candle. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Candlelight supports intimate facial modeling and readable eyes while the silence gesture carries the tension without optical distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement prayer room contains a cross, a small desk, chairs and a small harmonium, with the candle brought downstairs providing light. Charlie remains there in his old coat and hat; the church entrance is still shut before it is answered. 신부: He is preparing to ascend the stairs, wearing his clerical collar and carrying the candle brought down earlier.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 신부가 검지손가락을 자신의 입술에 바짝 가져다 댄 채 매서운 눈빛을 보내는 동작.\n\nLOCATION (lock): At the foot of the stairway inside a church's sparse basement prayer room, lit by the priest's candle. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Candlelight supports intimate facial modeling and readable eyes while the silence gesture carries the tension without optical distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement prayer room contains a cross, a small desk, chairs and a small harmonium, with the candle brought downstairs providing light. Charlie remains there in his old coat and hat; the church entrance is still shut before it is answered. 신부: He is preparing to ascend the stairs, wearing his clerical collar and carrying the candle brought down earlier.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S29sh8__bgfirst_bg.png",
     "asset_id": "4b90ee72-e082-487f-92a0-660e8b0ece7e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S29sh8.png",
     "asset_id": "5fb34c39-5546-49d4-be4e-ba102f8e791b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:672084>",
     "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L35B03.png",
     "asset_id": "d5f56d4b-9f05-4a32-becb-1f0727ee98b9",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:672084>",
     "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "신부가 앞쪽의 남성을 향해 매서운 눈빛을 보내며 검지를 입술에 대고 있음.",
    "built_space": "계단과 난간이 있으나 우측에 넓은 방 공간이 임의로 추가되어 원본 좁은 계단실 구조와 다름.",
    "entities": "신부는 레퍼런스와 잘 일치하나, 지침에 어긋나게 모자를 쓴 남성이 전면에 등장함.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 없는 인물(모자와 코트를 입은 뒷모습) 임의 추가",
     "[gpt-high] 신부만 등장할 수 있는 숏에 모자 쓴 남성의 머리와 상체를 전경 인물로 추가했다."
    ],
    "physics": "신부가 오른손으로 촛대를 쥐고 있어 자연스럽게 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "신부가 정면을 매섭게 응시하며 검지를 입술에 바짝 대고 있음.",
    "built_space": "좁은 계단실과 좌측 난간이 레퍼런스 이미지의 공간 구조와 완벽히 일치함.",
    "entities": "신부만 단독으로 등장하며 외양 묘사가 레퍼런스와 일치함.",
    "hard_violations": [
     "[gemini-pro] 지지하는 손이나 물체 없이 허공에 떠 있는 양초"
    ],
    "physics": "화면 좌측 하단에 위치한 양초를 잡고 있는 손이나 놓여 있는 표면이 전혀 없어 허공에 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "신부 단독 클로즈업 및 표정 연기는 우수하나, 아무런 지지대 없이 공중에 떠 있는 양초가 치명적인 물리적 오류를 발생시킴."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷 텍스트에 없는 추가 인물을 프레임 전면에 배치하여 엄격한 인물 제한 지침을 심각하게 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "신부가 앞쪽의 남성을 향해 매서운 눈빛을 보내며 검지를 입술에 대고 있음.",
        "built_space": "계단과 난간이 있으나 우측에 넓은 방 공간이 임의로 추가되어 원본 좁은 계단실 구조와 다름.",
        "entities": "신부는 레퍼런스와 잘 일치하나, 지침에 어긋나게 모자를 쓴 남성이 전면에 등장함.",
        "hard_violations": [
         "샷 텍스트에 없는 인물(모자와 코트를 입은 뒷모습) 임의 추가"
        ],
        "physics": "신부가 오른손으로 촛대를 쥐고 있어 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "신부가 정면을 매섭게 응시하며 검지를 입술에 바짝 대고 있음.",
        "built_space": "좁은 계단실과 좌측 난간이 레퍼런스 이미지의 공간 구조와 완벽히 일치함.",
        "entities": "신부만 단독으로 등장하며 외양 묘사가 레퍼런스와 일치함.",
        "hard_violations": [
         "지지하는 손이나 물체 없이 허공에 떠 있는 양초"
        ],
        "physics": "화면 좌측 하단에 위치한 양초를 잡고 있는 손이나 놓여 있는 표면이 전혀 없어 허공에 떠 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "신부 단독 클로즈업 및 표정 연기는 우수하나, 아무런 지지대 없이 공중에 떠 있는 양초가 치명적인 물리적 오류를 발생시킴."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷 텍스트에 없는 추가 인물을 프레임 전면에 배치하여 엄격한 인물 제한 지침을 심각하게 위반함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "신부가 앞쪽의 남성을 향해 매서운 눈빛을 보내며 검지를 입술에 대고 있음.",
        "built_space": "계단과 난간이 있으나 우측에 넓은 방 공간이 임의로 추가되어 원본 좁은 계단실 구조와 다름.",
        "entities": "신부는 레퍼런스와 잘 일치하나, 지침에 어긋나게 모자를 쓴 남성이 전면에 등장함.",
        "hard_violations": [
         "샷 텍스트에 없는 인물(모자와 코트를 입은 뒷모습) 임의 추가"
        ],
        "physics": "신부가 오른손으로 촛대를 쥐고 있어 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "신부가 정면을 매섭게 응시하며 검지를 입술에 바짝 대고 있음.",
        "built_space": "좁은 계단실과 좌측 난간이 레퍼런스 이미지의 공간 구조와 완벽히 일치함.",
        "entities": "신부만 단독으로 등장하며 외양 묘사가 레퍼런스와 일치함.",
        "hard_violations": [
         "지지하는 손이나 물체 없이 허공에 떠 있는 양초"
        ],
        "physics": "화면 좌측 하단에 위치한 양초를 잡고 있는 손이나 놓여 있는 표면이 전혀 없어 허공에 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "신부만 담은 밀착 클로즈업으로 입술에 댄 검지와 매서운 눈빛을 정확히 구현하고, 계단 아래 공간과 촛불 조명도 유지한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "침묵 동작과 시선은 명확하지만, 등장할 수 없는 모자 쓴 인물을 전경에 추가하고 어깨너머 구도로 넓혀 핵심 인물 제한과 프레이밍을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "검지를 수직으로 세워 자신의 입술 중앙에 밀착했으며 손끝은 코밑까지 올라온다. 눈은 카메라 바로 앞의 화면 밖 상대를 향하는 듯하고, 찌푸린 미간과 집중된 시선으로 매서운 표정을 만든다. 화면 안에 시선의 상대는 없다.",
        "built_space": "신부 뒤로 위쪽으로 올라가는 계단 한 줄과 왼쪽 벽에 고정된 금속 난간 하나가 보인다. 양쪽의 낡은 회벽과 어두운 상부 통로는 장소 참조와 부합하며, 신부는 계단 아래에 위치한다. 십자가·책상·의자·풍금은 밀착 구도 밖이므로 누락으로 판단하지 않는다. 반사는 없다.",
        "entities": "참조처럼 밝은 피부의 60대 남성 신부 한 명이며, 뒤로 물러난 짧은 회색 머리와 노년의 얼굴 특징이 대체로 일치한다. 검은 성직복과 흰 성직자 칼라가 보인다. 왼쪽 아래에는 불붙은 초 하나가 있다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "들어 올린 손은 손목과 팔에 자연스럽게 연결되고 검지가 입술에 닿아 있어 수행 가능한 동작이다. 하체와 바닥 접촉은 클로즈업 밖이다. 초는 화면 아래로 이어져 밑부분과 잡은 손이 잘렸으므로 지지 방식을 직접 확인할 수 없지만, 공중에 떠 있다고 볼 근거도 없다."
       },
       {
        "label": "B",
        "direction": "신부의 검지는 자기 입술에 수직으로 닿아 있고, 날카로운 시선은 왼쪽 전경의 모자 쓴 남자를 향한다. 그 남자의 머리도 신부 쪽으로 돌아가 있다. 침묵을 요구하는 관계는 명확하지만 그 상대를 화면에 등장시키는 것은 허용되지 않았다.",
        "built_space": "뒤쪽으로 올라가는 계단 한 줄, 왼쪽 벽 난간 하나와 상부 출입구가 보인다. 오른쪽에는 벽 십자가 하나, 작은 책상 하나, 일부 가려진 등받이 의자 하나, 풍금 하나와 그 앞의 작은 의자 하나가 보인다. 신부는 계단 아래에 있으나, 방의 비품과 전경 인물까지 포함하면서 요구된 얼굴 중심 클로즈업보다 넓은 어깨너머 구도가 되었다. 반사는 없다.",
        "entities": "회색 머리와 밝은 피부의 노년 남성 신부는 참조 인물과 대체로 닮았으며 검은 성직복과 흰 칼라를 착용했다. 다른 손에는 받침 위의 불붙은 초 하나가 있다. 왼쪽 전경에 모자와 외투를 입은 남성이 추가되었는데, 찰리를 의도했더라도 이 숏에서 허용된 인물은 신부뿐이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "신부만 등장할 수 있는 숏에 모자 쓴 남성의 머리와 상체를 전경 인물로 추가했다."
        ],
        "physics": "입술에 댄 검지와 굽힌 팔은 자연스럽게 연결되어 있다. 반대쪽 손은 촛대 받침 아래를 잡아 초를 지지하고 있으며, 초와 불꽃도 수직으로 선다. 두 인물의 하체는 화면 밖이지만 상체가 떠 있거나 지지 없이 놓인 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "신부만 담은 밀착 클로즈업으로 입술에 댄 검지와 매서운 눈빛을 정확히 구현하고, 계단 아래 공간과 촛불 조명도 유지한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "침묵 동작과 시선은 명확하지만, 등장할 수 없는 모자 쓴 인물을 전경에 추가하고 어깨너머 구도로 넓혀 핵심 인물 제한과 프레이밍을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "검지를 수직으로 세워 자신의 입술 중앙에 밀착했으며 손끝은 코밑까지 올라온다. 눈은 카메라 바로 앞의 화면 밖 상대를 향하는 듯하고, 찌푸린 미간과 집중된 시선으로 매서운 표정을 만든다. 화면 안에 시선의 상대는 없다.",
        "built_space": "신부 뒤로 위쪽으로 올라가는 계단 한 줄과 왼쪽 벽에 고정된 금속 난간 하나가 보인다. 양쪽의 낡은 회벽과 어두운 상부 통로는 장소 참조와 부합하며, 신부는 계단 아래에 위치한다. 십자가·책상·의자·풍금은 밀착 구도 밖이므로 누락으로 판단하지 않는다. 반사는 없다.",
        "entities": "참조처럼 밝은 피부의 60대 남성 신부 한 명이며, 뒤로 물러난 짧은 회색 머리와 노년의 얼굴 특징이 대체로 일치한다. 검은 성직복과 흰 성직자 칼라가 보인다. 왼쪽 아래에는 불붙은 초 하나가 있다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "들어 올린 손은 손목과 팔에 자연스럽게 연결되고 검지가 입술에 닿아 있어 수행 가능한 동작이다. 하체와 바닥 접촉은 클로즈업 밖이다. 초는 화면 아래로 이어져 밑부분과 잡은 손이 잘렸으므로 지지 방식을 직접 확인할 수 없지만, 공중에 떠 있다고 볼 근거도 없다."
       },
       {
        "label": "A",
        "direction": "신부의 검지는 자기 입술에 수직으로 닿아 있고, 날카로운 시선은 왼쪽 전경의 모자 쓴 남자를 향한다. 그 남자의 머리도 신부 쪽으로 돌아가 있다. 침묵을 요구하는 관계는 명확하지만 그 상대를 화면에 등장시키는 것은 허용되지 않았다.",
        "built_space": "뒤쪽으로 올라가는 계단 한 줄, 왼쪽 벽 난간 하나와 상부 출입구가 보인다. 오른쪽에는 벽 십자가 하나, 작은 책상 하나, 일부 가려진 등받이 의자 하나, 풍금 하나와 그 앞의 작은 의자 하나가 보인다. 신부는 계단 아래에 있으나, 방의 비품과 전경 인물까지 포함하면서 요구된 얼굴 중심 클로즈업보다 넓은 어깨너머 구도가 되었다. 반사는 없다.",
        "entities": "회색 머리와 밝은 피부의 노년 남성 신부는 참조 인물과 대체로 닮았으며 검은 성직복과 흰 칼라를 착용했다. 다른 손에는 받침 위의 불붙은 초 하나가 있다. 왼쪽 전경에 모자와 외투를 입은 남성이 추가되었는데, 찰리를 의도했더라도 이 숏에서 허용된 인물은 신부뿐이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "신부만 등장할 수 있는 숏에 모자 쓴 남성의 머리와 상체를 전경 인물로 추가했다."
        ],
        "physics": "입술에 댄 검지와 굽힌 팔은 자연스럽게 연결되어 있다. 반대쪽 손은 촛대 받침 아래를 잡아 초를 지지하고 있으며, 초와 불꽃도 수직으로 선다. 두 인물의 하체는 화면 밖이지만 상체가 떠 있거나 지지 없이 놓인 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 샷 텍스트에 없는 인물(모자와 코트를 입은 뒷모습) 임의 추가",
     "[gpt-high] 신부만 등장할 수 있는 숏에 모자 쓴 남성의 머리와 상체를 전경 인물로 추가했다."
    ],
    "B": [
     "[gemini-pro] 지지하는 손이나 물체 없이 허공에 떠 있는 양초"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 1083
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "신부 단독 클로즈업 및 표정 연기는 우수하나, 아무런 지지대 없이 공중에 떠 있는 양초가 치명적인 물리적 오류를 발생시킴.  ★위반: [gemini-pro] 지지하는 손이나 물체 없이 허공에 떠 있는 양초"
   },
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "샷 텍스트에 없는 추가 인물을 프레임 전면에 배치하여 엄격한 인물 제한 지침을 심각하게 위반함.  ★위반: [gemini-pro] 샷 텍스트에 없는 인물(모자와 코트를 입은 뒷모습) 임의 추가 / [gpt-high] 신부만 등장할 수 있는 숏에 모자 쓴 남성의 머리와 상체를 전경 인물로 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L35B03.png",
    "asset_id": "d5f56d4b-9f05-4a32-becb-1f0727ee98b9",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:672084>",
    "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdab-99d0-74d5-815b-e32ad533112c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S29sh8__bgfirst_bg.png",
   "bg_asset_id": "4b90ee72-e082-487f-92a0-660e8b0ece7e",
   "bg_record_key": "S29sh8::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S29sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:31:21.698132+00:00",
  "fingerprint": "cf4d3d0b21f6a7dd8b2a4ac69ada2793bd0c5baa5f2b5349630f791e92c8570e",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S29sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S29sh8_sel.png",
  "source_sha256": "46e24c71200f2b21247d6b2ebe8c0febc0a63e951e1eed4c05ea27c2613b41d0",
  "file": "S29sh8_cine.png",
  "staged_sha256": "6eb3fed16900d5114044229d91b750de1d0ce690009ac29396c3d02ab05ea891",
  "latency_ms": 11266
 },
 "S29sh13::signage": {
  "fp": "91c7d0559bd0a82b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S29sh13::bgfirst_bg": {
  "input_fingerprint": "1a54320a724f78d6",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 굳게 닫힌 성당 문 밖, 구도환과 신부가 서로 마주 보고 선 대치 구도.\n\nLOCATION (lock): Outside the church's closed main entrance, on the nighttime approach immediately beyond the doors.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Church entrance door (Firmly closed) — The exterior face recedes at an oblique angle behind the two men; used as A shared threshold anchoring the exterior encounter, kept to a limited background portion.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient light maintains facial separation and restrained contrast without carrying the basement candlelight outdoors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 굳게 닫힌 성당 문 밖, 구도환과 신부가 서로 마주 보고 선 대치 구도.\n\nLOCATION (lock): Outside the church's closed main entrance, on the nighttime approach immediately beyond the doors.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Church entrance door (Firmly closed) — The exterior face recedes at an oblique angle behind the two men; used as A shared threshold anchoring the exterior encounter, kept to a limited background portion.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient light maintains facial separation and restrained contrast without carrying the basement candlelight outdoors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S29sh13__bgfirst_bg.png",
  "asset_id": "6ccda324-e781-46e9-92ef-7652e903451d",
  "input_asset_ids": [
   "0f7f9a8e-53e4-47a5-b4c4-2cf55c90fb71",
   "9567428d-cab9-40ce-a731-f482cf5d48d1"
  ]
 },
 "S29sh13": {
  "input_fingerprint": "1959ef5cc9d889b2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 굳게 닫힌 성당 문 밖, 구도환과 신부가 서로 마주 보고 선 대치 구도.\n\nLOCATION (lock): Outside the church's closed main entrance, on the nighttime approach immediately beyond the doors. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Church entrance door (Firmly closed) — The exterior face recedes at an oblique angle behind the two men; used as A shared threshold anchoring the exterior encounter, kept to a limited background portion.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient light maintains facial separation and restrained contrast without carrying the basement candlelight outdoors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The church entrance has been opened and its small viewing hatch opened; the text provides no subsequent closing. The basement retains its cross, small desk, chairs and harmonium, with Charlie waiting there in his old coat and hat. 신부: He is outside the church entrance, wearing his clerical collar and retaining the candle brought upstairs.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 굳게 닫힌 성당 문 밖, 구도환과 신부가 서로 마주 보고 선 대치 구도.\n\nLOCATION (lock): Outside the church's closed main entrance, on the nighttime approach immediately beyond the doors. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Church entrance door (Firmly closed) — The exterior face recedes at an oblique angle behind the two men; used as A shared threshold anchoring the exterior encounter, kept to a limited background portion.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient light maintains facial separation and restrained contrast without carrying the basement candlelight outdoors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The church entrance has been opened and its small viewing hatch opened; the text provides no subsequent closing. The basement retains its cross, small desk, chairs and harmonium, with Charlie waiting there in his old coat and hat. 신부: He is outside the church entrance, wearing his clerical collar and retaining the candle brought upstairs.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 굳게 닫힌 성당 문 밖, 구도환과 신부가 서로 마주 보고 선 대치 구도.\n\nLOCATION (lock): Outside the church's closed main entrance, on the nighttime approach immediately beyond the doors. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Church entrance door (Firmly closed) — The exterior face recedes at an oblique angle behind the two men; used as A shared threshold anchoring the exterior encounter, kept to a limited background portion.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient light maintains facial separation and restrained contrast without carrying the basement candlelight outdoors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The church entrance has been opened and its small viewing hatch opened; the text provides no subsequent closing. The basement retains its cross, small desk, chairs and harmonium, with Charlie waiting there in his old coat and hat. 신부: He is outside the church entrance, wearing his clerical collar and retaining the candle brought upstairs.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S29sh13__bgfirst_bg.png",
     "asset_id": "6ccda324-e781-46e9-92ef-7652e903451d",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S29sh13.png",
     "asset_id": "0f7f9a8e-53e4-47a5-b4c4-2cf55c90fb71",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:672084>",
     "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L35B01.png",
     "asset_id": "9567428d-cab9-40ce-a731-f482cf5d48d1",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:672084>",
     "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물이 서로를 향해 마주 보고 있으며, 시선이 상대방에게 정확히 향함.",
    "built_space": "성당 문 밖 야외. 참조된 벽돌 벽, 조명, 닫힌 문이 화면 좌측 배경에 올바르게 배치됨.",
    "entities": "신부(참조 이미지의 얼굴, 백발, 사제복 일치), 구도환(가죽 재킷을 입은 아시아계 남성) 모두 식별됨. 거대한 양초가 묘사됨.",
    "hard_violations": [],
    "physics": "두 사람 모두 지면에 안정적으로 서 있으며, 신부의 손이 양초를 자연스럽게 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "두 사람이 서로 마주 보고 있음.",
    "built_space": "성당 문 밖 야외 공간이 참조 이미지와 일치하게 배경에 배치됨.",
    "entities": "신부(사제복 및 외모 일치하나 촛불에 의한 묘사가 다소 인위적임), 구도환(아시아계 남성) 식별됨.",
    "hard_violations": [],
    "physics": "지면에 서 있으나 구도환의 자세가 비자연스럽게 경직되어 팔을 아래로 늘어뜨리고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 위치와 인물 참조를 잘 반영하였으며, 자연스러운 자세(특히 구도환)와 차분한 야간 조명을 성공적으로 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "프롬프트가 명시적으로 피하라고 한 '경직된 차렷 자세'를 구도환이 취하고 있으며, 촛불 조명이 과도하여 지정된 조명 무드에서 벗어남."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 서로를 향해 마주 보고 있으며, 시선이 상대방에게 정확히 향함.",
        "built_space": "성당 문 밖 야외. 참조된 벽돌 벽, 조명, 닫힌 문이 화면 좌측 배경에 올바르게 배치됨.",
        "entities": "신부(참조 이미지의 얼굴, 백발, 사제복 일치), 구도환(가죽 재킷을 입은 아시아계 남성) 모두 식별됨. 거대한 양초가 묘사됨.",
        "hard_violations": [],
        "physics": "두 사람 모두 지면에 안정적으로 서 있으며, 신부의 손이 양초를 자연스럽게 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "두 사람이 서로 마주 보고 있음.",
        "built_space": "성당 문 밖 야외 공간이 참조 이미지와 일치하게 배경에 배치됨.",
        "entities": "신부(사제복 및 외모 일치하나 촛불에 의한 묘사가 다소 인위적임), 구도환(아시아계 남성) 식별됨.",
        "hard_violations": [],
        "physics": "지면에 서 있으나 구도환의 자세가 비자연스럽게 경직되어 팔을 아래로 늘어뜨리고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 위치와 인물 참조를 잘 반영하였으며, 자연스러운 자세(특히 구도환)와 차분한 야간 조명을 성공적으로 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "프롬프트가 명시적으로 피하라고 한 '경직된 차렷 자세'를 구도환이 취하고 있으며, 촛불 조명이 과도하여 지정된 조명 무드에서 벗어남."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 서로를 향해 마주 보고 있으며, 시선이 상대방에게 정확히 향함.",
        "built_space": "성당 문 밖 야외. 참조된 벽돌 벽, 조명, 닫힌 문이 화면 좌측 배경에 올바르게 배치됨.",
        "entities": "신부(참조 이미지의 얼굴, 백발, 사제복 일치), 구도환(가죽 재킷을 입은 아시아계 남성) 모두 식별됨. 거대한 양초가 묘사됨.",
        "hard_violations": [],
        "physics": "두 사람 모두 지면에 안정적으로 서 있으며, 신부의 손이 양초를 자연스럽게 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "두 사람이 서로 마주 보고 있음.",
        "built_space": "성당 문 밖 야외 공간이 참조 이미지와 일치하게 배경에 배치됨.",
        "entities": "신부(사제복 및 외모 일치하나 촛불에 의한 묘사가 다소 인위적임), 구도환(아시아계 남성) 식별됨.",
        "hard_violations": [],
        "physics": "지면에 서 있으나 구도환의 자세가 비자연스럽게 경직되어 팔을 아래로 늘어뜨리고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "야간의 닫힌 성당 문 밖에서 서로 마주 선 전신 와이드 숏은 맞지만, 두 사람이 오른쪽에 몰려 출입문이 별도의 큰 배경으로 부각되고 신부 얼굴의 촛불색도 강하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "서로를 응시하는 두 사람과 비스듬히 물러나는 닫힌 출입문을 하나의 대치 공간으로 묶었으며, 신부의 정체성·복장·촛불 소지와 절제된 야간 얼굴 조명이 더 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽의 신부는 오른쪽 구도환의 얼굴을, 구도환은 신부의 얼굴을 바라본다. 몸도 서로를 향해 있어 대치 방향이 맞는다. 신부가 든 초는 수직으로 서 있다.",
        "built_space": "화면 왼쪽에 닫힌 양개 목문 한 세트와 손잡이 두 개, 그 위 십자가 하나, 왼쪽 벽등 하나가 보인다. 오른쪽에는 세로 창 하나와 배수관 하나, 가장자리 화단이 있다. 붉은 벽돌과 석재 문틀, 마모된 문 표면은 장소 참조와 부합한다. 두 사람은 모두 문 오른쪽의 외부 포장면에 서 있다. 외벽은 비스듬히 이어지지만 문 전체가 크게 노출되어 두 사람의 공통 문턱보다는 옆의 독립된 배경처럼 읽힌다. 열린 작은 확인창은 보이지 않는다.",
        "entities": "인물은 두 명뿐이다. 신부는 참조처럼 짧고 성긴 회색 머리의 백인으로 보이는 노년 남성이며, 긴 검은 수단과 흰 성직자 칼라를 착용했다. 얼굴 윤곽과 체형도 대체로 일치한다. 구도환은 짧은 검은 머리의 젊은 동아시아계로 보이는 남성으로 회색 재킷과 검은 바지를 입었다. 구도환의 별도 외모 참조는 없다. 신부 손에는 불붙은 가는 초 하나가 있다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "두 사람의 신발은 포장면에 닿아 있고 체중을 지탱한다. 신부는 손으로 초의 아래쪽을 쥐고 있으며 초가 떠 있지 않다. 구도환의 내려놓은 팔과 약간 벌린 발은 정지한 대치 자세로 가능하다. 문과 벽등, 십자가도 각각 경첩과 벽에 고정된 구조로 자연스럽다."
       },
       {
        "label": "B",
        "direction": "왼쪽 구도환은 오른쪽 신부의 얼굴을 똑바로 보고, 신부 역시 구도환을 바라본다. 두 사람의 몸 방향도 서로 마주한다. 신부가 두 손으로 잡은 초는 위로 곧게 서 있다.",
        "built_space": "왼쪽 배경의 닫힌 양개 목문 한 세트에 손잡이 두 개가 있고, 위쪽 십자가 하나와 왼쪽 벽등 하나가 보인다. 오른쪽으로 세로 창 하나, 배수관 하나와 화단이 이어진다. 벽돌 외벽, 석재 문틀과 기단, 낡은 목문의 재료와 배치는 참조 장소에 부합한다. 구도환은 문 앞 외부 포장면에, 신부는 그 오른쪽 접근로에 서 있어 문턱을 공유하는 대치로 읽힌다. 문 일부가 구도환 뒤에 가려지고 외벽이 비스듬히 이어져 배경 역할이 더 제한된다. 열린 작은 확인창은 보이지 않는다.",
        "entities": "구도환과 신부 두 명만 보인다. 오른쪽 신부는 짧은 회색 머리와 노년의 얼굴, 마른 체형을 가진 백인으로 보이는 남성으로 참조 인물과 잘 맞는다. 긴 검은 수단과 흰 성직자 칼라도 유지했다. 왼쪽 구도환은 검은 머리의 젊은 동아시아계로 보이는 남성으로 어두운 가죽 재킷과 검은 바지를 착용했다. 신부는 불붙은 굵은 초 하나를 들고 있다. 초의 굵기를 고정하는 참조는 없다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "두 사람 모두 신발 바닥이 지면에 닿아 체중을 받는다. 구도환은 발을 조금 어긋나게 두고 손을 주머니에 넣어 자연스럽게 버티며, 신부는 두 손으로 초를 감싸 지탱한다. 몸이나 소품이 떠 있지 않고, 옷자락과 그림자도 서 있는 자세에 맞는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "야간의 닫힌 성당 문 밖에서 서로 마주 선 전신 와이드 숏은 맞지만, 두 사람이 오른쪽에 몰려 출입문이 별도의 큰 배경으로 부각되고 신부 얼굴의 촛불색도 강하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "서로를 응시하는 두 사람과 비스듬히 물러나는 닫힌 출입문을 하나의 대치 공간으로 묶었으며, 신부의 정체성·복장·촛불 소지와 절제된 야간 얼굴 조명이 더 충실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽의 신부는 오른쪽 구도환의 얼굴을, 구도환은 신부의 얼굴을 바라본다. 몸도 서로를 향해 있어 대치 방향이 맞는다. 신부가 든 초는 수직으로 서 있다.",
        "built_space": "화면 왼쪽에 닫힌 양개 목문 한 세트와 손잡이 두 개, 그 위 십자가 하나, 왼쪽 벽등 하나가 보인다. 오른쪽에는 세로 창 하나와 배수관 하나, 가장자리 화단이 있다. 붉은 벽돌과 석재 문틀, 마모된 문 표면은 장소 참조와 부합한다. 두 사람은 모두 문 오른쪽의 외부 포장면에 서 있다. 외벽은 비스듬히 이어지지만 문 전체가 크게 노출되어 두 사람의 공통 문턱보다는 옆의 독립된 배경처럼 읽힌다. 열린 작은 확인창은 보이지 않는다.",
        "entities": "인물은 두 명뿐이다. 신부는 참조처럼 짧고 성긴 회색 머리의 백인으로 보이는 노년 남성이며, 긴 검은 수단과 흰 성직자 칼라를 착용했다. 얼굴 윤곽과 체형도 대체로 일치한다. 구도환은 짧은 검은 머리의 젊은 동아시아계로 보이는 남성으로 회색 재킷과 검은 바지를 입었다. 구도환의 별도 외모 참조는 없다. 신부 손에는 불붙은 가는 초 하나가 있다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "두 사람의 신발은 포장면에 닿아 있고 체중을 지탱한다. 신부는 손으로 초의 아래쪽을 쥐고 있으며 초가 떠 있지 않다. 구도환의 내려놓은 팔과 약간 벌린 발은 정지한 대치 자세로 가능하다. 문과 벽등, 십자가도 각각 경첩과 벽에 고정된 구조로 자연스럽다."
       },
       {
        "label": "A",
        "direction": "왼쪽 구도환은 오른쪽 신부의 얼굴을 똑바로 보고, 신부 역시 구도환을 바라본다. 두 사람의 몸 방향도 서로 마주한다. 신부가 두 손으로 잡은 초는 위로 곧게 서 있다.",
        "built_space": "왼쪽 배경의 닫힌 양개 목문 한 세트에 손잡이 두 개가 있고, 위쪽 십자가 하나와 왼쪽 벽등 하나가 보인다. 오른쪽으로 세로 창 하나, 배수관 하나와 화단이 이어진다. 벽돌 외벽, 석재 문틀과 기단, 낡은 목문의 재료와 배치는 참조 장소에 부합한다. 구도환은 문 앞 외부 포장면에, 신부는 그 오른쪽 접근로에 서 있어 문턱을 공유하는 대치로 읽힌다. 문 일부가 구도환 뒤에 가려지고 외벽이 비스듬히 이어져 배경 역할이 더 제한된다. 열린 작은 확인창은 보이지 않는다.",
        "entities": "구도환과 신부 두 명만 보인다. 오른쪽 신부는 짧은 회색 머리와 노년의 얼굴, 마른 체형을 가진 백인으로 보이는 남성으로 참조 인물과 잘 맞는다. 긴 검은 수단과 흰 성직자 칼라도 유지했다. 왼쪽 구도환은 검은 머리의 젊은 동아시아계로 보이는 남성으로 어두운 가죽 재킷과 검은 바지를 착용했다. 신부는 불붙은 굵은 초 하나를 들고 있다. 초의 굵기를 고정하는 참조는 없다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "두 사람 모두 신발 바닥이 지면에 닿아 체중을 받는다. 구도환은 발을 조금 어긋나게 두고 손을 주머니에 넣어 자연스럽게 버티며, 신부는 두 손으로 초를 감싸 지탱한다. 몸이나 소품이 떠 있지 않고, 옷자락과 그림자도 서 있는 자세에 맞는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.46
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.46
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1460
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 위치와 인물 참조를 잘 반영하였으며, 자연스러운 자세(특히 구도환)와 차분한 야간 조명을 성공적으로 구현함."
   },
   {
    "label": "B",
    "score": 1460,
    "verdict_ko": "프롬프트가 명시적으로 피하라고 한 '경직된 차렷 자세'를 구도환이 취하고 있으며, 촛불 조명이 과도하여 지정된 조명 무드에서 벗어남."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L35B01.png",
    "asset_id": "9567428d-cab9-40ce-a731-f482cf5d48d1",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:672084>",
    "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdac-549c-7369-b11d-8f36f4821c36",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S29sh13__bgfirst_bg.png",
   "bg_asset_id": "6ccda324-e781-46e9-92ef-7652e903451d",
   "bg_record_key": "S29sh13::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S29sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:31:33.373530+00:00",
  "fingerprint": "1f73867677efd687dc415492b025cb6c4db53022c5b9183953deea42cb4c1b6a",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S29sh13_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S29sh13_sel.png",
  "source_sha256": "bc1b7556b4cd4ce4085ec6a623194ffead73933e9428dd56fb17abe02c6f5b57",
  "file": "S29sh13_cine.png",
  "staged_sha256": "e170600d51c5b336df1035d92891b5836cc42c87b5288e12a32be006a884d537",
  "latency_ms": 12429
 },
 "S29sh14::signage": {
  "fp": "b3a4f7cb61501e43",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S29sh14": {
  "input_fingerprint": "a9e5c84ff399a99e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 위로 거대한 헬리콥터가 눈부신 서치라이트 불빛을 아래로 쏘아 내리는 웅장한 찰나.\n\nLOCATION (lock): In the night sky over the refugee settlement near the church, where a searching helicopter casts its beam downward. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Helicopter (Airborne during the search) — Its underside and an oblique side are visible from the offset ground viewpoint; used as Upper-frame scale anchor whose negative-space surroundings convey mechanical grandeur; Night sky (Dark); used as Open surrounding space reserving room for the subsequent pan toward the drones.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dazzling downward searchlight contrasts with the dark night sky, rendering the flashback with direct physical clarity rather than dreamlike haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Helicopters sweep the night with searchlights while multiple drones fly along their beams. The church entrance and viewing hatch have been opened, and the furnished basement remains occupied by Charlie in his coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 위로 거대한 헬리콥터가 눈부신 서치라이트 불빛을 아래로 쏘아 내리는 웅장한 찰나.\n\nLOCATION (lock): In the night sky over the refugee settlement near the church, where a searching helicopter casts its beam downward. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Helicopter (Airborne during the search) — Its underside and an oblique side are visible from the offset ground viewpoint; used as Upper-frame scale anchor whose negative-space surroundings convey mechanical grandeur; Night sky (Dark); used as Open surrounding space reserving room for the subsequent pan toward the drones.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dazzling downward searchlight contrasts with the dark night sky, rendering the flashback with direct physical clarity rather than dreamlike haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Helicopters sweep the night with searchlights while multiple drones fly along their beams. The church entrance and viewing hatch have been opened, and the furnished basement remains occupied by Charlie in his coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 위로 거대한 헬리콥터가 눈부신 서치라이트 불빛을 아래로 쏘아 내리는 웅장한 찰나.\n\nLOCATION (lock): In the night sky over the refugee settlement near the church, where a searching helicopter casts its beam downward. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Helicopter (Airborne during the search) — Its underside and an oblique side are visible from the offset ground viewpoint; used as Upper-frame scale anchor whose negative-space surroundings convey mechanical grandeur; Night sky (Dark); used as Open surrounding space reserving room for the subsequent pan toward the drones.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dazzling downward searchlight contrasts with the dark night sky, rendering the flashback with direct physical clarity rather than dreamlike haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Helicopters sweep the night with searchlights while multiple drones fly along their beams. The church entrance and viewing hatch have been opened, and the furnished basement remains occupied by Charlie in his coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "서치라이트 빔이 화면 좌측 하단을 향해 길게 뻗음.",
    "built_space": "화면 하단에 레퍼런스와 일치하는 교회 종탑과 우측에 일부 건축물이 보임.",
    "entities": "지시된 거대한 헬리콥터와 빔을 따라 비행하는 다수의 드론이 존재함.",
    "hard_violations": [],
    "physics": "헬리콥터와 드론들 모두 기체 특성으로 공중에 정상 체공 중임."
   },
   {
    "label": "B",
    "direction": "서치라이트 빔이 수직 아래 지면을 비춤.",
    "built_space": "좌측에 교회 건물, 우측에 텐트 등 난민촌 구조물이 넓게 배치됨.",
    "entities": "헬리콥터와 드론 외에 지상에 다수의 인물(전경의 남성 포함)이 보임.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 인물들을 화면에 임의로 추가함 (Invented people)",
     "[gpt-high] 숏 텍스트에 없는 사람들을 전경과 정착촌 마당에 여러 명 추가했습니다."
    ],
    "physics": "공중의 기체들은 비행 중이며, 지상의 인물들은 땅에 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 구도에 맞춰 밤하늘의 여백과 헬리콥터를 강조했으며, 샷 텍스트에 없는 인물을 추가하지 않아 요구사항을 잘 충족함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "샷 텍스트에 전혀 언급되지 않은 다수의 인물을 지상에 임의로 배치하여 치명적인 지시 위반이 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "서치라이트 빔이 화면 좌측 하단을 향해 길게 뻗음.",
        "built_space": "화면 하단에 레퍼런스와 일치하는 교회 종탑과 우측에 일부 건축물이 보임.",
        "entities": "지시된 거대한 헬리콥터와 빔을 따라 비행하는 다수의 드론이 존재함.",
        "hard_violations": [],
        "physics": "헬리콥터와 드론들 모두 기체 특성으로 공중에 정상 체공 중임."
       },
       {
        "label": "B",
        "direction": "서치라이트 빔이 수직 아래 지면을 비춤.",
        "built_space": "좌측에 교회 건물, 우측에 텐트 등 난민촌 구조물이 넓게 배치됨.",
        "entities": "헬리콥터와 드론 외에 지상에 다수의 인물(전경의 남성 포함)이 보임.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 인물들을 화면에 임의로 추가함 (Invented people)"
        ],
        "physics": "공중의 기체들은 비행 중이며, 지상의 인물들은 땅에 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 구도에 맞춰 밤하늘의 여백과 헬리콥터를 강조했으며, 샷 텍스트에 없는 인물을 추가하지 않아 요구사항을 잘 충족함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "샷 텍스트에 전혀 언급되지 않은 다수의 인물을 지상에 임의로 배치하여 치명적인 지시 위반이 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "서치라이트 빔이 화면 좌측 하단을 향해 길게 뻗음.",
        "built_space": "화면 하단에 레퍼런스와 일치하는 교회 종탑과 우측에 일부 건축물이 보임.",
        "entities": "지시된 거대한 헬리콥터와 빔을 따라 비행하는 다수의 드론이 존재함.",
        "hard_violations": [],
        "physics": "헬리콥터와 드론들 모두 기체 특성으로 공중에 정상 체공 중임."
       },
       {
        "label": "B",
        "direction": "서치라이트 빔이 수직 아래 지면을 비춤.",
        "built_space": "좌측에 교회 건물, 우측에 텐트 등 난민촌 구조물이 넓게 배치됨.",
        "entities": "헬리콥터와 드론 외에 지상에 다수의 인물(전경의 남성 포함)이 보임.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 인물들을 화면에 임의로 추가함 (Invented people)"
        ],
        "physics": "공중의 기체들은 비행 중이며, 지상의 인물들은 땅에 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "하향 서치라이트는 맞지만, 지시되지 않은 사람들을 추가했고 헬리콥터를 상단에서 잘릴 만큼 확대해 밤하늘의 여백 중심 와이드 구도를 놓쳤습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "상단 헬리콥터의 밑면과 비스듬한 측면, 넓은 밤하늘, 눈부신 하향 광선을 충실히 구현했으며 일부 드론이 광선 밖에 있는 점만 다소 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "헬리콥터 기수는 오른쪽을 향하고, 동체 아래 서치라이트는 오른쪽 아래 정착촌 마당을 비춥니다. 광선이 닿은 지면의 밝은 영역이 보입니다. 드론 네 대는 광선 안팎에 있으나 정지 화면에서 이동 방향은 확인되지 않습니다. 전경과 마당의 사람들은 얼굴이 어두워 정확한 시선 대상을 판별하기 어렵습니다.",
        "built_space": "하단에 벽돌 교회와 콘크리트 종탑 한 개, 종탑 위 십자가 한 개가 보이며 교회 박공에도 십자가가 있습니다. 교회 옆에는 밝은 출입구 한 곳이 보이고, 마당 양쪽에는 천막과 생활 물품이 늘어서 있습니다. 참고의 벽돌과 낡은 콘크리트 재질은 이어지지만, 카메라가 마당과 건물 전면을 크게 포함하여 요구된 하늘 중심 구도와 달라졌습니다. 전경 왼쪽과 오른쪽 가장자리, 마당 중앙에 사람들이 배치되어 있습니다.",
        "entities": "헬리콥터 한 대, 소형 드론 네 대, 어두운 밤하늘, 강한 서치라이트가 있습니다. 헬리콥터의 밑면과 측면은 보이지만 꼬리와 회전날개 일부가 상단에서 잘립니다. 여러 사람이 추가되어 있으며, 얼굴과 복장이 충분히 드러나지 않아 특정 인물이나 연령·민족 정체성을 확인할 수 없습니다. 이 숏에는 애초에 사람이 지시되지 않았습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "숏 텍스트에 없는 사람들을 전경과 정착촌 마당에 여러 명 추가했습니다."
        ],
        "physics": "헬리콥터에는 주회전날개가 있고 드론에도 회전날개 구조가 보여 동력 비행으로 해석할 수 있습니다. 서치라이트는 동체 아래에 부착되어 있으며 광선과 지면의 조명 방향이 연결됩니다. 사람들은 지면에 서 있고 천막과 생활 물품도 지면에 놓여 있습니다. 명백히 지지 없이 떠 있는 물체나 인체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "헬리콥터 기수는 왼쪽 아래, 꼬리는 오른쪽 위를 향합니다. 기체 전방 하부의 서치라이트가 왼쪽 아래 지상 방향으로 뻗으며, 실제 조사 지점은 화면 밖입니다. 이는 아래로 빛을 쏘라는 지시와 맞습니다. 드론 다섯 대가 광선 내부와 주변에 분포하지만 정확한 진행 방향은 확인되지 않습니다.",
        "built_space": "하단에 벽돌 교회 지붕과 콘크리트 종탑 한 개, 그 위 십자가 한 개가 작게 보입니다. 오른쪽 가장자리에는 인접 건물 일부, 교회 뒤에는 나무가 있습니다. 참고의 벽돌 교회와 낡은 종탑이라는 장소 특징을 유지하며, 지상에서 올려다보는 시점에 맞게 건축물이 아래에 놓였습니다. 출입구와 지하실은 프레임 밖이며, 사람이나 반사면은 보이지 않습니다.",
        "entities": "헬리콥터 한 대가 상단 오른쪽에 있으며 밑면과 비스듬한 측면이 보입니다. 주변에는 넓은 어두운 밤하늘이 남아 있고, 눈부신 서치라이트와 소형 드론 다섯 대가 있습니다. 일부 드론은 광선에서 떨어져 있지만 장면의 드론 비행 상태와는 대체로 일치합니다. 지시되지 않은 사람이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "헬리콥터의 회전 중인 주회전날개와 꼬리 회전날개가 보여 비행을 지탱하는 동력 구조가 확인됩니다. 드론들은 회전날개를 갖춘 소형 비행체로 보입니다. 서치라이트의 발광점은 기체에 연결되어 있고 광선은 그 지점에서 아래로 확장됩니다. 건물은 지면에 이어져 있으며 지지 없는 인체나 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "하향 서치라이트는 맞지만, 지시되지 않은 사람들을 추가했고 헬리콥터를 상단에서 잘릴 만큼 확대해 밤하늘의 여백 중심 와이드 구도를 놓쳤습니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "상단 헬리콥터의 밑면과 비스듬한 측면, 넓은 밤하늘, 눈부신 하향 광선을 충실히 구현했으며 일부 드론이 광선 밖에 있는 점만 다소 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "헬리콥터 기수는 오른쪽을 향하고, 동체 아래 서치라이트는 오른쪽 아래 정착촌 마당을 비춥니다. 광선이 닿은 지면의 밝은 영역이 보입니다. 드론 네 대는 광선 안팎에 있으나 정지 화면에서 이동 방향은 확인되지 않습니다. 전경과 마당의 사람들은 얼굴이 어두워 정확한 시선 대상을 판별하기 어렵습니다.",
        "built_space": "하단에 벽돌 교회와 콘크리트 종탑 한 개, 종탑 위 십자가 한 개가 보이며 교회 박공에도 십자가가 있습니다. 교회 옆에는 밝은 출입구 한 곳이 보이고, 마당 양쪽에는 천막과 생활 물품이 늘어서 있습니다. 참고의 벽돌과 낡은 콘크리트 재질은 이어지지만, 카메라가 마당과 건물 전면을 크게 포함하여 요구된 하늘 중심 구도와 달라졌습니다. 전경 왼쪽과 오른쪽 가장자리, 마당 중앙에 사람들이 배치되어 있습니다.",
        "entities": "헬리콥터 한 대, 소형 드론 네 대, 어두운 밤하늘, 강한 서치라이트가 있습니다. 헬리콥터의 밑면과 측면은 보이지만 꼬리와 회전날개 일부가 상단에서 잘립니다. 여러 사람이 추가되어 있으며, 얼굴과 복장이 충분히 드러나지 않아 특정 인물이나 연령·민족 정체성을 확인할 수 없습니다. 이 숏에는 애초에 사람이 지시되지 않았습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "숏 텍스트에 없는 사람들을 전경과 정착촌 마당에 여러 명 추가했습니다."
        ],
        "physics": "헬리콥터에는 주회전날개가 있고 드론에도 회전날개 구조가 보여 동력 비행으로 해석할 수 있습니다. 서치라이트는 동체 아래에 부착되어 있으며 광선과 지면의 조명 방향이 연결됩니다. 사람들은 지면에 서 있고 천막과 생활 물품도 지면에 놓여 있습니다. 명백히 지지 없이 떠 있는 물체나 인체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "헬리콥터 기수는 왼쪽 아래, 꼬리는 오른쪽 위를 향합니다. 기체 전방 하부의 서치라이트가 왼쪽 아래 지상 방향으로 뻗으며, 실제 조사 지점은 화면 밖입니다. 이는 아래로 빛을 쏘라는 지시와 맞습니다. 드론 다섯 대가 광선 내부와 주변에 분포하지만 정확한 진행 방향은 확인되지 않습니다.",
        "built_space": "하단에 벽돌 교회 지붕과 콘크리트 종탑 한 개, 그 위 십자가 한 개가 작게 보입니다. 오른쪽 가장자리에는 인접 건물 일부, 교회 뒤에는 나무가 있습니다. 참고의 벽돌 교회와 낡은 종탑이라는 장소 특징을 유지하며, 지상에서 올려다보는 시점에 맞게 건축물이 아래에 놓였습니다. 출입구와 지하실은 프레임 밖이며, 사람이나 반사면은 보이지 않습니다.",
        "entities": "헬리콥터 한 대가 상단 오른쪽에 있으며 밑면과 비스듬한 측면이 보입니다. 주변에는 넓은 어두운 밤하늘이 남아 있고, 눈부신 서치라이트와 소형 드론 다섯 대가 있습니다. 일부 드론은 광선에서 떨어져 있지만 장면의 드론 비행 상태와는 대체로 일치합니다. 지시되지 않은 사람이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "헬리콥터의 회전 중인 주회전날개와 꼬리 회전날개가 보여 비행을 지탱하는 동력 구조가 확인됩니다. 드론들은 회전날개를 갖춘 소형 비행체로 보입니다. 서치라이트의 발광점은 기체에 연결되어 있고 광선은 그 지점에서 아래로 확장됩니다. 건물은 지면에 이어져 있으며 지지 없는 인체나 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.651
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.401
   },
   "violations": {
    "B": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 인물들을 화면에 임의로 추가함 (Invented people)",
     "[gpt-high] 숏 텍스트에 없는 사람들을 전경과 정착촌 마당에 여러 명 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 401
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 구도에 맞춰 밤하늘의 여백과 헬리콥터를 강조했으며, 샷 텍스트에 없는 인물을 추가하지 않아 요구사항을 잘 충족함."
   },
   {
    "label": "B",
    "score": 401,
    "verdict_ko": "샷 텍스트에 전혀 언급되지 않은 다수의 인물을 지상에 임의로 배치하여 치명적인 지시 위반이 발생함.  ★위반: [gemini-pro] 샷 텍스트에 명시되지 않은 인물들을 화면에 임의로 추가함 (Invented people) / [gpt-high] 숏 텍스트에 없는 사람들을 전경과 정착촌 마당에 여러 명 추가했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L35B02.png",
    "asset_id": "301b39f1-059d-4870-80e8-1f7b6a07f7f5",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdad-222c-78b4-8bf9-c525ce55fc78",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S29sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:31:46.145936+00:00",
  "fingerprint": "f03c32df7b1ca736c30a4a74a57a19f0b3a5aa4efadb0cd3c685d4ad9360f5f2",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S29sh14_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S29sh14_sel.png",
  "source_sha256": "b49953b5cc1c65644cb7882390ac4c3b4c5d98045f1195c3b05f9712eae5b95e",
  "file": "S29sh14_cine.png",
  "staged_sha256": "25df01322c9aa3d807bda2e83e729089e6c849840e65a794ad4dacd7212d6177",
  "latency_ms": 10611
 },
 "S30sh6::signage": {
  "fp": "7808637c39d90a26",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S30sh6": {
  "input_fingerprint": "a6ee477e865d6413",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 빛을 받은 주변의 낡은 랜턴들이 일제히 밝은 불빛을 내며 켜진 환한 전경.\n\nLOCATION (lock): Inside the sparse church basement prayer room, around old lanterns newly shining brightly near a small reed organ and radio. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Surrounding lanterns (Distributed around the basement) — Seen individually from above at differing oblique angles; used as Separated practical objects establishing the spatial reach of 찰리's effect; Small organ (Beside 찰리 after his playing) — Its keyboard side is visible obliquely beneath the elevated camera; used as Midground scale reference and continuity anchor; Old radio (Previously broken and beginning to play) — Its front is angled into the room from the rear-left corner; used as A small secondary focal point anticipating the next camera move.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light radiating from 찰리's chest accompanies the lanterns' simultaneous brightening, opening the basement's tonal range without assigning an unsupported hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its cross, small desk, chairs and harmonium; its lanterns are now brightly lit and the formerly broken radio in the corner has activated. Charlie retains his old coat, hat and worn chest logo, with light emerging from his chest.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 빛을 받은 주변의 낡은 랜턴들이 일제히 밝은 불빛을 내며 켜진 환한 전경.\n\nLOCATION (lock): Inside the sparse church basement prayer room, around old lanterns newly shining brightly near a small reed organ and radio. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Surrounding lanterns (Distributed around the basement) — Seen individually from above at differing oblique angles; used as Separated practical objects establishing the spatial reach of 찰리's effect; Small organ (Beside 찰리 after his playing) — Its keyboard side is visible obliquely beneath the elevated camera; used as Midground scale reference and continuity anchor; Old radio (Previously broken and beginning to play) — Its front is angled into the room from the rear-left corner; used as A small secondary focal point anticipating the next camera move.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light radiating from 찰리's chest accompanies the lanterns' simultaneous brightening, opening the basement's tonal range without assigning an unsupported hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its cross, small desk, chairs and harmonium; its lanterns are now brightly lit and the formerly broken radio in the corner has activated. Charlie retains his old coat, hat and worn chest logo, with light emerging from his chest.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 빛을 받은 주변의 낡은 랜턴들이 일제히 밝은 불빛을 내며 켜진 환한 전경.\n\nLOCATION (lock): Inside the sparse church basement prayer room, around old lanterns newly shining brightly near a small reed organ and radio. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Surrounding lanterns (Distributed around the basement) — Seen individually from above at differing oblique angles; used as Separated practical objects establishing the spatial reach of 찰리's effect; Small organ (Beside 찰리 after his playing) — Its keyboard side is visible obliquely beneath the elevated camera; used as Midground scale reference and continuity anchor; Old radio (Previously broken and beginning to play) — Its front is angled into the room from the rear-left corner; used as A small secondary focal point anticipating the next camera move.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light radiating from 찰리's chest accompanies the lanterns' simultaneous brightening, opening the basement's tonal range without assigning an unsupported hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its cross, small desk, chairs and harmonium; its lanterns are now brightly lit and the formerly broken radio in the corner has activated. Charlie retains his old coat, hat and worn chest logo, with light emerging from his chest.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 위에서 내려다보며, 인물들의 시선은 찰리를 향함. 찰리는 정면 좌측을, 라디오는 방 안쪽을 향함.",
    "built_space": "콘크리트 벽면의 지하실. 뒤쪽 십자가와 책상, 좌측의 라디오가 있음. 중앙 풍금은 악보대가 카메라 쪽에 있고 건반이 역방향으로 배치된 모순된 구조임.",
    "entities": "찰리는 고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입 지시를 정확히 따름. 구도환과 이현우의 외형 특징도 부합함.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 풍금 구조 (악보대가 카메라 쪽 가장자리에 있고 건반이 역방향으로 배치됨)",
     "[gemini-pro] 인물 배치 오류 (연주가 끝난 후 풍금 옆에 있어야 하나 연주 중인 상태로 배치됨)",
     "[gpt-high] 숏 텍스트가 등장시키지 않은 사람 다섯 명을 추가했습니다."
    ],
    "physics": "찰리의 손이 건반에 놓여 있으나 형체가 뭉개짐. 랜턴들은 바닥과 가구 위에 놓여 무게를 지탱함."
   },
   {
    "label": "B",
    "direction": "카메라가 위에서 내려다봄. 찰리는 측면에서 건반을 향해 시선을 두고 연주하며, 사람들은 찰리를 바라봄.",
    "built_space": "돌벽과 나무 기둥이 늘어선 구조. 십자가와 라디오가 있으며, 중앙의 풍금 외에 우측 뒷벽에 피아노가 추가로 배치됨.",
    "entities": "찰리는 낡은 코트를 입었으나 지시된 흰 얼굴판과 기계적 이목구비가 아닌 사실적인 유인원 마스크를 착용함.",
    "hard_violations": [
     "[gemini-pro] 중복/발명된 사물 (공간에 명시되지 않은 두 번째 건반 악기인 피아노가 배경에 추가됨)",
     "[gemini-pro] 인물 배치 오류 (연주가 끝난 후 풍금 옆에 있어야 하나 연주 중인 상태로 배치됨)",
     "[gpt-high] 숏 텍스트가 등장시키지 않은 사람 다섯 명을 추가했습니다.",
     "[gpt-high] 찰리 앞의 오르간 외에 뒤쪽 벽에 별도의 건반악기를 추가했습니다."
    ],
    "physics": "랜턴들이 기둥에 매달려 지탱됨. 왼쪽 인물은 기둥에 상반신이 완벽히 가려져 하반신만 보임. 찰리의 오른손이 건반과 융합됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리의 기계 얼굴 묘사(흰 얼굴판, 점 형태의 눈)는 지시와 일치하나, 풍금의 물리적 구조가 불가능하며 연주 후 옆에 위치해야 한다는 배치를 어겼습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배경에 피아노가 추가되어 공간 구조를 위반했고, 찰리의 얼굴이 기계가 아닌 사실적 유인원이며, 연주 후 위치 지시를 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 위에서 내려다보며, 인물들의 시선은 찰리를 향함. 찰리는 정면 좌측을, 라디오는 방 안쪽을 향함.",
        "built_space": "콘크리트 벽면의 지하실. 뒤쪽 십자가와 책상, 좌측의 라디오가 있음. 중앙 풍금은 악보대가 카메라 쪽에 있고 건반이 역방향으로 배치된 모순된 구조임.",
        "entities": "찰리는 고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입 지시를 정확히 따름. 구도환과 이현우의 외형 특징도 부합함.",
        "hard_violations": [
         "물리적으로 불가능한 풍금 구조 (악보대가 카메라 쪽 가장자리에 있고 건반이 역방향으로 배치됨)",
         "인물 배치 오류 (연주가 끝난 후 풍금 옆에 있어야 하나 연주 중인 상태로 배치됨)"
        ],
        "physics": "찰리의 손이 건반에 놓여 있으나 형체가 뭉개짐. 랜턴들은 바닥과 가구 위에 놓여 무게를 지탱함."
       },
       {
        "label": "B",
        "direction": "카메라가 위에서 내려다봄. 찰리는 측면에서 건반을 향해 시선을 두고 연주하며, 사람들은 찰리를 바라봄.",
        "built_space": "돌벽과 나무 기둥이 늘어선 구조. 십자가와 라디오가 있으며, 중앙의 풍금 외에 우측 뒷벽에 피아노가 추가로 배치됨.",
        "entities": "찰리는 낡은 코트를 입었으나 지시된 흰 얼굴판과 기계적 이목구비가 아닌 사실적인 유인원 마스크를 착용함.",
        "hard_violations": [
         "중복/발명된 사물 (공간에 명시되지 않은 두 번째 건반 악기인 피아노가 배경에 추가됨)",
         "인물 배치 오류 (연주가 끝난 후 풍금 옆에 있어야 하나 연주 중인 상태로 배치됨)"
        ],
        "physics": "랜턴들이 기둥에 매달려 지탱됨. 왼쪽 인물은 기둥에 상반신이 완벽히 가려져 하반신만 보임. 찰리의 오른손이 건반과 융합됨."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리의 기계 얼굴 묘사(흰 얼굴판, 점 형태의 눈)는 지시와 일치하나, 풍금의 물리적 구조가 불가능하며 연주 후 옆에 위치해야 한다는 배치를 어겼습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배경에 피아노가 추가되어 공간 구조를 위반했고, 찰리의 얼굴이 기계가 아닌 사실적 유인원이며, 연주 후 위치 지시를 어겼습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 위에서 내려다보며, 인물들의 시선은 찰리를 향함. 찰리는 정면 좌측을, 라디오는 방 안쪽을 향함.",
        "built_space": "콘크리트 벽면의 지하실. 뒤쪽 십자가와 책상, 좌측의 라디오가 있음. 중앙 풍금은 악보대가 카메라 쪽에 있고 건반이 역방향으로 배치된 모순된 구조임.",
        "entities": "찰리는 고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입 지시를 정확히 따름. 구도환과 이현우의 외형 특징도 부합함.",
        "hard_violations": [
         "물리적으로 불가능한 풍금 구조 (악보대가 카메라 쪽 가장자리에 있고 건반이 역방향으로 배치됨)",
         "인물 배치 오류 (연주가 끝난 후 풍금 옆에 있어야 하나 연주 중인 상태로 배치됨)"
        ],
        "physics": "찰리의 손이 건반에 놓여 있으나 형체가 뭉개짐. 랜턴들은 바닥과 가구 위에 놓여 무게를 지탱함."
       },
       {
        "label": "B",
        "direction": "카메라가 위에서 내려다봄. 찰리는 측면에서 건반을 향해 시선을 두고 연주하며, 사람들은 찰리를 바라봄.",
        "built_space": "돌벽과 나무 기둥이 늘어선 구조. 십자가와 라디오가 있으며, 중앙의 풍금 외에 우측 뒷벽에 피아노가 추가로 배치됨.",
        "entities": "찰리는 낡은 코트를 입었으나 지시된 흰 얼굴판과 기계적 이목구비가 아닌 사실적인 유인원 마스크를 착용함.",
        "hard_violations": [
         "중복/발명된 사물 (공간에 명시되지 않은 두 번째 건반 악기인 피아노가 배경에 추가됨)",
         "인물 배치 오류 (연주가 끝난 후 풍금 옆에 있어야 하나 연주 중인 상태로 배치됨)"
        ],
        "physics": "랜턴들이 기둥에 매달려 지탱됨. 왼쪽 인물은 기둥에 상반신이 완벽히 가려져 하반신만 보임. 찰리의 오른손이 건반과 융합됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "높은 시점의 환한 전경은 구현했지만, 불필요한 인물들과 추가 건반악기를 넣었고 석조·목조 공간이 참조의 낡은 미장벽 지하 공간과 크게 다릅니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "추가 인물들 때문에 부적격이지만, 낡은 미장벽과 분산된 랜턴, 후방 왼쪽 라디오, 찰리의 흉부 발광과 얼굴 형태는 A보다 요구에 가깝습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 화면 오른쪽을 향하고 두 손은 자기 앞 건반에 놓여 있습니다. 오른쪽 청년과 뒤쪽에 앉은 남성은 찰리 쪽을 바라봅니다. 좌측 후방 라디오의 앞면은 방 안쪽으로 비스듬히 향하며, 높은 카메라에서 오르간 건반 면이 보이는 방향은 자연스럽습니다. 흉부와 주변 랜턴이 함께 빛나지만, 손의 위치는 연주를 마친 순간보다 아직 연주 중인 모습에 가깝습니다.",
        "built_space": "큰 십자가 하나, 작은 책상 하나, 전경 오르간 하나, 뒤쪽 벽의 별도 건반악기 하나, 좌측 후방 라디오 하나가 보입니다. 랜턴은 약 12개로 대부분 목재 기둥이나 벽에 걸려 있습니다. 여러 의자에 사람들이 앉고 중앙과 오른쪽에는 사람들이 서 있습니다. 카메라는 실내를 비스듬히 내려다보지만, 두드러진 석재 벽과 굵은 목재 구조는 참조에서 보이는 미장벽 및 금속 난간의 장소감을 유지하지 못합니다.",
        "entities": "찰리 외에 사람이 다섯 명 보이며, 이는 이 숏에서 요구하지 않은 군중입니다. 오른쪽에는 검은 헝클어진 머리의 젊은 동아시아계 남성, 뒤에는 중년 동아시아계 남성이 보이지만 나머지는 가림 때문에 나이와 얼굴을 확정하기 어렵습니다. 찰리는 낡은 코트와 모자, 흰 고릴라형 기계 얼굴을 갖췄으나 눈은 요청한 점 형태보다 복잡한 눈구멍 형태입니다. 흉부 발광은 보이지만 닳은 로고는 뚜렷하지 않습니다. 오래된 랜턴과 라디오, 십자가, 책상은 식별되며 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "숏 텍스트가 등장시키지 않은 사람 다섯 명을 추가했습니다.",
         "찰리 앞의 오르간 외에 뒤쪽 벽에 별도의 건반악기를 추가했습니다."
        ],
        "physics": "찰리는 양발로 바닥을 딛고 손을 건반에 얹고 있습니다. 앉은 사람들은 의자에, 서 있는 사람들은 바닥에 지지됩니다. 랜턴은 벽이나 기둥의 걸이로 지지되며, 라디오와 악기 및 책상은 가구나 바닥 위에 놓여 있습니다. 지지 없이 떠 있는 물체나 명백히 불가능한 신체 자세는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "찰리는 화면 오른쪽 앞을 향하며 두 손을 자신 쪽으로 향한 건반에 놓고 있습니다. 주변 사람들은 대체로 찰리를 바라봅니다. 후방 왼쪽 라디오의 앞면은 방 안쪽으로 향하고, 높은 카메라에서 건반과 흩어진 랜턴들을 서로 다른 사선 각도로 내려다볼 수 있습니다. 흉부의 강한 빛과 랜턴의 점등은 함께 보이지만, 연주가 끝났다는 행동상의 구분은 약합니다.",
        "built_space": "벽 십자가 하나, 중앙 뒤쪽의 작은 책상 하나, 찰리 앞 오르간 하나, 후방 왼쪽 라디오 하나가 보입니다. 랜턴은 여덟 개로 바닥 세 곳과 탁자·책상·오르간·의자 및 뒤쪽 가구에 분산되어 있습니다. 접이식 의자 여러 개와 오르간용 벤치가 있으며, 사람들은 악기 주변과 뒤쪽 바닥에 서 있습니다. 얼룩진 미장벽은 참조의 재질에 더 가깝지만 참조의 계단과 난간은 확인되지 않아 정확히 같은 장소인지는 입증되지 않습니다. 높은 실내 와이드 시점과 오르간 건반의 가시성은 요구에 맞습니다.",
        "entities": "찰리 외에 사람 다섯 명이 추가되어 있습니다. 오른쪽 청년은 검은 헝클어진 머리의 젊은 동아시아계 남성으로 보이고, 뒤쪽 짧은 머리 남성은 중년으로 보입니다. 나머지 세 사람은 요청되지 않은 인물이며 일부는 뒷모습이라 세부 정체성을 판단하기 어렵습니다. 찰리는 모자와 낡은 코트, 닳은 원형 가슴 표식, 흰 고릴라형 얼굴판과 점눈·선입을 갖췄습니다. 다만 머리 외장은 요청보다 둥글고 단순합니다. 라디오에는 켜진 표시부가 보이며 낡은 랜턴들도 빛납니다. 배경 종이와 표식에 판독 가능한 글자는 확인되지 않습니다.",
        "hard_violations": [
         "숏 텍스트가 등장시키지 않은 사람 다섯 명을 추가했습니다."
        ],
        "physics": "사람들은 바닥에 서 있고, 찰리는 오르간 앞에 서서 양손을 건반에 접촉시키고 있습니다. 찰리의 하체 일부는 악기에 가려져 있지만 공중에 떠 있다고 볼 근거는 없습니다. 랜턴들은 바닥이나 가구의 수평면에 놓여 있고 라디오는 받침 가구 위에 있습니다. 지지 없는 물체나 불가능한 자세는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "높은 시점의 환한 전경은 구현했지만, 불필요한 인물들과 추가 건반악기를 넣었고 석조·목조 공간이 참조의 낡은 미장벽 지하 공간과 크게 다릅니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "추가 인물들 때문에 부적격이지만, 낡은 미장벽과 분산된 랜턴, 후방 왼쪽 라디오, 찰리의 흉부 발광과 얼굴 형태는 A보다 요구에 가깝습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 화면 오른쪽을 향하고 두 손은 자기 앞 건반에 놓여 있습니다. 오른쪽 청년과 뒤쪽에 앉은 남성은 찰리 쪽을 바라봅니다. 좌측 후방 라디오의 앞면은 방 안쪽으로 비스듬히 향하며, 높은 카메라에서 오르간 건반 면이 보이는 방향은 자연스럽습니다. 흉부와 주변 랜턴이 함께 빛나지만, 손의 위치는 연주를 마친 순간보다 아직 연주 중인 모습에 가깝습니다.",
        "built_space": "큰 십자가 하나, 작은 책상 하나, 전경 오르간 하나, 뒤쪽 벽의 별도 건반악기 하나, 좌측 후방 라디오 하나가 보입니다. 랜턴은 약 12개로 대부분 목재 기둥이나 벽에 걸려 있습니다. 여러 의자에 사람들이 앉고 중앙과 오른쪽에는 사람들이 서 있습니다. 카메라는 실내를 비스듬히 내려다보지만, 두드러진 석재 벽과 굵은 목재 구조는 참조에서 보이는 미장벽 및 금속 난간의 장소감을 유지하지 못합니다.",
        "entities": "찰리 외에 사람이 다섯 명 보이며, 이는 이 숏에서 요구하지 않은 군중입니다. 오른쪽에는 검은 헝클어진 머리의 젊은 동아시아계 남성, 뒤에는 중년 동아시아계 남성이 보이지만 나머지는 가림 때문에 나이와 얼굴을 확정하기 어렵습니다. 찰리는 낡은 코트와 모자, 흰 고릴라형 기계 얼굴을 갖췄으나 눈은 요청한 점 형태보다 복잡한 눈구멍 형태입니다. 흉부 발광은 보이지만 닳은 로고는 뚜렷하지 않습니다. 오래된 랜턴과 라디오, 십자가, 책상은 식별되며 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "숏 텍스트가 등장시키지 않은 사람 다섯 명을 추가했습니다.",
         "찰리 앞의 오르간 외에 뒤쪽 벽에 별도의 건반악기를 추가했습니다."
        ],
        "physics": "찰리는 양발로 바닥을 딛고 손을 건반에 얹고 있습니다. 앉은 사람들은 의자에, 서 있는 사람들은 바닥에 지지됩니다. 랜턴은 벽이나 기둥의 걸이로 지지되며, 라디오와 악기 및 책상은 가구나 바닥 위에 놓여 있습니다. 지지 없이 떠 있는 물체나 명백히 불가능한 신체 자세는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "찰리는 화면 오른쪽 앞을 향하며 두 손을 자신 쪽으로 향한 건반에 놓고 있습니다. 주변 사람들은 대체로 찰리를 바라봅니다. 후방 왼쪽 라디오의 앞면은 방 안쪽으로 향하고, 높은 카메라에서 건반과 흩어진 랜턴들을 서로 다른 사선 각도로 내려다볼 수 있습니다. 흉부의 강한 빛과 랜턴의 점등은 함께 보이지만, 연주가 끝났다는 행동상의 구분은 약합니다.",
        "built_space": "벽 십자가 하나, 중앙 뒤쪽의 작은 책상 하나, 찰리 앞 오르간 하나, 후방 왼쪽 라디오 하나가 보입니다. 랜턴은 여덟 개로 바닥 세 곳과 탁자·책상·오르간·의자 및 뒤쪽 가구에 분산되어 있습니다. 접이식 의자 여러 개와 오르간용 벤치가 있으며, 사람들은 악기 주변과 뒤쪽 바닥에 서 있습니다. 얼룩진 미장벽은 참조의 재질에 더 가깝지만 참조의 계단과 난간은 확인되지 않아 정확히 같은 장소인지는 입증되지 않습니다. 높은 실내 와이드 시점과 오르간 건반의 가시성은 요구에 맞습니다.",
        "entities": "찰리 외에 사람 다섯 명이 추가되어 있습니다. 오른쪽 청년은 검은 헝클어진 머리의 젊은 동아시아계 남성으로 보이고, 뒤쪽 짧은 머리 남성은 중년으로 보입니다. 나머지 세 사람은 요청되지 않은 인물이며 일부는 뒷모습이라 세부 정체성을 판단하기 어렵습니다. 찰리는 모자와 낡은 코트, 닳은 원형 가슴 표식, 흰 고릴라형 얼굴판과 점눈·선입을 갖췄습니다. 다만 머리 외장은 요청보다 둥글고 단순합니다. 라디오에는 켜진 표시부가 보이며 낡은 랜턴들도 빛납니다. 배경 종이와 표식에 판독 가능한 글자는 확인되지 않습니다.",
        "hard_violations": [
         "숏 텍스트가 등장시키지 않은 사람 다섯 명을 추가했습니다."
        ],
        "physics": "사람들은 바닥에 서 있고, 찰리는 오르간 앞에 서서 양손을 건반에 접촉시키고 있습니다. 찰리의 하체 일부는 악기에 가려져 있지만 공중에 떠 있다고 볼 근거는 없습니다. 랜턴들은 바닥이나 가구의 수평면에 놓여 있고 라디오는 받침 가구 위에 있습니다. 지지 없는 물체나 불가능한 자세는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.25
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 풍금 구조 (악보대가 카메라 쪽 가장자리에 있고 건반이 역방향으로 배치됨)",
     "[gemini-pro] 인물 배치 오류 (연주가 끝난 후 풍금 옆에 있어야 하나 연주 중인 상태로 배치됨)",
     "[gpt-high] 숏 텍스트가 등장시키지 않은 사람 다섯 명을 추가했습니다."
    ],
    "B": [
     "[gemini-pro] 중복/발명된 사물 (공간에 명시되지 않은 두 번째 건반 악기인 피아노가 배경에 추가됨)",
     "[gemini-pro] 인물 배치 오류 (연주가 끝난 후 풍금 옆에 있어야 하나 연주 중인 상태로 배치됨)",
     "[gpt-high] 숏 텍스트가 등장시키지 않은 사람 다섯 명을 추가했습니다.",
     "[gpt-high] 찰리 앞의 오르간 외에 뒤쪽 벽에 별도의 건반악기를 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "찰리의 기계 얼굴 묘사(흰 얼굴판, 점 형태의 눈)는 지시와 일치하나, 풍금의 물리적 구조가 불가능하며 연주 후 옆에 위치해야 한다는 배치를 어겼습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 풍금 구조 (악보대가 카메라 쪽 가장자리에 있고 건반이 역방향으로 배치됨) / [gemini-pro] 인물 배치 오류 (연주가 끝난 후 풍금 옆에 있어야 하나 연주 중인 상태로 배치됨) / [gpt-high] 숏 텍스트가 등장시키지 않은 사람 다섯 명을 추가했습니다."
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "배경에 피아노가 추가되어 공간 구조를 위반했고, 찰리의 얼굴이 기계가 아닌 사실적 유인원이며, 연주 후 위치 지시를 어겼습니다.  ★위반: [gemini-pro] 중복/발명된 사물 (공간에 명시되지 않은 두 번째 건반 악기인 피아노가 배경에 추가됨) / [gemini-pro] 인물 배치 오류 (연주가 끝난 후 풍금 옆에 있어야 하나 연주 중인 상태로 배치됨) / [gpt-high] 숏 텍스트가 등장시키지 않은 사람 다섯 명을 추가했습니다. / [gpt-high] 찰리 앞의 오르간 외에 뒤쪽 벽에 별도의 건반악기를 추가했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S29sh8_sel.png",
    "asset_id": "b2e1d5c1-a2ca-4bc6-831f-7a24be935118",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdad-d119-7945-98f2-89202bad70a0",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S29sh8"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S30sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:31:57.078380+00:00",
  "fingerprint": "7e43cb1460e68e3f2d7f88472c5ad2562ece880a6253a12fff885ffb060e95e6",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S30sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S30sh6_sel.png",
  "source_sha256": "901e30db7904eb7a8395de5f421fe1bcd42ac8aef8c81c7f4ef0f32d0f2109db",
  "file": "S30sh6_cine.png",
  "staged_sha256": "536c269f4e0a15183b92cd0179159596879f6d8dabc6eb3a62b83be90f3b9e95",
  "latency_ms": 11540
 },
 "S30sh13::signage": {
  "fp": "3c156b3cbc1e024f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S30sh13": {
  "input_fingerprint": "6a80bb2b02df1431",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 구도환이 찰리의 거대한 금속 몸통을 두 팔로 와락 끌어안은 구도.\n\nLOCATION (lock): In the open floor area of the church basement prayer room near the stairway, illuminated by the brightened lanterns. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Small organ (Remaining near 찰리) — A partial side remains visible behind the pair; used as Soft continuity reference that does not compete with the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established lantern illumination reveals the human embrace against 찰리's metal body with gentle facial contrast and no new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the illuminated old lanterns, sparse furnishings, small organ, and basement finishes from the reference. EXCLUDE the transient flare associated with the lamps first switching on.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement lanterns remain lit and the old radio remains activated, with the cross, desk, chairs and harmonium unchanged. Charlie retains his old coat and hat, blue-lit eyes and worn chest logo. 구도환: He has entered the basement and stands with both arms extended in an embrace.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 구도환 right now, so 구도환's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 구도환: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 구도환이 찰리의 거대한 금속 몸통을 두 팔로 와락 끌어안은 구도.\n\nLOCATION (lock): In the open floor area of the church basement prayer room near the stairway, illuminated by the brightened lanterns. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Small organ (Remaining near 찰리) — A partial side remains visible behind the pair; used as Soft continuity reference that does not compete with the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established lantern illumination reveals the human embrace against 찰리's metal body with gentle facial contrast and no new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the illuminated old lanterns, sparse furnishings, small organ, and basement finishes from the reference. EXCLUDE the transient flare associated with the lamps first switching on.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement lanterns remain lit and the old radio remains activated, with the cross, desk, chairs and harmonium unchanged. Charlie retains his old coat and hat, blue-lit eyes and worn chest logo. 구도환: He has entered the basement and stands with both arms extended in an embrace.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 구도환 right now, so 구도환's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 구도환: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 구도환이 찰리의 거대한 금속 몸통을 두 팔로 와락 끌어안은 구도.\n\nLOCATION (lock): In the open floor area of the church basement prayer room near the stairway, illuminated by the brightened lanterns. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Small organ (Remaining near 찰리) — A partial side remains visible behind the pair; used as Soft continuity reference that does not compete with the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established lantern illumination reveals the human embrace against 찰리's metal body with gentle facial contrast and no new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the illuminated old lanterns, sparse furnishings, small organ, and basement finishes from the reference. EXCLUDE the transient flare associated with the lamps first switching on.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement lanterns remain lit and the old radio remains activated, with the cross, desk, chairs and harmonium unchanged. Charlie retains his old coat and hat, blue-lit eyes and worn chest logo. 구도환: He has entered the basement and stands with both arms extended in an embrace.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 구도환 right now, so 구도환's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 구도환: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "구도환은 찰리의 어깨 쪽에 머리를 기대고 허공을 향해 시선을 두며, 찰리는 무표정하게 정면을 응시함.",
    "built_space": "교회 지하실. 벽면에 십자가, 책상, 랜턴이 보임. 인물 뒤에 있어야 할 작은 오르간 대신 책상 형태의 다른 가구가 놓여 있음.",
    "entities": "구도환은 레퍼런스의 트렌치코트를 입었으나 얼굴이 다름. 찰리는 텍스트에 명시된 '하얀 얼굴판'과 점 형태의 눈을 가졌으며, 코트와 모자를 착용함.",
    "hard_violations": [
     "[gemini-pro] 찰리의 왼쪽 기계 손이 쥐고 있는, 맥락 없이 잘려나간 형태의 피아노 건반 뭉치 (존재할 수 없는 왜곡된 사물 생성)"
    ],
    "physics": "찰리가 손에 들고 있는 건반 뭉치는 원래 형태에서 뜯겨져 나온 물리적으로 불가능한 물체임. 구도환의 포옹 자세는 찰리의 몸통에 밀착되어 있음."
   },
   {
    "label": "B",
    "direction": "구도환은 찰리의 가슴에 얼굴을 완전히 파묻고 있으며, 찰리는 살짝 아래쪽을 향해 시선을 둠.",
    "built_space": "교회 지하실. 십자가와 배경 가구들이 보임. 프롬프트상 인물 뒤에 배치되어야 할 오르간이 화면 우측 전면을 크게 차지하며 인물들을 가리고 있음.",
    "entities": "구도환은 레퍼런스(코트 차림의 노인)가 아닌 이전 샷의 엑스트라(초록색 재킷 남성)로 잘못 그려짐. 찰리는 텍스트에 명시된 '하얀 얼굴판'을 완전히 무시하고 어두운 금속 얼굴과 푸른 빛의 눈동자를 가짐.",
    "hard_violations": [],
    "physics": "두 인물은 바닥에 안정적으로 서서 포옹하고 있으며, 전면의 오르간이나 랜턴 등 사물들의 중력과 접촉 상태는 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "치명적인 형태 왜곡은 피했으나, 텍스트에 명시된 찰리의 핵심 특징(하얀 얼굴판)을 무시하고 오르간을 인물 전면에 배치하여 프레이밍 지시를 완전히 위반했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트가 요구한 찰리의 하얀 얼굴판은 올바르게 구현했으나, 찰리의 손에 들린 정체불명의 잘려나간 건반 조각이라는 치명적인 환각 현상 때문에 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "구도환은 찰리의 어깨 쪽에 머리를 기대고 허공을 향해 시선을 두며, 찰리는 무표정하게 정면을 응시함.",
        "built_space": "교회 지하실. 벽면에 십자가, 책상, 랜턴이 보임. 인물 뒤에 있어야 할 작은 오르간 대신 책상 형태의 다른 가구가 놓여 있음.",
        "entities": "구도환은 레퍼런스의 트렌치코트를 입었으나 얼굴이 다름. 찰리는 텍스트에 명시된 '하얀 얼굴판'과 점 형태의 눈을 가졌으며, 코트와 모자를 착용함.",
        "hard_violations": [
         "찰리의 왼쪽 기계 손이 쥐고 있는, 맥락 없이 잘려나간 형태의 피아노 건반 뭉치 (존재할 수 없는 왜곡된 사물 생성)"
        ],
        "physics": "찰리가 손에 들고 있는 건반 뭉치는 원래 형태에서 뜯겨져 나온 물리적으로 불가능한 물체임. 구도환의 포옹 자세는 찰리의 몸통에 밀착되어 있음."
       },
       {
        "label": "B",
        "direction": "구도환은 찰리의 가슴에 얼굴을 완전히 파묻고 있으며, 찰리는 살짝 아래쪽을 향해 시선을 둠.",
        "built_space": "교회 지하실. 십자가와 배경 가구들이 보임. 프롬프트상 인물 뒤에 배치되어야 할 오르간이 화면 우측 전면을 크게 차지하며 인물들을 가리고 있음.",
        "entities": "구도환은 레퍼런스(코트 차림의 노인)가 아닌 이전 샷의 엑스트라(초록색 재킷 남성)로 잘못 그려짐. 찰리는 텍스트에 명시된 '하얀 얼굴판'을 완전히 무시하고 어두운 금속 얼굴과 푸른 빛의 눈동자를 가짐.",
        "hard_violations": [],
        "physics": "두 인물은 바닥에 안정적으로 서서 포옹하고 있으며, 전면의 오르간이나 랜턴 등 사물들의 중력과 접촉 상태는 자연스러움."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "치명적인 형태 왜곡은 피했으나, 텍스트에 명시된 찰리의 핵심 특징(하얀 얼굴판)을 무시하고 오르간을 인물 전면에 배치하여 프레이밍 지시를 완전히 위반했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트가 요구한 찰리의 하얀 얼굴판은 올바르게 구현했으나, 찰리의 손에 들린 정체불명의 잘려나간 건반 조각이라는 치명적인 환각 현상 때문에 실격입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "구도환은 찰리의 어깨 쪽에 머리를 기대고 허공을 향해 시선을 두며, 찰리는 무표정하게 정면을 응시함.",
        "built_space": "교회 지하실. 벽면에 십자가, 책상, 랜턴이 보임. 인물 뒤에 있어야 할 작은 오르간 대신 책상 형태의 다른 가구가 놓여 있음.",
        "entities": "구도환은 레퍼런스의 트렌치코트를 입었으나 얼굴이 다름. 찰리는 텍스트에 명시된 '하얀 얼굴판'과 점 형태의 눈을 가졌으며, 코트와 모자를 착용함.",
        "hard_violations": [
         "찰리의 왼쪽 기계 손이 쥐고 있는, 맥락 없이 잘려나간 형태의 피아노 건반 뭉치 (존재할 수 없는 왜곡된 사물 생성)"
        ],
        "physics": "찰리가 손에 들고 있는 건반 뭉치는 원래 형태에서 뜯겨져 나온 물리적으로 불가능한 물체임. 구도환의 포옹 자세는 찰리의 몸통에 밀착되어 있음."
       },
       {
        "label": "B",
        "direction": "구도환은 찰리의 가슴에 얼굴을 완전히 파묻고 있으며, 찰리는 살짝 아래쪽을 향해 시선을 둠.",
        "built_space": "교회 지하실. 십자가와 배경 가구들이 보임. 프롬프트상 인물 뒤에 배치되어야 할 오르간이 화면 우측 전면을 크게 차지하며 인물들을 가리고 있음.",
        "entities": "구도환은 레퍼런스(코트 차림의 노인)가 아닌 이전 샷의 엑스트라(초록색 재킷 남성)로 잘못 그려짐. 찰리는 텍스트에 명시된 '하얀 얼굴판'을 완전히 무시하고 어두운 금속 얼굴과 푸른 빛의 눈동자를 가짐.",
        "hard_violations": [],
        "physics": "두 인물은 바닥에 안정적으로 서서 포옹하고 있으며, 전면의 오르간이나 랜턴 등 사물들의 중력과 접촉 상태는 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "포옹은 전달되지만 무릎 부근까지 넓어진 구도와 전경의 큰 오르간이 지정된 미디엄 숏을 벗어나며, 찰리의 흰 얼굴판과 구도환의 인물 참조도 맞지 않는다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "상반신 중심의 포옹과 두 손의 접촉, 찰리의 흰 얼굴판을 더 정확히 구현했지만, 구도환의 참조 정체성과 오르간의 배치·방향은 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "구도환은 눈을 감고 얼굴을 찰리의 가슴 쪽에 붙인다. 보이는 팔은 찰리의 몸통 앞을 가로질러 반대쪽 어깨를 감싸며 손바닥이 금속 어깨에 닿는다. 다른 팔과 손의 접촉은 몸 뒤에 가려 확인하기 어렵다. 찰리의 얼굴은 약간 왼쪽 아래를 향해 포옹하는 남자 쪽으로 돌아가 있다.",
        "built_space": "왼쪽 벽에 십자가 하나, 뒤쪽에 책상 하나와 의자 하나, 오른쪽에 접이식 의자 하나, 오른쪽 전경에 오르간 하나와 벤치 일부가 보인다. 랜턴은 책상 위 하나, 바닥 두 개, 뒤쪽 작은 탁자 위 하나, 오르간 위 하나로 다섯 개가 보인다. 왼쪽에는 계단 난간 일부가 있다. 낡은 벽과 기둥은 장소의 재질을 이어가지만, 오르간이 두 사람 뒤의 작은 측면이 아니라 큰 전경 요소가 되었다. 인물도 무릎 부근까지 보여 지정된 미디엄 숏보다 넓다. 상단의 밝은 천장 등기구는 랜턴만으로 밝히라는 조건과 맞지 않는다.",
        "entities": "인물은 남자 한 명과 기계 찰리뿐이다. 찰리의 모자, 파란 눈, 마모된 금속 몸통은 맞지만, 얼굴은 흰 얼굴판과 점 눈·선 입 대신 금속 고릴라 얼굴이며 코트는 주로 뒤쪽으로 젖혀져 있다. 구도환은 검은 짧은 머리의 중년 동아시아계 남성처럼 보인다. 참조의 희끗한 수염, 길고 좁은 얼굴, 모자와 베이지색 트렌치코트 대신 이전 장면 인물과 비슷한 얼굴과 짧은 재킷을 사용했다. 라디오는 화면에서 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 인물의 하체는 화면 아래로 이어지고 몸통은 선 자세로 정렬되어 있어 공중에 떠 있다는 징후가 없다. 구도환은 찰리에게 상체를 기대고 손바닥으로 어깨를 감싸며, 보이는 팔의 굽힘과 접촉은 가능한 포옹 자세다. 발은 잘려 직접적인 바닥 접촉은 확인할 수 없다. 랜턴과 가구는 각각 바닥이나 상판에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "구도환은 찰리의 어깨에 뺨을 대고 시선을 아래로 둔다. 앞쪽 팔은 몸통을 가로질러 가슴 금속판을 누르고, 다른 팔은 뒤를 돌아 반대쪽 어깨에 손을 얹어 두 팔로 안는 관계가 분명하다. 찰리는 화면 오른쪽을 바라보며 카메라를 응시하지 않는다.",
        "built_space": "뒤 벽의 십자가 하나, 왼쪽 라디오 하나와 받침장 하나, 랜턴용 탁자 하나, 오른쪽 오르간 하나, 양쪽 가장자리의 의자 일부 두 개가 보인다. 랜턴은 왼쪽 탁자, 중앙 바닥, 오른쪽 오르간 위에 하나씩 총 세 개다. 낡은 투톤 벽과 오른쪽 기둥은 이전 장소와 유사하다. 두 인물의 상반신을 크게 잡아 A보다 미디엄 숏에 가깝다. 다만 오르간 건반부가 찰리의 오른팔 바로 밑에서 가파르게 기울어 보여 이전 장면의 수평 건반 배치를 잇지 못하며, 조용한 배경 측면으로도 읽히지 않는다. 상단에는 밝게 빛나는 천장 등기구가 보인다.",
        "entities": "남자 한 명과 찰리만 등장한다. 찰리는 각진 기계 머리, 흰 얼굴판, 작은 파란 눈과 선 모양 입, 낡은 모자와 코트, 마모된 금속 몸통을 갖춰 A보다 명시된 외형에 가깝다. 코트의 가슴 표식은 일부만 보여 판독되지 않는다. 구도환은 중년 남성이지만 검은 짧은 머리와 동아시아계로 보이는 얼굴, 낡은 짙은 외투를 사용해 참조의 얼굴·희끗한 수염·모자·베이지색 트렌치코트와 다르다. 오래된 라디오는 보이나 작동 상태는 확정하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "구도환의 두 손은 각각 찰리의 가슴과 반대쪽 어깨에 실제로 접촉하고, 팔은 몸통을 감싸는 경로를 이룬다. 두 인물은 선 자세로 화면 아래까지 이어지며 떠 있는 징후는 없다. 발은 프레임 밖이다. 기울어진 건반부에는 찰리의 금속 손이 걸쳐 있고 뒤쪽 목제 본체와 겹쳐 있어 무지지 부유라고 단정할 수는 없지만, 정상적으로 설치된 오르간의 건반 방향은 아니다. 랜턴은 탁자·바닥·목제 상판에 받쳐져 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "포옹은 전달되지만 무릎 부근까지 넓어진 구도와 전경의 큰 오르간이 지정된 미디엄 숏을 벗어나며, 찰리의 흰 얼굴판과 구도환의 인물 참조도 맞지 않는다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "상반신 중심의 포옹과 두 손의 접촉, 찰리의 흰 얼굴판을 더 정확히 구현했지만, 구도환의 참조 정체성과 오르간의 배치·방향은 어긋난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "구도환은 눈을 감고 얼굴을 찰리의 가슴 쪽에 붙인다. 보이는 팔은 찰리의 몸통 앞을 가로질러 반대쪽 어깨를 감싸며 손바닥이 금속 어깨에 닿는다. 다른 팔과 손의 접촉은 몸 뒤에 가려 확인하기 어렵다. 찰리의 얼굴은 약간 왼쪽 아래를 향해 포옹하는 남자 쪽으로 돌아가 있다.",
        "built_space": "왼쪽 벽에 십자가 하나, 뒤쪽에 책상 하나와 의자 하나, 오른쪽에 접이식 의자 하나, 오른쪽 전경에 오르간 하나와 벤치 일부가 보인다. 랜턴은 책상 위 하나, 바닥 두 개, 뒤쪽 작은 탁자 위 하나, 오르간 위 하나로 다섯 개가 보인다. 왼쪽에는 계단 난간 일부가 있다. 낡은 벽과 기둥은 장소의 재질을 이어가지만, 오르간이 두 사람 뒤의 작은 측면이 아니라 큰 전경 요소가 되었다. 인물도 무릎 부근까지 보여 지정된 미디엄 숏보다 넓다. 상단의 밝은 천장 등기구는 랜턴만으로 밝히라는 조건과 맞지 않는다.",
        "entities": "인물은 남자 한 명과 기계 찰리뿐이다. 찰리의 모자, 파란 눈, 마모된 금속 몸통은 맞지만, 얼굴은 흰 얼굴판과 점 눈·선 입 대신 금속 고릴라 얼굴이며 코트는 주로 뒤쪽으로 젖혀져 있다. 구도환은 검은 짧은 머리의 중년 동아시아계 남성처럼 보인다. 참조의 희끗한 수염, 길고 좁은 얼굴, 모자와 베이지색 트렌치코트 대신 이전 장면 인물과 비슷한 얼굴과 짧은 재킷을 사용했다. 라디오는 화면에서 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 인물의 하체는 화면 아래로 이어지고 몸통은 선 자세로 정렬되어 있어 공중에 떠 있다는 징후가 없다. 구도환은 찰리에게 상체를 기대고 손바닥으로 어깨를 감싸며, 보이는 팔의 굽힘과 접촉은 가능한 포옹 자세다. 발은 잘려 직접적인 바닥 접촉은 확인할 수 없다. 랜턴과 가구는 각각 바닥이나 상판에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "구도환은 찰리의 어깨에 뺨을 대고 시선을 아래로 둔다. 앞쪽 팔은 몸통을 가로질러 가슴 금속판을 누르고, 다른 팔은 뒤를 돌아 반대쪽 어깨에 손을 얹어 두 팔로 안는 관계가 분명하다. 찰리는 화면 오른쪽을 바라보며 카메라를 응시하지 않는다.",
        "built_space": "뒤 벽의 십자가 하나, 왼쪽 라디오 하나와 받침장 하나, 랜턴용 탁자 하나, 오른쪽 오르간 하나, 양쪽 가장자리의 의자 일부 두 개가 보인다. 랜턴은 왼쪽 탁자, 중앙 바닥, 오른쪽 오르간 위에 하나씩 총 세 개다. 낡은 투톤 벽과 오른쪽 기둥은 이전 장소와 유사하다. 두 인물의 상반신을 크게 잡아 A보다 미디엄 숏에 가깝다. 다만 오르간 건반부가 찰리의 오른팔 바로 밑에서 가파르게 기울어 보여 이전 장면의 수평 건반 배치를 잇지 못하며, 조용한 배경 측면으로도 읽히지 않는다. 상단에는 밝게 빛나는 천장 등기구가 보인다.",
        "entities": "남자 한 명과 찰리만 등장한다. 찰리는 각진 기계 머리, 흰 얼굴판, 작은 파란 눈과 선 모양 입, 낡은 모자와 코트, 마모된 금속 몸통을 갖춰 A보다 명시된 외형에 가깝다. 코트의 가슴 표식은 일부만 보여 판독되지 않는다. 구도환은 중년 남성이지만 검은 짧은 머리와 동아시아계로 보이는 얼굴, 낡은 짙은 외투를 사용해 참조의 얼굴·희끗한 수염·모자·베이지색 트렌치코트와 다르다. 오래된 라디오는 보이나 작동 상태는 확정하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "구도환의 두 손은 각각 찰리의 가슴과 반대쪽 어깨에 실제로 접촉하고, 팔은 몸통을 감싸는 경로를 이룬다. 두 인물은 선 자세로 화면 아래까지 이어지며 떠 있는 징후는 없다. 발은 프레임 밖이다. 기울어진 건반부에는 찰리의 금속 손이 걸쳐 있고 뒤쪽 목제 본체와 겹쳐 있어 무지지 부유라고 단정할 수는 없지만, 정상적으로 설치된 오르간의 건반 방향은 아니다. 랜턴은 탁자·바닥·목제 상판에 받쳐져 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.571
   },
   "violations": {
    "A": [
     "[gemini-pro] 찰리의 왼쪽 기계 손이 쥐고 있는, 맥락 없이 잘려나간 형태의 피아노 건반 뭉치 (존재할 수 없는 왜곡된 사물 생성)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1571,
   "A": 1417
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "치명적인 형태 왜곡은 피했으나, 텍스트에 명시된 찰리의 핵심 특징(하얀 얼굴판)을 무시하고 오르간을 인물 전면에 배치하여 프레이밍 지시를 완전히 위반했습니다."
   },
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "프롬프트가 요구한 찰리의 하얀 얼굴판은 올바르게 구현했으나, 찰리의 손에 들린 정체불명의 잘려나간 건반 조각이라는 치명적인 환각 현상 때문에 실격입니다.  ★위반: [gemini-pro] 찰리의 왼쪽 기계 손이 쥐고 있는, 맥락 없이 잘려나간 형태의 피아노 건반 뭉치 (존재할 수 없는 왜곡된 사물 생성)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S30sh6_sel.png",
    "asset_id": "47ad9bb1-44ed-44c8-9b93-06103124f7da",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1597480>",
    "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929233>",
    "asset_id": "1297d7df-4792-4e39-bdf8-1c0f5b59585c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdae-8f18-746e-8e71-8e6f067afbaf",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S30sh6"
  }
 },
 "S30sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:32:08.955071+00:00",
  "fingerprint": "77c1753d5f1aa5c61bf48dbb9185b89931b5a9c57557a8bc6982c6e64c7988fd",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S30sh13_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S30sh13_sel.png",
  "source_sha256": "beb3d43da663c18df97336e793e473d3c906f34fcbc3409b3d195aaa113477f8",
  "file": "S30sh13_cine.png",
  "staged_sha256": "21b643711b4b7993175d86cc27873de8b9431e09e2b22acb00dcc801594a9e4e",
  "latency_ms": 11310
 },
 "S30sh17::signage": {
  "fp": "e99863cdf37ae494",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S30sh17": {
  "input_fingerprint": "7af295eb77a3497c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 거대한 금속 팔을 한 손으로 꽉 움켜쥔 채 자기 쪽으로 강하게 끌어당기느라 상체가 뒤로 젖혀진, 힘이 실린 mid-action 동작.\n\nLOCATION (lock): In the lantern-lit open area of the church basement prayer room, beside its sparse furniture. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the middle-right of the frame, foreground; 찰리 in the middle-left of the frame, midground; 구도환 beyond the grip in the middle-center of the frame, background.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established lantern-lit basement exposure, allowing the strained gesture rather than a lighting shift to create conflict.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the basement finishes, lit lanterns, small organ, and sparse furniture from the reference. EXCLUDE any transient activation glow rather than treating it as a permanent light source.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The lanterns remain brightly lit and the old radio remains functional in the basement, whose cross, desk, chairs and harmonium are unchanged. Charlie remains in his old coat and hat, with blue-lit eyes and the worn chest logo. 이현우: He stands in the basement with an openly distrustful expression, retaining his facial bruises and leg wound. His outer garment remains removed and his contact card stays inside his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 거대한 금속 팔을 한 손으로 꽉 움켜쥔 채 자기 쪽으로 강하게 끌어당기느라 상체가 뒤로 젖혀진, 힘이 실린 mid-action 동작.\n\nLOCATION (lock): In the lantern-lit open area of the church basement prayer room, beside its sparse furniture. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the middle-right of the frame, foreground; 찰리 in the middle-left of the frame, midground; 구도환 beyond the grip in the middle-center of the frame, background.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established lantern-lit basement exposure, allowing the strained gesture rather than a lighting shift to create conflict.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the basement finishes, lit lanterns, small organ, and sparse furniture from the reference. EXCLUDE any transient activation glow rather than treating it as a permanent light source.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The lanterns remain brightly lit and the old radio remains functional in the basement, whose cross, desk, chairs and harmonium are unchanged. Charlie remains in his old coat and hat, with blue-lit eyes and the worn chest logo. 이현우: He stands in the basement with an openly distrustful expression, retaining his facial bruises and leg wound. His outer garment remains removed and his contact card stays inside his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 거대한 금속 팔을 한 손으로 꽉 움켜쥔 채 자기 쪽으로 강하게 끌어당기느라 상체가 뒤로 젖혀진, 힘이 실린 mid-action 동작.\n\nLOCATION (lock): In the lantern-lit open area of the church basement prayer room, beside its sparse furniture. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the middle-right of the frame, foreground; 찰리 in the middle-left of the frame, midground; 구도환 beyond the grip in the middle-center of the frame, background.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established lantern-lit basement exposure, allowing the strained gesture rather than a lighting shift to create conflict.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the basement finishes, lit lanterns, small organ, and sparse furniture from the reference. EXCLUDE any transient activation glow rather than treating it as a permanent light source.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The lanterns remain brightly lit and the old radio remains functional in the basement, whose cross, desk, chairs and harmonium are unchanged. Charlie remains in his old coat and hat, with blue-lit eyes and the worn chest logo. 이현우: He stands in the basement with an openly distrustful expression, retaining his facial bruises and leg wound. His outer garment remains removed and his contact card stays inside his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 찰리를 주시하고, 찰리는 잡힌 팔 쪽을 내려다봄. 배경의 노인은 두 사람을 향해 시선을 둠.",
    "built_space": "지하실 벽면의 십자가(1), 피아노(1), 랜턴, 가구 등 참조된 공간 요소가 올바르게 배치됨.",
    "entities": "이현우(상의 탈의, 상처)와 찰리(모자, 외투 착용)의 묘사가 정확함. 제외 규칙과 레이아웃 지시가 충돌한 구도환이 배경에 등장함.",
    "hard_violations": [
     "[gpt-high] 이전 스틸에서 제외하라고 지정한 중년 남성의 얼굴과 갈색 재킷을 중앙 배경 인물에 재사용했다.",
     "[gpt-high] 기존의 왼쪽 바닥 등불 두 개와 가구 위 등불들을 유지하면서 중앙 오른쪽 바닥에 등불 하나를 추가해, 고정된 장소와 소품의 연속성을 바꿨다."
    ],
    "physics": "이현우는 두 발로 바닥을 딛고 뒤로 몸을 젖혀 당기는 힘을 잘 표현하며, 찰리는 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "이현우와 찰리가 서로를 마주 봄. 배경의 노인은 두 사람을 바라보며 다가옴.",
    "built_space": "지하실 공간 내 십자가, 피아노, 랜턴이 올바른 위치와 비율로 나타남.",
    "entities": "찰리가 지정된 낡은 외투를 입지 않음. 이현우의 신체 일부가 기계 구조물로 변형됨.",
    "hard_violations": [
     "[gemini-pro] 이현우의 왼팔 피부가 찰리의 기계 팔로 이어져 융합되는 물리적으로 불가능한 해부학적 오류",
     "[gpt-high] 이전 스틸에서 찰리를 안았던 남성의 얼굴과 갈색 재킷을 중앙 배경 인물에 재사용했다. 해당 인물의 얼굴·몸·의상을 이번 인물에게 옮기지 말라는 명시적 제외 조건에 어긋난다."
    ],
    "physics": "이현우가 바닥에 서서 당기는 자세를 취하나, 팔이 상대방의 금속 팔에 불가능한 형태로 결합되어 지지가 성립하지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리의 외투와 지하실 배경을 훌륭히 구현했으나, 한 손 지정과 달리 두 손으로 잡았고 프롬프트 제외 규칙에 어긋나는 노인이 등장함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 필수 의상인 외투가 누락되었으며, 이현우의 팔이 기계와 융합되는 치명적인 시각적 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 찰리를 주시하고, 찰리는 잡힌 팔 쪽을 내려다봄. 배경의 노인은 두 사람을 향해 시선을 둠.",
        "built_space": "지하실 벽면의 십자가(1), 피아노(1), 랜턴, 가구 등 참조된 공간 요소가 올바르게 배치됨.",
        "entities": "이현우(상의 탈의, 상처)와 찰리(모자, 외투 착용)의 묘사가 정확함. 제외 규칙과 레이아웃 지시가 충돌한 구도환이 배경에 등장함.",
        "hard_violations": [],
        "physics": "이현우는 두 발로 바닥을 딛고 뒤로 몸을 젖혀 당기는 힘을 잘 표현하며, 찰리는 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "이현우와 찰리가 서로를 마주 봄. 배경의 노인은 두 사람을 바라보며 다가옴.",
        "built_space": "지하실 공간 내 십자가, 피아노, 랜턴이 올바른 위치와 비율로 나타남.",
        "entities": "찰리가 지정된 낡은 외투를 입지 않음. 이현우의 신체 일부가 기계 구조물로 변형됨.",
        "hard_violations": [
         "이현우의 왼팔 피부가 찰리의 기계 팔로 이어져 융합되는 물리적으로 불가능한 해부학적 오류"
        ],
        "physics": "이현우가 바닥에 서서 당기는 자세를 취하나, 팔이 상대방의 금속 팔에 불가능한 형태로 결합되어 지지가 성립하지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리의 외투와 지하실 배경을 훌륭히 구현했으나, 한 손 지정과 달리 두 손으로 잡았고 프롬프트 제외 규칙에 어긋나는 노인이 등장함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 필수 의상인 외투가 누락되었으며, 이현우의 팔이 기계와 융합되는 치명적인 시각적 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 찰리를 주시하고, 찰리는 잡힌 팔 쪽을 내려다봄. 배경의 노인은 두 사람을 향해 시선을 둠.",
        "built_space": "지하실 벽면의 십자가(1), 피아노(1), 랜턴, 가구 등 참조된 공간 요소가 올바르게 배치됨.",
        "entities": "이현우(상의 탈의, 상처)와 찰리(모자, 외투 착용)의 묘사가 정확함. 제외 규칙과 레이아웃 지시가 충돌한 구도환이 배경에 등장함.",
        "hard_violations": [],
        "physics": "이현우는 두 발로 바닥을 딛고 뒤로 몸을 젖혀 당기는 힘을 잘 표현하며, 찰리는 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "이현우와 찰리가 서로를 마주 봄. 배경의 노인은 두 사람을 바라보며 다가옴.",
        "built_space": "지하실 공간 내 십자가, 피아노, 랜턴이 올바른 위치와 비율로 나타남.",
        "entities": "찰리가 지정된 낡은 외투를 입지 않음. 이현우의 신체 일부가 기계 구조물로 변형됨.",
        "hard_violations": [
         "이현우의 왼팔 피부가 찰리의 기계 팔로 이어져 융합되는 물리적으로 불가능한 해부학적 오류"
        ],
        "physics": "이현우가 바닥에 서서 당기는 자세를 취하나, 팔이 상대방의 금속 팔에 불가능한 형태로 결합되어 지지가 성립하지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "금속 팔을 잡고 뒤로 젖힌 이현우의 동작과 상대적으로 가까운 구도는 더 적합하지만, 제외된 이전 인물의 외형을 재사용했고 찰리의 낡은 코트도 누락했다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "지면을 버티며 당기는 힘은 보이지만, 한 손 지시를 양손으로 바꾸고 미디엄 숏을 거의 전신 구도로 넓혔으며 제외된 이전 인물까지 재사용했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 왼쪽의 찰리 얼굴을 바라보며 상체를 화면 오른쪽 뒤로 젖힌다. 찰리는 고개를 이현우와 팔의 접촉 지점 쪽으로 숙인다. 이현우의 한 손은 금속 전완 윗부분에 걸려 있고, 팔을 자기 몸 쪽으로 당기는 방향이다. 배경 남성은 두 사람의 접촉 쪽을 바라본다.",
        "built_space": "왼쪽 벽의 십자가 하나, 왼쪽 계단 난간, 오른쪽 콘크리트 기둥 하나, 오른쪽 하모니움 하나와 연주용 벤치가 보인다. 일반 의자는 왼쪽 벽, 중앙 뒤, 오른쪽 기둥 옆에 각각 하나씩 보인다. 등불은 왼쪽 바닥 두 개, 중앙 뒤 하나, 오른쪽 뒤 작은 하나, 하모니움 위 하나로 다섯 개가 보인다. 낡은 투톤 벽과 따뜻한 지하실 노출은 이어진다. 이현우는 오른쪽, 찰리는 왼쪽, 남성은 손잡이 너머 중앙에 있지만 찰리도 상당히 전경에 가깝고, 허벅지까지 담아 미디엄 숏보다 조금 넓다.",
        "entities": "이현우는 헝클어진 검은 머리의 젊은 한국계 남성으로 보이고 얼굴 멍과 청바지가 있다. 상체는 완전히 벗었으며 다리 상처와 신발 속 카드는 프레임에서 확인할 수 없다. 찰리는 모자, 파란 눈, 마모된 고릴라형 금속 얼굴과 거대한 금속 팔을 갖췄지만 이전 장면의 긴 코트가 보이지 않는다. 흰 얼굴판보다는 이전 스틸의 금속 얼굴에 가깝다. 중앙 남성은 이전 스틸에서 찰리를 안았던 중년 남성의 얼굴, 짧은 머리, 갈색 재킷을 재사용한 모습이다. 라디오는 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 스틸에서 찰리를 안았던 남성의 얼굴과 갈색 재킷을 중앙 배경 인물에 재사용했다. 해당 인물의 얼굴·몸·의상을 이번 인물에게 옮기지 말라는 명시적 제외 조건에 어긋난다."
        ],
        "physics": "이현우의 손과 손목이 금속 팔 위에 실제로 접촉하고 찰리의 금속 손도 이현우의 전완에 닿아 있다. 한 손의 감김은 일부 가려져 강한 움켜쥠이 아주 명료하지는 않다. 뒤로 젖힌 몸은 프레임 아래로 이어지는 다리와 팔의 접촉으로 버티는 자세이며, 발이 잘렸다는 이유만으로 공중 부양으로 볼 수는 없다. 찰리의 팔은 어깨와 팔꿈치에 연결되어 있다. 배경 남성은 한 발을 바닥에 딛고 다른 발을 내딛는 중이며, 가구와 등불도 바닥이나 가구 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 금속 손목을 자기 쪽으로 끌어당기며 오른쪽 뒤로 크게 젖혀지지만 얼굴과 시선은 찰리가 아니라 위쪽을 향한다. 찰리는 붙잡힌 팔과 이현우 쪽으로 고개를 돌린다. 배경 남성은 두 사람의 손 부근을 내려다본다. 당기는 방향은 맞지만 금속 팔을 양손으로 잡고 있어 한 손 동작 지시와 다르다.",
        "built_space": "왼쪽 십자가 하나와 계단 난간, 오른쪽 기둥 하나, 오른쪽 하모니움 하나와 벤치가 보인다. 일반 의자는 왼쪽, 중앙 책상 앞, 오른쪽에 세 개가 보인다. 등불은 왼쪽 바닥 두 개, 중앙 책상 위 하나, 중앙 오른쪽 바닥 하나, 오른쪽 뒤 작은 하나, 하모니움 위 하나로 여섯 개가 보이며 이전 스틸보다 바닥 등불이 추가된 구성이다. 벽과 콘크리트의 재질 및 따뜻한 조명은 유사하다. 인물의 좌우 순서는 맞지만 찰리의 발과 이현우의 다리 대부분까지 보여 미디엄 숏 대신 거의 전신 구도이며, 배경 남성도 정중앙보다 오른쪽에 있다.",
        "entities": "이현우는 검은 헝클어진 머리, 얼굴 상처, 맨상체와 청바지를 가진 젊은 한국계 남성으로 보인다. 고개를 심하게 젖혀 참조 얼굴과의 세부 비교는 제한된다. 찰리는 낡은 모자와 긴 코트, 파란 눈, 금속 고릴라 얼굴, 원형 가슴 장식을 유지한다. 배경 남성은 이전 스틸의 중년 남성과 같은 얼굴 및 갈색 재킷을 사용한 모습이다. 라디오와 신발 속 카드는 확인되지 않고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 스틸에서 제외하라고 지정한 중년 남성의 얼굴과 갈색 재킷을 중앙 배경 인물에 재사용했다.",
         "기존의 왼쪽 바닥 등불 두 개와 가구 위 등불들을 유지하면서 중앙 오른쪽 바닥에 등불 하나를 추가해, 고정된 장소와 소품의 연속성을 바꿨다."
        ],
        "physics": "이현우의 양손이 찰리의 금속 손목을 감싸고 있어 접촉과 힘 전달은 분명하다. 무릎을 굽혀 다리를 벌렸고 오른쪽 신발이 바닥에 닿아 있어 뒤로 젖힌 상체를 버틸 지지가 있다. 찰리도 두 발을 넓게 벌려 바닥을 딛고 있으며 금속 팔은 어깨부터 손까지 연결된다. 떠 있는 인체나 지지 없는 물체는 보이지 않는다. 다만 이 자세는 한 손으로 잡아당기는 요청을 양손으로 버티는 동작으로 바꾼 것이다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "금속 팔을 잡고 뒤로 젖힌 이현우의 동작과 상대적으로 가까운 구도는 더 적합하지만, 제외된 이전 인물의 외형을 재사용했고 찰리의 낡은 코트도 누락했다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지면을 버티며 당기는 힘은 보이지만, 한 손 지시를 양손으로 바꾸고 미디엄 숏을 거의 전신 구도로 넓혔으며 제외된 이전 인물까지 재사용했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 왼쪽의 찰리 얼굴을 바라보며 상체를 화면 오른쪽 뒤로 젖힌다. 찰리는 고개를 이현우와 팔의 접촉 지점 쪽으로 숙인다. 이현우의 한 손은 금속 전완 윗부분에 걸려 있고, 팔을 자기 몸 쪽으로 당기는 방향이다. 배경 남성은 두 사람의 접촉 쪽을 바라본다.",
        "built_space": "왼쪽 벽의 십자가 하나, 왼쪽 계단 난간, 오른쪽 콘크리트 기둥 하나, 오른쪽 하모니움 하나와 연주용 벤치가 보인다. 일반 의자는 왼쪽 벽, 중앙 뒤, 오른쪽 기둥 옆에 각각 하나씩 보인다. 등불은 왼쪽 바닥 두 개, 중앙 뒤 하나, 오른쪽 뒤 작은 하나, 하모니움 위 하나로 다섯 개가 보인다. 낡은 투톤 벽과 따뜻한 지하실 노출은 이어진다. 이현우는 오른쪽, 찰리는 왼쪽, 남성은 손잡이 너머 중앙에 있지만 찰리도 상당히 전경에 가깝고, 허벅지까지 담아 미디엄 숏보다 조금 넓다.",
        "entities": "이현우는 헝클어진 검은 머리의 젊은 한국계 남성으로 보이고 얼굴 멍과 청바지가 있다. 상체는 완전히 벗었으며 다리 상처와 신발 속 카드는 프레임에서 확인할 수 없다. 찰리는 모자, 파란 눈, 마모된 고릴라형 금속 얼굴과 거대한 금속 팔을 갖췄지만 이전 장면의 긴 코트가 보이지 않는다. 흰 얼굴판보다는 이전 스틸의 금속 얼굴에 가깝다. 중앙 남성은 이전 스틸에서 찰리를 안았던 중년 남성의 얼굴, 짧은 머리, 갈색 재킷을 재사용한 모습이다. 라디오는 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 스틸에서 찰리를 안았던 남성의 얼굴과 갈색 재킷을 중앙 배경 인물에 재사용했다. 해당 인물의 얼굴·몸·의상을 이번 인물에게 옮기지 말라는 명시적 제외 조건에 어긋난다."
        ],
        "physics": "이현우의 손과 손목이 금속 팔 위에 실제로 접촉하고 찰리의 금속 손도 이현우의 전완에 닿아 있다. 한 손의 감김은 일부 가려져 강한 움켜쥠이 아주 명료하지는 않다. 뒤로 젖힌 몸은 프레임 아래로 이어지는 다리와 팔의 접촉으로 버티는 자세이며, 발이 잘렸다는 이유만으로 공중 부양으로 볼 수는 없다. 찰리의 팔은 어깨와 팔꿈치에 연결되어 있다. 배경 남성은 한 발을 바닥에 딛고 다른 발을 내딛는 중이며, 가구와 등불도 바닥이나 가구 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "이현우는 금속 손목을 자기 쪽으로 끌어당기며 오른쪽 뒤로 크게 젖혀지지만 얼굴과 시선은 찰리가 아니라 위쪽을 향한다. 찰리는 붙잡힌 팔과 이현우 쪽으로 고개를 돌린다. 배경 남성은 두 사람의 손 부근을 내려다본다. 당기는 방향은 맞지만 금속 팔을 양손으로 잡고 있어 한 손 동작 지시와 다르다.",
        "built_space": "왼쪽 십자가 하나와 계단 난간, 오른쪽 기둥 하나, 오른쪽 하모니움 하나와 벤치가 보인다. 일반 의자는 왼쪽, 중앙 책상 앞, 오른쪽에 세 개가 보인다. 등불은 왼쪽 바닥 두 개, 중앙 책상 위 하나, 중앙 오른쪽 바닥 하나, 오른쪽 뒤 작은 하나, 하모니움 위 하나로 여섯 개가 보이며 이전 스틸보다 바닥 등불이 추가된 구성이다. 벽과 콘크리트의 재질 및 따뜻한 조명은 유사하다. 인물의 좌우 순서는 맞지만 찰리의 발과 이현우의 다리 대부분까지 보여 미디엄 숏 대신 거의 전신 구도이며, 배경 남성도 정중앙보다 오른쪽에 있다.",
        "entities": "이현우는 검은 헝클어진 머리, 얼굴 상처, 맨상체와 청바지를 가진 젊은 한국계 남성으로 보인다. 고개를 심하게 젖혀 참조 얼굴과의 세부 비교는 제한된다. 찰리는 낡은 모자와 긴 코트, 파란 눈, 금속 고릴라 얼굴, 원형 가슴 장식을 유지한다. 배경 남성은 이전 스틸의 중년 남성과 같은 얼굴 및 갈색 재킷을 사용한 모습이다. 라디오와 신발 속 카드는 확인되지 않고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 스틸에서 제외하라고 지정한 중년 남성의 얼굴과 갈색 재킷을 중앙 배경 인물에 재사용했다.",
         "기존의 왼쪽 바닥 등불 두 개와 가구 위 등불들을 유지하면서 중앙 오른쪽 바닥에 등불 하나를 추가해, 고정된 장소와 소품의 연속성을 바꿨다."
        ],
        "physics": "이현우의 양손이 찰리의 금속 손목을 감싸고 있어 접촉과 힘 전달은 분명하다. 무릎을 굽혀 다리를 벌렸고 오른쪽 신발이 바닥에 닿아 있어 뒤로 젖힌 상체를 버틸 지지가 있다. 찰리도 두 발을 넓게 벌려 바닥을 딛고 있으며 금속 팔은 어깨부터 손까지 연결된다. 떠 있는 인체나 지지 없는 물체는 보이지 않는다. 다만 이 자세는 한 손으로 잡아당기는 요청을 양손으로 버티는 동작으로 바꾼 것이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.179
   },
   "violations": {
    "B": [
     "[gemini-pro] 이현우의 왼팔 피부가 찰리의 기계 팔로 이어져 융합되는 물리적으로 불가능한 해부학적 오류",
     "[gpt-high] 이전 스틸에서 찰리를 안았던 남성의 얼굴과 갈색 재킷을 중앙 배경 인물에 재사용했다. 해당 인물의 얼굴·몸·의상을 이번 인물에게 옮기지 말라는 명시적 제외 조건에 어긋난다."
    ],
    "A": [
     "[gpt-high] 이전 스틸에서 제외하라고 지정한 중년 남성의 얼굴과 갈색 재킷을 중앙 배경 인물에 재사용했다.",
     "[gpt-high] 기존의 왼쪽 바닥 등불 두 개와 가구 위 등불들을 유지하면서 중앙 오른쪽 바닥에 등불 하나를 추가해, 고정된 장소와 소품의 연속성을 바꿨다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "찰리의 외투와 지하실 배경을 훌륭히 구현했으나, 한 손 지정과 달리 두 손으로 잡았고 프롬프트 제외 규칙에 어긋나는 노인이 등장함.  ★위반: [gpt-high] 이전 스틸에서 제외하라고 지정한 중년 남성의 얼굴과 갈색 재킷을 중앙 배경 인물에 재사용했다. / [gpt-high] 기존의 왼쪽 바닥 등불 두 개와 가구 위 등불들을 유지하면서 중앙 오른쪽 바닥에 등불 하나를 추가해, 고정된 장소와 소품의 연속성을 바꿨다."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "찰리의 필수 의상인 외투가 누락되었으며, 이현우의 팔이 기계와 융합되는 치명적인 시각적 오류가 발생함.  ★위반: [gemini-pro] 이현우의 왼팔 피부가 찰리의 기계 팔로 이어져 융합되는 물리적으로 불가능한 해부학적 오류 / [gpt-high] 이전 스틸에서 찰리를 안았던 남성의 얼굴과 갈색 재킷을 중앙 배경 인물에 재사용했다. 해당 인물의 얼굴·몸·의상을 이번 인물에게 옮기지 말라는 명시적 제외 조건에 어긋난다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S30sh13_sel.png",
    "asset_id": "7996f16e-a373-4b7a-a1d7-af8d426c1c68",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1597480>",
    "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdaf-4990-717d-b4e4-bfe541107079",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S30sh13"
  }
 },
 "S30sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:32:20.610129+00:00",
  "fingerprint": "f67785066b2e7982ec668567246c95759c1291299ca51a95b7ebcb66b5005860",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S30sh17_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S30sh17_sel.png",
  "source_sha256": "1f6ed9d7288f9606f993fcb1e785b49bf7dd65f4dffab69d69a5e47485110215",
  "file": "S30sh17_cine.png",
  "staged_sha256": "13d0a7390de3d5e115b2423cd5d23cb36108769c2e0f5d772cb5eaa4cec1440c",
  "latency_ms": 22385
 },
 "S31sh2::signage": {
  "fp": "63ba97697f6751f7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::4d415a3b25e628e5": {
  "subjects": [],
  "subject_text": "인천 난민촌 관리사무소 내부 미연의 감금실\n관리 건물 3층의 닫힌 방. 실내에 의자가 놓여 있으며 복도와 연결된 문에는 손잡이와 잠금장치가 달려 있다.",
  "identity": "canonical",
  "scope_id": "L44",
  "scope_role": "location_interior",
  "scope_sha": "0667757a617702b8"
 },
 "S31sh2::bgfirst_bg": {
  "input_fingerprint": "9c23a7e94c40b628",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 얼굴이 퉁퉁 붓고 피멍이 든 미연이 의자에 앉아 고통스러운 표정으로 앞을 응시하는 얼굴.\n\nLOCATION (lock): At the interrogation chair inside a third-floor room of the brightly lit refugee-settlement administration building.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 미연's chair (Occupied) — Only a small supporting edge is visible beneath her shoulder; used as Minimal evidence of her seated position without widening away from her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brightly lit nighttime office exposes 미연's swelling and bruises with controlled facial contrast rather than concealing them in stylized darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 얼굴이 퉁퉁 붓고 피멍이 든 미연이 의자에 앉아 고통스러운 표정으로 앞을 응시하는 얼굴.\n\nLOCATION (lock): At the interrogation chair inside a third-floor room of the brightly lit refugee-settlement administration building.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 미연's chair (Occupied) — Only a small supporting edge is visible beneath her shoulder; used as Minimal evidence of her seated position without widening away from her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brightly lit nighttime office exposes 미연's swelling and bruises with controlled facial contrast rather than concealing them in stylized darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S31sh2__bgfirst_bg.png",
  "asset_id": "758b1bb9-5bdf-499c-8525-da54edfcb964",
  "input_asset_ids": [
   "7ee00446-d130-4c9e-8fbc-177baed0bf3e",
   "fb1c05ef-2a65-46b1-8703-de2401f7e330"
  ]
 },
 "S31sh2": {
  "input_fingerprint": "bc7b1780b902fd0d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 얼굴이 퉁퉁 붓고 피멍이 든 미연이 의자에 앉아 고통스러운 표정으로 앞을 응시하는 얼굴.\n\nLOCATION (lock): At the interrogation chair inside a third-floor room of the brightly lit refugee-settlement administration building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 미연's chair (Occupied) — Only a small supporting edge is visible beneath her shoulder; used as Minimal evidence of her seated position without widening away from her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brightly lit nighttime office exposes 미연's swelling and bruises with controlled facial contrast rather than concealing them in stylized darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management building remains brightly illuminated, with the interrogation taking place on the third floor. 미연: She remains detained in the third-floor interrogation room, her face badly swollen from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 얼굴이 퉁퉁 붓고 피멍이 든 미연이 의자에 앉아 고통스러운 표정으로 앞을 응시하는 얼굴.\n\nLOCATION (lock): At the interrogation chair inside a third-floor room of the brightly lit refugee-settlement administration building. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 미연's chair (Occupied) — Only a small supporting edge is visible beneath her shoulder; used as Minimal evidence of her seated position without widening away from her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brightly lit nighttime office exposes 미연's swelling and bruises with controlled facial contrast rather than concealing them in stylized darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management building remains brightly illuminated, with the interrogation taking place on the third floor. 미연: She remains detained in the third-floor interrogation room, her face badly swollen from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 얼굴이 퉁퉁 붓고 피멍이 든 미연이 의자에 앉아 고통스러운 표정으로 앞을 응시하는 얼굴.\n\nLOCATION (lock): At the interrogation chair inside a third-floor room of the brightly lit refugee-settlement administration building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 미연's chair (Occupied) — Only a small supporting edge is visible beneath her shoulder; used as Minimal evidence of her seated position without widening away from her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brightly lit nighttime office exposes 미연's swelling and bruises with controlled facial contrast rather than concealing them in stylized darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management building remains brightly illuminated, with the interrogation taking place on the third floor. 미연: She remains detained in the third-floor interrogation room, her face badly swollen from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S31sh2__bgfirst_bg.png",
     "asset_id": "758b1bb9-5bdf-499c-8525-da54edfcb964",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S31sh2.png",
     "asset_id": "7ee00446-d130-4c9e-8fbc-177baed0bf3e",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:878964>",
     "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L44B03.png",
     "asset_id": "fb1c05ef-2a65-46b1-8703-de2401f7e330",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:878964>",
     "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "미연의 시선은 프레임 우측 전경에 있는 남성을 향함.",
    "built_space": "참조된 취조실 내부(우측의 철문, 투톤 벽면)가 보임.",
    "entities": "얼굴에 멍이 든 미연, 회색 티셔츠, 프롬프트에 없는 정장 차림의 남성 뒷모습(우측 전경).",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 없는 인물 추가 (우측 전경의 정장 입은 남성)",
     "[gpt-high] 미연만 보여야 하는 장면에 다른 인물의 머리·목·어깨를 오른쪽 전경에 추가했다."
    ],
    "physics": "미연은 의자에 앉아 있으며, 특별한 지지력 문제는 없음."
   },
   {
    "label": "B",
    "direction": "미연의 시선은 화면 정면(카메라)을 응시함.",
    "built_space": "참조된 취조실 내부(우측의 철문, 투톤 벽면, 천장 조명)가 보임.",
    "entities": "피멍이 들고 심하게 부은 얼굴의 미연, 회색 티셔츠, 검은색 가죽 쿠션과 금속 테두리가 있는 의자 등받이.",
    "hard_violations": [],
    "physics": "미연은 화면 하단에 보이는 의자에 앉아 자세를 유지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 인물만을 정확한 클로즈업으로 담아냈으며, 퉁퉁 붓고 피멍이 든 얼굴과 지정된 배경 공간을 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 인물의 멍든 얼굴은 잘 표현했으나, 프롬프트에 명시되지 않은 인물이 화면에 등장하여 치명적인 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "미연의 시선은 화면 정면(카메라)을 응시함.",
        "built_space": "참조된 취조실 내부(우측의 철문, 투톤 벽면, 천장 조명)가 보임.",
        "entities": "피멍이 들고 심하게 부은 얼굴의 미연, 회색 티셔츠, 검은색 가죽 쿠션과 금속 테두리가 있는 의자 등받이.",
        "hard_violations": [],
        "physics": "미연은 화면 하단에 보이는 의자에 앉아 자세를 유지하고 있음."
       },
       {
        "label": "A",
        "direction": "미연의 시선은 프레임 우측 전경에 있는 남성을 향함.",
        "built_space": "참조된 취조실 내부(우측의 철문, 투톤 벽면)가 보임.",
        "entities": "얼굴에 멍이 든 미연, 회색 티셔츠, 프롬프트에 없는 정장 차림의 남성 뒷모습(우측 전경).",
        "hard_violations": [
         "프롬프트에 없는 인물 추가 (우측 전경의 정장 입은 남성)"
        ],
        "physics": "미연은 의자에 앉아 있으며, 특별한 지지력 문제는 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 인물만을 정확한 클로즈업으로 담아냈으며, 퉁퉁 붓고 피멍이 든 얼굴과 지정된 배경 공간을 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 인물의 멍든 얼굴은 잘 표현했으나, 프롬프트에 명시되지 않은 인물이 화면에 등장하여 치명적인 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "미연의 시선은 화면 정면(카메라)을 응시함.",
        "built_space": "참조된 취조실 내부(우측의 철문, 투톤 벽면, 천장 조명)가 보임.",
        "entities": "피멍이 들고 심하게 부은 얼굴의 미연, 회색 티셔츠, 검은색 가죽 쿠션과 금속 테두리가 있는 의자 등받이.",
        "hard_violations": [],
        "physics": "미연은 화면 하단에 보이는 의자에 앉아 자세를 유지하고 있음."
       },
       {
        "label": "A",
        "direction": "미연의 시선은 프레임 우측 전경에 있는 남성을 향함.",
        "built_space": "참조된 취조실 내부(우측의 철문, 투톤 벽면)가 보임.",
        "entities": "얼굴에 멍이 든 미연, 회색 티셔츠, 프롬프트에 없는 정장 차림의 남성 뒷모습(우측 전경).",
        "hard_violations": [
         "프롬프트에 없는 인물 추가 (우측 전경의 정장 입은 남성)"
        ],
        "physics": "미연은 의자에 앉아 있으며, 특별한 지지력 문제는 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "부어오르고 피멍 든 얼굴의 정면 클로즈업과 고통스러운 앞쪽 응시를 충실히 구현했으나, 의자 등받이가 요구보다 많이 보이고 참조의 모자·겉옷·머리 모양은 일치하지 않는다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "허용되지 않은 상대 인물을 전경에 추가한 실격 사례이며, 얼굴 단독 클로즈업 대신 상반신이 넓게 보이는 어깨 너머 구도를 만들고 얼굴의 심한 부기도 약하게 표현했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연의 얼굴과 양눈은 카메라 바로 앞쪽을 향한다. 화면 밖 앞쪽을 고통스럽고 지친 표정으로 응시하여 지문과 맞는다. 방향을 확인할 무기나 손에 든 물건은 없다.",
        "built_space": "얼굴 중심의 클로즈업이다. 뒤에는 얼룩진 밝은 벽과 짙은 하부 띠, 오른쪽의 작은 창이 달린 금속문 하나, 천장 조명 하나가 보인다. 장소 참조의 재료와 주요 설비에 부합한다. 미연 뒤로 검은 패딩과 금속 테두리의 의자 등받이 하나가 보이지만, 어깨 밑 작은 가장자리만 보이라는 요구보다 노출 면적이 크다. 다른 의자와 바닥 대부분은 프레임 밖이며 반사는 없다. 층수와 밤 여부는 직접 확인되지 않지만 실내 조명은 얼굴의 상처를 충분히 드러낸다.",
        "entities": "보이는 사람은 중년의 한국인 여성 설정에 부합하는 미연 한 명뿐이다. 눈 밑과 이마의 넓은 피멍, 부은 눈꺼풀과 볼, 입술 상처가 보이며 눈의 해부학은 자연스럽다. 얼굴에는 참조 인물과 유사한 특징이 있으나 검은 머리는 참조의 풀어 내린 단발보다 뒤로 모인 형태다. 회색 티셔츠는 맞지만, 프레임에 보여야 할 참조의 모자와 갈색 겉옷은 없다. 읽을 수 있는 글자나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 자연스럽게 이어지고 몸통 뒤에 의자 등받이가 있어 착석 상태를 뒷받침한다. 엉덩이와 좌판 접촉은 클로즈업 밖이라 확인할 수 없지만 공중에 떠 있는 신체나 물체는 없다. 멍과 피부의 광택은 얼굴 곡면과 조명을 따른다."
       },
       {
        "label": "B",
        "direction": "미연은 얼굴을 화면 오른쪽으로 조금 돌리고 오른쪽 전경 인물의 얼굴을 올려다본다. 전경 인물도 미연 쪽으로 머리를 향하지만 눈은 보이지 않는다. 미연의 시선 목표가 추가된 상대 인물로 명확히 설정되어 있다.",
        "built_space": "미연의 머리부터 상반신까지 보이는 어깨 너머 구도로, 요구한 얼굴 중심 클로즈업보다 넓다. 밝은 벽과 짙은 하부 띠, 오른쪽의 작은 창이 달린 금속문 하나, 천장 조명 하나가 보여 장소 참조와 대체로 맞는다. 의자 하나의 검은 등받이와 금속 가장자리가 미연의 왼쪽 어깨 아래에 보인다. 오른쪽 전경의 다른 사람이 화면을 크게 차지한다. 반사나 중복 설비는 보이지 않으며 층수와 밤 여부는 직접 확인되지 않는다.",
        "entities": "미연은 참조와 유사한 검은 단발의 중년 한국인 여성으로 보인다. 회색 티셔츠는 일치하지만 참조의 모자와 갈색 겉옷은 없다. 눈 밑 부기와 볼·이마의 멍은 있으나 얼굴 전체가 퉁퉁 부었다는 요구에는 약하다. 오른쪽에는 짧은 검은 머리와 어두운 옷을 입은 남성으로 보이는 인물의 귀·목·어깨가 추가되어 있다. 이 인물의 나이와 민족적 정체성은 보이는 부분만으로 확정할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "미연만 보여야 하는 장면에 다른 인물의 머리·목·어깨를 오른쪽 전경에 추가했다."
        ],
        "physics": "미연은 의자 등받이 앞에 앉아 상체를 약간 앞으로 기울인 모습으로, 목과 어깨의 연결 및 착석 자세가 물리적으로 가능하다. 좌판과 하체는 프레임 밖이다. 추가 인물도 머리·목·어깨가 연결된 정상적인 상체 일부로 보이며, 떠 있는 신체나 지지 없는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "부어오르고 피멍 든 얼굴의 정면 클로즈업과 고통스러운 앞쪽 응시를 충실히 구현했으나, 의자 등받이가 요구보다 많이 보이고 참조의 모자·겉옷·머리 모양은 일치하지 않는다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "허용되지 않은 상대 인물을 전경에 추가한 실격 사례이며, 얼굴 단독 클로즈업 대신 상반신이 넓게 보이는 어깨 너머 구도를 만들고 얼굴의 심한 부기도 약하게 표현했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "미연의 얼굴과 양눈은 카메라 바로 앞쪽을 향한다. 화면 밖 앞쪽을 고통스럽고 지친 표정으로 응시하여 지문과 맞는다. 방향을 확인할 무기나 손에 든 물건은 없다.",
        "built_space": "얼굴 중심의 클로즈업이다. 뒤에는 얼룩진 밝은 벽과 짙은 하부 띠, 오른쪽의 작은 창이 달린 금속문 하나, 천장 조명 하나가 보인다. 장소 참조의 재료와 주요 설비에 부합한다. 미연 뒤로 검은 패딩과 금속 테두리의 의자 등받이 하나가 보이지만, 어깨 밑 작은 가장자리만 보이라는 요구보다 노출 면적이 크다. 다른 의자와 바닥 대부분은 프레임 밖이며 반사는 없다. 층수와 밤 여부는 직접 확인되지 않지만 실내 조명은 얼굴의 상처를 충분히 드러낸다.",
        "entities": "보이는 사람은 중년의 한국인 여성 설정에 부합하는 미연 한 명뿐이다. 눈 밑과 이마의 넓은 피멍, 부은 눈꺼풀과 볼, 입술 상처가 보이며 눈의 해부학은 자연스럽다. 얼굴에는 참조 인물과 유사한 특징이 있으나 검은 머리는 참조의 풀어 내린 단발보다 뒤로 모인 형태다. 회색 티셔츠는 맞지만, 프레임에 보여야 할 참조의 모자와 갈색 겉옷은 없다. 읽을 수 있는 글자나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 자연스럽게 이어지고 몸통 뒤에 의자 등받이가 있어 착석 상태를 뒷받침한다. 엉덩이와 좌판 접촉은 클로즈업 밖이라 확인할 수 없지만 공중에 떠 있는 신체나 물체는 없다. 멍과 피부의 광택은 얼굴 곡면과 조명을 따른다."
       },
       {
        "label": "A",
        "direction": "미연은 얼굴을 화면 오른쪽으로 조금 돌리고 오른쪽 전경 인물의 얼굴을 올려다본다. 전경 인물도 미연 쪽으로 머리를 향하지만 눈은 보이지 않는다. 미연의 시선 목표가 추가된 상대 인물로 명확히 설정되어 있다.",
        "built_space": "미연의 머리부터 상반신까지 보이는 어깨 너머 구도로, 요구한 얼굴 중심 클로즈업보다 넓다. 밝은 벽과 짙은 하부 띠, 오른쪽의 작은 창이 달린 금속문 하나, 천장 조명 하나가 보여 장소 참조와 대체로 맞는다. 의자 하나의 검은 등받이와 금속 가장자리가 미연의 왼쪽 어깨 아래에 보인다. 오른쪽 전경의 다른 사람이 화면을 크게 차지한다. 반사나 중복 설비는 보이지 않으며 층수와 밤 여부는 직접 확인되지 않는다.",
        "entities": "미연은 참조와 유사한 검은 단발의 중년 한국인 여성으로 보인다. 회색 티셔츠는 일치하지만 참조의 모자와 갈색 겉옷은 없다. 눈 밑 부기와 볼·이마의 멍은 있으나 얼굴 전체가 퉁퉁 부었다는 요구에는 약하다. 오른쪽에는 짧은 검은 머리와 어두운 옷을 입은 남성으로 보이는 인물의 귀·목·어깨가 추가되어 있다. 이 인물의 나이와 민족적 정체성은 보이는 부분만으로 확정할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "미연만 보여야 하는 장면에 다른 인물의 머리·목·어깨를 오른쪽 전경에 추가했다."
        ],
        "physics": "미연은 의자 등받이 앞에 앉아 상체를 약간 앞으로 기울인 모습으로, 목과 어깨의 연결 및 착석 자세가 물리적으로 가능하다. 좌판과 하체는 프레임 밖이다. 추가 인물도 머리·목·어깨가 연결된 정상적인 상체 일부로 보이며, 떠 있는 신체나 지지 없는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.679,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.429,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 없는 인물 추가 (우측 전경의 정장 입은 남성)",
     "[gpt-high] 미연만 보여야 하는 장면에 다른 인물의 머리·목·어깨를 오른쪽 전경에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 429
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 인물만을 정확한 클로즈업으로 담아냈으며, 퉁퉁 붓고 피멍이 든 얼굴과 지정된 배경 공간을 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 429,
    "verdict_ko": "지정된 인물의 멍든 얼굴은 잘 표현했으나, 프롬프트에 명시되지 않은 인물이 화면에 등장하여 치명적인 위반이 발생했습니다.  ★위반: [gemini-pro] 프롬프트에 없는 인물 추가 (우측 전경의 정장 입은 남성) / [gpt-high] 미연만 보여야 하는 장면에 다른 인물의 머리·목·어깨를 오른쪽 전경에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L44B03.png",
    "asset_id": "fb1c05ef-2a65-46b1-8703-de2401f7e330",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:878964>",
    "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb0-b5cc-760c-b3f5-0d9528d85a80",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S31sh2__bgfirst_bg.png",
   "bg_asset_id": "758b1bb9-5bdf-499c-8525-da54edfcb964",
   "bg_record_key": "S31sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S31sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:32:43.456884+00:00",
  "fingerprint": "d31d6434d598f411bccc4e94c74abb0e5d80667bba33a7f7080ccfa99438fb4f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S31sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S31sh2_sel.png",
  "source_sha256": "070ec4280361fbb7489a863e5d4f2e3a8400465a313308bd18e7a85524b4778d",
  "file": "S31sh2_cine.png",
  "staged_sha256": "098d3f0bc0ff0644919b2905f06547e77bb15defce993eb31aaa7eee3defcc90",
  "latency_ms": 12414
 },
 "S31sh5::signage": {
  "fp": "36cfc7fc674c3844",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S31sh5": {
  "input_fingerprint": "113bbb40265a7a3f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 상체를 미연 쪽으로 불쑥 내민 채 눈을 부릅뜨고 매섭게 쏘아보는 위압적인 시선.\n\nLOCATION (lock): In the face-to-face interrogation area inside the administration building's lit third-floor detention room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination and controlled facial contrast emphasize the pressure without introducing a specific light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management building remains brightly lit, and the third-floor interrogation room is unchanged. 박철진: He remains seated in the interrogation position.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 상체를 미연 쪽으로 불쑥 내민 채 눈을 부릅뜨고 매섭게 쏘아보는 위압적인 시선.\n\nLOCATION (lock): In the face-to-face interrogation area inside the administration building's lit third-floor detention room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination and controlled facial contrast emphasize the pressure without introducing a specific light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management building remains brightly lit, and the third-floor interrogation room is unchanged. 박철진: He remains seated in the interrogation position.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 상체를 미연 쪽으로 불쑥 내민 채 눈을 부릅뜨고 매섭게 쏘아보는 위압적인 시선.\n\nLOCATION (lock): In the face-to-face interrogation area inside the administration building's lit third-floor detention room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination and controlled facial contrast emphasize the pressure without introducing a specific light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management building remains brightly lit, and the third-floor interrogation room is unchanged. 박철진: He remains seated in the interrogation position.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 카메라(상대방의 위치)를 향해 강렬하게 고정됨.",
    "built_space": "콘크리트 벽면과 천장 형광등이 보이며, 배경 왼쪽에 빈 의자가 배치되어 역방향 앵글을 적절히 구현함.",
    "entities": "박철진(중년 남성, 군복, 완장)이 명확히 묘사됨. 옷깃에 참조에 없는 표식이 추가됨.",
    "hard_violations": [],
    "physics": "상체를 앞으로 깊게 기울인 자세이며, 화면 밖의 의자와 하체에 의해 안정적으로 지탱됨."
   },
   {
    "label": "B",
    "direction": "시선은 카메라를 향해 정면으로 고정됨.",
    "built_space": "오른쪽의 문과 벽면 모서리 등 이전 샷의 배경 구조를 거의 그대로 복제하여 공간적 방향성이 어긋남.",
    "entities": "박철진(중년 남성, 군복, 완장)이 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 이전 샷의 카메라 프레이밍과 배경을 그대로 복사하여, 마주보는 구도임에도 인물을 이전 샷의 인물과 물리적으로 동일한 위치에 잘못 배치함.",
     "[gpt-high] 완장에 판독 가능한 한글 '한'이 노출되어 화면 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
    ],
    "physics": "상체를 앞으로 기울이고 있으며, 보이지 않는 하체에 의해 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "마주보는 심문 구도에 맞게 배경 앵글을 변경하고, 상체를 내민 위압적인 자세와 시선을 지시문대로 훌륭하게 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물의 표정과 자세는 좋으나, 이전 샷의 배경 프레이밍을 그대로 복사하여 마주보는 구도의 공간적 개연성을 무시함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 카메라(상대방의 위치)를 향해 강렬하게 고정됨.",
        "built_space": "콘크리트 벽면과 천장 형광등이 보이며, 배경 왼쪽에 빈 의자가 배치되어 역방향 앵글을 적절히 구현함.",
        "entities": "박철진(중년 남성, 군복, 완장)이 명확히 묘사됨. 옷깃에 참조에 없는 표식이 추가됨.",
        "hard_violations": [],
        "physics": "상체를 앞으로 깊게 기울인 자세이며, 화면 밖의 의자와 하체에 의해 안정적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "시선은 카메라를 향해 정면으로 고정됨.",
        "built_space": "오른쪽의 문과 벽면 모서리 등 이전 샷의 배경 구조를 거의 그대로 복제하여 공간적 방향성이 어긋남.",
        "entities": "박철진(중년 남성, 군복, 완장)이 묘사됨.",
        "hard_violations": [
         "이전 샷의 카메라 프레이밍과 배경을 그대로 복사하여, 마주보는 구도임에도 인물을 이전 샷의 인물과 물리적으로 동일한 위치에 잘못 배치함."
        ],
        "physics": "상체를 앞으로 기울이고 있으며, 보이지 않는 하체에 의해 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "마주보는 심문 구도에 맞게 배경 앵글을 변경하고, 상체를 내민 위압적인 자세와 시선을 지시문대로 훌륭하게 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물의 표정과 자세는 좋으나, 이전 샷의 배경 프레이밍을 그대로 복사하여 마주보는 구도의 공간적 개연성을 무시함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 카메라(상대방의 위치)를 향해 강렬하게 고정됨.",
        "built_space": "콘크리트 벽면과 천장 형광등이 보이며, 배경 왼쪽에 빈 의자가 배치되어 역방향 앵글을 적절히 구현함.",
        "entities": "박철진(중년 남성, 군복, 완장)이 명확히 묘사됨. 옷깃에 참조에 없는 표식이 추가됨.",
        "hard_violations": [],
        "physics": "상체를 앞으로 깊게 기울인 자세이며, 화면 밖의 의자와 하체에 의해 안정적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "시선은 카메라를 향해 정면으로 고정됨.",
        "built_space": "오른쪽의 문과 벽면 모서리 등 이전 샷의 배경 구조를 거의 그대로 복제하여 공간적 방향성이 어긋남.",
        "entities": "박철진(중년 남성, 군복, 완장)이 묘사됨.",
        "hard_violations": [
         "이전 샷의 카메라 프레이밍과 배경을 그대로 복사하여, 마주보는 구도임에도 인물을 이전 샷의 인물과 물리적으로 동일한 위치에 잘못 배치함."
        ],
        "physics": "상체를 앞으로 기울이고 있으며, 보이지 않는 하체에 의해 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "정면으로 상체를 내밀며 노려보는 클로즈업과 공간 재현은 충실하지만, 완장에 읽을 수 있는 한글이 노출되어 명시적 문자 금지 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "눈을 부릅뜨고 상대 쪽으로 상체를 내미는 위압적인 클로즈업을 잘 구현하며, 다만 참조의 모자가 없고 착석 상태는 화면 안에서 확인되지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 상체가 카메라 정면으로 돌출되고 두 눈도 렌즈 방향을 응시한다. 화면 밖 미연의 위치를 카메라에 둔 대면 구도로 읽히며, 미연 자체는 보이지 않는다. 무기나 방향을 확인할 소품은 없다.",
        "built_space": "오른쪽에 회색 금속문 한 개, 뒤와 양옆에 낡은 밝은 벽과 짙은 하단 띠가 보인다. 참조 공간의 재료와 문 배치에 가깝다. 의자와 좌판은 프레임 밖이므로 착석 여부는 확인할 수 없다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "중년 남성 한 명만 보이며 짧은 검회색 머리, 희끗한 콧수염과 턱수염, 얼굴 윤곽과 올리브색 제복은 인물 참조에 대체로 부합한다. 참조에서 모자로 가려졌던 머리 형태는 직접 대조할 수 없고, 모자는 착용하지 않았다. 오른쪽 아래 검은 완장에는 읽을 수 있는 '한' 글자가 보인다. 이전 사진의 여성이나 다른 인물은 없다.",
        "hard_violations": [
         "완장에 판독 가능한 한글 '한'이 노출되어 화면 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "어깨와 몸통을 앞으로 숙인 자세는 실제 사람이 앉아서도 취할 수 있다. 하체, 좌판, 손의 접촉점은 잘려 있어 지지 방식은 직접 확인되지 않지만, 몸이 공중에 떠 있거나 해부학적으로 불가능하다는 증거는 없다."
       },
       {
        "label": "B",
        "direction": "고개와 상체를 카메라 쪽으로 내밀고 눈을 크게 뜬 채 렌즈 가까운 상대 위치를 매섭게 응시한다. 화면 밖 미연을 향한 시선으로 성립한다. 조준할 무기나 손에 든 물체는 없다.",
        "built_space": "왼쪽 위에 천장 형광등 설비 한 개, 왼쪽 아래에 금속 테두리의 검은 의자 등받이 한 개, 뒤에 얼룩진 밝은 벽과 짙은 하단 띠가 보인다. 재료와 조명은 참조 공간에 부합하나 벽 얼룩은 더 두드러진다. 문은 이 구도에서 보이지 않는다. 왼쪽 의자는 인물의 몸 바로 뒤가 아니므로 그의 착석을 입증하지는 않는다. 중복 설비나 불가능한 반사는 없다.",
        "entities": "중년 남성 한 명의 얼굴과 상체만 보인다. 얼굴 윤곽, 검회색 머리와 수염, 올리브색 제복 및 칼라 표식이 인물 참조와 대체로 일치한다. 참조의 모자는 없다. 눈은 정상적인 홍채와 동공을 유지한다. 완장 대부분은 잘려 있으며 판독 가능한 문구는 보이지 않는다. 다른 사람이나 손에 든 소품은 없다.",
        "hard_violations": [],
        "physics": "몸통을 앞으로 기울이고 화면 왼쪽 팔을 굽힌 자세다. 손과 골반이 화면 밖이라 좌판이나 팔의 지지 접촉은 확인할 수 없지만, 착석한 채 몸을 내미는 동작으로 가능하다. 왼쪽 빈 의자만으로 이 인물이 서 있다고 단정할 수 없으며, 부유나 불가능한 관절 형태는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "정면으로 상체를 내밀며 노려보는 클로즈업과 공간 재현은 충실하지만, 완장에 읽을 수 있는 한글이 노출되어 명시적 문자 금지 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "눈을 부릅뜨고 상대 쪽으로 상체를 내미는 위압적인 클로즈업을 잘 구현하며, 다만 참조의 모자가 없고 착석 상태는 화면 안에서 확인되지 않는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 상체가 카메라 정면으로 돌출되고 두 눈도 렌즈 방향을 응시한다. 화면 밖 미연의 위치를 카메라에 둔 대면 구도로 읽히며, 미연 자체는 보이지 않는다. 무기나 방향을 확인할 소품은 없다.",
        "built_space": "오른쪽에 회색 금속문 한 개, 뒤와 양옆에 낡은 밝은 벽과 짙은 하단 띠가 보인다. 참조 공간의 재료와 문 배치에 가깝다. 의자와 좌판은 프레임 밖이므로 착석 여부는 확인할 수 없다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "중년 남성 한 명만 보이며 짧은 검회색 머리, 희끗한 콧수염과 턱수염, 얼굴 윤곽과 올리브색 제복은 인물 참조에 대체로 부합한다. 참조에서 모자로 가려졌던 머리 형태는 직접 대조할 수 없고, 모자는 착용하지 않았다. 오른쪽 아래 검은 완장에는 읽을 수 있는 '한' 글자가 보인다. 이전 사진의 여성이나 다른 인물은 없다.",
        "hard_violations": [
         "완장에 판독 가능한 한글 '한'이 노출되어 화면 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "어깨와 몸통을 앞으로 숙인 자세는 실제 사람이 앉아서도 취할 수 있다. 하체, 좌판, 손의 접촉점은 잘려 있어 지지 방식은 직접 확인되지 않지만, 몸이 공중에 떠 있거나 해부학적으로 불가능하다는 증거는 없다."
       },
       {
        "label": "A",
        "direction": "고개와 상체를 카메라 쪽으로 내밀고 눈을 크게 뜬 채 렌즈 가까운 상대 위치를 매섭게 응시한다. 화면 밖 미연을 향한 시선으로 성립한다. 조준할 무기나 손에 든 물체는 없다.",
        "built_space": "왼쪽 위에 천장 형광등 설비 한 개, 왼쪽 아래에 금속 테두리의 검은 의자 등받이 한 개, 뒤에 얼룩진 밝은 벽과 짙은 하단 띠가 보인다. 재료와 조명은 참조 공간에 부합하나 벽 얼룩은 더 두드러진다. 문은 이 구도에서 보이지 않는다. 왼쪽 의자는 인물의 몸 바로 뒤가 아니므로 그의 착석을 입증하지는 않는다. 중복 설비나 불가능한 반사는 없다.",
        "entities": "중년 남성 한 명의 얼굴과 상체만 보인다. 얼굴 윤곽, 검회색 머리와 수염, 올리브색 제복 및 칼라 표식이 인물 참조와 대체로 일치한다. 참조의 모자는 없다. 눈은 정상적인 홍채와 동공을 유지한다. 완장 대부분은 잘려 있으며 판독 가능한 문구는 보이지 않는다. 다른 사람이나 손에 든 소품은 없다.",
        "hard_violations": [],
        "physics": "몸통을 앞으로 기울이고 화면 왼쪽 팔을 굽힌 자세다. 손과 골반이 화면 밖이라 좌판이나 팔의 지지 접촉은 확인할 수 없지만, 착석한 채 몸을 내미는 동작으로 가능하다. 왼쪽 빈 의자만으로 이 인물이 서 있다고 단정할 수 없으며, 부유나 불가능한 관절 형태는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.929
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.679
   },
   "violations": {
    "B": [
     "[gemini-pro] 이전 샷의 카메라 프레이밍과 배경을 그대로 복사하여, 마주보는 구도임에도 인물을 이전 샷의 인물과 물리적으로 동일한 위치에 잘못 배치함.",
     "[gpt-high] 완장에 판독 가능한 한글 '한'이 노출되어 화면 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 679
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "마주보는 심문 구도에 맞게 배경 앵글을 변경하고, 상체를 내민 위압적인 자세와 시선을 지시문대로 훌륭하게 구현함."
   },
   {
    "label": "B",
    "score": 679,
    "verdict_ko": "인물의 표정과 자세는 좋으나, 이전 샷의 배경 프레이밍을 그대로 복사하여 마주보는 구도의 공간적 개연성을 무시함.  ★위반: [gemini-pro] 이전 샷의 카메라 프레이밍과 배경을 그대로 복사하여, 마주보는 구도임에도 인물을 이전 샷의 인물과 물리적으로 동일한 위치에 잘못 배치함. / [gpt-high] 완장에 판독 가능한 한글 '한'이 노출되어 화면 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S31sh2_sel.png",
    "asset_id": "c4633ce2-fdcc-4ff0-8fc9-260786f3d692",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb1-8326-7a65-a904-6f9046ed9571",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S31sh2"
  }
 },
 "S31sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:32:56.205501+00:00",
  "fingerprint": "32bb68c5b6aad477a5135a340aa0554cb0cb9fe15f03b1b108ae44dda43e5dd6",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S31sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S31sh5_sel.png",
  "source_sha256": "7bc172ab142b23f4dae7365e819a8de8a934db00722284724715319dd968b6d3",
  "file": "S31sh5_cine.png",
  "staged_sha256": "d90953617309795339c81548659a7e2b57774ed591ca17364bd2ea5073d6ea45",
  "latency_ms": 14724
 },
 "S31sh9::signage": {
  "fp": "787ff5328f1d3235",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S31sh9": {
  "input_fingerprint": "3f208c0fbe18dbdf",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진의 수하 1이 몸을 움츠린 채 박철진의 귓가로 상체를 바짝 숙인 상태로 속삭이는 자세.\n\nLOCATION (lock): Near the doorway inside the administration building's illuminated third-floor interrogation room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the interrogation's restrained interior illumination and consistent contrast as the exchange contracts into a whisper.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor interrogation room remains brightly lit. Its door has just been opened for the incoming report, with no subsequent closing specified. 박철진: He remains at the interrogation position, visibly angry during the interrupted questioning. 박철진의 수하 1: He has entered the room and lowered his voice to deliver a private report.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상); 박철진의 수하 1 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진의 수하 1이 몸을 움츠린 채 박철진의 귓가로 상체를 바짝 숙인 상태로 속삭이는 자세.\n\nLOCATION (lock): Near the doorway inside the administration building's illuminated third-floor interrogation room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the interrogation's restrained interior illumination and consistent contrast as the exchange contracts into a whisper.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor interrogation room remains brightly lit. Its door has just been opened for the incoming report, with no subsequent closing specified. 박철진: He remains at the interrogation position, visibly angry during the interrupted questioning. 박철진의 수하 1: He has entered the room and lowered his voice to deliver a private report.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상); 박철진의 수하 1 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진의 수하 1이 몸을 움츠린 채 박철진의 귓가로 상체를 바짝 숙인 상태로 속삭이는 자세.\n\nLOCATION (lock): Near the doorway inside the administration building's illuminated third-floor interrogation room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the interrogation's restrained interior illumination and consistent contrast as the exchange contracts into a whisper.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor interrogation room remains brightly lit. Its door has just been opened for the incoming report, with no subsequent closing specified. 박철진: He remains at the interrogation position, visibly angry during the interrupted questioning. 박철진의 수하 1: He has entered the room and lowered his voice to deliver a private report.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상); 박철진의 수하 1 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "수하 1은 박철진의 귀를 향해 시선을 두고 속삭임. 박철진은 정면 아래를 응시함.",
    "built_space": "취조실 내부. 뒤쪽에 열린 문과 의자가 보이며 인물들은 방 안쪽에 치우쳐 위치함.",
    "entities": "박철진은 참조 이미지와 유사하나, 수하 1은 삭발로 묘사되어 참조 인물과 일치하지 않음.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥에 발을 딛고 안정적으로 자세를 유지하고 있음."
   },
   {
    "label": "B",
    "direction": "수하 1은 박철진의 귀에 밀착하여 시선을 향하고 속삭임. 박철진은 앞쪽을 응시함.",
    "built_space": "취조실 출입문 근처. 열린 문과 문틀이 보이며 인물들이 지시된 위치인 문가에 정확히 서 있음.",
    "entities": "박철진과 수하 1 모두 참조 이미지의 얼굴, 머리 모양, 복장과 매우 높게 일치함.",
    "hard_violations": [
     "[gpt-high] 박철진의 완장에 판독 가능한 한글과 영문자가 크게 노출되어, 읽을 수 있는 문자를 전혀 허용하지 않는 조건을 위반한다."
    ],
    "physics": "두 인물 모두 바닥에 발을 딛고 자연스러운 무게 중심으로 서서 기대고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 미디엄 샷 프레이밍을 정확히 따랐으며, 수하 1의 외모를 포함한 두 인물의 참조 일치도가 매우 높고 문가의 위치도 잘 구현되었습니다(단, 글자 비가독성 지침은 일부 위반됨)."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프레이밍이 미디엄 샷보다 지나치게 넓게 잡혔고, 수하 1이 삭발로 묘사되어 참조 이미지의 인물 정체성을 크게 훼손했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수하 1은 박철진의 귀를 향해 시선을 두고 속삭임. 박철진은 정면 아래를 응시함.",
        "built_space": "취조실 내부. 뒤쪽에 열린 문과 의자가 보이며 인물들은 방 안쪽에 치우쳐 위치함.",
        "entities": "박철진은 참조 이미지와 유사하나, 수하 1은 삭발로 묘사되어 참조 인물과 일치하지 않음.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 발을 딛고 안정적으로 자세를 유지하고 있음."
       },
       {
        "label": "B",
        "direction": "수하 1은 박철진의 귀에 밀착하여 시선을 향하고 속삭임. 박철진은 앞쪽을 응시함.",
        "built_space": "취조실 출입문 근처. 열린 문과 문틀이 보이며 인물들이 지시된 위치인 문가에 정확히 서 있음.",
        "entities": "박철진과 수하 1 모두 참조 이미지의 얼굴, 머리 모양, 복장과 매우 높게 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 발을 딛고 자연스러운 무게 중심으로 서서 기대고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 미디엄 샷 프레이밍을 정확히 따랐으며, 수하 1의 외모를 포함한 두 인물의 참조 일치도가 매우 높고 문가의 위치도 잘 구현되었습니다(단, 글자 비가독성 지침은 일부 위반됨)."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프레이밍이 미디엄 샷보다 지나치게 넓게 잡혔고, 수하 1이 삭발로 묘사되어 참조 이미지의 인물 정체성을 크게 훼손했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "수하 1은 박철진의 귀를 향해 시선을 두고 속삭임. 박철진은 정면 아래를 응시함.",
        "built_space": "취조실 내부. 뒤쪽에 열린 문과 의자가 보이며 인물들은 방 안쪽에 치우쳐 위치함.",
        "entities": "박철진은 참조 이미지와 유사하나, 수하 1은 삭발로 묘사되어 참조 인물과 일치하지 않음.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 발을 딛고 안정적으로 자세를 유지하고 있음."
       },
       {
        "label": "B",
        "direction": "수하 1은 박철진의 귀에 밀착하여 시선을 향하고 속삭임. 박철진은 앞쪽을 응시함.",
        "built_space": "취조실 출입문 근처. 열린 문과 문틀이 보이며 인물들이 지시된 위치인 문가에 정확히 서 있음.",
        "entities": "박철진과 수하 1 모두 참조 이미지의 얼굴, 머리 모양, 복장과 매우 높게 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 발을 딛고 자연스러운 무게 중심으로 서서 기대고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "미디엄 숏과 인물 외형은 비교적 잘 맞지만, 수하의 입이 귀보다 얼굴 정면을 향하고 완장의 선명한 글자가 문자 금지 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "수하가 몸을 낮춰 박철진의 귓가에 속삭이는 동작과 열린 문을 잘 구현했지만, 무릎 부근까지 넓어진 구도와 수하의 달라진 외형은 아쉽다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 수하는 오른쪽 박철진에게 상체를 기울이지만, 입은 보이는 귀가 아니라 코와 입 옆에 가깝다. 수하의 시선도 박철진의 눈 부근을 향해 귓속말보다 얼굴을 맞댄 대화로 읽힌다. 박철진은 수하 쪽을 곁눈질한다. 겨누는 무기나 이동하는 물체는 없다.",
        "built_space": "오른쪽에 열린 회색 문 한 개와 문틀, 왼쪽 뒤에 검은 의자 한 개, 천장에 형광등 기구 한 개, 오른쪽 벽에 스위치 한 개가 보인다. 얼룩진 밝은 벽과 짙은 하부 띠는 이전 장면과 유사하다. 두 사람은 문 바로 안쪽에 있으며 공간적으로 들어설 수 있는 배치다. 반사는 없다.",
        "entities": "성인 남성 두 명만 보인다. 박철진의 중년 얼굴, 뒤로 넘긴 검은색·회색 머리, 희끗한 수염과 올리브색 군복은 이전 장면에 가깝다. 수하의 짧은 검은 머리와 옅은 수염, 얼굴 윤곽은 인물 참고와 비교적 유사하지만 모자는 없고 군복 표식이 달라졌다. 박철진의 완장에는 큰 한글과 영문자가 명확히 보이며 수하 가슴에도 글자 표식이 있다.",
        "hard_violations": [
         "박철진의 완장에 판독 가능한 한글과 영문자가 크게 노출되어, 읽을 수 있는 문자를 전혀 허용하지 않는 조건을 위반한다."
        ],
        "physics": "두 사람의 하체는 화면 아래로 이어지며 서 있는 자세로 읽힌다. 발 접촉은 프레임 밖이다. 수하는 허리와 목을 굽혀 접근하고, 얼굴 사이에 올린 손은 굽힌 팔과 연결되어 있다. 몸이나 물건이 허공에 떠 있거나 관절이 불가능하게 꺾인 모습은 없다."
       },
       {
        "label": "B",
        "direction": "오른쪽 수하의 입과 입을 가린 손이 왼쪽 박철진의 귀 바로 옆을 향한다. 수하는 박철진의 옆얼굴 쪽을 보고, 박철진은 미간을 찌푸린 채 화면 앞쪽 아래를 응시한다. 귓가에 밀착해 비밀 보고를 하는 목표 관계가 분명하다. 무기나 이동 물체는 없다.",
        "built_space": "오른쪽에 열린 회색 문 한 개와 문틀, 왼쪽 뒤에 검은 의자 한 개, 오른쪽 벽에 스위치 한 개가 있다. 천장에는 켜진 형광등 기구 한 개와 직사각형 패널 두 개가 보인다. 문 너머에는 수납장과 사무기기가 일부 보인다. 얼룩진 벽과 짙은 하부 띠, 의자 재질은 이전 장면과 연결된다. 두 사람은 문 안쪽에서 의자 앞에 서 있으며 서로나 문과 충돌하지 않는다. 다만 인물을 무릎 부근까지 담아 지정된 미디엄 숏보다 넓다.",
        "entities": "성인 남성 두 명만 등장한다. 박철진은 중년의 얼굴, 뒤로 넘긴 머리, 회색 수염과 올리브색 군복을 유지하며 화난 표정도 보인다. 수하는 참고보다 머리를 훨씬 짧게 밀었고 얼굴 윤곽도 달라 동일 인물 재현은 약하다. 수하의 올리브색 군복, 검은 벨트와 허리 파우치는 참고와 대체로 맞는다. 작은 소매 표식은 있으나 확실히 판독되는 문구는 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "박철진은 다리를 벌린 상태에서 상체를 앞으로 기울이고, 수하는 무릎과 허리를 굽혀 귀 높이에 입을 맞춘다. 발은 프레임 밖이지만 보이는 다리와 골반의 연결은 서서 체중을 지탱하는 자세에 부합한다. 입을 가린 손은 굽힌 팔에 자연스럽게 연결되고 허리 파우치는 벨트에 고정되어 있다. 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "미디엄 숏과 인물 외형은 비교적 잘 맞지만, 수하의 입이 귀보다 얼굴 정면을 향하고 완장의 선명한 글자가 문자 금지 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "수하가 몸을 낮춰 박철진의 귓가에 속삭이는 동작과 열린 문을 잘 구현했지만, 무릎 부근까지 넓어진 구도와 수하의 달라진 외형은 아쉽다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 수하는 오른쪽 박철진에게 상체를 기울이지만, 입은 보이는 귀가 아니라 코와 입 옆에 가깝다. 수하의 시선도 박철진의 눈 부근을 향해 귓속말보다 얼굴을 맞댄 대화로 읽힌다. 박철진은 수하 쪽을 곁눈질한다. 겨누는 무기나 이동하는 물체는 없다.",
        "built_space": "오른쪽에 열린 회색 문 한 개와 문틀, 왼쪽 뒤에 검은 의자 한 개, 천장에 형광등 기구 한 개, 오른쪽 벽에 스위치 한 개가 보인다. 얼룩진 밝은 벽과 짙은 하부 띠는 이전 장면과 유사하다. 두 사람은 문 바로 안쪽에 있으며 공간적으로 들어설 수 있는 배치다. 반사는 없다.",
        "entities": "성인 남성 두 명만 보인다. 박철진의 중년 얼굴, 뒤로 넘긴 검은색·회색 머리, 희끗한 수염과 올리브색 군복은 이전 장면에 가깝다. 수하의 짧은 검은 머리와 옅은 수염, 얼굴 윤곽은 인물 참고와 비교적 유사하지만 모자는 없고 군복 표식이 달라졌다. 박철진의 완장에는 큰 한글과 영문자가 명확히 보이며 수하 가슴에도 글자 표식이 있다.",
        "hard_violations": [
         "박철진의 완장에 판독 가능한 한글과 영문자가 크게 노출되어, 읽을 수 있는 문자를 전혀 허용하지 않는 조건을 위반한다."
        ],
        "physics": "두 사람의 하체는 화면 아래로 이어지며 서 있는 자세로 읽힌다. 발 접촉은 프레임 밖이다. 수하는 허리와 목을 굽혀 접근하고, 얼굴 사이에 올린 손은 굽힌 팔과 연결되어 있다. 몸이나 물건이 허공에 떠 있거나 관절이 불가능하게 꺾인 모습은 없다."
       },
       {
        "label": "A",
        "direction": "오른쪽 수하의 입과 입을 가린 손이 왼쪽 박철진의 귀 바로 옆을 향한다. 수하는 박철진의 옆얼굴 쪽을 보고, 박철진은 미간을 찌푸린 채 화면 앞쪽 아래를 응시한다. 귓가에 밀착해 비밀 보고를 하는 목표 관계가 분명하다. 무기나 이동 물체는 없다.",
        "built_space": "오른쪽에 열린 회색 문 한 개와 문틀, 왼쪽 뒤에 검은 의자 한 개, 오른쪽 벽에 스위치 한 개가 있다. 천장에는 켜진 형광등 기구 한 개와 직사각형 패널 두 개가 보인다. 문 너머에는 수납장과 사무기기가 일부 보인다. 얼룩진 벽과 짙은 하부 띠, 의자 재질은 이전 장면과 연결된다. 두 사람은 문 안쪽에서 의자 앞에 서 있으며 서로나 문과 충돌하지 않는다. 다만 인물을 무릎 부근까지 담아 지정된 미디엄 숏보다 넓다.",
        "entities": "성인 남성 두 명만 등장한다. 박철진은 중년의 얼굴, 뒤로 넘긴 머리, 회색 수염과 올리브색 군복을 유지하며 화난 표정도 보인다. 수하는 참고보다 머리를 훨씬 짧게 밀었고 얼굴 윤곽도 달라 동일 인물 재현은 약하다. 수하의 올리브색 군복, 검은 벨트와 허리 파우치는 참고와 대체로 맞는다. 작은 소매 표식은 있으나 확실히 판독되는 문구는 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "박철진은 다리를 벌린 상태에서 상체를 앞으로 기울이고, 수하는 무릎과 허리를 굽혀 귀 높이에 입을 맞춘다. 발은 프레임 밖이지만 보이는 다리와 골반의 연결은 서서 체중을 지탱하는 자세에 부합한다. 입을 가린 손은 굽힌 팔에 자연스럽게 연결되고 허리 파우치는 벨트에 고정되어 있다. 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.286
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.036
   },
   "violations": {
    "B": [
     "[gpt-high] 박철진의 완장에 판독 가능한 한글과 영문자가 크게 노출되어, 읽을 수 있는 문자를 전혀 허용하지 않는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1036,
   "A": 1571
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1036,
    "verdict_ko": "지정된 미디엄 샷 프레이밍을 정확히 따랐으며, 수하 1의 외모를 포함한 두 인물의 참조 일치도가 매우 높고 문가의 위치도 잘 구현되었습니다(단, 글자 비가독성 지침은 일부 위반됨).  ★위반: [gpt-high] 박철진의 완장에 판독 가능한 한글과 영문자가 크게 노출되어, 읽을 수 있는 문자를 전혀 허용하지 않는 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "프레이밍이 미디엄 샷보다 지나치게 넓게 잡혔고, 수하 1이 삭발로 묘사되어 참조 이미지의 인물 정체성을 크게 훼손했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S31sh5_sel.png",
    "asset_id": "6eb2c4bd-11ac-4f8e-87ac-f36214c1d835",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진의 수하 1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:967362>",
    "asset_id": "49efd036-7f14-493a-b65e-cec03d0dc53c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb2-7415-7208-bdda-1efa118c841c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S31sh5"
  }
 },
 "S31sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:33:11.265408+00:00",
  "fingerprint": "62ef5e257f14ad1a7edc67f449147ecbff34d548f0df8d77aa6d97c6a2421446",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S31sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S31sh9_sel.png",
  "source_sha256": "d0ba9e3bff9a6cf3ec9b84a05fa3e7be16a8580b3c4c5c32c6f71eb557f42193",
  "file": "S31sh9_cine.png",
  "staged_sha256": "2edcfea5ee0ed2e5b4a42061b4d560e4ea9e7b375261a4f460ffda49bc9dc9d3",
  "latency_ms": 12029
 },
 "S32sh1::signage": {
  "fp": "8e2006b7a7947fef",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S32sh1": {
  "input_fingerprint": "b31e9e9c4d15c107",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 낡은 십자가가 걸린 지하실 안, 팔짱을 낀 이현우가 벽에 등을 기댄 채 구도환을 매섭게 노려보는 구도.\n\nLOCATION (lock): Against a wall of the church's sparse basement prayer room, with a cross nearby and the previously brightened lanterns providing light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Basement wall (이현우 rests his back against it) — Seen diagonally behind 이현우; used as Holds his resistant posture apart from the table conversation; Table (구도환 is seated beside it) — A portion of the top and near edge enters from the lower-left; used as Establishes the table-side camera position without overwhelming the human scale; Old cross (Hanging on the basement wall) — Its recognizable cross shape is visible obliquely against the wall; used as Provides a restrained setting cue above the confrontation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient interior illumination with readable faces and subdued contrast, without specifying a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the same basement finishes, lantern lighting, small organ, and sparse furniture. EXCLUDE transient lamp-activation flares and any objects from the upstairs entrance.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its lit lanterns, cross, small desk, chairs and harmonium. Charlie sits in his old coat and hat, with the activated radio available in the room; it is in his hands by his later spoken response. 이현우: He stands against the wall with folded arms, retaining facial bruising and the leg wound. His outer garment remains removed, and his contact card stays concealed in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 낡은 십자가가 걸린 지하실 안, 팔짱을 낀 이현우가 벽에 등을 기댄 채 구도환을 매섭게 노려보는 구도.\n\nLOCATION (lock): Against a wall of the church's sparse basement prayer room, with a cross nearby and the previously brightened lanterns providing light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Basement wall (이현우 rests his back against it) — Seen diagonally behind 이현우; used as Holds his resistant posture apart from the table conversation; Table (구도환 is seated beside it) — A portion of the top and near edge enters from the lower-left; used as Establishes the table-side camera position without overwhelming the human scale; Old cross (Hanging on the basement wall) — Its recognizable cross shape is visible obliquely against the wall; used as Provides a restrained setting cue above the confrontation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient interior illumination with readable faces and subdued contrast, without specifying a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the same basement finishes, lantern lighting, small organ, and sparse furniture. EXCLUDE transient lamp-activation flares and any objects from the upstairs entrance.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its lit lanterns, cross, small desk, chairs and harmonium. Charlie sits in his old coat and hat, with the activated radio available in the room; it is in his hands by his later spoken response. 이현우: He stands against the wall with folded arms, retaining facial bruising and the leg wound. His outer garment remains removed, and his contact card stays concealed in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 낡은 십자가가 걸린 지하실 안, 팔짱을 낀 이현우가 벽에 등을 기댄 채 구도환을 매섭게 노려보는 구도.\n\nLOCATION (lock): Against a wall of the church's sparse basement prayer room, with a cross nearby and the previously brightened lanterns providing light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Basement wall (이현우 rests his back against it) — Seen diagonally behind 이현우; used as Holds his resistant posture apart from the table conversation; Table (구도환 is seated beside it) — A portion of the top and near edge enters from the lower-left; used as Establishes the table-side camera position without overwhelming the human scale; Old cross (Hanging on the basement wall) — Its recognizable cross shape is visible obliquely against the wall; used as Provides a restrained setting cue above the confrontation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient interior illumination with readable faces and subdued contrast, without specifying a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the same basement finishes, lantern lighting, small organ, and sparse furniture. EXCLUDE transient lamp-activation flares and any objects from the upstairs entrance.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its lit lanterns, cross, small desk, chairs and harmonium. Charlie sits in his old coat and hat, with the activated radio available in the room; it is in his hands by his later spoken response. 이현우: He stands against the wall with folded arms, retaining facial bruising and the leg wound. His outer garment remains removed, and his contact card stays concealed in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 테이블에 앉은 구도환을 향해 시선을 던지고 있으며, 구도환도 이현우를 바라봄. 뒤에 앉은 유인원 사이보그는 정면 방향으로 고개를 돌리고 있음.",
    "built_space": "콘크리트 기둥과 벽으로 이루어진 지하실. 좌측 하단에 테이블이 위치하고, 왼쪽 벽면에 십자가가 걸려 있음. 뒤편에 작은 책상과 의자가 배치됨.",
    "entities": "이현우(레퍼런스와 일치하는 상의 탈의 상태)와 중년 남성인 구도환이 묘사됨. 샷 텍스트에 지시되지 않은 유인원 사이보그(찰리)가 무전기를 쥔 채 배경에 등장함.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 등장인물(유인원 사이보그) 포함",
     "[gpt-high] 숏 텍스트가 요구하지 않은 모자 쓴 제삼자를 중앙에 추가했고, 명시적인 제외 지시에도 이전 스틸의 비인간 얼굴을 다시 사용했다.",
     "[gpt-high] 모든 등장인물을 실제 인간으로 표현하라는 지시와 달리 중앙 인물의 얼굴이 금속성 비인간 형태다.",
     "[gpt-high] 이전 스틸의 다른 인물을 가져오지 말라는 지시와 달리 전경 중년 남자에게 이전 남자의 외형과 의상을 재사용했다.",
     "[gpt-high] 이현우를 지정된 벽이 아닌 별도의 지지 기둥에 기대어 배치했다."
    ],
    "physics": "이현우는 바닥을 딛고 서서 체중을 콘크리트 기둥에 기대어 지탱함. 구도환과 유인원은 의자에 앉아 안정적으로 무게를 지지함."
   },
   {
    "label": "B",
    "direction": "이현우는 팔짱을 낀 채 테이블 곁의 구도환을 매섭게 노려보고 있음. 구도환의 시선은 이현우가 아닌 카메라 좌측 프레임 밖을 향함.",
    "built_space": "낡은 지하실 기도실 공간. 좌측 하단에서 테이블이 진입하며, 벽면 중앙 윗부분에 십자가가 걸려 있음. 우측에는 지시문이 요구한 소형 풍금과 랜턴이 정확히 배치됨.",
    "entities": "이현우(상의 탈의, 찢어진 청바지, 다리 상처 묘사)와 구도환(테이블에 앉은 중년 남성)이 샷 텍스트에 맞게 등장함. 불필요한 추가 인물 없음.",
    "hard_violations": [
     "[gpt-high] 이전 스틸에서 이현우 외 인물의 얼굴·몸·의상을 가져오지 말라는 지시와 달리, 중앙 중년 남자의 외형을 탁자 옆 인물에 재사용했다.",
     "[gpt-high] 화면 맨 왼쪽에 숏 텍스트가 요구하지 않은 제삼자의 머리·어깨로 보이는 전경 인물이 추가되었다."
    ],
    "physics": "이현우는 두 발로 바닥을 딛고 벽에 등을 기대어 자연스럽게 서 있음. 구도환은 의자에 앉아 오른팔을 테이블 위에 올려 체중을 분산시킴."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "샷 텍스트가 요구한 카메라 구도와 지정된 인물(이현우, 구도환)만을 정확히 배치했으며, 이전 샷에서 가져와야 할 배경 요소(풍금, 십자가)를 훌륭하게 반영했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷 텍스트와 프레임 구성 지시문에 없는 유인원 사이보그를 화면에 포함시켜 '지정되지 않은 인물 추가 금지' 규칙을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 테이블에 앉은 구도환을 향해 시선을 던지고 있으며, 구도환도 이현우를 바라봄. 뒤에 앉은 유인원 사이보그는 정면 방향으로 고개를 돌리고 있음.",
        "built_space": "콘크리트 기둥과 벽으로 이루어진 지하실. 좌측 하단에 테이블이 위치하고, 왼쪽 벽면에 십자가가 걸려 있음. 뒤편에 작은 책상과 의자가 배치됨.",
        "entities": "이현우(레퍼런스와 일치하는 상의 탈의 상태)와 중년 남성인 구도환이 묘사됨. 샷 텍스트에 지시되지 않은 유인원 사이보그(찰리)가 무전기를 쥔 채 배경에 등장함.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 등장인물(유인원 사이보그) 포함"
        ],
        "physics": "이현우는 바닥을 딛고 서서 체중을 콘크리트 기둥에 기대어 지탱함. 구도환과 유인원은 의자에 앉아 안정적으로 무게를 지지함."
       },
       {
        "label": "B",
        "direction": "이현우는 팔짱을 낀 채 테이블 곁의 구도환을 매섭게 노려보고 있음. 구도환의 시선은 이현우가 아닌 카메라 좌측 프레임 밖을 향함.",
        "built_space": "낡은 지하실 기도실 공간. 좌측 하단에서 테이블이 진입하며, 벽면 중앙 윗부분에 십자가가 걸려 있음. 우측에는 지시문이 요구한 소형 풍금과 랜턴이 정확히 배치됨.",
        "entities": "이현우(상의 탈의, 찢어진 청바지, 다리 상처 묘사)와 구도환(테이블에 앉은 중년 남성)이 샷 텍스트에 맞게 등장함. 불필요한 추가 인물 없음.",
        "hard_violations": [],
        "physics": "이현우는 두 발로 바닥을 딛고 벽에 등을 기대어 자연스럽게 서 있음. 구도환은 의자에 앉아 오른팔을 테이블 위에 올려 체중을 분산시킴."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "샷 텍스트가 요구한 카메라 구도와 지정된 인물(이현우, 구도환)만을 정확히 배치했으며, 이전 샷에서 가져와야 할 배경 요소(풍금, 십자가)를 훌륭하게 반영했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷 텍스트와 프레임 구성 지시문에 없는 유인원 사이보그를 화면에 포함시켜 '지정되지 않은 인물 추가 금지' 규칙을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 테이블에 앉은 구도환을 향해 시선을 던지고 있으며, 구도환도 이현우를 바라봄. 뒤에 앉은 유인원 사이보그는 정면 방향으로 고개를 돌리고 있음.",
        "built_space": "콘크리트 기둥과 벽으로 이루어진 지하실. 좌측 하단에 테이블이 위치하고, 왼쪽 벽면에 십자가가 걸려 있음. 뒤편에 작은 책상과 의자가 배치됨.",
        "entities": "이현우(레퍼런스와 일치하는 상의 탈의 상태)와 중년 남성인 구도환이 묘사됨. 샷 텍스트에 지시되지 않은 유인원 사이보그(찰리)가 무전기를 쥔 채 배경에 등장함.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 등장인물(유인원 사이보그) 포함"
        ],
        "physics": "이현우는 바닥을 딛고 서서 체중을 콘크리트 기둥에 기대어 지탱함. 구도환과 유인원은 의자에 앉아 안정적으로 무게를 지지함."
       },
       {
        "label": "B",
        "direction": "이현우는 팔짱을 낀 채 테이블 곁의 구도환을 매섭게 노려보고 있음. 구도환의 시선은 이현우가 아닌 카메라 좌측 프레임 밖을 향함.",
        "built_space": "낡은 지하실 기도실 공간. 좌측 하단에서 테이블이 진입하며, 벽면 중앙 윗부분에 십자가가 걸려 있음. 우측에는 지시문이 요구한 소형 풍금과 랜턴이 정확히 배치됨.",
        "entities": "이현우(상의 탈의, 찢어진 청바지, 다리 상처 묘사)와 구도환(테이블에 앉은 중년 남성)이 샷 텍스트에 맞게 등장함. 불필요한 추가 인물 없음.",
        "hard_violations": [],
        "physics": "이현우는 두 발로 바닥을 딛고 벽에 등을 기대어 자연스럽게 서 있음. 구도환은 의자에 앉아 오른팔을 테이블 위에 올려 체중을 분산시킴."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "벽에 등을 댄 팔짱 자세와 전신 와이드 구도는 더 충실하지만, 이전 장면 인물의 재사용과 왼쪽 추가 인물, 불명확한 대치 시선이 중대한 위반이다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "두 사람의 대치 시선은 명확하지만, 금지된 비인간 인물을 다시 등장시키고 벽 대신 기둥에 기대게 하며 요구된 와이드 구도보다 타이트하게 담았다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 얼굴과 눈을 화면 왼쪽 전경으로 향한다. 시선이 탁자 옆 남자의 얼굴에 정확히 닿는지는 불명확하며, 그보다 카메라 쪽을 보는 인상이다. 앉은 남자도 이현우를 돌아보지 않고 카메라 왼쪽 전경을 본다. 무기나 이동하는 물체는 없다.",
        "built_space": "때 묻은 두 색 벽이 이현우 뒤로 비스듬히 이어지고, 십자가 한 개가 왼쪽 위에 걸려 있다. 탁자 한 개의 상판과 앞 모서리가 왼쪽 아래를 차지하지만 상판의 비중이 상당히 크다. 켜진 랜턴 두 개, 오른쪽 소형 오르간 한 대와 일부 걸상, 앉은 남자 뒤의 갈색 의자 등받이, 왼쪽 뒤 라디오 한 대가 보인다. 이전 장면의 낡은 재질과 조명 분위기는 대체로 이어진다.",
        "entities": "이현우는 젊은 한국계 남성으로 보이며 검은 헝클어진 머리, 마른 체격, 벗은 상체, 얼굴 멍, 낡은 청바지와 신발이 이전 모습에 대체로 부합한다. 다리에는 피 묻은 붕대가 보인다. 탁자 옆 중년 남자는 이전 장면 중앙 남자의 얼굴과 갈색 겉옷을 그대로 가져온 모습이다. 화면 맨 왼쪽에는 별도 인물의 흐릿한 머리와 어깨로 보이는 검은 전경이 추가되어 있다. 십자가와 오르간은 알아볼 수 있으며 판독 가능한 글자는 없다.",
        "hard_violations": [
         "이전 스틸에서 이현우 외 인물의 얼굴·몸·의상을 가져오지 말라는 지시와 달리, 중앙 중년 남자의 외형을 탁자 옆 인물에 재사용했다.",
         "화면 맨 왼쪽에 숏 텍스트가 요구하지 않은 제삼자의 머리·어깨로 보이는 전경 인물이 추가되었다."
        ],
        "physics": "이현우는 등과 어깨를 벽에 대고 두 신발을 바닥에 둔 채 다리를 약간 교차한다. 벽과 바닥이 체중을 지지하며 팔짱도 해부학적으로 가능하다. 중년 남자는 의자에 앉아 손을 탁자 위에 얹고 있다. 랜턴과 라디오는 가구 위에 놓여 있고 십자가는 벽에 고정되어 있다. 공중에 지지 없이 떠 있는 물체나 몸은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽 아래의 탁자 옆 중년 남자를 노려보고, 중년 남자는 고개를 들어 이현우를 바라본다. 두 사람의 대치 시선은 명확하다. 뒤쪽 모자 쓴 인물은 손에 든 안테나 달린 기기 쪽으로 얼굴을 숙이고 있다.",
        "built_space": "이현우는 연속된 지하실 벽이 아니라 화면 오른쪽의 독립된 사각 기둥에 등을 댄다. 왼쪽 벽에는 십자가 한 개가 있다. 전경 탁자 한 개가 왼쪽 아래로 들어오며, 뒤에는 작은 책상과 수납 가구가 보인다. 의자는 전경 남자의 자리, 중앙 인물의 자리, 왼쪽 뒤 빈 자리로 세 개가 식별된다. 랜턴 두 개가 켜져 있고 작은 오르간은 확인되지 않는다. 콘크리트 기둥과 벽의 재질은 이전 장소와 유사하지만 주인공의 벽 배치와 와이드 숏 요구에는 덜 맞는다.",
        "entities": "이현우는 검은 머리의 젊은 한국계 남성으로 보이며 상체를 벗고 팔짱을 끼고 있다. 얼굴 멍과 피 묻은 다리 붕대, 낡은 청바지가 보인다. 전경 중년 남자의 머리와 갈색 겉옷은 이전 장면의 다른 남자를 재사용한 모습이다. 중앙에는 이전 스틸의 모자 쓴 금속성 비인간 얼굴을 가진 인물이 코트를 입고 추가 등장한다. 그 손에는 소형 무전기처럼 보이는 기기가 있고, 전경 탁자에는 책 두 권이 놓여 있다. 판독 가능한 글자는 없다.",
        "hard_violations": [
         "숏 텍스트가 요구하지 않은 모자 쓴 제삼자를 중앙에 추가했고, 명시적인 제외 지시에도 이전 스틸의 비인간 얼굴을 다시 사용했다.",
         "모든 등장인물을 실제 인간으로 표현하라는 지시와 달리 중앙 인물의 얼굴이 금속성 비인간 형태다.",
         "이전 스틸의 다른 인물을 가져오지 말라는 지시와 달리 전경 중년 남자에게 이전 남자의 외형과 의상을 재사용했다.",
         "이현우를 지정된 벽이 아닌 별도의 지지 기둥에 기대어 배치했다."
        ],
        "physics": "이현우의 등과 어깨는 기둥에 닿고 다리는 아래로 뻗는다. 발은 프레임 밖이므로 접지 상태를 직접 확인할 수 없지만 공중에 떠 있는 자세는 아니다. 전경 남자와 중앙 인물은 각각 의자에 앉아 있으며, 중앙 인물의 기기는 손이 쥐고 있다. 책과 랜턴은 가구에 놓여 있다. 보이는 범위에서 지지 없는 부유나 불가능한 관절 자세는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "벽에 등을 댄 팔짱 자세와 전신 와이드 구도는 더 충실하지만, 이전 장면 인물의 재사용과 왼쪽 추가 인물, 불명확한 대치 시선이 중대한 위반이다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "두 사람의 대치 시선은 명확하지만, 금지된 비인간 인물을 다시 등장시키고 벽 대신 기둥에 기대게 하며 요구된 와이드 구도보다 타이트하게 담았다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 얼굴과 눈을 화면 왼쪽 전경으로 향한다. 시선이 탁자 옆 남자의 얼굴에 정확히 닿는지는 불명확하며, 그보다 카메라 쪽을 보는 인상이다. 앉은 남자도 이현우를 돌아보지 않고 카메라 왼쪽 전경을 본다. 무기나 이동하는 물체는 없다.",
        "built_space": "때 묻은 두 색 벽이 이현우 뒤로 비스듬히 이어지고, 십자가 한 개가 왼쪽 위에 걸려 있다. 탁자 한 개의 상판과 앞 모서리가 왼쪽 아래를 차지하지만 상판의 비중이 상당히 크다. 켜진 랜턴 두 개, 오른쪽 소형 오르간 한 대와 일부 걸상, 앉은 남자 뒤의 갈색 의자 등받이, 왼쪽 뒤 라디오 한 대가 보인다. 이전 장면의 낡은 재질과 조명 분위기는 대체로 이어진다.",
        "entities": "이현우는 젊은 한국계 남성으로 보이며 검은 헝클어진 머리, 마른 체격, 벗은 상체, 얼굴 멍, 낡은 청바지와 신발이 이전 모습에 대체로 부합한다. 다리에는 피 묻은 붕대가 보인다. 탁자 옆 중년 남자는 이전 장면 중앙 남자의 얼굴과 갈색 겉옷을 그대로 가져온 모습이다. 화면 맨 왼쪽에는 별도 인물의 흐릿한 머리와 어깨로 보이는 검은 전경이 추가되어 있다. 십자가와 오르간은 알아볼 수 있으며 판독 가능한 글자는 없다.",
        "hard_violations": [
         "이전 스틸에서 이현우 외 인물의 얼굴·몸·의상을 가져오지 말라는 지시와 달리, 중앙 중년 남자의 외형을 탁자 옆 인물에 재사용했다.",
         "화면 맨 왼쪽에 숏 텍스트가 요구하지 않은 제삼자의 머리·어깨로 보이는 전경 인물이 추가되었다."
        ],
        "physics": "이현우는 등과 어깨를 벽에 대고 두 신발을 바닥에 둔 채 다리를 약간 교차한다. 벽과 바닥이 체중을 지지하며 팔짱도 해부학적으로 가능하다. 중년 남자는 의자에 앉아 손을 탁자 위에 얹고 있다. 랜턴과 라디오는 가구 위에 놓여 있고 십자가는 벽에 고정되어 있다. 공중에 지지 없이 떠 있는 물체나 몸은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우는 왼쪽 아래의 탁자 옆 중년 남자를 노려보고, 중년 남자는 고개를 들어 이현우를 바라본다. 두 사람의 대치 시선은 명확하다. 뒤쪽 모자 쓴 인물은 손에 든 안테나 달린 기기 쪽으로 얼굴을 숙이고 있다.",
        "built_space": "이현우는 연속된 지하실 벽이 아니라 화면 오른쪽의 독립된 사각 기둥에 등을 댄다. 왼쪽 벽에는 십자가 한 개가 있다. 전경 탁자 한 개가 왼쪽 아래로 들어오며, 뒤에는 작은 책상과 수납 가구가 보인다. 의자는 전경 남자의 자리, 중앙 인물의 자리, 왼쪽 뒤 빈 자리로 세 개가 식별된다. 랜턴 두 개가 켜져 있고 작은 오르간은 확인되지 않는다. 콘크리트 기둥과 벽의 재질은 이전 장소와 유사하지만 주인공의 벽 배치와 와이드 숏 요구에는 덜 맞는다.",
        "entities": "이현우는 검은 머리의 젊은 한국계 남성으로 보이며 상체를 벗고 팔짱을 끼고 있다. 얼굴 멍과 피 묻은 다리 붕대, 낡은 청바지가 보인다. 전경 중년 남자의 머리와 갈색 겉옷은 이전 장면의 다른 남자를 재사용한 모습이다. 중앙에는 이전 스틸의 모자 쓴 금속성 비인간 얼굴을 가진 인물이 코트를 입고 추가 등장한다. 그 손에는 소형 무전기처럼 보이는 기기가 있고, 전경 탁자에는 책 두 권이 놓여 있다. 판독 가능한 글자는 없다.",
        "hard_violations": [
         "숏 텍스트가 요구하지 않은 모자 쓴 제삼자를 중앙에 추가했고, 명시적인 제외 지시에도 이전 스틸의 비인간 얼굴을 다시 사용했다.",
         "모든 등장인물을 실제 인간으로 표현하라는 지시와 달리 중앙 인물의 얼굴이 금속성 비인간 형태다.",
         "이전 스틸의 다른 인물을 가져오지 말라는 지시와 달리 전경 중년 남자에게 이전 남자의 외형과 의상을 재사용했다.",
         "이현우를 지정된 벽이 아닌 별도의 지지 기둥에 기대어 배치했다."
        ],
        "physics": "이현우의 등과 어깨는 기둥에 닿고 다리는 아래로 뻗는다. 발은 프레임 밖이므로 접지 상태를 직접 확인할 수 없지만 공중에 떠 있는 자세는 아니다. 전경 남자와 중앙 인물은 각각 의자에 앉아 있으며, 중앙 인물의 기기는 손이 쥐고 있다. 책과 랜턴은 가구에 놓여 있다. 보이는 범위에서 지지 없는 부유나 불가능한 관절 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.929,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.679,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 등장인물(유인원 사이보그) 포함",
     "[gpt-high] 숏 텍스트가 요구하지 않은 모자 쓴 제삼자를 중앙에 추가했고, 명시적인 제외 지시에도 이전 스틸의 비인간 얼굴을 다시 사용했다.",
     "[gpt-high] 모든 등장인물을 실제 인간으로 표현하라는 지시와 달리 중앙 인물의 얼굴이 금속성 비인간 형태다.",
     "[gpt-high] 이전 스틸의 다른 인물을 가져오지 말라는 지시와 달리 전경 중년 남자에게 이전 남자의 외형과 의상을 재사용했다.",
     "[gpt-high] 이현우를 지정된 벽이 아닌 별도의 지지 기둥에 기대어 배치했다."
    ],
    "B": [
     "[gpt-high] 이전 스틸에서 이현우 외 인물의 얼굴·몸·의상을 가져오지 말라는 지시와 달리, 중앙 중년 남자의 외형을 탁자 옆 인물에 재사용했다.",
     "[gpt-high] 화면 맨 왼쪽에 숏 텍스트가 요구하지 않은 제삼자의 머리·어깨로 보이는 전경 인물이 추가되었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 679
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "샷 텍스트가 요구한 카메라 구도와 지정된 인물(이현우, 구도환)만을 정확히 배치했으며, 이전 샷에서 가져와야 할 배경 요소(풍금, 십자가)를 훌륭하게 반영했습니다.  ★위반: [gpt-high] 이전 스틸에서 이현우 외 인물의 얼굴·몸·의상을 가져오지 말라는 지시와 달리, 중앙 중년 남자의 외형을 탁자 옆 인물에 재사용했다. / [gpt-high] 화면 맨 왼쪽에 숏 텍스트가 요구하지 않은 제삼자의 머리·어깨로 보이는 전경 인물이 추가되었다."
   },
   {
    "label": "A",
    "score": 679,
    "verdict_ko": "샷 텍스트와 프레임 구성 지시문에 없는 유인원 사이보그를 화면에 포함시켜 '지정되지 않은 인물 추가 금지' 규칙을 심각하게 위반했습니다.  ★위반: [gemini-pro] 샷 텍스트에 명시되지 않은 등장인물(유인원 사이보그) 포함 / [gpt-high] 숏 텍스트가 요구하지 않은 모자 쓴 제삼자를 중앙에 추가했고, 명시적인 제외 지시에도 이전 스틸의 비인간 얼굴을 다시 사용했다. / [gpt-high] 모든 등장인물을 실제 인간으로 표현하라는 지시와 달리 중앙 인물의 얼굴이 금속성 비인간 형태다. / [gpt-high] 이전 스틸의 다른 인물을 가져오지 말라는 지시와 달리 전경 중년 남자에게 이전 남자의 외형과 의상을 재사용했다. / [gpt-high] 이현우를 지정된 벽이 아닌 별도의 지지 기둥에 기대어 배치했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S30sh17_sel.png",
    "asset_id": "240e4910-1de1-46fb-856b-c29dbf83e886",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb3-39f6-7a1e-8011-c064fc18b7ae",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S30sh17"
  }
 },
 "S32sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:33:23.632495+00:00",
  "fingerprint": "f60d73219135495bba740367eab1f0106ad72eea58f205ae094d377d5a86dcdb",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S32sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S32sh1_sel.png",
  "source_sha256": "a8232708066781cc88bcdb5f4dc54d931357e2f722d2ac4f1a55431b576290f5",
  "file": "S32sh1_cine.png",
  "staged_sha256": "e27b697bf0cfd7de77812668fd7254bbf6c54c598aafd5d7c355180bfbe2141e",
  "latency_ms": 12133
 },
 "S32sh11::signage": {
  "fp": "fd592fd05c39dd67",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S32sh11": {
  "input_fingerprint": "31e930c5af24a003",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 신부의 양어깨를 두 손으로 단단히 움켜쥔 채 절박한 눈빛으로 호소하는 얼굴 클로즈업.\n\nLOCATION (lock): Near the exit from the lantern-lit church basement prayer room, before the concealed escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the basement's ambient illumination, using gentle facial separation to make the urgency readable without a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the basement wall finishes, cross, furnishings, and established nighttime lighting. EXCLUDE the technician and the girl from this private exchange, and do not introduce passageway fixtures.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement furnishings and lit lanterns remain unchanged. Charlie retains his old coat and hat and now holds the radio. 신부: He is on his feet with a flashlight, ready to lead the way into the secret passage, still wearing his clerical collar. 이현우: He has left his folded-arm position at the wall and moved to the exit, with facial bruises and the wounded leg. His outer garment remains removed and his contact card stays in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 신부의 양어깨를 두 손으로 단단히 움켜쥔 채 절박한 눈빛으로 호소하는 얼굴 클로즈업.\n\nLOCATION (lock): Near the exit from the lantern-lit church basement prayer room, before the concealed escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the basement's ambient illumination, using gentle facial separation to make the urgency readable without a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the basement wall finishes, cross, furnishings, and established nighttime lighting. EXCLUDE the technician and the girl from this private exchange, and do not introduce passageway fixtures.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement furnishings and lit lanterns remain unchanged. Charlie retains his old coat and hat and now holds the radio. 신부: He is on his feet with a flashlight, ready to lead the way into the secret passage, still wearing his clerical collar. 이현우: He has left his folded-arm position at the wall and moved to the exit, with facial bruises and the wounded leg. His outer garment remains removed and his contact card stays in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 신부의 양어깨를 두 손으로 단단히 움켜쥔 채 절박한 눈빛으로 호소하는 얼굴 클로즈업.\n\nLOCATION (lock): Near the exit from the lantern-lit church basement prayer room, before the concealed escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the basement's ambient illumination, using gentle facial separation to make the urgency readable without a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the basement wall finishes, cross, furnishings, and established nighttime lighting. EXCLUDE the technician and the girl from this private exchange, and do not introduce passageway fixtures.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement furnishings and lit lanterns remain unchanged. Charlie retains his old coat and hat and now holds the radio. 신부: He is on his feet with a flashlight, ready to lead the way into the secret passage, still wearing his clerical collar. 이현우: He has left his folded-arm position at the wall and moved to the exit, with facial bruises and the wounded leg. His outer garment remains removed and his contact card stays in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선이 신부의 얼굴을 똑바로 향하며, 신부의 손전등은 앞을 비추고 있음.",
    "built_space": "지하실 벽면과 십자가는 구현되었으나, 왼쪽에 있어야 할 탁자 대신 오른쪽과 동일한 오르간과 랜턴이 중복 생성됨.",
    "entities": "두 인물의 얼굴은 레퍼런스와 일치하나, 이현우가 탈의 상태가 아닌 셔츠를 입고 있음.",
    "hard_violations": [
     "[gemini-pro] 중복된 가구 (오르간이 양쪽에 두 개 생성됨)",
     "[gpt-high] 참고 장소에 한 대만 있던 오르간을 좌우 두 대로 복제하여 고정 가구의 수와 공간 구조를 바꿨다."
    ],
    "physics": "이현우가 신부의 어깨를 단단히 잡고 서 있는 자세와 체중 실림이 자연스럽게 지탱됨."
   },
   {
    "label": "B",
    "direction": "이현우가 신부를 바라보고 있으며, 신부가 든 손전등은 신부 자신의 얼굴을 향해 위로 비추고 있음.",
    "built_space": "십자가와 왼쪽 탁자 위의 라디오 등 레퍼런스의 공간 구성 요소와 위치를 정확히 유지함.",
    "entities": "인물들의 외형은 일치하나 이현우가 재킷을 입고 있음.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (신부의 손이 3개로 생성됨)"
    ],
    "physics": "신부가 양손으로 이현우의 어깨를 잡고 있음과 동시에 손전등을 들고 있어 손이 3개가 되는 물리적 모순이 발생함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "이현우가 신부의 어깨를 쥐고 호소하는 제시문의 핵심 행동을 잘 묘사했으나, 배경에 오르간이 중복 생성되고 탈의 상태가 유지되지 않은 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배경 공간의 디테일은 레퍼런스를 잘 따랐으나, 행동의 주체가 뒤바뀌었고 신부의 팔과 손이 3개로 묘사되는 치명적인 해부학적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 신부의 얼굴을 똑바로 향하며, 신부의 손전등은 앞을 비추고 있음.",
        "built_space": "지하실 벽면과 십자가는 구현되었으나, 왼쪽에 있어야 할 탁자 대신 오른쪽과 동일한 오르간과 랜턴이 중복 생성됨.",
        "entities": "두 인물의 얼굴은 레퍼런스와 일치하나, 이현우가 탈의 상태가 아닌 셔츠를 입고 있음.",
        "hard_violations": [
         "중복된 가구 (오르간이 양쪽에 두 개 생성됨)"
        ],
        "physics": "이현우가 신부의 어깨를 단단히 잡고 서 있는 자세와 체중 실림이 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "이현우가 신부를 바라보고 있으며, 신부가 든 손전등은 신부 자신의 얼굴을 향해 위로 비추고 있음.",
        "built_space": "십자가와 왼쪽 탁자 위의 라디오 등 레퍼런스의 공간 구성 요소와 위치를 정확히 유지함.",
        "entities": "인물들의 외형은 일치하나 이현우가 재킷을 입고 있음.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (신부의 손이 3개로 생성됨)"
        ],
        "physics": "신부가 양손으로 이현우의 어깨를 잡고 있음과 동시에 손전등을 들고 있어 손이 3개가 되는 물리적 모순이 발생함."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "이현우가 신부의 어깨를 쥐고 호소하는 제시문의 핵심 행동을 잘 묘사했으나, 배경에 오르간이 중복 생성되고 탈의 상태가 유지되지 않은 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배경 공간의 디테일은 레퍼런스를 잘 따랐으나, 행동의 주체가 뒤바뀌었고 신부의 팔과 손이 3개로 묘사되는 치명적인 해부학적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 신부의 얼굴을 똑바로 향하며, 신부의 손전등은 앞을 비추고 있음.",
        "built_space": "지하실 벽면과 십자가는 구현되었으나, 왼쪽에 있어야 할 탁자 대신 오른쪽과 동일한 오르간과 랜턴이 중복 생성됨.",
        "entities": "두 인물의 얼굴은 레퍼런스와 일치하나, 이현우가 탈의 상태가 아닌 셔츠를 입고 있음.",
        "hard_violations": [
         "중복된 가구 (오르간이 양쪽에 두 개 생성됨)"
        ],
        "physics": "이현우가 신부의 어깨를 단단히 잡고 서 있는 자세와 체중 실림이 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "이현우가 신부를 바라보고 있으며, 신부가 든 손전등은 신부 자신의 얼굴을 향해 위로 비추고 있음.",
        "built_space": "십자가와 왼쪽 탁자 위의 라디오 등 레퍼런스의 공간 구성 요소와 위치를 정확히 유지함.",
        "entities": "인물들의 외형은 일치하나 이현우가 재킷을 입고 있음.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (신부의 손이 3개로 생성됨)"
        ],
        "physics": "신부가 양손으로 이현우의 어깨를 잡고 있음과 동시에 손전등을 들고 있어 손이 3개가 되는 물리적 모순이 발생함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "얼굴 클로즈업과 절박한 시선은 더 충실하지만, 신부가 이현우의 양어깨를 잡는 역방향 동작과 다시 입힌 겉옷이 명확한 불일치입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "양손으로 신부의 양어깨를 잡는 동작은 맞지만, 오르간을 두 대로 복제한 공간 오류가 치명적이며 얼굴 클로즈업보다 넓게 촬영했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 오른쪽의 신부 얼굴을 올려다보며 입을 벌려 호소하고, 신부도 이현우 쪽으로 얼굴과 시선을 향한다. 그러나 검은 소매에서 나온 신부의 두 손이 이현우의 양어깨를 잡고 있다. 이현우의 손은 화면 오른쪽 신부 어깨 한 곳에서만 명확하게 확인된다. 손전등의 발광면은 왼쪽 아래, 이현우 쪽을 향한다.",
        "built_space": "뒤쪽 왼편에 벽 십자가 하나, 켜진 랜턴 하나, 라디오 하나, 이를 받치는 가구와 의자 일부가 보인다. 낡은 회벽과 따뜻한 지하실 조명은 참고 장소와 대체로 이어진다. 출구는 이 구도에서 확인되지 않으며, 오르간 등 나머지 가구는 화면 밖이다. 중복된 고정 설비나 반사면은 보이지 않는다.",
        "entities": "젊은 한국계 남성으로 보이는 이현우와 회색 머리의 백인 노년 남성 신부, 두 사람만 보인다. 이현우의 헝클어진 검은 머리와 얼굴 상처는 맞지만, 이전 장면의 맨상체와 달리 티셔츠와 지퍼 달린 겉옷을 입었다. 신부의 검은 성직복과 흰 칼라는 맞는다. 손전등은 보이며, 다리 부상과 신발 속 카드는 적절히 화면 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "어깨를 잡은 손들은 실제 몸과 옷에 접촉하고 있으며, 공중에 뜬 몸이나 물체는 보이지 않는다. 손전등 아래에는 받치는 손이 보이지만 턱과 팔에 가려져 어느 팔로 이어지는지는 명확하지 않다. 하체와 바닥은 클로즈업 밖이므로 발의 지지 상태는 판단할 수 없다. 손의 배치는 요청한 행동과 다르지만 확정적인 해부학적 불가능으로 보기는 어렵다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 신부의 눈을 바라보고, 두 팔을 뻗어 가까운 어깨와 반대편 어깨를 각각 잡는다. 신부 역시 이현우를 마주 본다. 요구한 상호 시선과 양어깨를 붙잡는 방향은 맞는다. 신부가 아래쪽에 든 손전등은 왼쪽 위, 이현우의 몸통 쪽을 비춘다.",
        "built_space": "벽 십자가 하나, 출입구 하나, 켜진 랜턴 두 개가 보인다. 그러나 건반과 목제 몸체를 갖춘 오르간이 화면 왼쪽과 오른쪽에 각각 한 대씩, 총 두 대 배치되어 있다. 참고 장소의 오른쪽 오르간 한 대를 복제한 공간이다. 왼쪽 오르간 앞에는 벤치도 보인다. 두 인물의 몸통과 주변 가구까지 드러나 얼굴 클로즈업보다 넓다.",
        "entities": "젊은 한국계 남성 이현우와 짧은 백발의 백인 노년 남성 신부만 등장한다. 이현우의 검은 머리와 얼굴 멍, 신부의 검은 성직복과 흰 칼라는 부합한다. 다만 이현우가 회색 긴소매 셔츠를 입어 이전 장면의 맨상체 상태를 유지하지 못했다. 신부의 손전등은 확인된다. 다리와 신발 속 카드가 보이지 않는 것은 이 구도의 결점이 아니다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참고 장소에 한 대만 있던 오르간을 좌우 두 대로 복제하여 고정 가구의 수와 공간 구조를 바꿨다."
        ],
        "physics": "이현우의 두 손이 신부의 양어깨에 닿고, 팔은 몸통에서 자연스럽게 이어진다. 신부의 손은 손전등 몸체를 잡아 지지한다. 인물들은 직립한 상체로 보이며 공중에 떠 있다는 증거는 없다. 발은 화면 밖이다. 랜턴들은 각각 가구 윗면에 놓여 있고, 벤치와 오르간도 정상적으로 바닥에 놓인 형태다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "얼굴 클로즈업과 절박한 시선은 더 충실하지만, 신부가 이현우의 양어깨를 잡는 역방향 동작과 다시 입힌 겉옷이 명확한 불일치입니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "양손으로 신부의 양어깨를 잡는 동작은 맞지만, 오르간을 두 대로 복제한 공간 오류가 치명적이며 얼굴 클로즈업보다 넓게 촬영했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 오른쪽의 신부 얼굴을 올려다보며 입을 벌려 호소하고, 신부도 이현우 쪽으로 얼굴과 시선을 향한다. 그러나 검은 소매에서 나온 신부의 두 손이 이현우의 양어깨를 잡고 있다. 이현우의 손은 화면 오른쪽 신부 어깨 한 곳에서만 명확하게 확인된다. 손전등의 발광면은 왼쪽 아래, 이현우 쪽을 향한다.",
        "built_space": "뒤쪽 왼편에 벽 십자가 하나, 켜진 랜턴 하나, 라디오 하나, 이를 받치는 가구와 의자 일부가 보인다. 낡은 회벽과 따뜻한 지하실 조명은 참고 장소와 대체로 이어진다. 출구는 이 구도에서 확인되지 않으며, 오르간 등 나머지 가구는 화면 밖이다. 중복된 고정 설비나 반사면은 보이지 않는다.",
        "entities": "젊은 한국계 남성으로 보이는 이현우와 회색 머리의 백인 노년 남성 신부, 두 사람만 보인다. 이현우의 헝클어진 검은 머리와 얼굴 상처는 맞지만, 이전 장면의 맨상체와 달리 티셔츠와 지퍼 달린 겉옷을 입었다. 신부의 검은 성직복과 흰 칼라는 맞는다. 손전등은 보이며, 다리 부상과 신발 속 카드는 적절히 화면 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "어깨를 잡은 손들은 실제 몸과 옷에 접촉하고 있으며, 공중에 뜬 몸이나 물체는 보이지 않는다. 손전등 아래에는 받치는 손이 보이지만 턱과 팔에 가려져 어느 팔로 이어지는지는 명확하지 않다. 하체와 바닥은 클로즈업 밖이므로 발의 지지 상태는 판단할 수 없다. 손의 배치는 요청한 행동과 다르지만 확정적인 해부학적 불가능으로 보기는 어렵다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 신부의 눈을 바라보고, 두 팔을 뻗어 가까운 어깨와 반대편 어깨를 각각 잡는다. 신부 역시 이현우를 마주 본다. 요구한 상호 시선과 양어깨를 붙잡는 방향은 맞는다. 신부가 아래쪽에 든 손전등은 왼쪽 위, 이현우의 몸통 쪽을 비춘다.",
        "built_space": "벽 십자가 하나, 출입구 하나, 켜진 랜턴 두 개가 보인다. 그러나 건반과 목제 몸체를 갖춘 오르간이 화면 왼쪽과 오른쪽에 각각 한 대씩, 총 두 대 배치되어 있다. 참고 장소의 오른쪽 오르간 한 대를 복제한 공간이다. 왼쪽 오르간 앞에는 벤치도 보인다. 두 인물의 몸통과 주변 가구까지 드러나 얼굴 클로즈업보다 넓다.",
        "entities": "젊은 한국계 남성 이현우와 짧은 백발의 백인 노년 남성 신부만 등장한다. 이현우의 검은 머리와 얼굴 멍, 신부의 검은 성직복과 흰 칼라는 부합한다. 다만 이현우가 회색 긴소매 셔츠를 입어 이전 장면의 맨상체 상태를 유지하지 못했다. 신부의 손전등은 확인된다. 다리와 신발 속 카드가 보이지 않는 것은 이 구도의 결점이 아니다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참고 장소에 한 대만 있던 오르간을 좌우 두 대로 복제하여 고정 가구의 수와 공간 구조를 바꿨다."
        ],
        "physics": "이현우의 두 손이 신부의 양어깨에 닿고, 팔은 몸통에서 자연스럽게 이어진다. 신부의 손은 손전등 몸체를 잡아 지지한다. 인물들은 직립한 상체로 보이며 공중에 떠 있다는 증거는 없다. 발은 화면 밖이다. 랜턴들은 각각 가구 윗면에 놓여 있고, 벤치와 오르간도 정상적으로 바닥에 놓인 형태다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.6
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.35
   },
   "violations": {
    "A": [
     "[gemini-pro] 중복된 가구 (오르간이 양쪽에 두 개 생성됨)",
     "[gpt-high] 참고 장소에 한 대만 있던 오르간을 좌우 두 대로 복제하여 고정 가구의 수와 공간 구조를 바꿨다."
    ],
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (신부의 손이 3개로 생성됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1350
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "이현우가 신부의 어깨를 쥐고 호소하는 제시문의 핵심 행동을 잘 묘사했으나, 배경에 오르간이 중복 생성되고 탈의 상태가 유지되지 않은 점이 감점 요인입니다.  ★위반: [gemini-pro] 중복된 가구 (오르간이 양쪽에 두 개 생성됨) / [gpt-high] 참고 장소에 한 대만 있던 오르간을 좌우 두 대로 복제하여 고정 가구의 수와 공간 구조를 바꿨다."
   },
   {
    "label": "B",
    "score": 1350,
    "verdict_ko": "배경 공간의 디테일은 레퍼런스를 잘 따랐으나, 행동의 주체가 뒤바뀌었고 신부의 팔과 손이 3개로 묘사되는 치명적인 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (신부의 손이 3개로 생성됨)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S32sh1_sel.png",
    "asset_id": "5f3fb20c-cb69-4f2a-8be1-cb534cb53d2f",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:672084>",
    "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb4-016e-72da-9c86-81dbb5875488",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S32sh1"
  }
 },
 "S32sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:33:36.099750+00:00",
  "fingerprint": "1b5bd3c5f851bf2c0a591db0fe686ec474ef609a6f2cbb9558b47fd661e7c980",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S32sh11_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S32sh11_sel.png",
  "source_sha256": "5b6caf8a7ccf28e8ccf83f13bc5194fa3e9409427c22115b0afe7064a78bbbed",
  "file": "S32sh11_cine.png",
  "staged_sha256": "5faa77543f4271a1fc2fed2ff9af335f7d039d640f216b74080a5ee08b8dd983",
  "latency_ms": 14484
 },
 "S32sh13::signage": {
  "fp": "8f608daac4eed9db",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S32sh13": {
  "input_fingerprint": "9789659c07d96f30",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 지하실 문 밖을 향해 한쪽 발로 바닥을 차고 맹렬히 뛰어오른 이현우의 역동적인 뒷모습.\n\nLOCATION (lock): At the interior doorway leading out of the dark church basement prayer room, with the group's flashlight available for the escape. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Basement doorway ahead of 이현우 in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Basement doorway (Open for 이현우's departure) — The opening is seen from inside along his route; used as Provides the destination above his advancing figure; Basement floor (Under his pushing foot and departure route); used as Grounds the push-off and connects it continuously to the threshold.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the basement dark as described while retaining sufficient tonal separation to read his back, driving leg, and exit opening.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the prayer-room walls, cross, sparse furniture, and existing lighting. EXCLUDE the priest, the girl, and the technician as they depart by the separate escape route.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its cross, desk, chairs, harmonium and lit lanterns as the secret passage is used for evacuation. Charlie retains his old coat and hat and the radio already in his hands. 이현우: He is departing to reach the management office, with facial bruises and an unhealed leg wound. His outer garment remains removed and his contact card stays inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 지하실 문 밖을 향해 한쪽 발로 바닥을 차고 맹렬히 뛰어오른 이현우의 역동적인 뒷모습.\n\nLOCATION (lock): At the interior doorway leading out of the dark church basement prayer room, with the group's flashlight available for the escape. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Basement doorway ahead of 이현우 in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Basement doorway (Open for 이현우's departure) — The opening is seen from inside along his route; used as Provides the destination above his advancing figure; Basement floor (Under his pushing foot and departure route); used as Grounds the push-off and connects it continuously to the threshold.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the basement dark as described while retaining sufficient tonal separation to read his back, driving leg, and exit opening.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the prayer-room walls, cross, sparse furniture, and existing lighting. EXCLUDE the priest, the girl, and the technician as they depart by the separate escape route.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its cross, desk, chairs, harmonium and lit lanterns as the secret passage is used for evacuation. Charlie retains his old coat and hat and the radio already in his hands. 이현우: He is departing to reach the management office, with facial bruises and an unhealed leg wound. His outer garment remains removed and his contact card stays inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 지하실 문 밖을 향해 한쪽 발로 바닥을 차고 맹렬히 뛰어오른 이현우의 역동적인 뒷모습.\n\nLOCATION (lock): At the interior doorway leading out of the dark church basement prayer room, with the group's flashlight available for the escape. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Basement doorway ahead of 이현우 in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Basement doorway (Open for 이현우's departure) — The opening is seen from inside along his route; used as Provides the destination above his advancing figure; Basement floor (Under his pushing foot and departure route); used as Grounds the push-off and connects it continuously to the threshold.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the basement dark as described while retaining sufficient tonal separation to read his back, driving leg, and exit opening.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the prayer-room walls, cross, sparse furniture, and existing lighting. EXCLUDE the priest, the girl, and the technician as they depart by the separate escape route.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its cross, desk, chairs, harmonium and lit lanterns as the secret passage is used for evacuation. Charlie retains his old coat and hat and the radio already in his hands. 이현우: He is departing to reach the management office, with facial bruises and an unhealed leg wound. His outer garment remains removed and his contact card stays inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물은 정면의 열린 문을 향해 나아가는 자세를 취하고 있으며, 오른손에 쥐어진 손전등의 불빛이 문 너머 계단을 정확히 비추고 있다.",
    "built_space": "레퍼런스와 동일한 투톤 색상의 콘크리트 벽면, 나무 십자가, 라디오와 랜턴이 놓인 책상이 완벽히 재현되었다. 프롬프트에 명시된 하모늄(풍금)과 랜턴이 올려진 책상들이 적절히 배치되었으며, 열린 문 밖으로는 탈출 경로인 계단이 위로 향해 있다.",
    "entities": "헝클어진 검은 머리에 티셔츠와 청바지를 입은 이현우의 뒷모습으로, 겉옷을 벗었다는 설정 및 인물 레퍼런스와 훌륭하게 일치한다.",
    "hard_violations": [
     "[gemini-pro] 바닥을 지탱하거나 밀어내는 접촉점 없이 인물이 공중에 수평으로 떠 있음 (물리적 지지 기반 부재)"
    ],
    "physics": "인물이 역동적으로 달리는 자세를 취하고 있으나, 바닥을 차고 있어야 할 오른발 끝이 바닥에서 완전히 떨어져 그림자와 분리되어 있다. 발사나 착지의 궤적이 느껴지지 않은 채 허공에 떠 있으므로 아무것도 인물을 지탱하지 않는다 (nothing supports it)."
   },
   {
    "label": "B",
    "direction": "인물은 정면의 열린 문을 향해 이동하는 방향성을 띠고 있다.",
    "built_space": "벽면이 레퍼런스의 콘크리트가 아닌 나무 판자로 덮여 있어 이전 샷의 장소 특징을 완전히 무시했다. 십자가와 책상은 존재하나 하모늄이나 추가 랜턴 등의 세부 요소가 누락되었다.",
    "entities": "헝클어진 검은 머리에 어두운 색 긴팔 셔츠와 청바지를 입은 이현우의 뒷모습이다.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스와 전혀 다른 나무 판자 벽면 (장소 잠금 위반)",
     "[gemini-pro] 바닥을 지탱하거나 밀어내는 접촉점 없이 인물이 공중에 떠 있음 (물리적 지지 기반 부재)"
    ],
    "physics": "인물의 양발이 모두 바닥에서 높이 떨어져 있어 아무것도 인물을 지탱하지 않는다 (nothing supports it). 쥐고 있어야 할 손전등은 인물의 손을 벗어나 바닥에 떨어져 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "레퍼런스의 장소 묘사와 프롬프트가 요구한 소품(하모늄, 랜턴 등)을 훌륭하게 구현했으나, 바닥을 밀어내야 할 발이 허공에 수평으로 떠 있는 치명적인 물리 연출 위반으로 인해 최종 탈락함."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "인물이 지지대 없이 완전히 허공에 떠 있을 뿐만 아니라, 지하실 벽면 재질이 나무 판자로 바뀌어 장소 일관성을 심각하게 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 정면의 열린 문을 향해 나아가는 자세를 취하고 있으며, 오른손에 쥐어진 손전등의 불빛이 문 너머 계단을 정확히 비추고 있다.",
        "built_space": "레퍼런스와 동일한 투톤 색상의 콘크리트 벽면, 나무 십자가, 라디오와 랜턴이 놓인 책상이 완벽히 재현되었다. 프롬프트에 명시된 하모늄(풍금)과 랜턴이 올려진 책상들이 적절히 배치되었으며, 열린 문 밖으로는 탈출 경로인 계단이 위로 향해 있다.",
        "entities": "헝클어진 검은 머리에 티셔츠와 청바지를 입은 이현우의 뒷모습으로, 겉옷을 벗었다는 설정 및 인물 레퍼런스와 훌륭하게 일치한다.",
        "hard_violations": [
         "바닥을 지탱하거나 밀어내는 접촉점 없이 인물이 공중에 수평으로 떠 있음 (물리적 지지 기반 부재)"
        ],
        "physics": "인물이 역동적으로 달리는 자세를 취하고 있으나, 바닥을 차고 있어야 할 오른발 끝이 바닥에서 완전히 떨어져 그림자와 분리되어 있다. 발사나 착지의 궤적이 느껴지지 않은 채 허공에 떠 있으므로 아무것도 인물을 지탱하지 않는다 (nothing supports it)."
       },
       {
        "label": "B",
        "direction": "인물은 정면의 열린 문을 향해 이동하는 방향성을 띠고 있다.",
        "built_space": "벽면이 레퍼런스의 콘크리트가 아닌 나무 판자로 덮여 있어 이전 샷의 장소 특징을 완전히 무시했다. 십자가와 책상은 존재하나 하모늄이나 추가 랜턴 등의 세부 요소가 누락되었다.",
        "entities": "헝클어진 검은 머리에 어두운 색 긴팔 셔츠와 청바지를 입은 이현우의 뒷모습이다.",
        "hard_violations": [
         "레퍼런스와 전혀 다른 나무 판자 벽면 (장소 잠금 위반)",
         "바닥을 지탱하거나 밀어내는 접촉점 없이 인물이 공중에 떠 있음 (물리적 지지 기반 부재)"
        ],
        "physics": "인물의 양발이 모두 바닥에서 높이 떨어져 있어 아무것도 인물을 지탱하지 않는다 (nothing supports it). 쥐고 있어야 할 손전등은 인물의 손을 벗어나 바닥에 떨어져 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "레퍼런스의 장소 묘사와 프롬프트가 요구한 소품(하모늄, 랜턴 등)을 훌륭하게 구현했으나, 바닥을 밀어내야 할 발이 허공에 수평으로 떠 있는 치명적인 물리 연출 위반으로 인해 최종 탈락함."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "인물이 지지대 없이 완전히 허공에 떠 있을 뿐만 아니라, 지하실 벽면 재질이 나무 판자로 바뀌어 장소 일관성을 심각하게 위반함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 정면의 열린 문을 향해 나아가는 자세를 취하고 있으며, 오른손에 쥐어진 손전등의 불빛이 문 너머 계단을 정확히 비추고 있다.",
        "built_space": "레퍼런스와 동일한 투톤 색상의 콘크리트 벽면, 나무 십자가, 라디오와 랜턴이 놓인 책상이 완벽히 재현되었다. 프롬프트에 명시된 하모늄(풍금)과 랜턴이 올려진 책상들이 적절히 배치되었으며, 열린 문 밖으로는 탈출 경로인 계단이 위로 향해 있다.",
        "entities": "헝클어진 검은 머리에 티셔츠와 청바지를 입은 이현우의 뒷모습으로, 겉옷을 벗었다는 설정 및 인물 레퍼런스와 훌륭하게 일치한다.",
        "hard_violations": [
         "바닥을 지탱하거나 밀어내는 접촉점 없이 인물이 공중에 수평으로 떠 있음 (물리적 지지 기반 부재)"
        ],
        "physics": "인물이 역동적으로 달리는 자세를 취하고 있으나, 바닥을 차고 있어야 할 오른발 끝이 바닥에서 완전히 떨어져 그림자와 분리되어 있다. 발사나 착지의 궤적이 느껴지지 않은 채 허공에 떠 있으므로 아무것도 인물을 지탱하지 않는다 (nothing supports it)."
       },
       {
        "label": "B",
        "direction": "인물은 정면의 열린 문을 향해 이동하는 방향성을 띠고 있다.",
        "built_space": "벽면이 레퍼런스의 콘크리트가 아닌 나무 판자로 덮여 있어 이전 샷의 장소 특징을 완전히 무시했다. 십자가와 책상은 존재하나 하모늄이나 추가 랜턴 등의 세부 요소가 누락되었다.",
        "entities": "헝클어진 검은 머리에 어두운 색 긴팔 셔츠와 청바지를 입은 이현우의 뒷모습이다.",
        "hard_violations": [
         "레퍼런스와 전혀 다른 나무 판자 벽면 (장소 잠금 위반)",
         "바닥을 지탱하거나 밀어내는 접촉점 없이 인물이 공중에 떠 있음 (물리적 지지 기반 부재)"
        ],
        "physics": "인물의 양발이 모두 바닥에서 높이 떨어져 있어 아무것도 인물을 지탱하지 않는다 (nothing supports it). 쥐고 있어야 할 손전등은 인물의 손을 벗어나 바닥에 떨어져 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "출구로 도약하는 뒷모습과 와이드 구도는 맞지만, 기도실의 회벽을 목재 벽으로 바꾸고 기존 등불과 의상 연속성을 충분히 살리지 못했다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "상단 중앙 출구를 향한 강한 도약과 연속된 이동 바닥을 보여주며, 낡은 회벽·십자가·등불·풍금 및 겉옷을 벗은 상태를 더 충실히 유지했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리와 등, 몸의 진행 방향이 상단 중앙의 열린 문을 향한다. 얼굴과 눈은 보이지 않아 실제 시선은 확인할 수 없다. 문 오른쪽 바닥에 놓인 손전등은 출구 쪽 벽면을 비추며, 손에 들고 이동 경로를 조명하는 상태는 아니다.",
        "built_space": "실내의 낮은 시점에서 열린 출입구 하나를 바라본다. 왼쪽 책상 하나, 양쪽 가장자리의 의자 네 개가 일부 포함되고, 문 오른쪽 벽에 십자가 하나가 있다. 바닥은 문턱까지 끊김 없이 연결된다. 다만 참고의 얼룩진 밝은 회벽과 회색 하단 도장 대신 벽 전체가 세로 목재 판벽으로 바뀌었다. 풍금과 켜진 등불은 보이지 않는다.",
        "entities": "인물은 한 명뿐이며 헝클어진 검은 머리와 마른 청년 체형은 이현우의 참고에 부합한다. 뒷모습이므로 한국계 얼굴과 정확한 나이, 얼굴 멍은 확인할 수 없다. 어두운 긴소매 상의와 청바지를 입었고 겉재킷은 없다. 긴소매 상의는 이전 사진에서 드러난 안쪽 티셔츠와 차이가 있다. 다리의 미치유 상처와 신발 속 카드는 보이지 않는다. 바닥에 손전등 하나가 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 발 모두 바닥에서 떨어져 있다. 한쪽 다리는 뒤로 굽고 다른 쪽은 출구 방향으로 나가 있어, 바로 앞 바닥을 차고 문턱으로 향하는 짧은 도약으로 성립한다. 실제 발의 지면 접촉 순간은 없지만 추진 직후의 자세와 착지 가능한 문 안쪽 바닥이 있어 근거 없는 부유로 보이지는 않는다. 손전등은 바닥에, 책상과 의자는 각 다리로 바닥에 지지된다."
       },
       {
        "label": "B",
        "direction": "머리와 상체가 상단 중앙의 열린 출입구 및 그 너머 계단을 향한다. 오른손의 손전등도 앞으로 향해 계단을 비춘다. 눈은 뒷모습에 가려져 있지만 몸의 이동 목표는 출구로 분명하다.",
        "built_space": "기도실 안에서 출입구 하나를 정면에 가깝게 본다. 문은 오른쪽으로 열려 있고 그 너머 계단은 위로 이어진다. 실내 바닥에서 문 앞 낮은 단과 계단까지 이동 경로가 연결된다. 왼쪽 벽의 십자가 하나, 왼쪽 책상 하나와 풍금 하나, 오른쪽 긴 책상 하나, 양쪽의 의자 네 개가 보이며 등불 두 개가 켜져 있다. 얼룩진 회벽과 회색 하단 도장, 따뜻한 등불은 이전 사진의 장소 특성을 잘 유지한다.",
        "entities": "검은 헝클어진 머리와 마른 청년 체형의 인물 한 명만 등장한다. 얼굴이 가려져 한국계 얼굴의 세부와 얼굴 멍은 판정할 수 없다. 겉옷 없이 낡은 어두운 반소매 티셔츠, 해진 청바지, 운동화를 착용했다. 굽힌 다리의 찢어진 무릎 부근에 붉은 피부가 드러난다. 손전등은 오른손에 있고, 십자가·풍금·책상·의자·등불도 보인다. 왼쪽 책상에는 라디오가 있다. 오른쪽 위 게시물에는 작은 글줄 모양이 있으나 내용을 판독할 수 없으며, 신발 속 카드는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 발이 공중에 있지만, 아래로 길게 뻗은 다리는 바닥을 밀어낸 직후의 추진 다리로 읽히고 반대쪽 다리는 무릎을 강하게 접었다. 앞으로 기울어진 상체와 벌어진 팔이 출구로 뛰어오르는 동작에 맞는다. 출발 바닥과 착지할 문 앞 단이 모두 보여 현실적인 도약 궤적이 성립한다. 손전등은 오른손이 직접 쥐고 있으며 가구와 등불은 바닥 또는 책상에 지지되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "출구로 도약하는 뒷모습과 와이드 구도는 맞지만, 기도실의 회벽을 목재 벽으로 바꾸고 기존 등불과 의상 연속성을 충분히 살리지 못했다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "상단 중앙 출구를 향한 강한 도약과 연속된 이동 바닥을 보여주며, 낡은 회벽·십자가·등불·풍금 및 겉옷을 벗은 상태를 더 충실히 유지했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "머리와 등, 몸의 진행 방향이 상단 중앙의 열린 문을 향한다. 얼굴과 눈은 보이지 않아 실제 시선은 확인할 수 없다. 문 오른쪽 바닥에 놓인 손전등은 출구 쪽 벽면을 비추며, 손에 들고 이동 경로를 조명하는 상태는 아니다.",
        "built_space": "실내의 낮은 시점에서 열린 출입구 하나를 바라본다. 왼쪽 책상 하나, 양쪽 가장자리의 의자 네 개가 일부 포함되고, 문 오른쪽 벽에 십자가 하나가 있다. 바닥은 문턱까지 끊김 없이 연결된다. 다만 참고의 얼룩진 밝은 회벽과 회색 하단 도장 대신 벽 전체가 세로 목재 판벽으로 바뀌었다. 풍금과 켜진 등불은 보이지 않는다.",
        "entities": "인물은 한 명뿐이며 헝클어진 검은 머리와 마른 청년 체형은 이현우의 참고에 부합한다. 뒷모습이므로 한국계 얼굴과 정확한 나이, 얼굴 멍은 확인할 수 없다. 어두운 긴소매 상의와 청바지를 입었고 겉재킷은 없다. 긴소매 상의는 이전 사진에서 드러난 안쪽 티셔츠와 차이가 있다. 다리의 미치유 상처와 신발 속 카드는 보이지 않는다. 바닥에 손전등 하나가 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 발 모두 바닥에서 떨어져 있다. 한쪽 다리는 뒤로 굽고 다른 쪽은 출구 방향으로 나가 있어, 바로 앞 바닥을 차고 문턱으로 향하는 짧은 도약으로 성립한다. 실제 발의 지면 접촉 순간은 없지만 추진 직후의 자세와 착지 가능한 문 안쪽 바닥이 있어 근거 없는 부유로 보이지는 않는다. 손전등은 바닥에, 책상과 의자는 각 다리로 바닥에 지지된다."
       },
       {
        "label": "A",
        "direction": "머리와 상체가 상단 중앙의 열린 출입구 및 그 너머 계단을 향한다. 오른손의 손전등도 앞으로 향해 계단을 비춘다. 눈은 뒷모습에 가려져 있지만 몸의 이동 목표는 출구로 분명하다.",
        "built_space": "기도실 안에서 출입구 하나를 정면에 가깝게 본다. 문은 오른쪽으로 열려 있고 그 너머 계단은 위로 이어진다. 실내 바닥에서 문 앞 낮은 단과 계단까지 이동 경로가 연결된다. 왼쪽 벽의 십자가 하나, 왼쪽 책상 하나와 풍금 하나, 오른쪽 긴 책상 하나, 양쪽의 의자 네 개가 보이며 등불 두 개가 켜져 있다. 얼룩진 회벽과 회색 하단 도장, 따뜻한 등불은 이전 사진의 장소 특성을 잘 유지한다.",
        "entities": "검은 헝클어진 머리와 마른 청년 체형의 인물 한 명만 등장한다. 얼굴이 가려져 한국계 얼굴의 세부와 얼굴 멍은 판정할 수 없다. 겉옷 없이 낡은 어두운 반소매 티셔츠, 해진 청바지, 운동화를 착용했다. 굽힌 다리의 찢어진 무릎 부근에 붉은 피부가 드러난다. 손전등은 오른손에 있고, 십자가·풍금·책상·의자·등불도 보인다. 왼쪽 책상에는 라디오가 있다. 오른쪽 위 게시물에는 작은 글줄 모양이 있으나 내용을 판독할 수 없으며, 신발 속 카드는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 발이 공중에 있지만, 아래로 길게 뻗은 다리는 바닥을 밀어낸 직후의 추진 다리로 읽히고 반대쪽 다리는 무릎을 강하게 접었다. 앞으로 기울어진 상체와 벌어진 팔이 출구로 뛰어오르는 동작에 맞는다. 출발 바닥과 착지할 문 앞 단이 모두 보여 현실적인 도약 궤적이 성립한다. 손전등은 오른손이 직접 쥐고 있으며 가구와 등불은 바닥 또는 책상에 지지되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.889
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.639
   },
   "violations": {
    "A": [
     "[gemini-pro] 바닥을 지탱하거나 밀어내는 접촉점 없이 인물이 공중에 수평으로 떠 있음 (물리적 지지 기반 부재)"
    ],
    "B": [
     "[gemini-pro] 레퍼런스와 전혀 다른 나무 판자 벽면 (장소 잠금 위반)",
     "[gemini-pro] 바닥을 지탱하거나 밀어내는 접촉점 없이 인물이 공중에 떠 있음 (물리적 지지 기반 부재)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 639
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "레퍼런스의 장소 묘사와 프롬프트가 요구한 소품(하모늄, 랜턴 등)을 훌륭하게 구현했으나, 바닥을 밀어내야 할 발이 허공에 수평으로 떠 있는 치명적인 물리 연출 위반으로 인해 최종 탈락함.  ★위반: [gemini-pro] 바닥을 지탱하거나 밀어내는 접촉점 없이 인물이 공중에 수평으로 떠 있음 (물리적 지지 기반 부재)"
   },
   {
    "label": "B",
    "score": 639,
    "verdict_ko": "인물이 지지대 없이 완전히 허공에 떠 있을 뿐만 아니라, 지하실 벽면 재질이 나무 판자로 바뀌어 장소 일관성을 심각하게 위반함.  ★위반: [gemini-pro] 레퍼런스와 전혀 다른 나무 판자 벽면 (장소 잠금 위반) / [gemini-pro] 바닥을 지탱하거나 밀어내는 접촉점 없이 인물이 공중에 떠 있음 (물리적 지지 기반 부재)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S32sh11_sel.png",
    "asset_id": "b8b5e747-a468-4398-b1ec-4f019bf18317",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb4-ee89-7b62-8725-e9aa8905c851",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S32sh11"
  }
 },
 "S32sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:33:50.917626+00:00",
  "fingerprint": "2868e34eb1e029bf379ba08521f823150c7db386e93b60a2c7db9018d080209f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S32sh13_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S32sh13_sel.png",
  "source_sha256": "c8ca6a0415a335055af100fc1560ea2fb6f571610b5f43cfbcbe96526729dc40",
  "file": "S32sh13_cine.png",
  "staged_sha256": "e192cb21d9a223442997e21f3e880619b93ee93110b4be546af3a721762f1123",
  "latency_ms": 11043
 },
 "S33sh1::signage": {
  "fp": "7aba4eb2a3d2a4f9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S33sh1": {
  "input_fingerprint": "f143a831c6832e51",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 위, 거대한 전투 헬기가 눈부신 서치라이트를 난민촌 아래로 강하게 쏘아 내리는 웅장한 전경.\n\nLOCATION (lock): In the open night sky above the refugee settlement and its church, overlooking the streets below. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Combat helicopter (Airborne above the refugee camp) — Its underside and one side are visible obliquely from the road; used as Provides the airborne threat while remaining proportionate to the surrounding sky; Refugee camp and internal road (Below the hovering helicopter) — The road-level foreground leads toward the settlement beneath the aircraft; used as Supplies a grounded scale reference across the bottom of the image; Night sky (Dark); used as Provides negative space around the aircraft.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dazzling downward searchlight contrasts sharply with the dark night sky and the camp below.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A combat helicopter is above the settlement, while militia vehicles approach the church and a separate escaping convoy travels along the seawall road. The seawall is still standing but retains its severe cracks and seepage; the management building remains brightly lit.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 위, 거대한 전투 헬기가 눈부신 서치라이트를 난민촌 아래로 강하게 쏘아 내리는 웅장한 전경.\n\nLOCATION (lock): In the open night sky above the refugee settlement and its church, overlooking the streets below. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Combat helicopter (Airborne above the refugee camp) — Its underside and one side are visible obliquely from the road; used as Provides the airborne threat while remaining proportionate to the surrounding sky; Refugee camp and internal road (Below the hovering helicopter) — The road-level foreground leads toward the settlement beneath the aircraft; used as Supplies a grounded scale reference across the bottom of the image; Night sky (Dark); used as Provides negative space around the aircraft.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dazzling downward searchlight contrasts sharply with the dark night sky and the camp below.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A combat helicopter is above the settlement, while militia vehicles approach the church and a separate escaping convoy travels along the seawall road. The seawall is still standing but retains its severe cracks and seepage; the management building remains brightly lit.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 위, 거대한 전투 헬기가 눈부신 서치라이트를 난민촌 아래로 강하게 쏘아 내리는 웅장한 전경.\n\nLOCATION (lock): In the open night sky above the refugee settlement and its church, overlooking the streets below. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Combat helicopter (Airborne above the refugee camp) — Its underside and one side are visible obliquely from the road; used as Provides the airborne threat while remaining proportionate to the surrounding sky; Refugee camp and internal road (Below the hovering helicopter) — The road-level foreground leads toward the settlement beneath the aircraft; used as Supplies a grounded scale reference across the bottom of the image; Night sky (Dark); used as Provides negative space around the aircraft.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dazzling downward searchlight contrasts sharply with the dark night sky and the camp below.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A combat helicopter is above the settlement, while militia vehicles approach the church and a separate escaping convoy travels along the seawall road. The seawall is still standing but retains its severe cracks and seepage; the management building remains brightly lit.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "공중의 헬기에서 뻗어나온 서치라이트가 아래의 난민촌과 교회 주변을 정확히 쏘아 내리고 있다.",
    "built_space": "전면에 넓은 내부 도로가 교회와 난민촌으로 이어지며, 우측에는 심하게 갈라진 방파제와 그 위를 지나는 차량 행렬이 배치되어 있다.",
    "entities": "공격형 전투 헬기, 첨탑이 있는 교회, 난민촌의 판잣집들, 균열이 있는 방파제, 이동 중인 차량 행렬이 확인된다.",
    "hard_violations": [
     "[gpt-high] 사람이 등장하지 않는 샷 텍스트와 인물 추가 금지 지시를 어기고, 교회 주변 및 난민촌 도로에 여러 인물을 임의로 추가했습니다."
    ],
    "physics": "헬기는 정상적으로 공중에 체공 중이며, 서치라이트 빛줄기가 물리적으로 자연스럽게 지면을 향한다."
   },
   {
    "label": "B",
    "direction": "헬기의 서치라이트가 난민촌 중앙의 좁은 길과 텐트들을 향해 비추고 있다.",
    "built_space": "화면 중앙에 난민촌이 밀집해 있고, 좌측에 갈라진 방파제, 우측에 밝게 켜진 사각형 건물이 있으며 교회는 원경에 작게 배치되었다.",
    "entities": "대형 군용 수송 헬기, 난민촌 텐트들, 방파제, 이동하는 사람들, 멀리 위치한 교회가 확인된다.",
    "hard_violations": [
     "[gpt-high] 사람이 등장하지 않는 샷 텍스트와 인물 추가 금지 지시를 어기고, 전경 도로에 최소 여덟 명의 인물을 임의로 추가했습니다."
    ],
    "physics": "헬기는 허공에 떠 있고, 사람들은 지면에 발을 딛고 안정적으로 걷고 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 '전투 헬기'의 외형과 서치라이트가 돋보이는 웅장한 전경, 방파제 위의 차량 행렬 및 프레이밍 구도를 매우 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "밝게 빛나는 건물과 방파제는 묘사되었으나, 헬기가 수송기에 가깝고 내부 도로가 이끄는 웅장한 구도의 지시를 충분히 살리지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "공중의 헬기에서 뻗어나온 서치라이트가 아래의 난민촌과 교회 주변을 정확히 쏘아 내리고 있다.",
        "built_space": "전면에 넓은 내부 도로가 교회와 난민촌으로 이어지며, 우측에는 심하게 갈라진 방파제와 그 위를 지나는 차량 행렬이 배치되어 있다.",
        "entities": "공격형 전투 헬기, 첨탑이 있는 교회, 난민촌의 판잣집들, 균열이 있는 방파제, 이동 중인 차량 행렬이 확인된다.",
        "hard_violations": [],
        "physics": "헬기는 정상적으로 공중에 체공 중이며, 서치라이트 빛줄기가 물리적으로 자연스럽게 지면을 향한다."
       },
       {
        "label": "B",
        "direction": "헬기의 서치라이트가 난민촌 중앙의 좁은 길과 텐트들을 향해 비추고 있다.",
        "built_space": "화면 중앙에 난민촌이 밀집해 있고, 좌측에 갈라진 방파제, 우측에 밝게 켜진 사각형 건물이 있으며 교회는 원경에 작게 배치되었다.",
        "entities": "대형 군용 수송 헬기, 난민촌 텐트들, 방파제, 이동하는 사람들, 멀리 위치한 교회가 확인된다.",
        "hard_violations": [],
        "physics": "헬기는 허공에 떠 있고, 사람들은 지면에 발을 딛고 안정적으로 걷고 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 '전투 헬기'의 외형과 서치라이트가 돋보이는 웅장한 전경, 방파제 위의 차량 행렬 및 프레이밍 구도를 매우 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "밝게 빛나는 건물과 방파제는 묘사되었으나, 헬기가 수송기에 가깝고 내부 도로가 이끄는 웅장한 구도의 지시를 충분히 살리지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "공중의 헬기에서 뻗어나온 서치라이트가 아래의 난민촌과 교회 주변을 정확히 쏘아 내리고 있다.",
        "built_space": "전면에 넓은 내부 도로가 교회와 난민촌으로 이어지며, 우측에는 심하게 갈라진 방파제와 그 위를 지나는 차량 행렬이 배치되어 있다.",
        "entities": "공격형 전투 헬기, 첨탑이 있는 교회, 난민촌의 판잣집들, 균열이 있는 방파제, 이동 중인 차량 행렬이 확인된다.",
        "hard_violations": [],
        "physics": "헬기는 정상적으로 공중에 체공 중이며, 서치라이트 빛줄기가 물리적으로 자연스럽게 지면을 향한다."
       },
       {
        "label": "B",
        "direction": "헬기의 서치라이트가 난민촌 중앙의 좁은 길과 텐트들을 향해 비추고 있다.",
        "built_space": "화면 중앙에 난민촌이 밀집해 있고, 좌측에 갈라진 방파제, 우측에 밝게 켜진 사각형 건물이 있으며 교회는 원경에 작게 배치되었다.",
        "entities": "대형 군용 수송 헬기, 난민촌 텐트들, 방파제, 이동하는 사람들, 멀리 위치한 교회가 확인된다.",
        "hard_violations": [],
        "physics": "헬기는 허공에 떠 있고, 사람들은 지면에 발을 딛고 안정적으로 걷고 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "서치라이트는 난민촌 아래를 제대로 비추지만, 지시하지 않은 전경 인물들이 추가되었고 내려다보는 구도와 잘린 헬기가 요구한 도로 기준 전경에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "임의 인물 추가로 부적격이지만, 전투 헬기의 측면·하부와 난민촌으로 이어지는 도로, 하향 서치라이트의 관계는 A보다 충실합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "헬기의 기수는 화면 왼쪽을 향하고, 하부 서치라이트는 왼쪽 아래로 내려가 난민촌 중앙 통로와 인접 지붕을 비춥니다. 빛의 목표는 요청한 난민촌과 일치합니다. 전경 인물들은 도로를 좌우로 이동하며, 얼굴의 시선 방향은 어두워 판별하기 어렵습니다. 방조제 위 차량의 진행 방향은 작은 전조등만으로 확정하기 어렵습니다.",
        "built_space": "중앙 뒤편에 작은 교회 한 채, 왼쪽에 이어진 방조제 한 줄, 오른쪽 뒤편에 창문이 밝은 사각 건물 한 채가 보입니다. 판재·천막·골함석 주거지가 중앙의 좁은 통로 양옆을 채우고, 화면 아래에는 가로 도로가 있습니다. 카메라는 전경 도로보다 높은 곳에서 지붕들을 내려다봅니다. 따라서 요청한 도로 높이의 전경보다는 높은 관찰점에 가깝고, 가로 도로는 정착지 안으로 직접 이어지는 주된 시선축이 아닙니다.",
        "entities": "헬기는 한 대이며 군용 대형 수송헬기에 가까운 형태입니다. 금속 동체의 측면과 하부는 보이지만 꼬리 상단과 회전날개 일부가 프레임 밖으로 잘립니다. 밤하늘, 난민촌, 교회, 균열이 있는 방조제와 밝은 관리동으로 해석 가능한 건물은 보입니다. 방조제의 누수와 서로 구분되는 두 차량 행렬은 확인하기 어렵습니다. 전경에 최소 여덟 명의 사람이 추가되어 있으며, 샷 텍스트는 이들을 등장시키지 않습니다. 이들의 나이·성별·민족적 정체성은 확정할 수 없고, 이현우로 확인되는 인물도 없습니다. 읽을 수 있는 문구는 보이지 않습니다.",
        "hard_violations": [
         "사람이 등장하지 않는 샷 텍스트와 인물 추가 금지 지시를 어기고, 전경 도로에 최소 여덟 명의 인물을 임의로 추가했습니다."
        ],
        "physics": "헬기는 회전 중인 주회전날개가 보여 양력에 의한 비행으로 읽히며, 서치라이트는 동체 하부에 부착되어 있습니다. 전경 인물들은 도로에 발을 딛거나 보행 중 한쪽 다리를 들고 있어 지지 없는 공중 부양은 아닙니다. 차량은 방조제 위 도로에 놓여 있고 주거지와 교회도 지면에 지지됩니다. 명백하게 지지체가 없는 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "헬기 기수와 기수 아래 총열은 화면 왼쪽을 향하며, 특정 사람을 겨누는 관계는 보이지 않습니다. 서치라이트는 왼쪽 아래의 교회 뒤편과 인접 난민촌을 향해 내려갑니다. 오른쪽 방조제 도로의 차량들은 전조등이 화면 오른쪽 앞을 향해 보여 그 방향으로 접근하는 것으로 읽힙니다. 내부 도로 차량들이 교회를 향해 이동하는지는 정지 화면에서 명확하지 않습니다. 작은 인물들의 시선도 식별하기 어렵습니다.",
        "built_space": "왼쪽에 첨탑이 있는 교회 한 채, 오른쪽에 방조제와 그 위 도로 한 줄이 있습니다. 넓은 내부 도로가 화면 아래에서 중앙 뒤편 난민촌으로 이어지고 양옆에 임시 주거지가 배치됩니다. 교회 입구와 계단, 도로 가장자리에 인물들이 서 있습니다. 지붕과 도로를 상당히 내려다보므로 카메라는 요청한 도로 높이보다 높습니다. 다만 전경 도로가 헬기 아래 정착지로 이어지는 구성은 A보다 분명합니다. 밝은 관리동으로 확정할 건물은 보이지 않습니다.",
        "entities": "한 대의 공격형 전투 헬기가 회전날개, 무장, 착륙장치와 함께 보이고 측면과 하부가 모두 드러납니다. 어두운 하늘, 난민촌, 교회, 균열이 있는 방조제와 차량 여러 대가 있습니다. 방조제 위 차량 행렬과 내부 도로 차량은 공간적으로 구분되지만, 탈출 및 교회 접근이라는 목적까지는 확인되지 않습니다. 교회 주변과 도로 곳곳에 지시되지 않은 인물들이 여러 명 추가되어 있습니다. 이들은 작아서 나이·성별·민족적 정체성을 판별하기 어렵습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "사람이 등장하지 않는 샷 텍스트와 인물 추가 금지 지시를 어기고, 교회 주변 및 난민촌 도로에 여러 인물을 임의로 추가했습니다."
        ],
        "physics": "헬기는 회전하는 주회전날개와 꼬리회전날개를 갖추고 있어 비행을 유지하는 물리적 장치가 보입니다. 무장과 서치라이트, 착륙장치는 동체에 연결되어 있습니다. 차량들은 각 도로 표면에 놓여 있고, 인물들도 도로 또는 교회 계단에 발을 디딘 것으로 보입니다. 빛은 공기 중 산란으로 드러나는 광선이며 떠 있는 고체로 묘사되지 않습니다. 명백한 무지지 부양이나 불가능한 자세는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "서치라이트는 난민촌 아래를 제대로 비추지만, 지시하지 않은 전경 인물들이 추가되었고 내려다보는 구도와 잘린 헬기가 요구한 도로 기준 전경에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "임의 인물 추가로 부적격이지만, 전투 헬기의 측면·하부와 난민촌으로 이어지는 도로, 하향 서치라이트의 관계는 A보다 충실합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "헬기의 기수는 화면 왼쪽을 향하고, 하부 서치라이트는 왼쪽 아래로 내려가 난민촌 중앙 통로와 인접 지붕을 비춥니다. 빛의 목표는 요청한 난민촌과 일치합니다. 전경 인물들은 도로를 좌우로 이동하며, 얼굴의 시선 방향은 어두워 판별하기 어렵습니다. 방조제 위 차량의 진행 방향은 작은 전조등만으로 확정하기 어렵습니다.",
        "built_space": "중앙 뒤편에 작은 교회 한 채, 왼쪽에 이어진 방조제 한 줄, 오른쪽 뒤편에 창문이 밝은 사각 건물 한 채가 보입니다. 판재·천막·골함석 주거지가 중앙의 좁은 통로 양옆을 채우고, 화면 아래에는 가로 도로가 있습니다. 카메라는 전경 도로보다 높은 곳에서 지붕들을 내려다봅니다. 따라서 요청한 도로 높이의 전경보다는 높은 관찰점에 가깝고, 가로 도로는 정착지 안으로 직접 이어지는 주된 시선축이 아닙니다.",
        "entities": "헬기는 한 대이며 군용 대형 수송헬기에 가까운 형태입니다. 금속 동체의 측면과 하부는 보이지만 꼬리 상단과 회전날개 일부가 프레임 밖으로 잘립니다. 밤하늘, 난민촌, 교회, 균열이 있는 방조제와 밝은 관리동으로 해석 가능한 건물은 보입니다. 방조제의 누수와 서로 구분되는 두 차량 행렬은 확인하기 어렵습니다. 전경에 최소 여덟 명의 사람이 추가되어 있으며, 샷 텍스트는 이들을 등장시키지 않습니다. 이들의 나이·성별·민족적 정체성은 확정할 수 없고, 이현우로 확인되는 인물도 없습니다. 읽을 수 있는 문구는 보이지 않습니다.",
        "hard_violations": [
         "사람이 등장하지 않는 샷 텍스트와 인물 추가 금지 지시를 어기고, 전경 도로에 최소 여덟 명의 인물을 임의로 추가했습니다."
        ],
        "physics": "헬기는 회전 중인 주회전날개가 보여 양력에 의한 비행으로 읽히며, 서치라이트는 동체 하부에 부착되어 있습니다. 전경 인물들은 도로에 발을 딛거나 보행 중 한쪽 다리를 들고 있어 지지 없는 공중 부양은 아닙니다. 차량은 방조제 위 도로에 놓여 있고 주거지와 교회도 지면에 지지됩니다. 명백하게 지지체가 없는 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "헬기 기수와 기수 아래 총열은 화면 왼쪽을 향하며, 특정 사람을 겨누는 관계는 보이지 않습니다. 서치라이트는 왼쪽 아래의 교회 뒤편과 인접 난민촌을 향해 내려갑니다. 오른쪽 방조제 도로의 차량들은 전조등이 화면 오른쪽 앞을 향해 보여 그 방향으로 접근하는 것으로 읽힙니다. 내부 도로 차량들이 교회를 향해 이동하는지는 정지 화면에서 명확하지 않습니다. 작은 인물들의 시선도 식별하기 어렵습니다.",
        "built_space": "왼쪽에 첨탑이 있는 교회 한 채, 오른쪽에 방조제와 그 위 도로 한 줄이 있습니다. 넓은 내부 도로가 화면 아래에서 중앙 뒤편 난민촌으로 이어지고 양옆에 임시 주거지가 배치됩니다. 교회 입구와 계단, 도로 가장자리에 인물들이 서 있습니다. 지붕과 도로를 상당히 내려다보므로 카메라는 요청한 도로 높이보다 높습니다. 다만 전경 도로가 헬기 아래 정착지로 이어지는 구성은 A보다 분명합니다. 밝은 관리동으로 확정할 건물은 보이지 않습니다.",
        "entities": "한 대의 공격형 전투 헬기가 회전날개, 무장, 착륙장치와 함께 보이고 측면과 하부가 모두 드러납니다. 어두운 하늘, 난민촌, 교회, 균열이 있는 방조제와 차량 여러 대가 있습니다. 방조제 위 차량 행렬과 내부 도로 차량은 공간적으로 구분되지만, 탈출 및 교회 접근이라는 목적까지는 확인되지 않습니다. 교회 주변과 도로 곳곳에 지시되지 않은 인물들이 여러 명 추가되어 있습니다. 이들은 작아서 나이·성별·민족적 정체성을 판별하기 어렵습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "사람이 등장하지 않는 샷 텍스트와 인물 추가 금지 지시를 어기고, 교회 주변 및 난민촌 도로에 여러 인물을 임의로 추가했습니다."
        ],
        "physics": "헬기는 회전하는 주회전날개와 꼬리회전날개를 갖추고 있어 비행을 유지하는 물리적 장치가 보입니다. 무장과 서치라이트, 착륙장치는 동체에 연결되어 있습니다. 차량들은 각 도로 표면에 놓여 있고, 인물들도 도로 또는 교회 계단에 발을 디딘 것으로 보입니다. 빛은 공기 중 산란으로 드러나는 광선이며 떠 있는 고체로 묘사되지 않습니다. 명백한 무지지 부양이나 불가능한 자세는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.464
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.214
   },
   "violations": {
    "B": [
     "[gpt-high] 사람이 등장하지 않는 샷 텍스트와 인물 추가 금지 지시를 어기고, 전경 도로에 최소 여덟 명의 인물을 임의로 추가했습니다."
    ],
    "A": [
     "[gpt-high] 사람이 등장하지 않는 샷 텍스트와 인물 추가 금지 지시를 어기고, 교회 주변 및 난민촌 도로에 여러 인물을 임의로 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1214
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "요구된 '전투 헬기'의 외형과 서치라이트가 돋보이는 웅장한 전경, 방파제 위의 차량 행렬 및 프레이밍 구도를 매우 충실하게 구현했습니다.  ★위반: [gpt-high] 사람이 등장하지 않는 샷 텍스트와 인물 추가 금지 지시를 어기고, 교회 주변 및 난민촌 도로에 여러 인물을 임의로 추가했습니다."
   },
   {
    "label": "B",
    "score": 1214,
    "verdict_ko": "밝게 빛나는 건물과 방파제는 묘사되었으나, 헬기가 수송기에 가깝고 내부 도로가 이끄는 웅장한 구도의 지시를 충분히 살리지 못했습니다.  ★위반: [gpt-high] 사람이 등장하지 않는 샷 텍스트와 인물 추가 금지 지시를 어기고, 전경 도로에 최소 여덟 명의 인물을 임의로 추가했습니다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb5-a47c-7be6-a000-a67abaa38b74",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S33sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:34:02.287879+00:00",
  "fingerprint": "a139a52fd97241091a417af7f9601164baf3977c5795d0ab0e894fa655896322",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S33sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S33sh1_sel.png",
  "source_sha256": "4d38fc3732406cce405129eacf794ed37db322f9ac23de15a6c2822ca6df9f85",
  "file": "S33sh1_cine.png",
  "staged_sha256": "47fe70114b7f22ac8ca3f73b1de486393744de9eed7739a759955892716f3ad5",
  "latency_ms": 11503
 },
 "S33sh10::signage": {
  "fp": "0314b59aa6b0b8a6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::80b0918370c74a36": {
  "subjects": [],
  "subject_text": "인천 난민촌과 성당 상공\n밤하늘 아래 성당 지붕과 밀집한 컨테이너 주거지가 펼쳐진 넓은 조망. 도로와 인공제방이 주거 구역 사이를 길게 가른다.",
  "identity": "canonical",
  "scope_id": "L39",
  "scope_role": "location_exterior",
  "scope_sha": "4766eaf7b5730a84"
 },
 "S33sh10::bgfirst_bg": {
  "input_fingerprint": "e65ded79bd1889f7",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 환하게 불이 켜진 관리사무소의 외부 계단을 향해 상체를 웅크린 채 한쪽 뒷발로 땅을 강하게 밀어내며 전력 질주하는 뒷모습.\n\nLOCATION (lock): In the administration building's outdoor yard, on the approach to its exterior stairs at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Exterior stairs ahead of 이현우 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Exterior office stairs (Ahead along 이현우's approach) — The foot of the stairs faces diagonally toward the approaching camera; used as Makes his immediate destination legible beyond his trailing foot; Management-office exterior (Adjoining the stairs) — Only the section bordering the stair route is retained; used as Anchors the stairs to the building without distracting from the sprint; Office yard ground (Crossed by 이현우); used as Retains the contact point of the pushing foot and the route toward the stairs.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brightly lit management office provides a clear destination against the night setting, without specifying visible lighting fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 환하게 불이 켜진 관리사무소의 외부 계단을 향해 상체를 웅크린 채 한쪽 뒷발로 땅을 강하게 밀어내며 전력 질주하는 뒷모습.\n\nLOCATION (lock): In the administration building's outdoor yard, on the approach to its exterior stairs at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Exterior stairs ahead of 이현우 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Exterior office stairs (Ahead along 이현우's approach) — The foot of the stairs faces diagonally toward the approaching camera; used as Makes his immediate destination legible beyond his trailing foot; Management-office exterior (Adjoining the stairs) — Only the section bordering the stair route is retained; used as Anchors the stairs to the building without distracting from the sprint; Office yard ground (Crossed by 이현우); used as Retains the contact point of the pushing foot and the route toward the stairs.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brightly lit management office provides a clear destination against the night setting, without specifying visible lighting fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S33sh10__bgfirst_bg.png",
  "asset_id": "b3e8aaef-486f-4746-8fdf-82c3f1db183c",
  "input_asset_ids": [
   "d075bc89-3d0f-442f-bf0f-022fc1aac3ad",
   "df4f6cc4-e0e9-495c-bf77-aa7acebc593c"
  ]
 },
 "S33sh10": {
  "input_fingerprint": "f1cb74b4a827eb11",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 환하게 불이 켜진 관리사무소의 외부 계단을 향해 상체를 웅크린 채 한쪽 뒷발로 땅을 강하게 밀어내며 전력 질주하는 뒷모습.\n\nLOCATION (lock): In the administration building's outdoor yard, on the approach to its exterior stairs at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Exterior stairs ahead of 이현우 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Exterior office stairs (Ahead along 이현우's approach) — The foot of the stairs faces diagonally toward the approaching camera; used as Makes his immediate destination legible beyond his trailing foot; Management-office exterior (Adjoining the stairs) — Only the section bordering the stair route is retained; used as Anchors the stairs to the building without distracting from the sprint; Office yard ground (Crossed by 이현우); used as Retains the contact point of the pushing foot and the route toward the stairs.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brightly lit management office provides a clear destination against the night setting, without specifying visible lighting fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The manhole beside the management building is now open, and the building remains brightly illuminated; its third-floor detention room is still locked. The church's basement entrance has been discovered, while the cracked, leaking seawall has not yet been breached. 이현우: He has emerged from the manhole and is ascending the office stairs, retaining facial bruising and the injured leg. His outer garment remains removed and his contact card stays in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 환하게 불이 켜진 관리사무소의 외부 계단을 향해 상체를 웅크린 채 한쪽 뒷발로 땅을 강하게 밀어내며 전력 질주하는 뒷모습.\n\nLOCATION (lock): In the administration building's outdoor yard, on the approach to its exterior stairs at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Exterior stairs ahead of 이현우 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Exterior office stairs (Ahead along 이현우's approach) — The foot of the stairs faces diagonally toward the approaching camera; used as Makes his immediate destination legible beyond his trailing foot; Management-office exterior (Adjoining the stairs) — Only the section bordering the stair route is retained; used as Anchors the stairs to the building without distracting from the sprint; Office yard ground (Crossed by 이현우); used as Retains the contact point of the pushing foot and the route toward the stairs.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brightly lit management office provides a clear destination against the night setting, without specifying visible lighting fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The manhole beside the management building is now open, and the building remains brightly illuminated; its third-floor detention room is still locked. The church's basement entrance has been discovered, while the cracked, leaking seawall has not yet been breached. 이현우: He has emerged from the manhole and is ascending the office stairs, retaining facial bruising and the injured leg. His outer garment remains removed and his contact card stays in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 환하게 불이 켜진 관리사무소의 외부 계단을 향해 상체를 웅크린 채 한쪽 뒷발로 땅을 강하게 밀어내며 전력 질주하는 뒷모습.\n\nLOCATION (lock): In the administration building's outdoor yard, on the approach to its exterior stairs at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Exterior stairs ahead of 이현우 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Exterior office stairs (Ahead along 이현우's approach) — The foot of the stairs faces diagonally toward the approaching camera; used as Makes his immediate destination legible beyond his trailing foot; Management-office exterior (Adjoining the stairs) — Only the section bordering the stair route is retained; used as Anchors the stairs to the building without distracting from the sprint; Office yard ground (Crossed by 이현우); used as Retains the contact point of the pushing foot and the route toward the stairs.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brightly lit management office provides a clear destination against the night setting, without specifying visible lighting fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The manhole beside the management building is now open, and the building remains brightly illuminated; its third-floor detention room is still locked. The church's basement entrance has been discovered, while the cracked, leaking seawall has not yet been breached. 이현우: He has emerged from the manhole and is ascending the office stairs, retaining facial bruising and the injured leg. His outer garment remains removed and his contact card stays in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S33sh10__bgfirst_bg.png",
     "asset_id": "b3e8aaef-486f-4746-8fdf-82c3f1db183c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S33sh10.png",
     "asset_id": "d075bc89-3d0f-442f-bf0f-022fc1aac3ad",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_camp_church_airspace_sel.png",
     "asset_id": "df4f6cc4-e0e9-495c-bf77-aa7acebc593c",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물이 상체를 웅크린 채 밝게 조명이 켜진 우측의 외부 계단을 향해 전력 질주하고 있음.",
    "built_space": "왼쪽에 창문으로 불빛이 새어나오는 관리사무소 건물이 있고, 우측면에 외부 계단이 위치함. 건물 옆 왼쪽 자갈밭에 열린 맨홀이 보이며, 우측 배경으로 방파제와 포장도로가 레퍼런스와 일치하게 배치됨.",
    "entities": "이현우(뒷모습, 검은 머리)가 겉옷을 벗은 상태로 반팔 티셔츠와 청바지를 입고 있음.",
    "hard_violations": [],
    "physics": "왼쪽 발끝이 지면에 확실히 닿아 땅을 밀어내는 달리기 자세가 올바르게 지탱됨."
   },
   {
    "label": "B",
    "direction": "인물이 상체를 웅크린 채 우측의 외부 계단을 향해 달려가고 있음.",
    "built_space": "왼쪽에 관리사무소 건물과 외부 계단이 있고, 우측 배경에 교회가 보임. 자갈밭 우측에 열린 맨홀이 배치됨.",
    "entities": "이현우(뒷모습, 검은 머리)가 겉옷을 벗은 상태로 어두운 반팔 티셔츠와 청바지를 입고 있음.",
    "hard_violations": [
     "[gemini-pro] 땅을 강하게 밀어내야 할 인물의 왼쪽 발끝이 지면(자갈밭)과 닿지 않고 명백히 허공에 떠 있어, 몸을 지탱하거나 추진력을 만들어내는 지지 기반이 전혀 없음."
    ],
    "physics": "달리는 포즈를 취하고 있으나 지면과 발 사이에 그림자와의 간격이 존재하여 몸이 허공에 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시문의 구도와 레퍼런스의 배경 구조를 정확히 반영했으며, 땅을 박차고 나가는 질주 동작이 자연스럽게 구현되었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배경과 조명 분위기는 적절히 연출되었으나, 땅을 밀어내야 할 발이 지면에 닿지 않고 공중에 떠 있는 치명적인 물리적 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물이 상체를 웅크린 채 밝게 조명이 켜진 우측의 외부 계단을 향해 전력 질주하고 있음.",
        "built_space": "왼쪽에 창문으로 불빛이 새어나오는 관리사무소 건물이 있고, 우측면에 외부 계단이 위치함. 건물 옆 왼쪽 자갈밭에 열린 맨홀이 보이며, 우측 배경으로 방파제와 포장도로가 레퍼런스와 일치하게 배치됨.",
        "entities": "이현우(뒷모습, 검은 머리)가 겉옷을 벗은 상태로 반팔 티셔츠와 청바지를 입고 있음.",
        "hard_violations": [],
        "physics": "왼쪽 발끝이 지면에 확실히 닿아 땅을 밀어내는 달리기 자세가 올바르게 지탱됨."
       },
       {
        "label": "B",
        "direction": "인물이 상체를 웅크린 채 우측의 외부 계단을 향해 달려가고 있음.",
        "built_space": "왼쪽에 관리사무소 건물과 외부 계단이 있고, 우측 배경에 교회가 보임. 자갈밭 우측에 열린 맨홀이 배치됨.",
        "entities": "이현우(뒷모습, 검은 머리)가 겉옷을 벗은 상태로 어두운 반팔 티셔츠와 청바지를 입고 있음.",
        "hard_violations": [
         "땅을 강하게 밀어내야 할 인물의 왼쪽 발끝이 지면(자갈밭)과 닿지 않고 명백히 허공에 떠 있어, 몸을 지탱하거나 추진력을 만들어내는 지지 기반이 전혀 없음."
        ],
        "physics": "달리는 포즈를 취하고 있으나 지면과 발 사이에 그림자와의 간격이 존재하여 몸이 허공에 떠 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시문의 구도와 레퍼런스의 배경 구조를 정확히 반영했으며, 땅을 박차고 나가는 질주 동작이 자연스럽게 구현되었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배경과 조명 분위기는 적절히 연출되었으나, 땅을 밀어내야 할 발이 지면에 닿지 않고 공중에 떠 있는 치명적인 물리적 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물이 상체를 웅크린 채 밝게 조명이 켜진 우측의 외부 계단을 향해 전력 질주하고 있음.",
        "built_space": "왼쪽에 창문으로 불빛이 새어나오는 관리사무소 건물이 있고, 우측면에 외부 계단이 위치함. 건물 옆 왼쪽 자갈밭에 열린 맨홀이 보이며, 우측 배경으로 방파제와 포장도로가 레퍼런스와 일치하게 배치됨.",
        "entities": "이현우(뒷모습, 검은 머리)가 겉옷을 벗은 상태로 반팔 티셔츠와 청바지를 입고 있음.",
        "hard_violations": [],
        "physics": "왼쪽 발끝이 지면에 확실히 닿아 땅을 밀어내는 달리기 자세가 올바르게 지탱됨."
       },
       {
        "label": "B",
        "direction": "인물이 상체를 웅크린 채 우측의 외부 계단을 향해 달려가고 있음.",
        "built_space": "왼쪽에 관리사무소 건물과 외부 계단이 있고, 우측 배경에 교회가 보임. 자갈밭 우측에 열린 맨홀이 배치됨.",
        "entities": "이현우(뒷모습, 검은 머리)가 겉옷을 벗은 상태로 어두운 반팔 티셔츠와 청바지를 입고 있음.",
        "hard_violations": [
         "땅을 강하게 밀어내야 할 인물의 왼쪽 발끝이 지면(자갈밭)과 닿지 않고 명백히 허공에 떠 있어, 몸을 지탱하거나 추진력을 만들어내는 지지 기반이 전혀 없음."
        ],
        "physics": "달리는 포즈를 취하고 있으나 지면과 발 사이에 그림자와의 간격이 존재하여 몸이 허공에 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "계단을 향한 뒷모습과 뒷발의 지면 밀기는 충실하지만, 계단과 건물이 너무 크게 차지하며 참조에 없는 실외기 등 장소 세부의 변형이 두드러집니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "우상단 배경의 계단을 향해 웅크리고 뒷발로 지면을 미는 와이드 숏을 더 정확히 구현하지만, 맨홀 위치와 포장도로 위의 질주 동선은 참조의 자갈 마당 배치와 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 등과 뒤통수가 보이며 머리와 기울어진 몸통은 화면 오른쪽 위의 외부 계단 하단을 향합니다. 진행 경로가 계단 입구로 이어지고 맨홀은 그 경로 오른쪽에 있습니다. 눈은 보이지 않아 실제 시선은 확인할 수 없고, 무기나 손에 든 물체는 없습니다.",
        "built_space": "자갈 마당에서 단일 직선 계단 한 개가 오른쪽 위 출입문과 상부 참으로 올라갑니다. 계단 양쪽 난간과 상부 참의 난간, 계단 아래 문 한 개 및 망형 문짝이 보입니다. 건물 긴 벽에는 상하로 여러 창과 실외기 세 대가 보여 참조보다 외벽 설비가 많습니다. 열린 맨홀 한 개와 분리된 뚜껑 한 개가 계단 오른쪽 앞에 있습니다. 계단 하단은 카메라 쪽으로 비스듬히 열려 있지만 계단과 건물의 화면 점유율이 커서 배경 목적지라는 배치가 다소 약합니다.",
        "entities": "인물은 한 명뿐이며 헝클어진 검은 머리, 마른 젊은 남성 체형, 청바지와 낡은 운동화가 보입니다. 얼굴 대부분이 숨겨져 한국계 얼굴의 정확한 동일성, 18세의 얼굴 및 멍은 확인할 수 없습니다. 후드 겉옷 없이 때 묻은 반팔 상의를 입어 겉옷 제거 상태와 맞습니다. 다리 부상과 신발 속 카드는 확인할 수 없습니다. 밝게 켜진 관리사무소, 외부 계단, 열린 맨홀이 있으며 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "뒤로 뻗은 다리의 신발 앞부분이 자갈 지면에 닿아 몸을 밀어내고, 반대쪽 발은 들려 있습니다. 앞으로 숙인 몸통과 굽힌 팔은 달리기 추진 동작으로 설명됩니다. 인물이 근거 없이 떠 있지는 않습니다. 맨홀 뚜껑은 구멍 옆 지면에 놓여 있고 계단과 난간은 건물 및 콘크리트 구조에 지지됩니다."
       },
       {
        "label": "B",
        "direction": "인물의 뒤통수와 등이 카메라를 향하고 머리, 몸통 및 달리는 방향은 우상단 계단의 첫 단을 향합니다. 인물 앞에 진입로가 열려 있어 목적지가 명확합니다. 계단 하단도 접근하는 카메라를 향해 비스듬히 놓였습니다. 눈은 가려져 직접적인 시선 확인은 불가능하며 손에 든 물체는 없습니다.",
        "built_space": "외부 직선 계단 한 개, 양쪽 난간, 상부 참 한 개, 밝은 상층 출입문 한 개와 그 위 차양, 계단 아래 문 한 개가 보입니다. 계단 옆 벽에는 상하층 창과 수직 배수관이 있으며 참조의 밝은 콘크리트 외벽과 계단 구조를 대체로 유지합니다. 계단이 우상단 배경에 자리하고 마당과 접근 지면이 충분히 남아 와이드 숏에 더 가깝습니다. 다만 인물은 자갈 마당보다 포장도로 위에 있고, 열린 맨홀 한 개와 뚜껑 한 개는 참조의 계단 오른쪽이 아니라 왼쪽으로 이동했습니다. 오른쪽에는 해안도로와 방호기둥이 이어집니다.",
        "entities": "한 명의 젊고 마른 남성이 검은 헝클어진 머리, 밝은 반팔 상의, 청바지와 운동화를 착용하고 있습니다. 겉옷은 제거되어 있습니다. 뒷모습이라 참조의 한국계 얼굴, 정확한 나이 인상과 얼굴 멍은 검증할 수 없습니다. 바지에 가려진 다리 부상과 신발 속 카드 역시 확인할 수 없습니다. 관리사무소와 외부 계단, 열린 맨홀이 보이며 밤 배경에 사무소 출입구와 창이 밝게 빛납니다. 추가 인물이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "뒤로 뻗은 발의 앞부분이 포장 지면에 닿고 뒤꿈치는 들려 있어 한쪽 뒷발로 밀어내는 순간이 읽힙니다. 반대쪽 무릎과 발은 들려 있으며 앞으로 기운 몸통과 팔의 움직임이 추진 동작을 뒷받침합니다. 공중에 뜬 발은 달리기 동작으로 설명되고 몸에는 지면 접점이 있습니다. 기울어진 맨홀 뚜껑은 아래 가장자리와 맨홀 테두리에 기대어 지지됩니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "계단을 향한 뒷모습과 뒷발의 지면 밀기는 충실하지만, 계단과 건물이 너무 크게 차지하며 참조에 없는 실외기 등 장소 세부의 변형이 두드러집니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "우상단 배경의 계단을 향해 웅크리고 뒷발로 지면을 미는 와이드 숏을 더 정확히 구현하지만, 맨홀 위치와 포장도로 위의 질주 동선은 참조의 자갈 마당 배치와 다릅니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "인물의 등과 뒤통수가 보이며 머리와 기울어진 몸통은 화면 오른쪽 위의 외부 계단 하단을 향합니다. 진행 경로가 계단 입구로 이어지고 맨홀은 그 경로 오른쪽에 있습니다. 눈은 보이지 않아 실제 시선은 확인할 수 없고, 무기나 손에 든 물체는 없습니다.",
        "built_space": "자갈 마당에서 단일 직선 계단 한 개가 오른쪽 위 출입문과 상부 참으로 올라갑니다. 계단 양쪽 난간과 상부 참의 난간, 계단 아래 문 한 개 및 망형 문짝이 보입니다. 건물 긴 벽에는 상하로 여러 창과 실외기 세 대가 보여 참조보다 외벽 설비가 많습니다. 열린 맨홀 한 개와 분리된 뚜껑 한 개가 계단 오른쪽 앞에 있습니다. 계단 하단은 카메라 쪽으로 비스듬히 열려 있지만 계단과 건물의 화면 점유율이 커서 배경 목적지라는 배치가 다소 약합니다.",
        "entities": "인물은 한 명뿐이며 헝클어진 검은 머리, 마른 젊은 남성 체형, 청바지와 낡은 운동화가 보입니다. 얼굴 대부분이 숨겨져 한국계 얼굴의 정확한 동일성, 18세의 얼굴 및 멍은 확인할 수 없습니다. 후드 겉옷 없이 때 묻은 반팔 상의를 입어 겉옷 제거 상태와 맞습니다. 다리 부상과 신발 속 카드는 확인할 수 없습니다. 밝게 켜진 관리사무소, 외부 계단, 열린 맨홀이 있으며 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "뒤로 뻗은 다리의 신발 앞부분이 자갈 지면에 닿아 몸을 밀어내고, 반대쪽 발은 들려 있습니다. 앞으로 숙인 몸통과 굽힌 팔은 달리기 추진 동작으로 설명됩니다. 인물이 근거 없이 떠 있지는 않습니다. 맨홀 뚜껑은 구멍 옆 지면에 놓여 있고 계단과 난간은 건물 및 콘크리트 구조에 지지됩니다."
       },
       {
        "label": "A",
        "direction": "인물의 뒤통수와 등이 카메라를 향하고 머리, 몸통 및 달리는 방향은 우상단 계단의 첫 단을 향합니다. 인물 앞에 진입로가 열려 있어 목적지가 명확합니다. 계단 하단도 접근하는 카메라를 향해 비스듬히 놓였습니다. 눈은 가려져 직접적인 시선 확인은 불가능하며 손에 든 물체는 없습니다.",
        "built_space": "외부 직선 계단 한 개, 양쪽 난간, 상부 참 한 개, 밝은 상층 출입문 한 개와 그 위 차양, 계단 아래 문 한 개가 보입니다. 계단 옆 벽에는 상하층 창과 수직 배수관이 있으며 참조의 밝은 콘크리트 외벽과 계단 구조를 대체로 유지합니다. 계단이 우상단 배경에 자리하고 마당과 접근 지면이 충분히 남아 와이드 숏에 더 가깝습니다. 다만 인물은 자갈 마당보다 포장도로 위에 있고, 열린 맨홀 한 개와 뚜껑 한 개는 참조의 계단 오른쪽이 아니라 왼쪽으로 이동했습니다. 오른쪽에는 해안도로와 방호기둥이 이어집니다.",
        "entities": "한 명의 젊고 마른 남성이 검은 헝클어진 머리, 밝은 반팔 상의, 청바지와 운동화를 착용하고 있습니다. 겉옷은 제거되어 있습니다. 뒷모습이라 참조의 한국계 얼굴, 정확한 나이 인상과 얼굴 멍은 검증할 수 없습니다. 바지에 가려진 다리 부상과 신발 속 카드 역시 확인할 수 없습니다. 관리사무소와 외부 계단, 열린 맨홀이 보이며 밤 배경에 사무소 출입구와 창이 밝게 빛납니다. 추가 인물이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "뒤로 뻗은 발의 앞부분이 포장 지면에 닿고 뒤꿈치는 들려 있어 한쪽 뒷발로 밀어내는 순간이 읽힙니다. 반대쪽 무릎과 발은 들려 있으며 앞으로 기운 몸통과 팔의 움직임이 추진 동작을 뒷받침합니다. 공중에 뜬 발은 달리기 동작으로 설명되고 몸에는 지면 접점이 있습니다. 기울어진 맨홀 뚜껑은 아래 가장자리와 맨홀 테두리에 기대어 지지됩니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.25
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.0
   },
   "violations": {
    "B": [
     "[gemini-pro] 땅을 강하게 밀어내야 할 인물의 왼쪽 발끝이 지면(자갈밭)과 닿지 않고 명백히 허공에 떠 있어, 몸을 지탱하거나 추진력을 만들어내는 지지 기반이 전혀 없음."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시문의 구도와 레퍼런스의 배경 구조를 정확히 반영했으며, 땅을 박차고 나가는 질주 동작이 자연스럽게 구현되었습니다."
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "배경과 조명 분위기는 적절히 연출되었으나, 땅을 밀어내야 할 발이 지면에 닿지 않고 공중에 떠 있는 치명적인 물리적 오류가 있습니다.  ★위반: [gemini-pro] 땅을 강하게 밀어내야 할 인물의 왼쪽 발끝이 지면(자갈밭)과 닿지 않고 명백히 허공에 떠 있어, 몸을 지탱하거나 추진력을 만들어내는 지지 기반이 전혀 없음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_camp_church_airspace_sel.png",
    "asset_id": "df4f6cc4-e0e9-495c-bf77-aa7acebc593c",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb6-6226-7a01-8bab-3db15b015311",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S33sh10__bgfirst_bg.png",
   "bg_asset_id": "b3e8aaef-486f-4746-8fdf-82c3f1db183c",
   "bg_record_key": "S33sh10::bgfirst_bg",
   "chain_winner": true,
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S33sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:34:14.223419+00:00",
  "fingerprint": "a6791575ac39e3f008afc26260409009e99161983ae0a7a131eedefda77ef78b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S33sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S33sh10_sel.png",
  "source_sha256": "58f24ff0d65a7af0307472729bae8ee4cae2045da07a803fa53297728d08838d",
  "file": "S33sh10_cine.png",
  "staged_sha256": "a21998f189e019329d13af20aa13f36d2af883cbf457d8aeabe48487361f38a4",
  "latency_ms": 13303
 },
 "S33sh14::signage": {
  "fp": "90089cd8b3ed6288",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S33sh14": {
  "input_fingerprint": "544a80c0d011f149",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 무너져 내린 제방 틈새로 검고 거친 바닷물이 폭포수처럼 쏟아져 들어오며 중장비들을 거세게 덮치는 재난 구도.\n\nLOCATION (lock): At the breached section of the refugee settlement's seawall, above the exposed construction equipment being struck by seawater. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Breach where seawater enters in the upper-center of the frame, background; Machinery below the breach on the inland side in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Breached artificial embankment (An upper section has broken away) — The damaged crest and inland face are visible from high above the inland side; used as Establishes the origin and direction of the inundation; Incoming seawater (Dark, turbulent water pours through the breach and sweeps inland); used as Connects the upper breach to the machinery being engulfed below; Repair and construction machinery (Being engulfed and swept away by seawater) — Separate machines present different oblique top and side views within the current; used as Provides distributed scale references and visible evidence of the flood's force.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained night illumination and controlled contrast distinguish the black, turbulent seawater from the broken embankment and engulfed machinery without adding a light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Explosive drones have blown out an upper section of the already cracked seawall, leaving a breach through which seawater pours and sweeps away the repair machinery. The nearby containers and convoy are under attack, while the management building remains lit with its yard manhole open.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 무너져 내린 제방 틈새로 검고 거친 바닷물이 폭포수처럼 쏟아져 들어오며 중장비들을 거세게 덮치는 재난 구도.\n\nLOCATION (lock): At the breached section of the refugee settlement's seawall, above the exposed construction equipment being struck by seawater. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Breach where seawater enters in the upper-center of the frame, background; Machinery below the breach on the inland side in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Breached artificial embankment (An upper section has broken away) — The damaged crest and inland face are visible from high above the inland side; used as Establishes the origin and direction of the inundation; Incoming seawater (Dark, turbulent water pours through the breach and sweeps inland); used as Connects the upper breach to the machinery being engulfed below; Repair and construction machinery (Being engulfed and swept away by seawater) — Separate machines present different oblique top and side views within the current; used as Provides distributed scale references and visible evidence of the flood's force.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained night illumination and controlled contrast distinguish the black, turbulent seawater from the broken embankment and engulfed machinery without adding a light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Explosive drones have blown out an upper section of the already cracked seawall, leaving a breach through which seawater pours and sweeps away the repair machinery. The nearby containers and convoy are under attack, while the management building remains lit with its yard manhole open.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 무너져 내린 제방 틈새로 검고 거친 바닷물이 폭포수처럼 쏟아져 들어오며 중장비들을 거세게 덮치는 재난 구도.\n\nLOCATION (lock): At the breached section of the refugee settlement's seawall, above the exposed construction equipment being struck by seawater. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Breach where seawater enters in the upper-center of the frame, background; Machinery below the breach on the inland side in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Breached artificial embankment (An upper section has broken away) — The damaged crest and inland face are visible from high above the inland side; used as Establishes the origin and direction of the inundation; Incoming seawater (Dark, turbulent water pours through the breach and sweeps inland); used as Connects the upper breach to the machinery being engulfed below; Repair and construction machinery (Being engulfed and swept away by seawater) — Separate machines present different oblique top and side views within the current; used as Provides distributed scale references and visible evidence of the flood's force.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained night illumination and controlled contrast distinguish the black, turbulent seawater from the broken embankment and engulfed machinery without adding a light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Explosive drones have blown out an upper section of the already cracked seawall, leaving a breach through which seawater pours and sweeps away the repair machinery. The nearby containers and convoy are under attack, while the management building remains lit with its yard manhole open.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "물이 내륙(왼쪽)에서 바다 쪽(오른쪽)으로 흐르며, 프롬프트의 해수 유입 방향과 반대임.",
    "built_space": "방파제와 건물의 형태는 참조와 같으나, 카메라는 바다 쪽에 위치하여 내륙 쪽 뷰를 요구한 지시를 어김.",
    "entities": "중장비가 묘사되었으나 내륙이 아닌 바다 쪽 바위 위에 배치됨. 인물은 없음.",
    "hard_violations": [
     "[gemini-pro] 지시된 카메라 위치 및 방향 위반 (내륙이 아닌 바다 쪽에서 촬영)",
     "[gemini-pro] 공간 구조 및 장비 배치 오류 (물이 내륙에서 바다로 흐름)",
     "[gpt-high] 명시된 내륙 쪽 고가 시점 대신 바다 쪽 사석 호안 위에서 육지를 보는 시점으로 배치했습니다.",
     "[gpt-high] 중장비를 내륙이 아닌 바다 쪽에 놓고, 내륙으로 들어와야 할 물을 육지에서 바다 쪽으로 낙하하게 만들어 핵심 공간 관계를 반대로 구현했습니다."
    ],
    "physics": "중장비는 바위 위에 지지되어 있으며, 물이 쏟아지는 물리적 표현 자체는 지면으로 자연스럽게 떨어짐."
   },
   {
    "label": "B",
    "direction": "바닷물이 배경의 무너진 방파제를 통해 전경(내륙)으로 정확하게 쏟아져 들어옴.",
    "built_space": "방파제의 수직 내륙면이 보이며, 카메라는 지시된 내륙 측에 위치함. 중장비도 내륙에 올바르게 배치됨.",
    "entities": "지시된 중장비(굴착기, 로더)가 존재하나, 프롬프트에 없는 인물 4명 이상이 장비 주변을 뛰어다님.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 인물 임의 추가 (invented people)",
     "[gpt-high] 인물을 보여주지 않는 쇼트에 적어도 다섯 명의 사람을 임의로 추가했습니다."
    ],
    "physics": "중장비는 지면에 고정되어 있고, 물은 중력에 따라 쏟아져 내리며 표면을 덮침."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "구도와 해수 유입 방향은 지시대로 구현했으나, 프롬프트에 없는 인물들을 다수 임의로 추가하여 치명적인 위반이 발생함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "내륙 측 시점 및 해수 유입 방향을 바다 쪽으로 완전히 반대로 연출하여 심각한 공간적, 구조적 구도 위반을 범함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "물이 내륙(왼쪽)에서 바다 쪽(오른쪽)으로 흐르며, 프롬프트의 해수 유입 방향과 반대임.",
        "built_space": "방파제와 건물의 형태는 참조와 같으나, 카메라는 바다 쪽에 위치하여 내륙 쪽 뷰를 요구한 지시를 어김.",
        "entities": "중장비가 묘사되었으나 내륙이 아닌 바다 쪽 바위 위에 배치됨. 인물은 없음.",
        "hard_violations": [
         "지시된 카메라 위치 및 방향 위반 (내륙이 아닌 바다 쪽에서 촬영)",
         "공간 구조 및 장비 배치 오류 (물이 내륙에서 바다로 흐름)"
        ],
        "physics": "중장비는 바위 위에 지지되어 있으며, 물이 쏟아지는 물리적 표현 자체는 지면으로 자연스럽게 떨어짐."
       },
       {
        "label": "B",
        "direction": "바닷물이 배경의 무너진 방파제를 통해 전경(내륙)으로 정확하게 쏟아져 들어옴.",
        "built_space": "방파제의 수직 내륙면이 보이며, 카메라는 지시된 내륙 측에 위치함. 중장비도 내륙에 올바르게 배치됨.",
        "entities": "지시된 중장비(굴착기, 로더)가 존재하나, 프롬프트에 없는 인물 4명 이상이 장비 주변을 뛰어다님.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 인물 임의 추가 (invented people)"
        ],
        "physics": "중장비는 지면에 고정되어 있고, 물은 중력에 따라 쏟아져 내리며 표면을 덮침."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "구도와 해수 유입 방향은 지시대로 구현했으나, 프롬프트에 없는 인물들을 다수 임의로 추가하여 치명적인 위반이 발생함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "내륙 측 시점 및 해수 유입 방향을 바다 쪽으로 완전히 반대로 연출하여 심각한 공간적, 구조적 구도 위반을 범함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "물이 내륙(왼쪽)에서 바다 쪽(오른쪽)으로 흐르며, 프롬프트의 해수 유입 방향과 반대임.",
        "built_space": "방파제와 건물의 형태는 참조와 같으나, 카메라는 바다 쪽에 위치하여 내륙 쪽 뷰를 요구한 지시를 어김.",
        "entities": "중장비가 묘사되었으나 내륙이 아닌 바다 쪽 바위 위에 배치됨. 인물은 없음.",
        "hard_violations": [
         "지시된 카메라 위치 및 방향 위반 (내륙이 아닌 바다 쪽에서 촬영)",
         "공간 구조 및 장비 배치 오류 (물이 내륙에서 바다로 흐름)"
        ],
        "physics": "중장비는 바위 위에 지지되어 있으며, 물이 쏟아지는 물리적 표현 자체는 지면으로 자연스럽게 떨어짐."
       },
       {
        "label": "B",
        "direction": "바닷물이 배경의 무너진 방파제를 통해 전경(내륙)으로 정확하게 쏟아져 들어옴.",
        "built_space": "방파제의 수직 내륙면이 보이며, 카메라는 지시된 내륙 측에 위치함. 중장비도 내륙에 올바르게 배치됨.",
        "entities": "지시된 중장비(굴착기, 로더)가 존재하나, 프롬프트에 없는 인물 4명 이상이 장비 주변을 뛰어다님.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 인물 임의 추가 (invented people)"
        ],
        "physics": "중장비는 지면에 고정되어 있고, 물은 중력에 따라 쏟아져 내리며 표면을 덮침."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "상단 중앙 파손부에서 하단 중장비로 쏟아지는 야간 재난 구도는 더 충실하지만, 금지된 인물들을 추가해 실사용 불가입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "참조의 건물과 사석 호안은 잘 보존했지만, 바다 쪽에서 육지를 바라보며 물이 바다 쪽 중장비로 떨어지게 해 핵심 공간 관계와 침수 방향을 뒤집었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "검은 물이 상단 중앙의 제방 파손부에서 아래로 낙하한 뒤 화면 하단의 로더와 굴착기 쪽으로 퍼집니다. 물줄기가 실제 중장비에 닿아 덮치므로 요청된 공격 대상과 흐름의 화면상 연결은 맞습니다. 추가된 인물들은 왼쪽 차량 주변에서 몸을 굽히거나 오른쪽 가장자리로 이동하며, 정확한 시선 목표는 어둠 때문에 판별하기 어렵습니다.",
        "built_space": "제방이 화면 상단을 가로지르고 중앙에 큰 파손부 하나가 있습니다. 상단에는 왼쪽 끝의 잘린 것을 포함해 황백색 방호 블록 아홉 개가 보입니다. 그 아래 작업면에 장비들이 놓여 있어 내륙 쪽 작업장을 내려다보는 구도로 읽힙니다. 다만 참조의 굽은 해안도로와 사석 호안은 확인되지 않고, 벽면도 참조보다 수직적인 콘크리트 벽으로 단순화됐습니다. 앞쪽 로더가 크게 잘려 들어와 장비들이 하단 중앙 중경에 분산된 구도보다는 전경에 치우칩니다.",
        "entities": "큰 로더 두 대, 굴착기 한 대, 소형 작업장비 한 대와 왼쪽 차량들이 보이며, 수리·건설 장비라는 종류는 대체로 맞습니다. 화면 왼쪽 차량 주변과 오른쪽 물가에 적어도 다섯 명의 사람이 추가됐습니다. 이 장면은 인물을 요구하지 않으며, 얼굴이 작고 어두워 이현우의 나이·외모 조건도 확인할 수 없습니다. 밤과 검은 난류는 표현됐고 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "인물을 보여주지 않는 쇼트에 적어도 다섯 명의 사람을 임의로 추가했습니다."
        ],
        "physics": "낙수는 파손부의 높은 수면에서 낮은 작업면으로 떨어지고, 장비 주변에서 포말과 흐름을 형성합니다. 장비의 바퀴와 궤도는 물에 잠긴 지면에 놓여 있으며 굴착기 버킷은 붐과 유압장치에 연결되어 있습니다. 사람들도 지면에 발을 딛거나 차량에 몸을 기대고 있어 명백한 무지지 부유는 없습니다. 다만 대부분의 장비가 똑바로 서 있어 거세게 휩쓸려가는 순간보다는 침수되기 시작한 순간에 가깝습니다."
       },
       {
        "label": "B",
        "direction": "교회와 관리동이 있는 육지 쪽의 물이 중앙 파손부를 넘어 화면 아래의 사석 호안과 중장비 쪽으로 떨어집니다. 참조에서 사석 호안은 바다 쪽이므로, 여기서는 물이 내륙으로 유입되는 것이 아니라 육지에서 바다 쪽으로 빠져나가는 방향입니다. 낙수가 중장비에 닿기는 하지만 그 장비들이 요구된 내륙 쪽에 있지 않습니다. 보이는 인물과 시선은 없습니다.",
        "built_space": "외부 계단 하나가 붙은 관리동 한 채가 왼쪽에, 교회 한 채가 상단 중앙에 있고, 황백색 방호 블록이 반복되는 제방과 바다 쪽 사석 호안이 이어집니다. 중앙 파손부 하나와 떨어진 블록들이 보입니다. 참조의 주요 건축 요소는 A보다 잘 남아 있지만, 카메라는 사석 호안 위 바다 쪽에서 육지를 바라봅니다. 굴착기·덤프트럭·불도저 세 대도 제방 바깥 사석 아래에 배치되어 내륙 중경이라는 위치 조건을 어깁니다.",
        "entities": "하단에는 굴착기 한 대, 덤프트럭 한 대, 불도저 한 대가 서로 다른 사선 각도로 보이며 장비 종류와 분산은 적합합니다. 인물은 없습니다. 배경에는 경광등을 켠 긴급차량 두 대와 작은 차량, 불꽃이 보입니다. 이는 공격받는 주변 지역을 암시하지만, 경광등과 불꽃은 광원을 추가하지 말라는 조명 조건과 맞지 않습니다. 관리동 창문은 뚜렷하게 밝혀져 있지 않고 열린 맨홀은 식별되지 않습니다. 읽을 수 있는 문구는 보이지 않습니다.",
        "hard_violations": [
         "명시된 내륙 쪽 고가 시점 대신 바다 쪽 사석 호안 위에서 육지를 보는 시점으로 배치했습니다.",
         "중장비를 내륙이 아닌 바다 쪽에 놓고, 내륙으로 들어와야 할 물을 육지에서 바다 쪽으로 낙하하게 만들어 핵심 공간 관계를 반대로 구현했습니다."
        ],
        "physics": "물 자체는 높은 배경 수면에서 낮은 전경으로 중력에 따라 떨어져 물리적으로 가능한 흐름입니다. 문제는 그 높은 수면이 육지 쪽이라는 점입니다. 굴착기 궤도는 사석과 잠긴 지면에 걸쳐 있고, 트럭과 불도저 하부는 물에 잠겨 있어 무지지 공중 부유로 보이지 않습니다. 굴착기 버킷도 붐으로 지지됩니다. 물이 장비를 둘러싸지만 장비가 실제로 떠밀리는 증거는 강하지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "상단 중앙 파손부에서 하단 중장비로 쏟아지는 야간 재난 구도는 더 충실하지만, 금지된 인물들을 추가해 실사용 불가입니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "참조의 건물과 사석 호안은 잘 보존했지만, 바다 쪽에서 육지를 바라보며 물이 바다 쪽 중장비로 떨어지게 해 핵심 공간 관계와 침수 방향을 뒤집었습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "검은 물이 상단 중앙의 제방 파손부에서 아래로 낙하한 뒤 화면 하단의 로더와 굴착기 쪽으로 퍼집니다. 물줄기가 실제 중장비에 닿아 덮치므로 요청된 공격 대상과 흐름의 화면상 연결은 맞습니다. 추가된 인물들은 왼쪽 차량 주변에서 몸을 굽히거나 오른쪽 가장자리로 이동하며, 정확한 시선 목표는 어둠 때문에 판별하기 어렵습니다.",
        "built_space": "제방이 화면 상단을 가로지르고 중앙에 큰 파손부 하나가 있습니다. 상단에는 왼쪽 끝의 잘린 것을 포함해 황백색 방호 블록 아홉 개가 보입니다. 그 아래 작업면에 장비들이 놓여 있어 내륙 쪽 작업장을 내려다보는 구도로 읽힙니다. 다만 참조의 굽은 해안도로와 사석 호안은 확인되지 않고, 벽면도 참조보다 수직적인 콘크리트 벽으로 단순화됐습니다. 앞쪽 로더가 크게 잘려 들어와 장비들이 하단 중앙 중경에 분산된 구도보다는 전경에 치우칩니다.",
        "entities": "큰 로더 두 대, 굴착기 한 대, 소형 작업장비 한 대와 왼쪽 차량들이 보이며, 수리·건설 장비라는 종류는 대체로 맞습니다. 화면 왼쪽 차량 주변과 오른쪽 물가에 적어도 다섯 명의 사람이 추가됐습니다. 이 장면은 인물을 요구하지 않으며, 얼굴이 작고 어두워 이현우의 나이·외모 조건도 확인할 수 없습니다. 밤과 검은 난류는 표현됐고 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "인물을 보여주지 않는 쇼트에 적어도 다섯 명의 사람을 임의로 추가했습니다."
        ],
        "physics": "낙수는 파손부의 높은 수면에서 낮은 작업면으로 떨어지고, 장비 주변에서 포말과 흐름을 형성합니다. 장비의 바퀴와 궤도는 물에 잠긴 지면에 놓여 있으며 굴착기 버킷은 붐과 유압장치에 연결되어 있습니다. 사람들도 지면에 발을 딛거나 차량에 몸을 기대고 있어 명백한 무지지 부유는 없습니다. 다만 대부분의 장비가 똑바로 서 있어 거세게 휩쓸려가는 순간보다는 침수되기 시작한 순간에 가깝습니다."
       },
       {
        "label": "A",
        "direction": "교회와 관리동이 있는 육지 쪽의 물이 중앙 파손부를 넘어 화면 아래의 사석 호안과 중장비 쪽으로 떨어집니다. 참조에서 사석 호안은 바다 쪽이므로, 여기서는 물이 내륙으로 유입되는 것이 아니라 육지에서 바다 쪽으로 빠져나가는 방향입니다. 낙수가 중장비에 닿기는 하지만 그 장비들이 요구된 내륙 쪽에 있지 않습니다. 보이는 인물과 시선은 없습니다.",
        "built_space": "외부 계단 하나가 붙은 관리동 한 채가 왼쪽에, 교회 한 채가 상단 중앙에 있고, 황백색 방호 블록이 반복되는 제방과 바다 쪽 사석 호안이 이어집니다. 중앙 파손부 하나와 떨어진 블록들이 보입니다. 참조의 주요 건축 요소는 A보다 잘 남아 있지만, 카메라는 사석 호안 위 바다 쪽에서 육지를 바라봅니다. 굴착기·덤프트럭·불도저 세 대도 제방 바깥 사석 아래에 배치되어 내륙 중경이라는 위치 조건을 어깁니다.",
        "entities": "하단에는 굴착기 한 대, 덤프트럭 한 대, 불도저 한 대가 서로 다른 사선 각도로 보이며 장비 종류와 분산은 적합합니다. 인물은 없습니다. 배경에는 경광등을 켠 긴급차량 두 대와 작은 차량, 불꽃이 보입니다. 이는 공격받는 주변 지역을 암시하지만, 경광등과 불꽃은 광원을 추가하지 말라는 조명 조건과 맞지 않습니다. 관리동 창문은 뚜렷하게 밝혀져 있지 않고 열린 맨홀은 식별되지 않습니다. 읽을 수 있는 문구는 보이지 않습니다.",
        "hard_violations": [
         "명시된 내륙 쪽 고가 시점 대신 바다 쪽 사석 호안 위에서 육지를 보는 시점으로 배치했습니다.",
         "중장비를 내륙이 아닌 바다 쪽에 놓고, 내륙으로 들어와야 할 물을 육지에서 바다 쪽으로 낙하하게 만들어 핵심 공간 관계를 반대로 구현했습니다."
        ],
        "physics": "물 자체는 높은 배경 수면에서 낮은 전경으로 중력에 따라 떨어져 물리적으로 가능한 흐름입니다. 문제는 그 높은 수면이 육지 쪽이라는 점입니다. 굴착기 궤도는 사석과 잠긴 지면에 걸쳐 있고, 트럭과 불도저 하부는 물에 잠겨 있어 무지지 공중 부유로 보이지 않습니다. 굴착기 버킷도 붐으로 지지됩니다. 물이 장비를 둘러싸지만 장비가 실제로 떠밀리는 증거는 강하지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 지시된 카메라 위치 및 방향 위반 (내륙이 아닌 바다 쪽에서 촬영)",
     "[gemini-pro] 공간 구조 및 장비 배치 오류 (물이 내륙에서 바다로 흐름)",
     "[gpt-high] 명시된 내륙 쪽 고가 시점 대신 바다 쪽 사석 호안 위에서 육지를 보는 시점으로 배치했습니다.",
     "[gpt-high] 중장비를 내륙이 아닌 바다 쪽에 놓고, 내륙으로 들어와야 할 물을 육지에서 바다 쪽으로 낙하하게 만들어 핵심 공간 관계를 반대로 구현했습니다."
    ],
    "B": [
     "[gemini-pro] 프롬프트에 명시되지 않은 인물 임의 추가 (invented people)",
     "[gpt-high] 인물을 보여주지 않는 쇼트에 적어도 다섯 명의 사람을 임의로 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 1417
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "구도와 해수 유입 방향은 지시대로 구현했으나, 프롬프트에 없는 인물들을 다수 임의로 추가하여 치명적인 위반이 발생함.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 인물 임의 추가 (invented people) / [gpt-high] 인물을 보여주지 않는 쇼트에 적어도 다섯 명의 사람을 임의로 추가했습니다."
   },
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "내륙 측 시점 및 해수 유입 방향을 바다 쪽으로 완전히 반대로 연출하여 심각한 공간적, 구조적 구도 위반을 범함.  ★위반: [gemini-pro] 지시된 카메라 위치 및 방향 위반 (내륙이 아닌 바다 쪽에서 촬영) / [gemini-pro] 공간 구조 및 장비 배치 오류 (물이 내륙에서 바다로 흐름) / [gpt-high] 명시된 내륙 쪽 고가 시점 대신 바다 쪽 사석 호안 위에서 육지를 보는 시점으로 배치했습니다. / [gpt-high] 중장비를 내륙이 아닌 바다 쪽에 놓고, 내륙으로 들어와야 할 물을 육지에서 바다 쪽으로 낙하하게 만들어 핵심 공간 관계를 반대로 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_camp_church_airspace_sel.png",
    "asset_id": "df4f6cc4-e0e9-495c-bf77-aa7acebc593c",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb7-3d34-73fa-8506-8825aab4b405",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S33sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:34:27.835274+00:00",
  "fingerprint": "c7096458f573f11c6fc146aea805248e8862a9f5f56d678cfa3d7b6bf024570b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S33sh14_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S33sh14_sel.png",
  "source_sha256": "77c5b372af9a06d9af65fff8f9929742d636980b6f7e8ae8565b52e4e334d526",
  "file": "S33sh14_cine.png",
  "staged_sha256": "3191372f65a2bfe5ea500f060d0642880cd84ea86f16197eb87de3dfd321cc87",
  "latency_ms": 12712
 },
 "S34sh2::signage": {
  "fp": "cee1f60174b26200",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S34sh2::bgfirst_bg": {
  "input_fingerprint": "56cc3419382c7a30",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 어깨 위로 번쩍 치켜든 붉은색 소화기 바닥이 낡은 철문 손잡이에 맹렬히 내리꽂히는 타격의 순간.\n\nLOCATION (lock): In the third-floor corridor outside the locked detention-room door of the illuminated administration building.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Red fire extinguisher (Its base is striking the door handle) — The base angles downward into the handle, with part of the body visible beside his hands; used as Defines the strike without obscuring the contact point or human scale; Old iron door and handle (The door remains closed as the handle receives the blow) — A cropped corridor-facing section and the projecting handle are seen obliquely; used as Keeps the impact target and resistant barrier legible together.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained interior illumination with readable impact detail, allowing the extinguisher's stated red color to remain the selective accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 어깨 위로 번쩍 치켜든 붉은색 소화기 바닥이 낡은 철문 손잡이에 맹렬히 내리꽂히는 타격의 순간.\n\nLOCATION (lock): In the third-floor corridor outside the locked detention-room door of the illuminated administration building.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Red fire extinguisher (Its base is striking the door handle) — The base angles downward into the handle, with part of the body visible beside his hands; used as Defines the strike without obscuring the contact point or human scale; Old iron door and handle (The door remains closed as the handle receives the blow) — A cropped corridor-facing section and the projecting handle are seen obliquely; used as Keeps the impact target and resistant barrier legible together.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained interior illumination with readable impact detail, allowing the extinguisher's stated red color to remain the selective accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S34sh2__bgfirst_bg.png",
  "asset_id": "257ba7fb-cad9-404a-9e68-e61f8727355a",
  "input_asset_ids": [
   "99714a35-51e9-4658-8319-a0fca25f2aa9",
   "1f859547-fe0c-4349-bcf6-82d452b30f86"
  ]
 },
 "S34sh2": {
  "input_fingerprint": "c6b6b3601c5a7715",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 어깨 위로 번쩍 치켜든 붉은색 소화기 바닥이 낡은 철문 손잡이에 맹렬히 내리꽂히는 타격의 순간.\n\nLOCATION (lock): In the third-floor corridor outside the locked detention-room door of the illuminated administration building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Red fire extinguisher (Its base is striking the door handle) — The base angles downward into the handle, with part of the body visible beside his hands; used as Defines the strike without obscuring the contact point or human scale; Old iron door and handle (The door remains closed as the handle receives the blow) — A cropped corridor-facing section and the projecting handle are seen obliquely; used as Keeps the impact target and resistant barrier legible together.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained interior illumination with readable impact detail, allowing the extinguisher's stated red color to remain the selective accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor detention-room door remains locked while its handle is being struck; the handle does not break away until the fifth or sixth attempt. The management building is still illuminated, and the fire extinguisher has been removed from its corner position. 이현우: He stands outside the locked room holding the fire extinguisher, with facial bruises and the unhealed leg wound. His outer garment remains removed and his contact card stays concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 어깨 위로 번쩍 치켜든 붉은색 소화기 바닥이 낡은 철문 손잡이에 맹렬히 내리꽂히는 타격의 순간.\n\nLOCATION (lock): In the third-floor corridor outside the locked detention-room door of the illuminated administration building. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Red fire extinguisher (Its base is striking the door handle) — The base angles downward into the handle, with part of the body visible beside his hands; used as Defines the strike without obscuring the contact point or human scale; Old iron door and handle (The door remains closed as the handle receives the blow) — A cropped corridor-facing section and the projecting handle are seen obliquely; used as Keeps the impact target and resistant barrier legible together.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained interior illumination with readable impact detail, allowing the extinguisher's stated red color to remain the selective accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor detention-room door remains locked while its handle is being struck; the handle does not break away until the fifth or sixth attempt. The management building is still illuminated, and the fire extinguisher has been removed from its corner position. 이현우: He stands outside the locked room holding the fire extinguisher, with facial bruises and the unhealed leg wound. His outer garment remains removed and his contact card stays concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 어깨 위로 번쩍 치켜든 붉은색 소화기 바닥이 낡은 철문 손잡이에 맹렬히 내리꽂히는 타격의 순간.\n\nLOCATION (lock): In the third-floor corridor outside the locked detention-room door of the illuminated administration building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Red fire extinguisher (Its base is striking the door handle) — The base angles downward into the handle, with part of the body visible beside his hands; used as Defines the strike without obscuring the contact point or human scale; Old iron door and handle (The door remains closed as the handle receives the blow) — A cropped corridor-facing section and the projecting handle are seen obliquely; used as Keeps the impact target and resistant barrier legible together.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained interior illumination with readable impact detail, allowing the extinguisher's stated red color to remain the selective accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor detention-room door remains locked while its handle is being struck; the handle does not break away until the fifth or sixth attempt. The management building is still illuminated, and the fire extinguisher has been removed from its corner position. 이현우: He stands outside the locked room holding the fire extinguisher, with facial bruises and the unhealed leg wound. His outer garment remains removed and his contact card stays concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S34sh2__bgfirst_bg.png",
     "asset_id": "257ba7fb-cad9-404a-9e68-e61f8727355a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S34sh2.png",
     "asset_id": "99714a35-51e9-4658-8319-a0fca25f2aa9",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L44B01.png",
     "asset_id": "1f859547-fe0c-4349-bcf6-82d452b30f86",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선과 타격 지점이 철문 손잡이를 향함.",
    "built_space": "참조 사진과 동일한 복도 및 철문 구조가 정확히 배치됨.",
    "entities": "이현우의 얼굴과 멍 자국이 일치하며, 지시대로 겉옷을 벗은 상태임.",
    "hard_violations": [
     "[gemini-pro] 소화기가 손에 쥐어지지 않고 왼쪽 팔뚝에 융합되어 허공에 떠 있는 물리적/해부학적 오류"
    ],
    "physics": "손에는 정체불명의 붉은 부품을 쥐고 있으며, 실제 타격 중인 소화기는 아무런 지지 없이 팔에 기형적으로 결합되어 있음."
   },
   {
    "label": "B",
    "direction": "시선과 소화기의 바닥 면이 문손잡이를 정확히 조준함.",
    "built_space": "참조 사진과 동일한 복도 구조와 철문 요소가 올바르게 구현됨.",
    "entities": "이현우의 인물 특징은 일치하나, 겉옷을 벗으라는 지시와 달리 후드티를 입고 있음.",
    "hard_violations": [],
    "physics": "양손으로 소화기 몸통을 단단히 쥐고 수평으로 타격하는 동작이 물리적으로 자연스럽게 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "겉옷을 벗은 설정은 지켰으나, 소화기가 손이 아닌 팔뚝에 융합되어 있는 치명적인 물리적 오류가 발생함."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "물리적 형태는 온전하나, 어깨 위에서 내리꽂는 자세가 아닌 수평 타격이며 겉옷을 입고 있어 지시문과 불일치함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선과 타격 지점이 철문 손잡이를 향함.",
        "built_space": "참조 사진과 동일한 복도 및 철문 구조가 정확히 배치됨.",
        "entities": "이현우의 얼굴과 멍 자국이 일치하며, 지시대로 겉옷을 벗은 상태임.",
        "hard_violations": [
         "소화기가 손에 쥐어지지 않고 왼쪽 팔뚝에 융합되어 허공에 떠 있는 물리적/해부학적 오류"
        ],
        "physics": "손에는 정체불명의 붉은 부품을 쥐고 있으며, 실제 타격 중인 소화기는 아무런 지지 없이 팔에 기형적으로 결합되어 있음."
       },
       {
        "label": "B",
        "direction": "시선과 소화기의 바닥 면이 문손잡이를 정확히 조준함.",
        "built_space": "참조 사진과 동일한 복도 구조와 철문 요소가 올바르게 구현됨.",
        "entities": "이현우의 인물 특징은 일치하나, 겉옷을 벗으라는 지시와 달리 후드티를 입고 있음.",
        "hard_violations": [],
        "physics": "양손으로 소화기 몸통을 단단히 쥐고 수평으로 타격하는 동작이 물리적으로 자연스럽게 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "겉옷을 벗은 설정은 지켰으나, 소화기가 손이 아닌 팔뚝에 융합되어 있는 치명적인 물리적 오류가 발생함."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "물리적 형태는 온전하나, 어깨 위에서 내리꽂는 자세가 아닌 수평 타격이며 겉옷을 입고 있어 지시문과 불일치함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선과 타격 지점이 철문 손잡이를 향함.",
        "built_space": "참조 사진과 동일한 복도 및 철문 구조가 정확히 배치됨.",
        "entities": "이현우의 얼굴과 멍 자국이 일치하며, 지시대로 겉옷을 벗은 상태임.",
        "hard_violations": [
         "소화기가 손에 쥐어지지 않고 왼쪽 팔뚝에 융합되어 허공에 떠 있는 물리적/해부학적 오류"
        ],
        "physics": "손에는 정체불명의 붉은 부품을 쥐고 있으며, 실제 타격 중인 소화기는 아무런 지지 없이 팔에 기형적으로 결합되어 있음."
       },
       {
        "label": "B",
        "direction": "시선과 소화기의 바닥 면이 문손잡이를 정확히 조준함.",
        "built_space": "참조 사진과 동일한 복도 구조와 철문 요소가 올바르게 구현됨.",
        "entities": "이현우의 인물 특징은 일치하나, 겉옷을 벗으라는 지시와 달리 후드티를 입고 있음.",
        "hard_violations": [],
        "physics": "양손으로 소화기 몸통을 단단히 쥐고 수평으로 타격하는 동작이 물리적으로 자연스럽게 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "소화기 바닥이 닫힌 철문 손잡이에 닿는 타격과 양손 지지는 명확하지만, 어깨 위에서 내리꽂는 동작보다 낮게 밀어치는 자세이며 벗어야 할 겉옷을 입고 있다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "겉옷을 벗은 상태와 야간 복도는 맞지만, 손잡이를 치는 부분이 소화기 바닥이 아니라 밸브 쪽이고 양손도 몸통 대신 호스 끝 부품을 잡아 핵심 타격 동작이 틀렸다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 오른쪽 아래 철문 손잡이를 향한다. 소화기는 왼쪽 위의 목 부분에서 오른쪽 아래의 바닥으로 기울어 있고, 바닥 테두리가 손잡이에 접촉한다. 타격 대상과 사용하는 끝은 맞지만, 팔과 소화기가 낮아 어깨 위에서 내리꽂는 동작보다는 전방으로 밀어치는 동작에 가깝다.",
        "built_space": "오른쪽에 닫힌 철문 한 개, 상단 관찰창 한 개, 잠금 걸쇠 한 개와 타격받는 손잡이 한 개가 보인다. 인물은 문 바깥 복도에 서 있다. 뒤로 출입구들과 켜진 천장 형광등이 보이며, 벽 하단의 어두운 띠와 낡은 도장면은 참조와 유사하다. 다만 바닥은 참조의 연속된 마감과 달리 격자형 타일로 보인다. 구도는 무릎 부근까지 포함해 요구된 미디엄 숏보다 다소 넓다.",
        "entities": "검은 헝클어진 머리의 젊은 동아시아계 남성 한 명이며, 얼굴과 체격은 이현우 참조에 대체로 가깝고 뺨의 멍도 보인다. 그러나 제거되어 있어야 할 어두운 후드 겉옷을 그대로 입었다. 낡은 청바지와 붉은 소화기 한 개, 오래된 철문이 보인다. 드러난 무릎에서는 미치유 상처가 명확하지 않으며 신발과 숨긴 카드는 화면 밖이다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "한 손은 소화기 목과 손잡이 부근을 잡고 다른 손은 몸통 아래를 받쳐 무게를 지지한다. 다리를 벌리고 상체를 문 쪽으로 기울인 자세이며 발은 화면 밖이다. 문 손잡이는 부착된 채 충격을 받고 주변에 작은 파편이 튄다. 물체가 지지 없이 떠 있지는 않지만, 현재의 낮은 양손 자세는 강한 머리 위 내려치기를 충분히 표현하지 못한다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 아래 문손잡이를 바라본다. 소화기의 붉은 몸통은 오른쪽 아래로 기울지만, 손잡이에 닿는 쪽은 바닥이 아니라 밸브와 검은 작동 레버가 있는 목 부분이다. 위로 든 양손은 문손잡이가 아니라 머리 위 호스 끝의 붉은 부품을 잡고 있어, 지정된 바닥 타격 방향과 다르다.",
        "built_space": "오른쪽에 닫힌 철문 한 개, 사각 관찰창 한 개, 손잡이와 아래쪽 열쇠구멍이 있는 세로 철판 한 개가 보인다. 인물은 복도 쪽에 서 있고, 뒤에는 두 개의 켜진 형광등과 복도 끝 창문 한 개가 보인다. 창밖은 어두워 야간 조건에 맞는다. 벽의 재료와 하단 띠, 바닥은 참조와 유사하나 문 잠금장치의 구성은 다르다. 무릎 부근까지 보이는 구도이고 들어 올린 팔 일부는 상단에서 잘린다.",
        "entities": "검은 헝클어진 머리와 뺨의 상처가 있는 젊은 동아시아계 남성 한 명으로, 참조 인물의 외모에 대체로 부합한다. 회색 반소매 티셔츠와 낡은 청바지는 겉옷을 벗었다는 조건에 맞는다. 붉은 소화기 몸통은 보이지만, 양손이 잡은 작은 붉은 원통형 부품과 몸통의 연결 및 용도는 불명확하다. 다리 상처와 신발 속 카드는 이 구도로 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "소화기 몸통 위쪽은 겨드랑이와 팔 부근에 닿아 있어 완전히 지지 없이 떠 있다고 단정할 수는 없다. 그러나 몸통이나 운반 손잡이를 잡는 손은 없고, 양손은 머리 위 호스 끝 부품에 놓여 있다. 이 지지 방식으로 몸통 바닥을 어깨 위에서 손잡이에 내려꽂는 동작은 표현되지 않는다. 벌린 다리와 기울어진 상체는 버티는 자세로 가능하며, 문손잡이는 떨어지지 않고 붙어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "소화기 바닥이 닫힌 철문 손잡이에 닿는 타격과 양손 지지는 명확하지만, 어깨 위에서 내리꽂는 동작보다 낮게 밀어치는 자세이며 벗어야 할 겉옷을 입고 있다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "겉옷을 벗은 상태와 야간 복도는 맞지만, 손잡이를 치는 부분이 소화기 바닥이 아니라 밸브 쪽이고 양손도 몸통 대신 호스 끝 부품을 잡아 핵심 타격 동작이 틀렸다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 시선은 오른쪽 아래 철문 손잡이를 향한다. 소화기는 왼쪽 위의 목 부분에서 오른쪽 아래의 바닥으로 기울어 있고, 바닥 테두리가 손잡이에 접촉한다. 타격 대상과 사용하는 끝은 맞지만, 팔과 소화기가 낮아 어깨 위에서 내리꽂는 동작보다는 전방으로 밀어치는 동작에 가깝다.",
        "built_space": "오른쪽에 닫힌 철문 한 개, 상단 관찰창 한 개, 잠금 걸쇠 한 개와 타격받는 손잡이 한 개가 보인다. 인물은 문 바깥 복도에 서 있다. 뒤로 출입구들과 켜진 천장 형광등이 보이며, 벽 하단의 어두운 띠와 낡은 도장면은 참조와 유사하다. 다만 바닥은 참조의 연속된 마감과 달리 격자형 타일로 보인다. 구도는 무릎 부근까지 포함해 요구된 미디엄 숏보다 다소 넓다.",
        "entities": "검은 헝클어진 머리의 젊은 동아시아계 남성 한 명이며, 얼굴과 체격은 이현우 참조에 대체로 가깝고 뺨의 멍도 보인다. 그러나 제거되어 있어야 할 어두운 후드 겉옷을 그대로 입었다. 낡은 청바지와 붉은 소화기 한 개, 오래된 철문이 보인다. 드러난 무릎에서는 미치유 상처가 명확하지 않으며 신발과 숨긴 카드는 화면 밖이다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "한 손은 소화기 목과 손잡이 부근을 잡고 다른 손은 몸통 아래를 받쳐 무게를 지지한다. 다리를 벌리고 상체를 문 쪽으로 기울인 자세이며 발은 화면 밖이다. 문 손잡이는 부착된 채 충격을 받고 주변에 작은 파편이 튄다. 물체가 지지 없이 떠 있지는 않지만, 현재의 낮은 양손 자세는 강한 머리 위 내려치기를 충분히 표현하지 못한다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 아래 문손잡이를 바라본다. 소화기의 붉은 몸통은 오른쪽 아래로 기울지만, 손잡이에 닿는 쪽은 바닥이 아니라 밸브와 검은 작동 레버가 있는 목 부분이다. 위로 든 양손은 문손잡이가 아니라 머리 위 호스 끝의 붉은 부품을 잡고 있어, 지정된 바닥 타격 방향과 다르다.",
        "built_space": "오른쪽에 닫힌 철문 한 개, 사각 관찰창 한 개, 손잡이와 아래쪽 열쇠구멍이 있는 세로 철판 한 개가 보인다. 인물은 복도 쪽에 서 있고, 뒤에는 두 개의 켜진 형광등과 복도 끝 창문 한 개가 보인다. 창밖은 어두워 야간 조건에 맞는다. 벽의 재료와 하단 띠, 바닥은 참조와 유사하나 문 잠금장치의 구성은 다르다. 무릎 부근까지 보이는 구도이고 들어 올린 팔 일부는 상단에서 잘린다.",
        "entities": "검은 헝클어진 머리와 뺨의 상처가 있는 젊은 동아시아계 남성 한 명으로, 참조 인물의 외모에 대체로 부합한다. 회색 반소매 티셔츠와 낡은 청바지는 겉옷을 벗었다는 조건에 맞는다. 붉은 소화기 몸통은 보이지만, 양손이 잡은 작은 붉은 원통형 부품과 몸통의 연결 및 용도는 불명확하다. 다리 상처와 신발 속 카드는 이 구도로 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "소화기 몸통 위쪽은 겨드랑이와 팔 부근에 닿아 있어 완전히 지지 없이 떠 있다고 단정할 수는 없다. 그러나 몸통이나 운반 손잡이를 잡는 손은 없고, 양손은 머리 위 호스 끝 부품에 놓여 있다. 이 지지 방식으로 몸통 바닥을 어깨 위에서 손잡이에 내려꽂는 동작은 표현되지 않는다. 벌린 다리와 기울어진 상체는 버티는 자세로 가능하며, 문손잡이는 떨어지지 않고 붙어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.1,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.85,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 소화기가 손에 쥐어지지 않고 왼쪽 팔뚝에 융합되어 허공에 떠 있는 물리적/해부학적 오류"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 850,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 850,
    "verdict_ko": "겉옷을 벗은 설정은 지켰으나, 소화기가 손이 아닌 팔뚝에 융합되어 있는 치명적인 물리적 오류가 발생함.  ★위반: [gemini-pro] 소화기가 손에 쥐어지지 않고 왼쪽 팔뚝에 융합되어 허공에 떠 있는 물리적/해부학적 오류"
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "물리적 형태는 온전하나, 어깨 위에서 내리꽂는 자세가 아닌 수평 타격이며 겉옷을 입고 있어 지시문과 불일치함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L44B01.png",
    "asset_id": "1f859547-fe0c-4349-bcf6-82d452b30f86",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb8-0e27-71cd-b63c-fcab2db5abd1",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S34sh2__bgfirst_bg.png",
   "bg_asset_id": "257ba7fb-cad9-404a-9e68-e61f8727355a",
   "bg_record_key": "S34sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S34sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:34:40.971148+00:00",
  "fingerprint": "14061d69998505fff0eca550d5a68bfc53619d1bc69633bf321de9f0d13649a1",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S34sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S34sh2_sel.png",
  "source_sha256": "091d9d6d8561a13d94c862b6bcdd4cae79e5236476d830b1307190e9eeb63d33",
  "file": "S34sh2_cine.png",
  "staged_sha256": "301b455e4314842f9c53922a0581bb96a8e422dc5819fdf4bd5c3b1641811e39",
  "latency_ms": 11234
 },
 "S34sh5::signage": {
  "fp": "5769583449758eed",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S34sh5": {
  "input_fingerprint": "4226b9c92250af9e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 미연의 어깨를 두 팔로 단단히 감싸 안으며 미간을 찌푸린 애절한 얼굴.\n\nLOCATION (lock): Just inside the forced-open detention room on the administration building's lit third floor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained interior illumination and gentle facial contrast, letting the emotional warmth come from the embrace rather than an invented warm light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The detention-room handle has broken off and the door is now open; the removed fire extinguisher must not reappear untouched in its original corner. The building remains illuminated, with no floodwater inside the room yet established. 이현우: He has entered the room and stands with his arms raised in an embrace, retaining facial bruising and the injured leg. His outer garment remains removed and his contact card stays concealed in his shoe. 미연: She is no longer separated by the locked door and still has severe facial swelling and visible injuries from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 미연의 어깨를 두 팔로 단단히 감싸 안으며 미간을 찌푸린 애절한 얼굴.\n\nLOCATION (lock): Just inside the forced-open detention room on the administration building's lit third floor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained interior illumination and gentle facial contrast, letting the emotional warmth come from the embrace rather than an invented warm light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The detention-room handle has broken off and the door is now open; the removed fire extinguisher must not reappear untouched in its original corner. The building remains illuminated, with no floodwater inside the room yet established. 이현우: He has entered the room and stands with his arms raised in an embrace, retaining facial bruising and the injured leg. His outer garment remains removed and his contact card stays concealed in his shoe. 미연: She is no longer separated by the locked door and still has severe facial swelling and visible injuries from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 미연의 어깨를 두 팔로 단단히 감싸 안으며 미간을 찌푸린 애절한 얼굴.\n\nLOCATION (lock): Just inside the forced-open detention room on the administration building's lit third floor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained interior illumination and gentle facial contrast, letting the emotional warmth come from the embrace rather than an invented warm light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The detention-room handle has broken off and the door is now open; the removed fire extinguisher must not reappear untouched in its original corner. The building remains illuminated, with no floodwater inside the room yet established. 이현우: He has entered the room and stands with his arms raised in an embrace, retaining facial bruising and the injured leg. His outer garment remains removed and his contact card stays concealed in his shoe. 미연: She is no longer separated by the locked door and still has severe facial swelling and visible injuries from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "이현우가 시선을 아래로 향한 채 미연을 안고 있음.",
    "built_space": "구금실 내부 벽면과 오른쪽에 문이 보임. 문이 약간 열려 있으나 부서져야 할 손잡이가 멀쩡히 달려 있음.",
    "entities": "이현우는 레퍼런스 얼굴과 일치하나 후드티를 그대로 입고 있음. 미연은 회색 티셔츠를 입고 멍든 얼굴이 보임.",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy (왼쪽에서 넘어오는 오른손이 2개로 그려짐)"
    ],
    "physics": "포옹하는 자세에서 이현우의 팔이 왼쪽에서 두 개가 뻗어나와 등을 감싸고 있어 구조적으로 불가능함."
   },
   {
    "label": "B",
    "direction": "이현우가 허공을 보며 미간을 찌푸리고 눈물을 흘리며 미연을 안고 있음.",
    "built_space": "구금실 내부 복도 형태이며 오른쪽에 문이 열려 있음.",
    "entities": "이현우는 상의를 탈의한 상태로 지시에 부합함. 미연은 뒷모습만 보이나 회색 티셔츠를 입고 있음.",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy (미연의 등을 감싸는 팔이 3개이며, 탈의 상태임에도 소매가 달린 팔이 존재함)"
    ],
    "physics": "미연의 등 뒤로 소매가 있는 팔 1개, 맨팔 1개가 왼쪽에서 넘어오고 오른쪽에서도 맨팔이 감싸고 있어 총 3개의 팔이 공존함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "애절한 표정과 상의 탈의 설정은 지시에 부합하나, 미연을 감싸는 팔이 3개로 그려지는 치명적인 해부학적 오류가 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "겉옷을 벗었다는 지시를 무시하고 후드티를 입고 있으며, 포옹하는 손이 모두 같은 방향에서 나오는 해부학적 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우가 시선을 아래로 향한 채 미연을 안고 있음.",
        "built_space": "구금실 내부 벽면과 오른쪽에 문이 보임. 문이 약간 열려 있으나 부서져야 할 손잡이가 멀쩡히 달려 있음.",
        "entities": "이현우는 레퍼런스 얼굴과 일치하나 후드티를 그대로 입고 있음. 미연은 회색 티셔츠를 입고 멍든 얼굴이 보임.",
        "hard_violations": [
         "physically impossible anatomy (왼쪽에서 넘어오는 오른손이 2개로 그려짐)"
        ],
        "physics": "포옹하는 자세에서 이현우의 팔이 왼쪽에서 두 개가 뻗어나와 등을 감싸고 있어 구조적으로 불가능함."
       },
       {
        "label": "B",
        "direction": "이현우가 허공을 보며 미간을 찌푸리고 눈물을 흘리며 미연을 안고 있음.",
        "built_space": "구금실 내부 복도 형태이며 오른쪽에 문이 열려 있음.",
        "entities": "이현우는 상의를 탈의한 상태로 지시에 부합함. 미연은 뒷모습만 보이나 회색 티셔츠를 입고 있음.",
        "hard_violations": [
         "physically impossible anatomy (미연의 등을 감싸는 팔이 3개이며, 탈의 상태임에도 소매가 달린 팔이 존재함)"
        ],
        "physics": "미연의 등 뒤로 소매가 있는 팔 1개, 맨팔 1개가 왼쪽에서 넘어오고 오른쪽에서도 맨팔이 감싸고 있어 총 3개의 팔이 공존함."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "애절한 표정과 상의 탈의 설정은 지시에 부합하나, 미연을 감싸는 팔이 3개로 그려지는 치명적인 해부학적 오류가 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "겉옷을 벗었다는 지시를 무시하고 후드티를 입고 있으며, 포옹하는 손이 모두 같은 방향에서 나오는 해부학적 오류가 있습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "이현우가 시선을 아래로 향한 채 미연을 안고 있음.",
        "built_space": "구금실 내부 벽면과 오른쪽에 문이 보임. 문이 약간 열려 있으나 부서져야 할 손잡이가 멀쩡히 달려 있음.",
        "entities": "이현우는 레퍼런스 얼굴과 일치하나 후드티를 그대로 입고 있음. 미연은 회색 티셔츠를 입고 멍든 얼굴이 보임.",
        "hard_violations": [
         "physically impossible anatomy (왼쪽에서 넘어오는 오른손이 2개로 그려짐)"
        ],
        "physics": "포옹하는 자세에서 이현우의 팔이 왼쪽에서 두 개가 뻗어나와 등을 감싸고 있어 구조적으로 불가능함."
       },
       {
        "label": "B",
        "direction": "이현우가 허공을 보며 미간을 찌푸리고 눈물을 흘리며 미연을 안고 있음.",
        "built_space": "구금실 내부 복도 형태이며 오른쪽에 문이 열려 있음.",
        "entities": "이현우는 상의를 탈의한 상태로 지시에 부합함. 미연은 뒷모습만 보이나 회색 티셔츠를 입고 있음.",
        "hard_violations": [
         "physically impossible anatomy (미연의 등을 감싸는 팔이 3개이며, 탈의 상태임에도 소매가 달린 팔이 존재함)"
        ],
        "physics": "미연의 등 뒤로 소매가 있는 팔 1개, 맨팔 1개가 왼쪽에서 넘어오고 오른쪽에서도 맨팔이 감싸고 있어 총 3개의 팔이 공존함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "클로즈업과 애절하게 찌푸린 표정은 충실하지만, 맨상체와 후드 소매가 뒤섞인 착장이 부자연스럽고 현우의 얼굴도 참조보다 나이 들어 보인다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "두 팔로 어깨를 감싼 동작, 현우의 얼굴 정체성, 미연의 부상과 공간 연속성이 더 명확하지만, 벗어 둬야 할 겉옷을 그대로 입었다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 미연의 머리 옆에서 아래쪽을 바라보며 미간을 깊게 찌푸린다. 미연은 카메라에 등을 보이고 현우 쪽으로 얼굴을 묻어 시선은 보이지 않는다. 위쪽 손은 미연의 반대편 어깨를, 아래쪽 손은 등을 감싸므로 포옹의 대상과 방향은 맞는다.",
        "built_space": "때 탄 밝은 벽과 하단의 짙은 띠, 천장 형광등 하나, 오른쪽 가장자리의 금속 문 또는 문틀 일부가 보인다. 이전 장면의 재료와 차가운 실내 조명은 이어진다. 두 사람은 방 안에서 밀착해 화면 대부분을 차지한다. 문 개방 상태와 부러진 손잡이는 잘려 확인할 수 없고, 의자도 화면에 없다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "인물은 젊은 한국계 남성으로 표현된 현우와 중년 한국인 여성으로 표현된 미연 두 명뿐이다. 현우의 헝클어진 검은 머리와 얼굴 멍은 맞지만 얼굴은 참조보다 성숙하고 윤곽도 다소 다르다. 현우는 상체와 한쪽 위팔을 드러낸 반면 두 손목 주변에는 참조 후드와 같은 낡은 소매가 남아 있어 겉옷을 완전히 벗은 상태가 명확하지 않다. 미연의 검은 묶은 머리와 회색 상의는 이전 장면과 부합한다. 미연의 얼굴은 가려져 부종을 평가할 수 없다. 다리, 신발 속 카드, 소화기는 프레임 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 손은 미연의 어깨와 등에 실제로 닿아 있고 팔은 몸을 둘러 감싼다. 추가 손이나 공중에 뜬 신체는 보이지 않는다. 하체는 잘려 지면 접촉을 확인할 수 없지만 상체가 부유한다고 볼 근거도 없다. 맨팔과 소매의 연결은 어색하나, 소매가 팔에 걸려 있는 해석이 가능하므로 불가능한 해부 구조로 단정하지 않는다."
       },
       {
        "label": "B",
        "direction": "현우는 미연의 어깨 쪽으로 시선을 낮추고 미간을 찌푸린다. 미연은 왼쪽 아래로 얼굴을 돌리고 눈을 거의 감은 채 현우에게 기대어 있다. 현우의 두 팔과 손은 미연의 어깨와 등으로 향해 실제로 감싸며, 카메라를 바라보는 인물은 없다.",
        "built_space": "밝은 오염 벽과 짙은 하단 띠, 위쪽에 일부 보이는 천장등 하나, 오른쪽의 열린 금속문 하나와 작은 사각 창 하나가 보인다. 문 뒤에는 어두운 틈이 있으며 손잡이 부근의 금속 부품은 흐려 정확한 파손 상태를 확인하기 어렵다. 두 사람은 문 안쪽에서 밀착하고, 배경 설비는 작은 면적으로 남아 클로즈업을 방해하지 않는다. 이전 장면의 공간 재료와 조명에 부합하며 중복 설비나 반사는 없다.",
        "entities": "현우와 미연 두 명만 보인다. 현우의 젊은 한국계 남성 얼굴, 검은 헝클어진 머리, 체격은 참조에 더 가깝고 얼굴 멍도 남아 있다. 다만 낡은 후드 겉옷을 온전히 입어 겉옷 제거 조건을 위반한다. 미연은 중년 한국인 여성으로 보이며 이전 장면의 회색 상의, 뒤로 모은 검은 머리, 심하게 멍들고 부은 얼굴을 유지한다. 다리 부상과 신발 속 카드는 프레임 밖이라 평가하지 않는다. 소화기나 추가 인물, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "현우의 양팔은 어깨에서 소매와 팔꿈치를 거쳐 손까지 자연스럽게 이어진다. 한 손은 미연의 어깨를 누르고 다른 손은 등을 받쳐 포옹의 압력과 접촉이 분명하다. 미연은 현우의 가슴과 팔에 기대어 있다. 발과 의자는 보이지 않아 하체 지지는 확인할 수 없지만, 공중에 뜨거나 지지 없이 놓인 신체와 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "클로즈업과 애절하게 찌푸린 표정은 충실하지만, 맨상체와 후드 소매가 뒤섞인 착장이 부자연스럽고 현우의 얼굴도 참조보다 나이 들어 보인다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 팔로 어깨를 감싼 동작, 현우의 얼굴 정체성, 미연의 부상과 공간 연속성이 더 명확하지만, 벗어 둬야 할 겉옷을 그대로 입었다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 미연의 머리 옆에서 아래쪽을 바라보며 미간을 깊게 찌푸린다. 미연은 카메라에 등을 보이고 현우 쪽으로 얼굴을 묻어 시선은 보이지 않는다. 위쪽 손은 미연의 반대편 어깨를, 아래쪽 손은 등을 감싸므로 포옹의 대상과 방향은 맞는다.",
        "built_space": "때 탄 밝은 벽과 하단의 짙은 띠, 천장 형광등 하나, 오른쪽 가장자리의 금속 문 또는 문틀 일부가 보인다. 이전 장면의 재료와 차가운 실내 조명은 이어진다. 두 사람은 방 안에서 밀착해 화면 대부분을 차지한다. 문 개방 상태와 부러진 손잡이는 잘려 확인할 수 없고, 의자도 화면에 없다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "인물은 젊은 한국계 남성으로 표현된 현우와 중년 한국인 여성으로 표현된 미연 두 명뿐이다. 현우의 헝클어진 검은 머리와 얼굴 멍은 맞지만 얼굴은 참조보다 성숙하고 윤곽도 다소 다르다. 현우는 상체와 한쪽 위팔을 드러낸 반면 두 손목 주변에는 참조 후드와 같은 낡은 소매가 남아 있어 겉옷을 완전히 벗은 상태가 명확하지 않다. 미연의 검은 묶은 머리와 회색 상의는 이전 장면과 부합한다. 미연의 얼굴은 가려져 부종을 평가할 수 없다. 다리, 신발 속 카드, 소화기는 프레임 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 손은 미연의 어깨와 등에 실제로 닿아 있고 팔은 몸을 둘러 감싼다. 추가 손이나 공중에 뜬 신체는 보이지 않는다. 하체는 잘려 지면 접촉을 확인할 수 없지만 상체가 부유한다고 볼 근거도 없다. 맨팔과 소매의 연결은 어색하나, 소매가 팔에 걸려 있는 해석이 가능하므로 불가능한 해부 구조로 단정하지 않는다."
       },
       {
        "label": "A",
        "direction": "현우는 미연의 어깨 쪽으로 시선을 낮추고 미간을 찌푸린다. 미연은 왼쪽 아래로 얼굴을 돌리고 눈을 거의 감은 채 현우에게 기대어 있다. 현우의 두 팔과 손은 미연의 어깨와 등으로 향해 실제로 감싸며, 카메라를 바라보는 인물은 없다.",
        "built_space": "밝은 오염 벽과 짙은 하단 띠, 위쪽에 일부 보이는 천장등 하나, 오른쪽의 열린 금속문 하나와 작은 사각 창 하나가 보인다. 문 뒤에는 어두운 틈이 있으며 손잡이 부근의 금속 부품은 흐려 정확한 파손 상태를 확인하기 어렵다. 두 사람은 문 안쪽에서 밀착하고, 배경 설비는 작은 면적으로 남아 클로즈업을 방해하지 않는다. 이전 장면의 공간 재료와 조명에 부합하며 중복 설비나 반사는 없다.",
        "entities": "현우와 미연 두 명만 보인다. 현우의 젊은 한국계 남성 얼굴, 검은 헝클어진 머리, 체격은 참조에 더 가깝고 얼굴 멍도 남아 있다. 다만 낡은 후드 겉옷을 온전히 입어 겉옷 제거 조건을 위반한다. 미연은 중년 한국인 여성으로 보이며 이전 장면의 회색 상의, 뒤로 모은 검은 머리, 심하게 멍들고 부은 얼굴을 유지한다. 다리 부상과 신발 속 카드는 프레임 밖이라 평가하지 않는다. 소화기나 추가 인물, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "현우의 양팔은 어깨에서 소매와 팔꿈치를 거쳐 손까지 자연스럽게 이어진다. 한 손은 미연의 어깨를 누르고 다른 손은 등을 받쳐 포옹의 압력과 접촉이 분명하다. 미연은 현우의 가슴과 팔에 기대어 있다. 발과 의자는 보이지 않아 하체 지지는 확인할 수 없지만, 공중에 뜨거나 지지 없이 놓인 신체와 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.464
   },
   "violations": {
    "A": [
     "[gemini-pro] physically impossible anatomy (왼쪽에서 넘어오는 오른손이 2개로 그려짐)"
    ],
    "B": [
     "[gemini-pro] physically impossible anatomy (미연의 등을 감싸는 팔이 3개이며, 탈의 상태임에도 소매가 달린 팔이 존재함)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1464,
   "A": 1500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1464,
    "verdict_ko": "애절한 표정과 상의 탈의 설정은 지시에 부합하나, 미연을 감싸는 팔이 3개로 그려지는 치명적인 해부학적 오류가 있습니다.  ★위반: [gemini-pro] physically impossible anatomy (미연의 등을 감싸는 팔이 3개이며, 탈의 상태임에도 소매가 달린 팔이 존재함)"
   },
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "겉옷을 벗었다는 지시를 무시하고 후드티를 입고 있으며, 포옹하는 손이 모두 같은 방향에서 나오는 해부학적 오류가 있습니다.  ★위반: [gemini-pro] physically impossible anatomy (왼쪽에서 넘어오는 오른손이 2개로 그려짐)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S31sh2_sel.png",
    "asset_id": "c4633ce2-fdcc-4ff0-8fc9-260786f3d692",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:878964>",
    "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb8-c89e-7e78-966b-774cd9d77607",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S31sh2"
  }
 },
 "S34sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:34:52.550270+00:00",
  "fingerprint": "646f6de8ed080e3d7604b81927573e7715509356bc0d952461a860f75e874ab4",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S34sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S34sh5_sel.png",
  "source_sha256": "16e2e976951bca99246abcb5e891fda44fa43169f6d0427ac18e2f554d2267bd",
  "file": "S34sh5_cine.png",
  "staged_sha256": "22d5807f31ec0651b2c95846c22eccca8c45ffa86756a77d71b3a4efb1732c9f",
  "latency_ms": 11360
 },
 "S34sh8::signage": {
  "fp": "2a321a53b829ebe7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S34sh8": {
  "input_fingerprint": "9f39a235b20b4d34",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 미연의 손목을 꽉 움켜쥔 채 열린 방 밖을 향해 몸을 절반쯤 돌려 옷자락이 펄럭이는 역동적인 뒷모습.\n\nLOCATION (lock): At the open doorway from the administration building's third-floor detention room into the illuminated corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Room doorway (Open after the door handle was broken) — Seen obliquely from inside the room, with the opening beyond the pair; used as Frames the escape route without blocking the wrist connection; Corridor beyond the room (Visible through the open doorway) — Recedes away from the camera beyond the turning figures; used as Provides destination depth for the beginning of the following movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination preserves the wrist grip and overlapping bodies without introducing a visible light source or a new color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor room door is open, with its handle broken off by repeated blows from a fire extinguisher. The management building remains brightly lit. 이현우: He is leaving the opened room, with facial bruising and an untreated dog-bite injury to his leg. A stiff contact card remains concealed inside his shoe. 미연: Her face remains badly swollen from the beating. She is leaving the room after being released.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 미연의 손목을 꽉 움켜쥔 채 열린 방 밖을 향해 몸을 절반쯤 돌려 옷자락이 펄럭이는 역동적인 뒷모습.\n\nLOCATION (lock): At the open doorway from the administration building's third-floor detention room into the illuminated corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Room doorway (Open after the door handle was broken) — Seen obliquely from inside the room, with the opening beyond the pair; used as Frames the escape route without blocking the wrist connection; Corridor beyond the room (Visible through the open doorway) — Recedes away from the camera beyond the turning figures; used as Provides destination depth for the beginning of the following movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination preserves the wrist grip and overlapping bodies without introducing a visible light source or a new color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor room door is open, with its handle broken off by repeated blows from a fire extinguisher. The management building remains brightly lit. 이현우: He is leaving the opened room, with facial bruising and an untreated dog-bite injury to his leg. A stiff contact card remains concealed inside his shoe. 미연: Her face remains badly swollen from the beating. She is leaving the room after being released.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 미연의 손목을 꽉 움켜쥔 채 열린 방 밖을 향해 몸을 절반쯤 돌려 옷자락이 펄럭이는 역동적인 뒷모습.\n\nLOCATION (lock): At the open doorway from the administration building's third-floor detention room into the illuminated corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Room doorway (Open after the door handle was broken) — Seen obliquely from inside the room, with the opening beyond the pair; used as Frames the escape route without blocking the wrist connection; Corridor beyond the room (Visible through the open doorway) — Recedes away from the camera beyond the turning figures; used as Provides destination depth for the beginning of the following movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination preserves the wrist grip and overlapping bodies without introducing a visible light source or a new color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor room door is open, with its handle broken off by repeated blows from a fire extinguisher. The management building remains brightly lit. 이현우: He is leaving the opened room, with facial bruising and an untreated dog-bite injury to his leg. A stiff contact card remains concealed inside his shoe. 미연: Her face remains badly swollen from the beating. She is leaving the room after being released.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 등진 채 복도 쪽을 향하고 있으며, 그의 오른손은 미연의 왼팔 손목을 정확히 향해 움켜쥐고 있다. 미연의 시선은 복도를 향한다.",
    "built_space": "수용실 내부에서 복도를 비스듬히 바라보는 구도이며, 왼쪽의 열린 문에는 부서진 손잡이 흔적이 사실적으로 묘사되어 있고 복도는 뒤로 깊이감을 주며 이어진다.",
    "entities": "이현우(18세 남성, 헝클어진 머리, 더러운 후드티, 얼굴의 멍)와 미연(40대 후반 여성, 부어오른 얼굴, 갈색 재킷) 모두 지정된 레퍼런스의 특징과 정확히 일치한다.",
    "hard_violations": [],
    "physics": "이현우는 왼발로 바닥을 딛고 체중을 복도 쪽으로 실어 몸을 돌리는 역동적인 자세를 취하고 있으며, 펄럭이는 옷자락이 움직임을 뒷받침하고 그의 오른손이 미연의 손목을 단단히 물리적으로 지탱하고 있다. 공중에 뜬 물체는 발견되지 않는다."
   },
   {
    "label": "B",
    "direction": "이현우와 미연 모두 정면 및 측면으로 복도를 향해 시선을 두고 있으며, 미연의 오른손이 이현우의 오른팔 소매를 쥐고 있다.",
    "built_space": "수용실 내부에서 복도를 바라보는 구도이며, 오른쪽에 열린 문과 심하게 뒤틀린 문손잡이가 있고 정중앙으로 복도가 이어진다.",
    "entities": "이현우와 미연 모두 레퍼런스의 의상, 외모 특징, 상처 등을 잘 반영하여 렌더링되었다.",
    "hard_violations": [
     "[gpt-high] 미연을 복도 쪽에, 현우를 방 안쪽으로 향하는 전경에 배치하여, 두 사람이 방에서 복도로 빠져나가는 지정 동선을 반대로 만들었다."
    ],
    "physics": "두 인물이 바닥에 서서 이동하려는 자세를 취하고 있으며 미연의 손이 이현우의 팔을 지탱하고 있으나, 옷자락이 펄럭이는 등의 역동적인 회전 움직임은 표현되지 않았다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "이현우가 미연의 손목을 꽉 쥐고 밖을 향해 몸을 돌린 역동적인 뒷모습과 펄럭이는 옷자락 등 프롬프트의 핵심 액션과 구도 지시를 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이현우의 뒷모습이 아닌 앞모습이 렌더링되었고, 이현우가 미연의 손목을 잡는 대신 미연이 이현우의 팔을 잡고 있어 샷 텍스트의 연출 지시를 근본적으로 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 등진 채 복도 쪽을 향하고 있으며, 그의 오른손은 미연의 왼팔 손목을 정확히 향해 움켜쥐고 있다. 미연의 시선은 복도를 향한다.",
        "built_space": "수용실 내부에서 복도를 비스듬히 바라보는 구도이며, 왼쪽의 열린 문에는 부서진 손잡이 흔적이 사실적으로 묘사되어 있고 복도는 뒤로 깊이감을 주며 이어진다.",
        "entities": "이현우(18세 남성, 헝클어진 머리, 더러운 후드티, 얼굴의 멍)와 미연(40대 후반 여성, 부어오른 얼굴, 갈색 재킷) 모두 지정된 레퍼런스의 특징과 정확히 일치한다.",
        "hard_violations": [],
        "physics": "이현우는 왼발로 바닥을 딛고 체중을 복도 쪽으로 실어 몸을 돌리는 역동적인 자세를 취하고 있으며, 펄럭이는 옷자락이 움직임을 뒷받침하고 그의 오른손이 미연의 손목을 단단히 물리적으로 지탱하고 있다. 공중에 뜬 물체는 발견되지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우와 미연 모두 정면 및 측면으로 복도를 향해 시선을 두고 있으며, 미연의 오른손이 이현우의 오른팔 소매를 쥐고 있다.",
        "built_space": "수용실 내부에서 복도를 바라보는 구도이며, 오른쪽에 열린 문과 심하게 뒤틀린 문손잡이가 있고 정중앙으로 복도가 이어진다.",
        "entities": "이현우와 미연 모두 레퍼런스의 의상, 외모 특징, 상처 등을 잘 반영하여 렌더링되었다.",
        "hard_violations": [],
        "physics": "두 인물이 바닥에 서서 이동하려는 자세를 취하고 있으며 미연의 손이 이현우의 팔을 지탱하고 있으나, 옷자락이 펄럭이는 등의 역동적인 회전 움직임은 표현되지 않았다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "이현우가 미연의 손목을 꽉 쥐고 밖을 향해 몸을 돌린 역동적인 뒷모습과 펄럭이는 옷자락 등 프롬프트의 핵심 액션과 구도 지시를 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이현우의 뒷모습이 아닌 앞모습이 렌더링되었고, 이현우가 미연의 손목을 잡는 대신 미연이 이현우의 팔을 잡고 있어 샷 텍스트의 연출 지시를 근본적으로 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 등진 채 복도 쪽을 향하고 있으며, 그의 오른손은 미연의 왼팔 손목을 정확히 향해 움켜쥐고 있다. 미연의 시선은 복도를 향한다.",
        "built_space": "수용실 내부에서 복도를 비스듬히 바라보는 구도이며, 왼쪽의 열린 문에는 부서진 손잡이 흔적이 사실적으로 묘사되어 있고 복도는 뒤로 깊이감을 주며 이어진다.",
        "entities": "이현우(18세 남성, 헝클어진 머리, 더러운 후드티, 얼굴의 멍)와 미연(40대 후반 여성, 부어오른 얼굴, 갈색 재킷) 모두 지정된 레퍼런스의 특징과 정확히 일치한다.",
        "hard_violations": [],
        "physics": "이현우는 왼발로 바닥을 딛고 체중을 복도 쪽으로 실어 몸을 돌리는 역동적인 자세를 취하고 있으며, 펄럭이는 옷자락이 움직임을 뒷받침하고 그의 오른손이 미연의 손목을 단단히 물리적으로 지탱하고 있다. 공중에 뜬 물체는 발견되지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우와 미연 모두 정면 및 측면으로 복도를 향해 시선을 두고 있으며, 미연의 오른손이 이현우의 오른팔 소매를 쥐고 있다.",
        "built_space": "수용실 내부에서 복도를 바라보는 구도이며, 오른쪽에 열린 문과 심하게 뒤틀린 문손잡이가 있고 정중앙으로 복도가 이어진다.",
        "entities": "이현우와 미연 모두 레퍼런스의 의상, 외모 특징, 상처 등을 잘 반영하여 렌더링되었다.",
        "hard_violations": [],
        "physics": "두 인물이 바닥에 서서 이동하려는 자세를 취하고 있으며 미연의 손이 이현우의 팔을 지탱하고 있으나, 옷자락이 펄럭이는 등의 역동적인 회전 움직임은 표현되지 않았다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "미연이 현우의 손목을 잡아 주체가 뒤바뀌었고, 현우가 복도를 등지고 방 안쪽으로 향해 탈출 동선과 반대다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "현우가 미연의 손목을 잡고 복도로 돌아서는 중간 크기 뒷모습과 펄럭이는 옷자락을 구현했으나, 미연의 이전 장면 의상과 광원 비노출 조건은 지키지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우의 등은 복도 쪽에 있고 얼굴과 앞선 팔은 화면 오른쪽 방 안쪽을 향한다. 미연은 복도 쪽에서 현우를 바라본다. 접촉부에서는 미연의 손가락이 현우의 손목을 감싸며 현우의 손은 아래로 내려가 있어, 지시된 손목 잡기의 주체가 반대다.",
        "built_space": "방 안에서 열린 출입문 하나와 그 너머 긴 복도를 본다. 오른쪽 문짝에는 도어클로저 하나와 파손된 잠금장치, 아래쪽 손잡이가 보인다. 낡은 벽과 금속문 재질은 참고와 유사하다. 복도 천장의 흰 조명 네 개와 끝의 녹색 발광부가 노출되어 광원을 보이지 말라는 조건을 어긴다. 미연은 문 너머, 현우는 방 안 전경에 있어 방에서 나가는 배치로 읽히지 않는다.",
        "entities": "인물은 젊은 동아시아계 남성 한 명과 중년 동아시아계 여성 한 명뿐이며, 검은 머리와 얼굴의 멍은 지정 인물에 대체로 부합한다. 현우의 낡은 짙은 회색 후드와 청바지는 참고와 맞는다. 미연은 갈색 겉옷을 입어, 이전 장면에서 확정된 회색 상의 차림과 다르다. 다리 상처와 신발 속 카드는 프레임 밖이므로 판단하지 않는다. 읽을 수 있는 문구는 확인되지 않는다.",
        "hard_violations": [
         "미연을 복도 쪽에, 현우를 방 안쪽으로 향하는 전경에 배치하여, 두 사람이 방에서 복도로 빠져나가는 지정 동선을 반대로 만들었다."
        ],
        "physics": "두 사람의 몸통은 아래쪽으로 이어지는 다리 위에 서 있으며 공중에 뜬 신체는 없다. 발은 잘려 있어 접지는 직접 확인할 수 없다. 손목 접촉 자체는 물리적으로 가능하지만 미연이 잡는 동작이다. 현우의 후드 밑단은 대부분 아래로 처져 있어 급회전으로 옷자락이 펄럭이는 순간은 약하다."
       },
       {
        "label": "B",
        "direction": "현우는 등을 카메라에 보이고 몸을 열린 복도로 돌리면서 고개만 오른쪽 뒤의 미연에게 향한다. 미연은 현우를 바라본다. 현우의 뻗은 손이 미연의 손목을 감싸며, 복도 쪽으로 이끄는 관계가 명확하다.",
        "built_space": "방 안에서 출입문 하나를 비스듬히 보며, 왼쪽 열린 금속 문짝에 작은 망입 유리창 하나와 도어클로저 하나, 파손된 잠금장치와 아래쪽 레버가 있다. 문 너머 복도는 끝의 문까지 이어지고 손목 연결을 가리지 않는다. 낡은 밝은 벽과 어두운 하부 띠는 참고의 재질과 이어진다. 흰 천장 조명 세 개와 먼 녹색 발광부가 보여 광원 비노출 조건에는 어긋난다. 방 쪽 바닥에는 소화기 하나가 놓여 있다.",
        "entities": "젊은 동아시아계 남성 현우와 중년 동아시아계 여성 미연만 보인다. 현우의 헝클어진 검은 머리, 얼굴 멍, 해진 짙은 후드와 청바지는 참고와 대체로 일치한다. 미연의 검은 머리와 얼굴 부상은 이어지지만 갈색 재킷은 이전 장면의 회색 상의 차림을 유지하지 못한다. 소화기는 문을 부순 경위에 명시된 물건이다. 다리 상처와 신발 속 카드는 보이지 않아 평가하지 않는다. 소화기 라벨에 문자 형태가 있으나 구체적인 문구의 판독 여부는 불확실하다.",
        "hard_violations": [],
        "physics": "현우의 골반과 다리는 문 쪽으로 진행하고 상체는 미연 쪽으로 비틀려 있어 회전 중인 자세가 가능하다. 발은 프레임 밖이지만 몸이 떠 있는 징후는 없다. 현우의 손은 미연의 손목에 실제로 접촉하고, 미연의 팔도 어깨까지 자연스럽게 이어진다. 후드 밑단이 몸에서 벌어진 모습은 회전 관성으로 설명된다. 문은 경첩으로, 소화기와 바닥의 물건은 바닥으로 지지된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미연이 현우의 손목을 잡아 주체가 뒤바뀌었고, 현우가 복도를 등지고 방 안쪽으로 향해 탈출 동선과 반대다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "현우가 미연의 손목을 잡고 복도로 돌아서는 중간 크기 뒷모습과 펄럭이는 옷자락을 구현했으나, 미연의 이전 장면 의상과 광원 비노출 조건은 지키지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우의 등은 복도 쪽에 있고 얼굴과 앞선 팔은 화면 오른쪽 방 안쪽을 향한다. 미연은 복도 쪽에서 현우를 바라본다. 접촉부에서는 미연의 손가락이 현우의 손목을 감싸며 현우의 손은 아래로 내려가 있어, 지시된 손목 잡기의 주체가 반대다.",
        "built_space": "방 안에서 열린 출입문 하나와 그 너머 긴 복도를 본다. 오른쪽 문짝에는 도어클로저 하나와 파손된 잠금장치, 아래쪽 손잡이가 보인다. 낡은 벽과 금속문 재질은 참고와 유사하다. 복도 천장의 흰 조명 네 개와 끝의 녹색 발광부가 노출되어 광원을 보이지 말라는 조건을 어긴다. 미연은 문 너머, 현우는 방 안 전경에 있어 방에서 나가는 배치로 읽히지 않는다.",
        "entities": "인물은 젊은 동아시아계 남성 한 명과 중년 동아시아계 여성 한 명뿐이며, 검은 머리와 얼굴의 멍은 지정 인물에 대체로 부합한다. 현우의 낡은 짙은 회색 후드와 청바지는 참고와 맞는다. 미연은 갈색 겉옷을 입어, 이전 장면에서 확정된 회색 상의 차림과 다르다. 다리 상처와 신발 속 카드는 프레임 밖이므로 판단하지 않는다. 읽을 수 있는 문구는 확인되지 않는다.",
        "hard_violations": [
         "미연을 복도 쪽에, 현우를 방 안쪽으로 향하는 전경에 배치하여, 두 사람이 방에서 복도로 빠져나가는 지정 동선을 반대로 만들었다."
        ],
        "physics": "두 사람의 몸통은 아래쪽으로 이어지는 다리 위에 서 있으며 공중에 뜬 신체는 없다. 발은 잘려 있어 접지는 직접 확인할 수 없다. 손목 접촉 자체는 물리적으로 가능하지만 미연이 잡는 동작이다. 현우의 후드 밑단은 대부분 아래로 처져 있어 급회전으로 옷자락이 펄럭이는 순간은 약하다."
       },
       {
        "label": "A",
        "direction": "현우는 등을 카메라에 보이고 몸을 열린 복도로 돌리면서 고개만 오른쪽 뒤의 미연에게 향한다. 미연은 현우를 바라본다. 현우의 뻗은 손이 미연의 손목을 감싸며, 복도 쪽으로 이끄는 관계가 명확하다.",
        "built_space": "방 안에서 출입문 하나를 비스듬히 보며, 왼쪽 열린 금속 문짝에 작은 망입 유리창 하나와 도어클로저 하나, 파손된 잠금장치와 아래쪽 레버가 있다. 문 너머 복도는 끝의 문까지 이어지고 손목 연결을 가리지 않는다. 낡은 밝은 벽과 어두운 하부 띠는 참고의 재질과 이어진다. 흰 천장 조명 세 개와 먼 녹색 발광부가 보여 광원 비노출 조건에는 어긋난다. 방 쪽 바닥에는 소화기 하나가 놓여 있다.",
        "entities": "젊은 동아시아계 남성 현우와 중년 동아시아계 여성 미연만 보인다. 현우의 헝클어진 검은 머리, 얼굴 멍, 해진 짙은 후드와 청바지는 참고와 대체로 일치한다. 미연의 검은 머리와 얼굴 부상은 이어지지만 갈색 재킷은 이전 장면의 회색 상의 차림을 유지하지 못한다. 소화기는 문을 부순 경위에 명시된 물건이다. 다리 상처와 신발 속 카드는 보이지 않아 평가하지 않는다. 소화기 라벨에 문자 형태가 있으나 구체적인 문구의 판독 여부는 불확실하다.",
        "hard_violations": [],
        "physics": "현우의 골반과 다리는 문 쪽으로 진행하고 상체는 미연 쪽으로 비틀려 있어 회전 중인 자세가 가능하다. 발은 프레임 밖이지만 몸이 떠 있는 징후는 없다. 현우의 손은 미연의 손목에 실제로 접촉하고, 미연의 팔도 어깨까지 자연스럽게 이어진다. 후드 밑단이 몸에서 벌어진 모습은 회전 관성으로 설명된다. 문은 경첩으로, 소화기와 바닥의 물건은 바닥으로 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.729
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.479
   },
   "violations": {
    "B": [
     "[gpt-high] 미연을 복도 쪽에, 현우를 방 안쪽으로 향하는 전경에 배치하여, 두 사람이 방에서 복도로 빠져나가는 지정 동선을 반대로 만들었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 479
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "이현우가 미연의 손목을 꽉 쥐고 밖을 향해 몸을 돌린 역동적인 뒷모습과 펄럭이는 옷자락 등 프롬프트의 핵심 액션과 구도 지시를 완벽하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 479,
    "verdict_ko": "이현우의 뒷모습이 아닌 앞모습이 렌더링되었고, 이현우가 미연의 손목을 잡는 대신 미연이 이현우의 팔을 잡고 있어 샷 텍스트의 연출 지시를 근본적으로 위반했습니다.  ★위반: [gpt-high] 미연을 복도 쪽에, 현우를 방 안쪽으로 향하는 전경에 배치하여, 두 사람이 방에서 복도로 빠져나가는 지정 동선을 반대로 만들었다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S34sh5_sel.png",
    "asset_id": "9ae88bca-c437-4a52-998b-9bf8218a3a3e",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:878964>",
    "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdb9-83c4-7f5d-9284-1cdcd2d59ffa",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S34sh5"
  }
 },
 "S34sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:35:04.246794+00:00",
  "fingerprint": "eefffae43cc3fb0de6435ad1414c9cb2b2c972d15610f3e5670d53f41026c68c",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S34sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S34sh8_sel.png",
  "source_sha256": "c0344f841c325f415f7c8ab1aec68c0f9770de6f250a7a0e88e7dd3eaadc087f",
  "file": "S34sh8_cine.png",
  "staged_sha256": "6242307d2b1faa2c3a964d206574b6e0bd35921e81da6b1f86f8560d0c669b6d",
  "latency_ms": 13026
 },
 "S35sh2::signage": {
  "fp": "316888f7224c8efb",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::a107ff0c37ce9460": {
  "subjects": [],
  "subject_text": "해일이 밀려드는 인천 난민촌 거리\n인공제방 안쪽으로 컨테이너 주택과 상점들이 이어진 거리. 좁은 골목과 도로가 연결되고 가로등과 방송 스피커가 설치되어 있다.",
  "identity": "canonical",
  "scope_id": "L46",
  "scope_role": "location_exterior",
  "scope_sha": "c743c9fef561ed04"
 },
 "S35sh2::bgfirst_bg": {
  "input_fingerprint": "19d8237d4541d097",
  "prompt": "Turn the attached storyboard sketch into a photorealistic film still of\nits LOCATION ONLY, keeping the figures exactly as they are.\n\nKEEP EXACTLY: the camera framing, the horizon, and where every element\nsits in the frame. Each mannequin figure stays a plain grey featureless\nmannequin standing in the very same spot, at the same size, in the same\npose, turned the same way — do not turn them into people, do not move,\nrotate, mirror or re-pose them, do not add or remove figures.\n\nBUILD PHOTOREALISTICALLY: everything that is not a figure — ground,\nsurfacing, structures, vegetation, sky, water, distance. The bare lines\nof the sketch are a layout guide; replace them with the real materials,\ndepth and lighting of the place described below.\n\nSHOT TEXT this background must serve (Korean): 난민촌 거리에 거대하게 밀려든 검은 바닷물이, 다급히 등을 돌린 채 한 발을 공중에 띄운 질주 자세인 아시아계 난민들의 뒤를 집어삼키기 직전의 맹렬한 찰나.\n\nLOCATION (lock): Along an exposed refugee-settlement street where a towering surge of seawater overtakes fleeing residents at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: escape route ahead of the refugees in the upper-right of the frame, background; seawater advancing behind the refugees in the lower-left of the frame, midground, moves toward escape route ahead of the refugees.\n- KEY BACKGROUND ELEMENTS: Refugee settlement street (Being overtaken by incoming seawater) — The escape route recedes diagonally toward the upper right; used as Keeps the runners' route and the water's pursuit within one continuous space; Advancing seawater (Black water closing immediately behind the fleeing refugees) — Its advancing edge crosses from the lower-left background route toward the runners; used as Makes the remaining gap, rather than a separate spectacle, the primary threat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient illumination and controlled contrast distinguish the fleeing bodies from the black seawater without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHINGS AT THIS PLACE:\nThe site is a continuous street passage within the refugee settlement. Its travel surface and lateral margins form a single connected outdoor corridor, with access continuing beyond both ends of the mapped segment. The street itself provides the camera-placement area, including positions offset toward its margins with views diagonally along the corridor. This circulation surface is the minimum inferred infrastructure for access through the settlement. No building footprints, intersections, shoreline, or embankment position are established by the scene.\n- passage: Refugee settlement street passage\n\nWORLD FACTS (creator-confirmed — always true):\n- Region (real-world reference): 대한민국 중심. 인천과 경기권, 전북 익산, 전남 해남·목포, 제주도가 주요 배경이며, 미국의 과거 재현과 중국 등 해외의 원격 영상이 포함된다. 다국적·다인종 난민 사회가 존재한다.\n- Era: 2069년의 디스토피아적 미래. 수십 년 전의 기억과 과거 재현, 상대적으로 이후인 시점이 함께 포함된다.\n- 현재의 미래 기술 수준을 기본으로 하되, 과거 재현에서는 해당 시점의 연령·의상·기기 상태를 적용한다. 오래된 아날로그 제품이나 내연기관 차량을 시대착오로 제거하지 않으며, 현재의 첨단 장비를 과거 이미지에 자동으로 추가하지 않는다.\n- 스마트폰·노트북·태블릿·화상통화와 고도화된 로봇 기술을 함께 허용한다. 자원 부족 구역의 재사용 장비와 첨단 시설의 정밀 장비를 구분하고, 모든 공간을 동일한 미래형 인터페이스와 새 제품으로 채우지 않는다.\n- 소등·정전·가동 여부를 개별 공간과 시점에 맞춰 유지한다. 에너지 전달이 명시된 경우에만 주변 전기 장치의 밝기나 작동 상태를 변화시키며, 빈곤한 공간 전체를 항상 암흑으로 고정하지 않는다.\n- 육중한 고릴라형 실루엣, 금속 외장, 기계 관절, 푸른 눈과 가슴의 링 구조를 일관되게 유지한다. 감정은 눈의 크기·표시 이모티콘·자세로 표현하며, 어린 정서나 음성을 이유로 인간 아동의 얼굴과 신체로 바꾸지 않는다.\n- 인간형 모델은 명시된 인간 유사 외형을 유지하고, 군사용 모델의 관자놀이 스위치처럼 근거가 있는 식별부만 추가한다. 금속 내부는 외피 개방이나 정비 상태가 명시된 경우에만 드러내며, 모든 인간형 로봇에 노출 골격이나 발광 눈을 부여하지 않는다.\n- 능력의 시각적 중심은 가슴에 장착된 링의 회전과 발광으로 둔다. 기본적인 푸른 에너지 표현과 극고출력의 백색 광선을 구분하고, 출력에 따른 주변 장치의 반응을 표현한다. 마법진·영혼·종교적 후광으로 해석하지 않는다.\n- 물체의 부양이나 궤도 변화는 능력 작동이 명시된 경우에만 표현하고, 구동부의 회전·발광과 연결한다. 임의의 촉수, 투명한 손, 인간형 초능력 시전자 또는 상시 보호막을 추가하지 않는다.\n- 기능별 크기와 기계적 구조를 구분하고, 나비형 드론도 자연 생물이 아닌 소형 장치로 표현한다. 생물학적 나비와 새에는 드론의 센서나 무장 요소를 전이하지 않는다.\n- 전자적 교란은 기계의 자세 변화, 표시부 이상, 작동 정지처럼 관찰 가능한 반응으로 표현한다. 음향 자체를 반드시 가시적인 파동으로 그리지 않으며, 단순히 음악이 재생된다는 이유만으로 주변 로봇을 오작동시키지 않는다.\n\nTHINGS THAT LIVE AT THIS PLACE (what they are, not where they go): the\nlist above names what permanently belongs to this site. Where each one\nsits in this frame is decided by the attached sketch alone — do not add\nor reposition anything on the strength of the list, build only what the\nsketch already has lines for, and do not invent facilities that are not\nlisted.\n\nNo readable writing anywhere: surfaces that would carry writing may be\npresent, but stage any wording out of legibility — an oblique angle,\ndistance, shallow focus. No captions, watermarks or overlay text,\nand none of the sketch's diagram symbols, lines or markers anywhere in\nthe image.",
  "effective_prompt": "Turn the attached storyboard sketch into a photorealistic film still of\nits LOCATION ONLY, keeping the figures exactly as they are.\n\nKEEP EXACTLY: the camera framing, the horizon, and where every element\nsits in the frame. Each mannequin figure stays a plain grey featureless\nmannequin standing in the very same spot, at the same size, in the same\npose, turned the same way — do not turn them into people, do not move,\nrotate, mirror or re-pose them, do not add or remove figures.\n\nBUILD PHOTOREALISTICALLY: everything that is not a figure — ground,\nsurfacing, structures, vegetation, sky, water, distance. The bare lines\nof the sketch are a layout guide; replace them with the real materials,\ndepth and lighting of the place described below.\n\nSHOT TEXT this background must serve (Korean): 난민촌 거리에 거대하게 밀려든 검은 바닷물이, 다급히 등을 돌린 채 한 발을 공중에 띄운 질주 자세인 아시아계 난민들의 뒤를 집어삼키기 직전의 맹렬한 찰나.\n\nLOCATION (lock): Along an exposed refugee-settlement street where a towering surge of seawater overtakes fleeing residents at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: escape route ahead of the refugees in the upper-right of the frame, background; seawater advancing behind the refugees in the lower-left of the frame, midground, moves toward escape route ahead of the refugees.\n- KEY BACKGROUND ELEMENTS: Refugee settlement street (Being overtaken by incoming seawater) — The escape route recedes diagonally toward the upper right; used as Keeps the runners' route and the water's pursuit within one continuous space; Advancing seawater (Black water closing immediately behind the fleeing refugees) — Its advancing edge crosses from the lower-left background route toward the runners; used as Makes the remaining gap, rather than a separate spectacle, the primary threat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient illumination and controlled contrast distinguish the fleeing bodies from the black seawater without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHINGS AT THIS PLACE:\nThe site is a continuous street passage within the refugee settlement. Its travel surface and lateral margins form a single connected outdoor corridor, with access continuing beyond both ends of the mapped segment. The street itself provides the camera-placement area, including positions offset toward its margins with views diagonally along the corridor. This circulation surface is the minimum inferred infrastructure for access through the settlement. No building footprints, intersections, shoreline, or embankment position are established by the scene.\n- passage: Refugee settlement street passage\n\nWORLD FACTS (creator-confirmed — always true):\n- Region (real-world reference): 대한민국 중심. 인천과 경기권, 전북 익산, 전남 해남·목포, 제주도가 주요 배경이며, 미국의 과거 재현과 중국 등 해외의 원격 영상이 포함된다. 다국적·다인종 난민 사회가 존재한다.\n- Era: 2069년의 디스토피아적 미래. 수십 년 전의 기억과 과거 재현, 상대적으로 이후인 시점이 함께 포함된다.\n- 현재의 미래 기술 수준을 기본으로 하되, 과거 재현에서는 해당 시점의 연령·의상·기기 상태를 적용한다. 오래된 아날로그 제품이나 내연기관 차량을 시대착오로 제거하지 않으며, 현재의 첨단 장비를 과거 이미지에 자동으로 추가하지 않는다.\n- 스마트폰·노트북·태블릿·화상통화와 고도화된 로봇 기술을 함께 허용한다. 자원 부족 구역의 재사용 장비와 첨단 시설의 정밀 장비를 구분하고, 모든 공간을 동일한 미래형 인터페이스와 새 제품으로 채우지 않는다.\n- 소등·정전·가동 여부를 개별 공간과 시점에 맞춰 유지한다. 에너지 전달이 명시된 경우에만 주변 전기 장치의 밝기나 작동 상태를 변화시키며, 빈곤한 공간 전체를 항상 암흑으로 고정하지 않는다.\n- 육중한 고릴라형 실루엣, 금속 외장, 기계 관절, 푸른 눈과 가슴의 링 구조를 일관되게 유지한다. 감정은 눈의 크기·표시 이모티콘·자세로 표현하며, 어린 정서나 음성을 이유로 인간 아동의 얼굴과 신체로 바꾸지 않는다.\n- 인간형 모델은 명시된 인간 유사 외형을 유지하고, 군사용 모델의 관자놀이 스위치처럼 근거가 있는 식별부만 추가한다. 금속 내부는 외피 개방이나 정비 상태가 명시된 경우에만 드러내며, 모든 인간형 로봇에 노출 골격이나 발광 눈을 부여하지 않는다.\n- 능력의 시각적 중심은 가슴에 장착된 링의 회전과 발광으로 둔다. 기본적인 푸른 에너지 표현과 극고출력의 백색 광선을 구분하고, 출력에 따른 주변 장치의 반응을 표현한다. 마법진·영혼·종교적 후광으로 해석하지 않는다.\n- 물체의 부양이나 궤도 변화는 능력 작동이 명시된 경우에만 표현하고, 구동부의 회전·발광과 연결한다. 임의의 촉수, 투명한 손, 인간형 초능력 시전자 또는 상시 보호막을 추가하지 않는다.\n- 기능별 크기와 기계적 구조를 구분하고, 나비형 드론도 자연 생물이 아닌 소형 장치로 표현한다. 생물학적 나비와 새에는 드론의 센서나 무장 요소를 전이하지 않는다.\n- 전자적 교란은 기계의 자세 변화, 표시부 이상, 작동 정지처럼 관찰 가능한 반응으로 표현한다. 음향 자체를 반드시 가시적인 파동으로 그리지 않으며, 단순히 음악이 재생된다는 이유만으로 주변 로봇을 오작동시키지 않는다.\n\nTHINGS THAT LIVE AT THIS PLACE (what they are, not where they go): the\nlist above names what permanently belongs to this site. Where each one\nsits in this frame is decided by the attached sketch alone — do not add\nor reposition anything on the strength of the list, build only what the\nsketch already has lines for, and do not invent facilities that are not\nlisted.\n\nNo readable writing anywhere: surfaces that would carry writing may be\npresent, but stage any wording out of legibility — an oblique angle,\ndistance, shallow focus. No captions, watermarks or overlay text,\nand none of the sketch's diagram symbols, lines or markers anywhere in\nthe image.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S35sh2__bgfirst_bg.png",
  "asset_id": "90dcc9c3-2cf9-4c54-827f-adfb21c4ce6a",
  "input_asset_ids": [
   "ddda9700-58b6-4059-bfb4-3efd1b1c41b6"
  ]
 },
 "S35sh2": {
  "input_fingerprint": "0e926cf405e6756a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 난민촌 거리에 거대하게 밀려든 검은 바닷물이, 다급히 등을 돌린 채 한 발을 공중에 띄운 질주 자세인 아시아계 난민들의 뒤를 집어삼키기 직전의 맹렬한 찰나.\n\nLOCATION (lock): Along an exposed refugee-settlement street where a towering surge of seawater overtakes fleeing residents at night. The shot takes place here — the attached STORYBOARD SKETCH fixes the staging, camera and figure placement of this exact place. No location photograph is attached — build the location itself strictly from the location text above and the shot text, inventing nothing beyond them.\n\nFRAMING SCALE (follow exactly — this alone decides how much of the frame the subject fills; the storyboard sketch supplies where the figures stand and which way they face, never this scale):\n- FRAMING SCALE: wide shot\nMatch this shot size exactly. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient illumination and controlled contrast distinguish the fleeing bodies from the black seawater without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seawater is pouring into the settlement through the breached seawall, with a rapidly advancing surge filling the streets. The ground is shaking.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Replace every grey or outlined mannequin figure in the FIRST attached\nimage with the real people the CHARACTER REFERENCE images show, and\noutput a photorealistic film still. Each mannequin becomes the character\nwhose POSE CANON and the shot text match that figure's pose and\nposition; never repeat one person across two mannequins. If no CHARACTER\nREFERENCE image is attached for a figure, still replace it with a\nplausible real person consistent with the text. Every mannequin becomes\na person — erasing a figure, or leaving one out, is not a replacement,\nand the number of figures never changes.\n\nKEEP EXACTLY: the background of the first image, the camera framing,\nand each mannequin's position, size, pose and the direction it is\nturned. A person must be exactly where their mannequin was, at the same\nscale, in the same pose, facing the same way — never mirrored, never\nre-staged, never straightened up, stood upright or otherwise re-posed.\nNo mannequin, grey figure or sketch line may remain anywhere in the\noutput.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 난민촌 거리에 거대하게 밀려든 검은 바닷물이, 다급히 등을 돌린 채 한 발을 공중에 띄운 질주 자세인 아시아계 난민들의 뒤를 집어삼키기 직전의 맹렬한 찰나.\n\nLOCATION (lock): Along an exposed refugee-settlement street where a towering surge of seawater overtakes fleeing residents at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: escape route ahead of the refugees in the upper-right of the frame, background; seawater advancing behind the refugees in the lower-left of the frame, midground, moves toward escape route ahead of the refugees.\n- KEY BACKGROUND ELEMENTS: Refugee settlement street (Being overtaken by incoming seawater) — The escape route recedes diagonally toward the upper right; used as Keeps the runners' route and the water's pursuit within one continuous space; Advancing seawater (Black water closing immediately behind the fleeing refugees) — Its advancing edge crosses from the lower-left background route toward the runners; used as Makes the remaining gap, rather than a separate spectacle, the primary threat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient illumination and controlled contrast distinguish the fleeing bodies from the black seawater without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seawater is pouring into the settlement through the breached seawall, with a rapidly advancing surge filling the streets. The ground is shaking.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Replace every grey or outlined mannequin figure in the FIRST attached\nimage with the real people the CHARACTER REFERENCE images show, and\noutput a photorealistic film still. Each mannequin becomes the character\nwhose POSE CANON and the shot text match that figure's pose and\nposition; never repeat one person across two mannequins. If no CHARACTER\nREFERENCE image is attached for a figure, still replace it with a\nplausible real person consistent with the text. Every mannequin becomes\na person — erasing a figure, or leaving one out, is not a replacement,\nand the number of figures never changes.\n\nKEEP EXACTLY: the background of the first image, the camera framing,\nand each mannequin's position, size, pose and the direction it is\nturned. A person must be exactly where their mannequin was, at the same\nscale, in the same pose, facing the same way — never mirrored, never\nre-staged, never straightened up, stood upright or otherwise re-posed.\nNo mannequin, grey figure or sketch line may remain anywhere in the\noutput.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 난민촌 거리에 거대하게 밀려든 검은 바닷물이, 다급히 등을 돌린 채 한 발을 공중에 띄운 질주 자세인 아시아계 난민들의 뒤를 집어삼키기 직전의 맹렬한 찰나.\n\nLOCATION (lock): Along an exposed refugee-settlement street where a towering surge of seawater overtakes fleeing residents at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: escape route ahead of the refugees in the upper-right of the frame, background; seawater advancing behind the refugees in the lower-left of the frame, midground, moves toward escape route ahead of the refugees.\n- KEY BACKGROUND ELEMENTS: Refugee settlement street (Being overtaken by incoming seawater) — The escape route recedes diagonally toward the upper right; used as Keeps the runners' route and the water's pursuit within one continuous space; Advancing seawater (Black water closing immediately behind the fleeing refugees) — Its advancing edge crosses from the lower-left background route toward the runners; used as Makes the remaining gap, rather than a separate spectacle, the primary threat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime ambient illumination and controlled contrast distinguish the fleeing bodies from the black seawater without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seawater is pouring into the settlement through the breached seawall, with a rapidly advancing surge filling the streets. The ground is shaking.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S35sh2__bgfirst_bg.png",
     "asset_id": "90dcc9c3-2cf9-4c54-827f-adfb21c4ce6a",
     "role": "bgfirst_bg"
    }
   ],
   "B": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S35sh2__bgfirst_bg.png",
     "asset_id": "90dcc9c3-2cf9-4c54-827f-adfb21c4ce6a",
     "role": "bgfirst_bg"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "세 명의 인물이 파도를 등지고 앞을 향해 달리고 있으며, 가장 먼 인물은 파도 쪽을 뒤돌아보고 있음.",
    "built_space": "오른쪽에 판자촌과 가로등이 있고 왼쪽에 거대한 파도가 덮치는 구조가 레퍼런스와 정확히 일치함.",
    "entities": "동양인 남성들, 획일적인 회색 작업복, 검은 바닷물.",
    "hard_violations": [],
    "physics": "앞의 두 명은 한 발을 딛고 한 발을 든 질주 상태, 가장 먼 인물은 두 발이 모두 땅에서 떨어진 도약 상태로 달리는 동작을 수행 중임."
   },
   {
    "label": "B",
    "direction": "세 인물 모두 파도를 피해 앞으로 달리고 있으며, 맨 뒤의 인물이 우측 어깨 너머로 뒤를 돌아봄.",
    "built_space": "판자촌 거리, 가로등 조명, 덮쳐오는 파도의 위치와 크기가 레퍼런스 스케치와 완벽하게 일치함.",
    "entities": "동양인 남성들, 점퍼와 카고 바지 등 개별적인 의상, 검은 바닷물.",
    "hard_violations": [],
    "physics": "세 명 모두 지면에 한 발을 딛고 다른 한 발은 공중에 띄운 정확한 질주 포즈를 유지하며 체중을 지탱하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스의 구도와 포즈를 완벽히 일치시켰으며, 의상과 젖은 질감 표현이 더 사실적이어 피난민의 느낌을 잘 살렸습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "레퍼런스를 충실히 따랐으나 인물들의 의상이 다소 획일적이며, 질감 묘사가 B에 비해 미흡합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 명의 인물이 파도를 등지고 앞을 향해 달리고 있으며, 가장 먼 인물은 파도 쪽을 뒤돌아보고 있음.",
        "built_space": "오른쪽에 판자촌과 가로등이 있고 왼쪽에 거대한 파도가 덮치는 구조가 레퍼런스와 정확히 일치함.",
        "entities": "동양인 남성들, 획일적인 회색 작업복, 검은 바닷물.",
        "hard_violations": [],
        "physics": "앞의 두 명은 한 발을 딛고 한 발을 든 질주 상태, 가장 먼 인물은 두 발이 모두 땅에서 떨어진 도약 상태로 달리는 동작을 수행 중임."
       },
       {
        "label": "B",
        "direction": "세 인물 모두 파도를 피해 앞으로 달리고 있으며, 맨 뒤의 인물이 우측 어깨 너머로 뒤를 돌아봄.",
        "built_space": "판자촌 거리, 가로등 조명, 덮쳐오는 파도의 위치와 크기가 레퍼런스 스케치와 완벽하게 일치함.",
        "entities": "동양인 남성들, 점퍼와 카고 바지 등 개별적인 의상, 검은 바닷물.",
        "hard_violations": [],
        "physics": "세 명 모두 지면에 한 발을 딛고 다른 한 발은 공중에 띄운 정확한 질주 포즈를 유지하며 체중을 지탱하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스의 구도와 포즈를 완벽히 일치시켰으며, 의상과 젖은 질감 표현이 더 사실적이어 피난민의 느낌을 잘 살렸습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "레퍼런스를 충실히 따랐으나 인물들의 의상이 다소 획일적이며, 질감 묘사가 B에 비해 미흡합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "세 명의 인물이 파도를 등지고 앞을 향해 달리고 있으며, 가장 먼 인물은 파도 쪽을 뒤돌아보고 있음.",
        "built_space": "오른쪽에 판자촌과 가로등이 있고 왼쪽에 거대한 파도가 덮치는 구조가 레퍼런스와 정확히 일치함.",
        "entities": "동양인 남성들, 획일적인 회색 작업복, 검은 바닷물.",
        "hard_violations": [],
        "physics": "앞의 두 명은 한 발을 딛고 한 발을 든 질주 상태, 가장 먼 인물은 두 발이 모두 땅에서 떨어진 도약 상태로 달리는 동작을 수행 중임."
       },
       {
        "label": "B",
        "direction": "세 인물 모두 파도를 피해 앞으로 달리고 있으며, 맨 뒤의 인물이 우측 어깨 너머로 뒤를 돌아봄.",
        "built_space": "판자촌 거리, 가로등 조명, 덮쳐오는 파도의 위치와 크기가 레퍼런스 스케치와 완벽하게 일치함.",
        "entities": "동양인 남성들, 점퍼와 카고 바지 등 개별적인 의상, 검은 바닷물.",
        "hard_violations": [],
        "physics": "세 명 모두 지면에 한 발을 딛고 다른 한 발은 공중에 띄운 정확한 질주 포즈를 유지하며 체중을 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "와이드 숏과 세 사람의 후면 대각선 배치, 한 발을 든 질주와 발뒤꿈치까지 덮치는 검은 해수를 충실히 구현하며, 선두의 얼굴도 B보다 덜 뒤로 돌아가 도주 방향이 더 일관된다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "공간과 파도의 접근, 세 사람의 질주 자세는 잘 맞지만, 선두가 카메라 쪽으로 더 뚜렷하게 돌아보아 등을 돌리고 달아나는 순간의 방향성이 A보다 조금 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 사람 모두 몸통과 달리는 다리를 화면 오른쪽 위 골목 안쪽으로 향한다. 가까운 두 사람은 얼굴을 거의 보이지 않고 진행 방향을 보며, 양팔을 든 선두는 오른쪽 옆과 뒤를 돌아본다. 검은 물은 화면 왼쪽과 전경에서 이들의 뒤쪽으로 밀려들어 마지막 사람의 발 바로 뒤까지 도달한다. 무기나 겨누는 소품은 없다.",
        "built_space": "지면 가까운 후방 시점의 와이드 숏이다. 세 사람은 가까운 중앙, 그보다 먼 오른쪽 중앙, 가장 먼 오른쪽 위에 차례로 놓인다. 왼쪽에는 해수와 낮은 울타리, 중앙과 오른쪽에는 판재·비닐·천막 주거지가 있으며 골목은 오른쪽 위로 이어진다. 오른쪽 가장자리에는 큰 파란 통 하나와 그 앞 회색 상자 하나, 더 뒤에는 작은 청록색 통이 보인다. 참고 이미지의 시설 배치와 통행 공간을 유지하며, 젖은 바닥의 빛 반사는 가능한 방향이다.",
        "entities": "참고의 세 인체 모형을 실제 사람 세 명으로 바꾸었다. 모두 성인 남성으로 보이고 보이는 얼굴과 머리 형태는 아시아계 설정과 부합한다. 가까운 두 사람의 얼굴은 부분적으로 가려져 세부 정체성은 판단하기 어렵다. 어두운 낡은 재킷과 긴 바지, 신발을 착용한다. 검고 광택 있는 해수와 흰 포말, 야간의 젖은 난민촌 골목이 보인다. 추가 인물이나 읽을 수 있는 글자는 없다. 방파제 파손과 지면 진동 자체는 이 프레임에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "가까운 두 사람은 오른발이 노면에 닿고 왼쪽 무릎을 뒤로 굽혀 발을 들어 올린 달리기 자세다. 선두는 한쪽 다리를 앞으로 내리고 다른 발을 뒤로 들어 올린 보행 주기의 순간으로, 착지할 노면이 바로 아래 있다. 추진과 착지 관계가 없는 부유 인체는 보이지 않는다. 파도는 거리와 연결된 연속적인 수면 위에서 솟아 깨지며, 분사된 물방울과 포말도 쇄파로 설명된다."
       },
       {
        "label": "B",
        "direction": "세 사람의 몸과 다리는 모두 오른쪽 위 골목으로 달아가는 방향이다. 가까운 두 사람은 진행 방향을 향하지만, 양팔을 든 선두는 얼굴을 카메라와 뒤쪽 해수 쪽으로 더 크게 돌린다. 물은 왼쪽과 전경에서 마지막 사람의 뒤를 따라 밀려온다. 도주와 추격 관계는 분명하지만 선두의 시선은 전진 방향과 다르다. 겨누는 물건은 없다.",
        "built_space": "A와 같은 후방 와이드 숏으로, 세 사람이 중앙 전경에서 오른쪽 위 원경까지 대각선으로 배치된다. 왼쪽 울타리와 해수, 중앙의 천막 주거지, 오른쪽 판재 벽과 골목이 참고 배치를 따른다. 오른쪽에는 큰 파란 통 하나, 그 앞 회색 상자 하나, 뒤쪽 작은 청록색 통이 보인다. 인물들은 시설 사이의 실제 통행 공간에 있으며 시설 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "실제 인간으로 묘사된 성인 남성 세 명이며, 보이는 외모는 아시아계 설정에 부합한다. 선두는 가까운 두 사람보다 나이가 들어 보이지만 지정된 연령 조건은 없다. 세 사람 모두 어두운 재킷과 바지, 신발을 착용한다. 야간 난민촌의 천막·판재와 검은 해수, 포말이 물리적인 재질로 표현되어 있다. 추가 인물과 읽을 수 있는 글자는 없다. 파손된 방파제와 땅의 흔들림은 화면만으로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "가까운 두 사람은 오른발을 지면에 두고 왼발을 뒤로 들어 달리는 자세이며, 굽힌 무릎과 팔 동작이 추진을 뒷받침한다. 선두도 아래로 뻗은 다리와 뒤로 들린 발이 있는 정상적인 달리기 순간으로 읽히며 바로 아래에 착지 노면이 있다. 지지나 운동 경로가 없는 부유는 없다. 해수는 지면을 타고 연속적으로 밀려오고, 솟은 포말과 물방울에는 파도 충돌이라는 원인이 보인다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "와이드 숏과 세 사람의 후면 대각선 배치, 한 발을 든 질주와 발뒤꿈치까지 덮치는 검은 해수를 충실히 구현하며, 선두의 얼굴도 B보다 덜 뒤로 돌아가 도주 방향이 더 일관된다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "공간과 파도의 접근, 세 사람의 질주 자세는 잘 맞지만, 선두가 카메라 쪽으로 더 뚜렷하게 돌아보아 등을 돌리고 달아나는 순간의 방향성이 A보다 조금 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "세 사람 모두 몸통과 달리는 다리를 화면 오른쪽 위 골목 안쪽으로 향한다. 가까운 두 사람은 얼굴을 거의 보이지 않고 진행 방향을 보며, 양팔을 든 선두는 오른쪽 옆과 뒤를 돌아본다. 검은 물은 화면 왼쪽과 전경에서 이들의 뒤쪽으로 밀려들어 마지막 사람의 발 바로 뒤까지 도달한다. 무기나 겨누는 소품은 없다.",
        "built_space": "지면 가까운 후방 시점의 와이드 숏이다. 세 사람은 가까운 중앙, 그보다 먼 오른쪽 중앙, 가장 먼 오른쪽 위에 차례로 놓인다. 왼쪽에는 해수와 낮은 울타리, 중앙과 오른쪽에는 판재·비닐·천막 주거지가 있으며 골목은 오른쪽 위로 이어진다. 오른쪽 가장자리에는 큰 파란 통 하나와 그 앞 회색 상자 하나, 더 뒤에는 작은 청록색 통이 보인다. 참고 이미지의 시설 배치와 통행 공간을 유지하며, 젖은 바닥의 빛 반사는 가능한 방향이다.",
        "entities": "참고의 세 인체 모형을 실제 사람 세 명으로 바꾸었다. 모두 성인 남성으로 보이고 보이는 얼굴과 머리 형태는 아시아계 설정과 부합한다. 가까운 두 사람의 얼굴은 부분적으로 가려져 세부 정체성은 판단하기 어렵다. 어두운 낡은 재킷과 긴 바지, 신발을 착용한다. 검고 광택 있는 해수와 흰 포말, 야간의 젖은 난민촌 골목이 보인다. 추가 인물이나 읽을 수 있는 글자는 없다. 방파제 파손과 지면 진동 자체는 이 프레임에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "가까운 두 사람은 오른발이 노면에 닿고 왼쪽 무릎을 뒤로 굽혀 발을 들어 올린 달리기 자세다. 선두는 한쪽 다리를 앞으로 내리고 다른 발을 뒤로 들어 올린 보행 주기의 순간으로, 착지할 노면이 바로 아래 있다. 추진과 착지 관계가 없는 부유 인체는 보이지 않는다. 파도는 거리와 연결된 연속적인 수면 위에서 솟아 깨지며, 분사된 물방울과 포말도 쇄파로 설명된다."
       },
       {
        "label": "A",
        "direction": "세 사람의 몸과 다리는 모두 오른쪽 위 골목으로 달아가는 방향이다. 가까운 두 사람은 진행 방향을 향하지만, 양팔을 든 선두는 얼굴을 카메라와 뒤쪽 해수 쪽으로 더 크게 돌린다. 물은 왼쪽과 전경에서 마지막 사람의 뒤를 따라 밀려온다. 도주와 추격 관계는 분명하지만 선두의 시선은 전진 방향과 다르다. 겨누는 물건은 없다.",
        "built_space": "A와 같은 후방 와이드 숏으로, 세 사람이 중앙 전경에서 오른쪽 위 원경까지 대각선으로 배치된다. 왼쪽 울타리와 해수, 중앙의 천막 주거지, 오른쪽 판재 벽과 골목이 참고 배치를 따른다. 오른쪽에는 큰 파란 통 하나, 그 앞 회색 상자 하나, 뒤쪽 작은 청록색 통이 보인다. 인물들은 시설 사이의 실제 통행 공간에 있으며 시설 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "실제 인간으로 묘사된 성인 남성 세 명이며, 보이는 외모는 아시아계 설정에 부합한다. 선두는 가까운 두 사람보다 나이가 들어 보이지만 지정된 연령 조건은 없다. 세 사람 모두 어두운 재킷과 바지, 신발을 착용한다. 야간 난민촌의 천막·판재와 검은 해수, 포말이 물리적인 재질로 표현되어 있다. 추가 인물과 읽을 수 있는 글자는 없다. 파손된 방파제와 땅의 흔들림은 화면만으로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "가까운 두 사람은 오른발을 지면에 두고 왼발을 뒤로 들어 달리는 자세이며, 굽힌 무릎과 팔 동작이 추진을 뒷받침한다. 선두도 아래로 뻗은 다리와 뒤로 들린 발이 있는 정상적인 달리기 순간으로 읽히며 바로 아래에 착지 노면이 있다. 지지나 운동 경로가 없는 부유는 없다. 해수는 지면을 타고 연속적으로 밀려오고, 솟은 포말과 물방울에는 파도 충돌이라는 원인이 보인다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.746,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.746,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1746
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "레퍼런스의 구도와 포즈를 완벽히 일치시켰으며, 의상과 젖은 질감 표현이 더 사실적이어 피난민의 느낌을 잘 살렸습니다."
   },
   {
    "label": "A",
    "score": 1746,
    "verdict_ko": "레퍼런스를 충실히 따랐으나 인물들의 의상이 다소 획일적이며, 질감 묘사가 B에 비해 미흡합니다."
   }
  ],
  "refs": [
   {
    "label": "SHOT BACKGROUND",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S35sh2__bgfirst_bg.png",
    "asset_id": "90dcc9c3-2cf9-4c54-827f-adfb21c4ce6a",
    "role": "bgfirst_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdba-59b0-762d-b530-2554d9090b5a",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S35sh2__bgfirst_bg.png",
   "bg_asset_id": "90dcc9c3-2cf9-4c54-827f-adfb21c4ce6a",
   "bg_record_key": "S35sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "lane_conti_only"
  },
  "ref_mode": "lane(map_marker): 스케치+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S35sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:35:17.688082+00:00",
  "fingerprint": "a80884c6f90929d14ca03b7179443aa72c11a1be03df26d19bf6559bd016997a",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S35sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S35sh2_sel.png",
  "source_sha256": "fbb6e546a19b2e8c5a6c3c9e140c73894a45195d26f3267e854dc081b72ad620",
  "file": "S35sh2_cine.png",
  "staged_sha256": "d380c5cf5f981b00ae59637e01e0e4248062bcf767c59f89e629a889193596bf",
  "latency_ms": 14478
 },
 "S36sh5::signage": {
  "fp": "2e0a58246cfb655d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S36sh5::bgfirst_bg": {
  "input_fingerprint": "eaa9d19900ad1c3d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 하수도 밖 검은 바닷물 위로 튕겨 나온 앰버가 턱을 하늘로 치켜들고 고통스럽게 입을 벌린 위태로운 얼굴.\n\nLOCATION (lock): At the floodwater surface outside the sewer outlet in the inundated refugee settlement at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater outside the sewer (Surrounding 앰버 as she struggles after being swept outside) — Seen directly at surface level beneath and beside her face; used as Maintains the immediate drowning context around the close facial framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination keeps 앰버's facial strain readable against the black seawater without adding a colored source or perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 하수도 밖 검은 바닷물 위로 튕겨 나온 앰버가 턱을 하늘로 치켜들고 고통스럽게 입을 벌린 위태로운 얼굴.\n\nLOCATION (lock): At the floodwater surface outside the sewer outlet in the inundated refugee settlement at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater outside the sewer (Surrounding 앰버 as she struggles after being swept outside) — Seen directly at surface level beneath and beside her face; used as Maintains the immediate drowning context around the close facial framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination keeps 앰버's facial strain readable against the black seawater without adding a colored source or perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S36sh5__bgfirst_bg.png",
  "asset_id": "8cf3b218-2f31-4804-9884-e14e7d360142",
  "input_asset_ids": [
   "1fd85d4d-97d5-4061-b50d-f17cbf5e8f19",
   "e9363d03-f2b0-4f3b-b05c-bf162e69594e"
  ]
 },
 "S36sh5": {
  "input_fingerprint": "afdf14048dc620ec",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 하수도 밖 검은 바닷물 위로 튕겨 나온 앰버가 턱을 하늘로 치켜들고 고통스럽게 입을 벌린 위태로운 얼굴.\n\nLOCATION (lock): At the floodwater surface outside the sewer outlet in the inundated refugee settlement at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater outside the sewer (Surrounding 앰버 as she struggles after being swept outside) — Seen directly at surface level beneath and beside her face; used as Maintains the immediate drowning context around the close facial framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination keeps 앰버's facial strain readable against the black seawater without adding a colored source or perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer has filled with surging seawater, which is forcing its contents outside. Charlie is submerged in the flood; his worn exterior, peeling UBIC chest logo, and previously acquired coat and hat have not undergone any stated repair or replacement. 앰버: She has been swept out of the sewer and is struggling in the seawater, with soaked clothing. Her severe coughing and breathing difficulty precede the inundation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 하수도 밖 검은 바닷물 위로 튕겨 나온 앰버가 턱을 하늘로 치켜들고 고통스럽게 입을 벌린 위태로운 얼굴.\n\nLOCATION (lock): At the floodwater surface outside the sewer outlet in the inundated refugee settlement at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater outside the sewer (Surrounding 앰버 as she struggles after being swept outside) — Seen directly at surface level beneath and beside her face; used as Maintains the immediate drowning context around the close facial framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination keeps 앰버's facial strain readable against the black seawater without adding a colored source or perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer has filled with surging seawater, which is forcing its contents outside. Charlie is submerged in the flood; his worn exterior, peeling UBIC chest logo, and previously acquired coat and hat have not undergone any stated repair or replacement. 앰버: She has been swept out of the sewer and is struggling in the seawater, with soaked clothing. Her severe coughing and breathing difficulty precede the inundation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 하수도 밖 검은 바닷물 위로 튕겨 나온 앰버가 턱을 하늘로 치켜들고 고통스럽게 입을 벌린 위태로운 얼굴.\n\nLOCATION (lock): At the floodwater surface outside the sewer outlet in the inundated refugee settlement at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater outside the sewer (Surrounding 앰버 as she struggles after being swept outside) — Seen directly at surface level beneath and beside her face; used as Maintains the immediate drowning context around the close facial framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination keeps 앰버's facial strain readable against the black seawater without adding a colored source or perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer has filled with surging seawater, which is forcing its contents outside. Charlie is submerged in the flood; his worn exterior, peeling UBIC chest logo, and previously acquired coat and hat have not undergone any stated repair or replacement. 앰버: She has been swept out of the sewer and is struggling in the seawater, with soaked clothing. Her severe coughing and breathing difficulty precede the inundation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S36sh5__bgfirst_bg.png",
     "asset_id": "8cf3b218-2f31-4804-9884-e14e7d360142",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S36sh5.png",
     "asset_id": "1fd85d4d-97d5-4061-b50d-f17cbf5e8f19",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L32B01.png",
     "asset_id": "e9363d03-f2b0-4f3b-b05c-bf162e69594e",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "앰버의 턱과 시선이 하늘을 향해 위로 들려 있다.",
    "built_space": "야간의 어두운 바다 수면이며, 배경에 콘크리트 하수구 및 방파제 구조물과 먼 불빛들이 보인다. 카메라는 수면 높이에 위치해 있다.",
    "entities": "레퍼런스와 일치하는 앰버가 물에 젖은 채 등장하지만, 프레임 우측에 지문에서 전혀 언급되지 않은 성인의 신체(얼굴 하관, 팔, 손)가 등장한다.",
    "hard_violations": [
     "[gemini-pro] 지문에 명시되지 않은 추가 인물 등장 (우측에서 앰버를 잡고 있는 사람)",
     "[gpt-high] 앰버만 등장해야 하는 장면에 성인 인물과 그가 앰버를 붙잡는 구조 행동을 추가했다."
    ],
    "physics": "앰버의 몸은 바닷물 속에 잠겨 부력을 받고 있으며, 우측의 미상 인물이 손으로 앰버의 팔을 꽉 잡고 지탱해주고 있다."
   },
   {
    "label": "B",
    "direction": "앰버의 시선이 화면 우측 상단 하늘을 향해 있다.",
    "built_space": "야간 배경의 콘크리트 구조물이 뒤로 보이나, 앰버는 바닷물 표면이 아닌 금속 바닥재와 파이프 난간이 있는 보트 또는 선착장 위에 위치해 있다.",
    "entities": "앰버가 등장하지만, 지문에 없는 성인 남성이 앰버를 안고 있으며 지문에 없는 파란색 담요로 몸을 감싸고 있다.",
    "hard_violations": [
     "[gemini-pro] 지문에 명시되지 않은 추가 인물 등장 (앰버를 안고 있는 성인 남성)",
     "[gemini-pro] 지문에 없는 사물 등장 (파란색 담요)",
     "[gemini-pro] 잘못된 스테이징 (바닷물 수면이 아닌 구조물 위에 앉아 있음)",
     "[gpt-high] 허용되지 않은 성인 남성의 몸통과 손, 얼굴 일부를 추가했다.",
     "[gpt-high] 지정되지 않은 담요·보온재 및 금속 바닥·난간을 추가해 구조 후 상황을 만들었다.",
     "[gpt-high] 앰버를 하수도 밖 바닷물 속이 아니라 바닥 위에서 성인에게 안긴 위치에 배치했다."
    ],
    "physics": "앰버는 물에 빠져 있지 않고 단단한 금속 구조물 위에 앉아 있으며, 뒤에 있는 남성의 몸과 팔에 기대어 지탱되고 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "지문에 명시되지 않은 미상의 인물이 프레임 우측에 등장하여 앰버를 붙잡고 있으며, 입을 고통스럽게 벌린 표정도 제대로 구현되지 않아 치명적인 규정 위반입니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "지문에 없는 성인 인물이 등장하고 파란색 담요가 추가되었으며, 바닷물 표면이라는 스테이징을 무시하고 금속 구조물 위에 앉아 있어 치명적인 규정 위반입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 턱과 시선이 하늘을 향해 위로 들려 있다.",
        "built_space": "야간의 어두운 바다 수면이며, 배경에 콘크리트 하수구 및 방파제 구조물과 먼 불빛들이 보인다. 카메라는 수면 높이에 위치해 있다.",
        "entities": "레퍼런스와 일치하는 앰버가 물에 젖은 채 등장하지만, 프레임 우측에 지문에서 전혀 언급되지 않은 성인의 신체(얼굴 하관, 팔, 손)가 등장한다.",
        "hard_violations": [
         "지문에 명시되지 않은 추가 인물 등장 (우측에서 앰버를 잡고 있는 사람)"
        ],
        "physics": "앰버의 몸은 바닷물 속에 잠겨 부력을 받고 있으며, 우측의 미상 인물이 손으로 앰버의 팔을 꽉 잡고 지탱해주고 있다."
       },
       {
        "label": "B",
        "direction": "앰버의 시선이 화면 우측 상단 하늘을 향해 있다.",
        "built_space": "야간 배경의 콘크리트 구조물이 뒤로 보이나, 앰버는 바닷물 표면이 아닌 금속 바닥재와 파이프 난간이 있는 보트 또는 선착장 위에 위치해 있다.",
        "entities": "앰버가 등장하지만, 지문에 없는 성인 남성이 앰버를 안고 있으며 지문에 없는 파란색 담요로 몸을 감싸고 있다.",
        "hard_violations": [
         "지문에 명시되지 않은 추가 인물 등장 (앰버를 안고 있는 성인 남성)",
         "지문에 없는 사물 등장 (파란색 담요)",
         "잘못된 스테이징 (바닷물 수면이 아닌 구조물 위에 앉아 있음)"
        ],
        "physics": "앰버는 물에 빠져 있지 않고 단단한 금속 구조물 위에 앉아 있으며, 뒤에 있는 남성의 몸과 팔에 기대어 지탱되고 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "지문에 명시되지 않은 미상의 인물이 프레임 우측에 등장하여 앰버를 붙잡고 있으며, 입을 고통스럽게 벌린 표정도 제대로 구현되지 않아 치명적인 규정 위반입니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "지문에 없는 성인 인물이 등장하고 파란색 담요가 추가되었으며, 바닷물 표면이라는 스테이징을 무시하고 금속 구조물 위에 앉아 있어 치명적인 규정 위반입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 턱과 시선이 하늘을 향해 위로 들려 있다.",
        "built_space": "야간의 어두운 바다 수면이며, 배경에 콘크리트 하수구 및 방파제 구조물과 먼 불빛들이 보인다. 카메라는 수면 높이에 위치해 있다.",
        "entities": "레퍼런스와 일치하는 앰버가 물에 젖은 채 등장하지만, 프레임 우측에 지문에서 전혀 언급되지 않은 성인의 신체(얼굴 하관, 팔, 손)가 등장한다.",
        "hard_violations": [
         "지문에 명시되지 않은 추가 인물 등장 (우측에서 앰버를 잡고 있는 사람)"
        ],
        "physics": "앰버의 몸은 바닷물 속에 잠겨 부력을 받고 있으며, 우측의 미상 인물이 손으로 앰버의 팔을 꽉 잡고 지탱해주고 있다."
       },
       {
        "label": "B",
        "direction": "앰버의 시선이 화면 우측 상단 하늘을 향해 있다.",
        "built_space": "야간 배경의 콘크리트 구조물이 뒤로 보이나, 앰버는 바닷물 표면이 아닌 금속 바닥재와 파이프 난간이 있는 보트 또는 선착장 위에 위치해 있다.",
        "entities": "앰버가 등장하지만, 지문에 없는 성인 남성이 앰버를 안고 있으며 지문에 없는 파란색 담요로 몸을 감싸고 있다.",
        "hard_violations": [
         "지문에 명시되지 않은 추가 인물 등장 (앰버를 안고 있는 성인 남성)",
         "지문에 없는 사물 등장 (파란색 담요)",
         "잘못된 스테이징 (바닷물 수면이 아닌 구조물 위에 앉아 있음)"
        ],
        "physics": "앰버는 물에 빠져 있지 않고 단단한 금속 구조물 위에 앉아 있으며, 뒤에 있는 남성의 몸과 팔에 기대어 지탱되고 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "얼굴 중심 구도와 밤은 맞지만, 앰버를 물속이 아닌 난간 있는 바닥 위에 두고 성인과 담요를 추가하여 익수 순간을 구조 후 장면으로 바꿨다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "수면 높이의 얼굴 클로즈업과 치켜든 턱은 더 가깝지만, 추가 인물이 앰버를 붙잡고 있으며 고통스럽게 입을 벌리는 연기도 부족해 두 후보 모두 부적합하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 턱을 조금 들고 화면 왼쪽 위의 프레임 밖을 바라본다. 입은 살짝 열렸지만 턱을 하늘로 크게 젖히며 고통스럽게 숨을 구하는 모습은 아니다. 뒤 성인은 눈이 잘려 시선 방향을 확인할 수 없다.",
        "built_space": "뒤에 직사각형 하수도 개구부 하나와 오염된 콘크리트 벽이 보인다. 전경에는 금속 바닥과 굽은 난간 하나가 있으며 앰버와 성인은 그 위쪽에 자리한다. 바닷물은 얼굴 바로 아래를 감싸지 않고 바닥 너머 배경에 있다. 장소에 새로운 전경 구조물을 도입하면서 지정된 수면 속 위치를 바꿨다.",
        "entities": "젖은 금발의 약 10세 여자아이 한 명은 연령과 머리색에서 앰버에 가깝지만, 참조의 갈색 계열 눈과 달리 푸른 눈으로 보이고 얼굴 일치도도 낮다. 한국계·백인 혼혈 여부는 외관만으로 확정할 수 없다. 회색 안옷과 녹색 겉옷 일부는 참조와 유사하다. 수염 있는 성인 남성의 턱·몸통·손이 추가되었고, 파란 담요와 은색 보온재도 등장한다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "허용되지 않은 성인 남성의 몸통과 손, 얼굴 일부를 추가했다.",
         "지정되지 않은 담요·보온재 및 금속 바닥·난간을 추가해 구조 후 상황을 만들었다.",
         "앰버를 하수도 밖 바닷물 속이 아니라 바닥 위에서 성인에게 안긴 위치에 배치했다."
        ],
        "physics": "성인의 팔과 손이 앰버의 몸을 감싸 지지하고, 하단에는 두 사람이 위치할 수 있는 금속 바닥이 보인다. 하체 접촉은 프레임 밖이지만 공중에 떠 있는 신체는 아니다. 담요는 어깨에 걸쳐져 있다. 물에 밀려 나와 몸부림치는 동작 대신 부축받는 정적인 자세다."
       },
       {
        "label": "B",
        "direction": "앰버는 목을 뒤로 젖혀 턱을 위로 향하지만 눈은 화면 오른쪽 위 성인의 얼굴 쪽을 향한다. 성인의 보이는 얼굴 부분은 앰버 쪽으로 향해 있으나 눈은 프레임 밖이다. 앰버의 입은 조금만 벌어져 있어 고통스러운 호흡 곤란보다 상대를 올려다보는 모습에 가깝다.",
        "built_space": "수면 높이에서 앰버의 얼굴과 어깨를 가까이 담았고, 검은 바닷물이 턱 아래 몸통과 양옆을 감싼다. 오른쪽 뒤에는 콘크리트 해안벽과 직사각형 하수도 개구부 하나가 보이며 그 아래로 물이 쏟아진다. 먼 방파제와 건물도 작게 배치되어 참조 장소의 구성에 가깝다. 앰버와 추가 성인은 함께 물속에 있다.",
        "entities": "젖은 금발과 큰 눈을 가진 약 10세 여자아이 한 명으로 연령·성별·머리색은 앰버 설정에 부합한다. 눈 색과 얼굴 형태는 A보다 참조에 가까우나 혼혈 정체성 자체는 외관만으로 확정할 수 없다. 젖은 녹색 겉옷과 어두운 안옷이 보이고 참조에 없는 붉은 어깨끈도 있다. 오른쪽에 성인의 얼굴 일부·몸통·팔·손이 추가되어 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "앰버만 등장해야 하는 장면에 성인 인물과 그가 앰버를 붙잡는 구조 행동을 추가했다."
        ],
        "physics": "앰버는 어깨와 가슴까지 잠겨 물의 부력을 받고 있으며 성인의 손이 어깨와 위팔을 붙잡아 지지한다. 성인 역시 몸통이 물에 잠겨 있다. 물방울은 뒤 하수도에서 쏟아지는 물과 파도 충돌에 연결되어 보인다. 무지지 공중 부양이나 명백한 불가능 자세는 없지만, 스스로 익수하며 몸부림치는 순간이 아니라 타인에게 지지받는 순간이다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "얼굴 중심 구도와 밤은 맞지만, 앰버를 물속이 아닌 난간 있는 바닥 위에 두고 성인과 담요를 추가하여 익수 순간을 구조 후 장면으로 바꿨다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "수면 높이의 얼굴 클로즈업과 치켜든 턱은 더 가깝지만, 추가 인물이 앰버를 붙잡고 있으며 고통스럽게 입을 벌리는 연기도 부족해 두 후보 모두 부적합하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 턱을 조금 들고 화면 왼쪽 위의 프레임 밖을 바라본다. 입은 살짝 열렸지만 턱을 하늘로 크게 젖히며 고통스럽게 숨을 구하는 모습은 아니다. 뒤 성인은 눈이 잘려 시선 방향을 확인할 수 없다.",
        "built_space": "뒤에 직사각형 하수도 개구부 하나와 오염된 콘크리트 벽이 보인다. 전경에는 금속 바닥과 굽은 난간 하나가 있으며 앰버와 성인은 그 위쪽에 자리한다. 바닷물은 얼굴 바로 아래를 감싸지 않고 바닥 너머 배경에 있다. 장소에 새로운 전경 구조물을 도입하면서 지정된 수면 속 위치를 바꿨다.",
        "entities": "젖은 금발의 약 10세 여자아이 한 명은 연령과 머리색에서 앰버에 가깝지만, 참조의 갈색 계열 눈과 달리 푸른 눈으로 보이고 얼굴 일치도도 낮다. 한국계·백인 혼혈 여부는 외관만으로 확정할 수 없다. 회색 안옷과 녹색 겉옷 일부는 참조와 유사하다. 수염 있는 성인 남성의 턱·몸통·손이 추가되었고, 파란 담요와 은색 보온재도 등장한다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "허용되지 않은 성인 남성의 몸통과 손, 얼굴 일부를 추가했다.",
         "지정되지 않은 담요·보온재 및 금속 바닥·난간을 추가해 구조 후 상황을 만들었다.",
         "앰버를 하수도 밖 바닷물 속이 아니라 바닥 위에서 성인에게 안긴 위치에 배치했다."
        ],
        "physics": "성인의 팔과 손이 앰버의 몸을 감싸 지지하고, 하단에는 두 사람이 위치할 수 있는 금속 바닥이 보인다. 하체 접촉은 프레임 밖이지만 공중에 떠 있는 신체는 아니다. 담요는 어깨에 걸쳐져 있다. 물에 밀려 나와 몸부림치는 동작 대신 부축받는 정적인 자세다."
       },
       {
        "label": "A",
        "direction": "앰버는 목을 뒤로 젖혀 턱을 위로 향하지만 눈은 화면 오른쪽 위 성인의 얼굴 쪽을 향한다. 성인의 보이는 얼굴 부분은 앰버 쪽으로 향해 있으나 눈은 프레임 밖이다. 앰버의 입은 조금만 벌어져 있어 고통스러운 호흡 곤란보다 상대를 올려다보는 모습에 가깝다.",
        "built_space": "수면 높이에서 앰버의 얼굴과 어깨를 가까이 담았고, 검은 바닷물이 턱 아래 몸통과 양옆을 감싼다. 오른쪽 뒤에는 콘크리트 해안벽과 직사각형 하수도 개구부 하나가 보이며 그 아래로 물이 쏟아진다. 먼 방파제와 건물도 작게 배치되어 참조 장소의 구성에 가깝다. 앰버와 추가 성인은 함께 물속에 있다.",
        "entities": "젖은 금발과 큰 눈을 가진 약 10세 여자아이 한 명으로 연령·성별·머리색은 앰버 설정에 부합한다. 눈 색과 얼굴 형태는 A보다 참조에 가까우나 혼혈 정체성 자체는 외관만으로 확정할 수 없다. 젖은 녹색 겉옷과 어두운 안옷이 보이고 참조에 없는 붉은 어깨끈도 있다. 오른쪽에 성인의 얼굴 일부·몸통·팔·손이 추가되어 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "앰버만 등장해야 하는 장면에 성인 인물과 그가 앰버를 붙잡는 구조 행동을 추가했다."
        ],
        "physics": "앰버는 어깨와 가슴까지 잠겨 물의 부력을 받고 있으며 성인의 손이 어깨와 위팔을 붙잡아 지지한다. 성인 역시 몸통이 물에 잠겨 있다. 물방울은 뒤 하수도에서 쏟아지는 물과 파도 충돌에 연결되어 보인다. 무지지 공중 부양이나 명백한 불가능 자세는 없지만, 스스로 익수하며 몸부림치는 순간이 아니라 타인에게 지지받는 순간이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 0.333
   },
   "adjusted": {
    "A": 0.75,
    "B": 0.083
   },
   "violations": {
    "A": [
     "[gemini-pro] 지문에 명시되지 않은 추가 인물 등장 (우측에서 앰버를 잡고 있는 사람)",
     "[gpt-high] 앰버만 등장해야 하는 장면에 성인 인물과 그가 앰버를 붙잡는 구조 행동을 추가했다."
    ],
    "B": [
     "[gemini-pro] 지문에 명시되지 않은 추가 인물 등장 (앰버를 안고 있는 성인 남성)",
     "[gemini-pro] 지문에 없는 사물 등장 (파란색 담요)",
     "[gemini-pro] 잘못된 스테이징 (바닷물 수면이 아닌 구조물 위에 앉아 있음)",
     "[gpt-high] 허용되지 않은 성인 남성의 몸통과 손, 얼굴 일부를 추가했다.",
     "[gpt-high] 지정되지 않은 담요·보온재 및 금속 바닥·난간을 추가해 구조 후 상황을 만들었다.",
     "[gpt-high] 앰버를 하수도 밖 바닷물 속이 아니라 바닥 위에서 성인에게 안긴 위치에 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 750,
   "B": 83
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "지문에 명시되지 않은 미상의 인물이 프레임 우측에 등장하여 앰버를 붙잡고 있으며, 입을 고통스럽게 벌린 표정도 제대로 구현되지 않아 치명적인 규정 위반입니다.  ★위반: [gemini-pro] 지문에 명시되지 않은 추가 인물 등장 (우측에서 앰버를 잡고 있는 사람) / [gpt-high] 앰버만 등장해야 하는 장면에 성인 인물과 그가 앰버를 붙잡는 구조 행동을 추가했다."
   },
   {
    "label": "B",
    "score": 83,
    "verdict_ko": "지문에 없는 성인 인물이 등장하고 파란색 담요가 추가되었으며, 바닷물 표면이라는 스테이징을 무시하고 금속 구조물 위에 앉아 있어 치명적인 규정 위반입니다.  ★위반: [gemini-pro] 지문에 명시되지 않은 추가 인물 등장 (앰버를 안고 있는 성인 남성) / [gemini-pro] 지문에 없는 사물 등장 (파란색 담요) / [gemini-pro] 잘못된 스테이징 (바닷물 수면이 아닌 구조물 위에 앉아 있음) / [gpt-high] 허용되지 않은 성인 남성의 몸통과 손, 얼굴 일부를 추가했다. / [gpt-high] 지정되지 않은 담요·보온재 및 금속 바닥·난간을 추가해 구조 후 상황을 만들었다. / [gpt-high] 앰버를 하수도 밖 바닷물 속이 아니라 바닥 위에서 성인에게 안긴 위치에 배치했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L32B01.png",
    "asset_id": "e9363d03-f2b0-4f3b-b05c-bf162e69594e",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdbb-47e7-7f1b-8951-f023a5916444",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S36sh5__bgfirst_bg.png",
   "bg_asset_id": "8cf3b218-2f31-4804-9884-e14e7d360142",
   "bg_record_key": "S36sh5::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S36sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:35:32.581893+00:00",
  "fingerprint": "fd786ac6e1457c2a27ee8b59792483ee7f9dcbbf425ec934bcde05d7ea1fac8f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S36sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S36sh5_sel.png",
  "source_sha256": "c9537bb16f088d5630ec46a543cc2842b33a6eca21dd2046b38bcf045aa48875",
  "file": "S36sh5_cine.png",
  "staged_sha256": "56ead00f1c7d88ef55d7c4e78d359fc78c767067716732b15cc0a5ef1dc8d2c2",
  "latency_ms": 12084
 },
 "S36sh6::signage": {
  "fp": "c8dafda4ca95c30e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S36sh6": {
  "input_fingerprint": "f0ae0d1f771a5139",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거친 바닷물 속에서 찰리의 굵은 금속 손이 앰버의 얇은 손을 빈틈없이 꽉 틀어쥔 손끝 클로즈업.\n\nLOCATION (lock): In the turbulent open floodwater just beyond the refugee settlement's sewer outlet. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater around the joined hands (Moving around their arms while 찰리 swims) — Seen directly from above rather than as a reflective image; used as Surrounds the contact point and preserves the physical context of the rescue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding shot's restrained nighttime illumination, using tonal separation to distinguish metal, skin, and moving seawater without introducing a new light cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seawater continues to surge outside the flooded sewer. Charlie is swimming with one metal hand extended in a firm grip; his unrepaired old body and peeling UBIC chest logo remain unchanged. 앰버: She remains immersed in the rough seawater, her clothing soaked and one hand extended. She is still struggling to stay afloat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거친 바닷물 속에서 찰리의 굵은 금속 손이 앰버의 얇은 손을 빈틈없이 꽉 틀어쥔 손끝 클로즈업.\n\nLOCATION (lock): In the turbulent open floodwater just beyond the refugee settlement's sewer outlet. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater around the joined hands (Moving around their arms while 찰리 swims) — Seen directly from above rather than as a reflective image; used as Surrounds the contact point and preserves the physical context of the rescue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding shot's restrained nighttime illumination, using tonal separation to distinguish metal, skin, and moving seawater without introducing a new light cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seawater continues to surge outside the flooded sewer. Charlie is swimming with one metal hand extended in a firm grip; his unrepaired old body and peeling UBIC chest logo remain unchanged. 앰버: She remains immersed in the rough seawater, her clothing soaked and one hand extended. She is still struggling to stay afloat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거친 바닷물 속에서 찰리의 굵은 금속 손이 앰버의 얇은 손을 빈틈없이 꽉 틀어쥔 손끝 클로즈업.\n\nLOCATION (lock): In the turbulent open floodwater just beyond the refugee settlement's sewer outlet. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater around the joined hands (Moving around their arms while 찰리 swims) — Seen directly from above rather than as a reflective image; used as Surrounds the contact point and preserves the physical context of the rescue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding shot's restrained nighttime illumination, using tonal separation to distinguish metal, skin, and moving seawater without introducing a new light cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seawater continues to surge outside the flooded sewer. Charlie is swimming with one metal hand extended in a firm grip; his unrepaired old body and peeling UBIC chest logo remain unchanged. 앰버: She remains immersed in the rough seawater, her clothing soaked and one hand extended. She is still struggling to stay afloat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 수면을 위에서 아래로 직접 내려다보고 있으며, 두 사람이 꽉 맞잡은 손을 정확히 향하고 있음.",
    "built_space": "지시된 클로즈업 구도에 맞게 거친 바닷물만 보이며, 그 외의 배경 구조물은 프레임에 없음.",
    "entities": "찰리의 금속 손과 녹색 소매를 입은 앰버의 손이 등장함. 하지만 금속 손은 왼쪽에 5개의 손가락, 오른쪽에 1개의 엄지가 있어 총 6개의 손가락을 가진 기형적인 모습임.",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy (금속 손의 손가락이 6개이며 사람 손과 기형적으로 융합됨)"
    ],
    "physics": "두 손이 서로 꽉 맞잡고 물 위에 떠 있으며 팔에 의해 지탱됨. 포즈 자체는 '빈틈없이 꽉 틀어쥔' 행동과 일치하지만 해부학적 형태가 불가능함."
   },
   {
    "label": "B",
    "direction": "카메라는 수면 높이에서 측면을 향해 두 사람의 팔과 배경의 하수구 배출구, 등대를 바라보고 있음.",
    "built_space": "오른쪽에 하수구 배출구 벽이 있고 멀리 등대가 보이며, 레퍼런스 이미지의 위치 및 구조와 일치함. 인물들은 구조물 앞 물속에 알맞게 위치함.",
    "entities": "레퍼런스와 질감이 일치하는 찰리의 낡은 금속 팔과 앰버의 녹색 재킷, 구명조끼 끈, 금발 머리가 보임.",
    "hard_violations": [],
    "physics": "앰버의 손이 찰리의 꽉 쥔 금속 주먹 위에 가볍게 얹혀 있으며 팔과 물에 의해 지탱됨. 이 포즈는 프롬프트가 지시한 '빈틈없이 꽉 틀어쥔' 행동으로 만들어질 수 없으므로 물리 축을 통과하지 못함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "치명적 위반은 없으나, 프롬프트의 핵심인 '손끝 클로즈업'과 '위에서 내려다보는 구도'를 완전히 무시했으며, 꽉 맞잡은 행동 대신 주먹에 손을 얹은 포즈를 묘사하여 지시를 실패했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지시된 구도와 맞잡은 행동을 따랐으나, 금속 손의 손가락이 6개로 묘사되고 사람 손과 융합되는 치명적 위반(해부학적 오류)이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 수면을 위에서 아래로 직접 내려다보고 있으며, 두 사람이 꽉 맞잡은 손을 정확히 향하고 있음.",
        "built_space": "지시된 클로즈업 구도에 맞게 거친 바닷물만 보이며, 그 외의 배경 구조물은 프레임에 없음.",
        "entities": "찰리의 금속 손과 녹색 소매를 입은 앰버의 손이 등장함. 하지만 금속 손은 왼쪽에 5개의 손가락, 오른쪽에 1개의 엄지가 있어 총 6개의 손가락을 가진 기형적인 모습임.",
        "hard_violations": [
         "physically impossible anatomy (금속 손의 손가락이 6개이며 사람 손과 기형적으로 융합됨)"
        ],
        "physics": "두 손이 서로 꽉 맞잡고 물 위에 떠 있으며 팔에 의해 지탱됨. 포즈 자체는 '빈틈없이 꽉 틀어쥔' 행동과 일치하지만 해부학적 형태가 불가능함."
       },
       {
        "label": "B",
        "direction": "카메라는 수면 높이에서 측면을 향해 두 사람의 팔과 배경의 하수구 배출구, 등대를 바라보고 있음.",
        "built_space": "오른쪽에 하수구 배출구 벽이 있고 멀리 등대가 보이며, 레퍼런스 이미지의 위치 및 구조와 일치함. 인물들은 구조물 앞 물속에 알맞게 위치함.",
        "entities": "레퍼런스와 질감이 일치하는 찰리의 낡은 금속 팔과 앰버의 녹색 재킷, 구명조끼 끈, 금발 머리가 보임.",
        "hard_violations": [],
        "physics": "앰버의 손이 찰리의 꽉 쥔 금속 주먹 위에 가볍게 얹혀 있으며 팔과 물에 의해 지탱됨. 이 포즈는 프롬프트가 지시한 '빈틈없이 꽉 틀어쥔' 행동으로 만들어질 수 없으므로 물리 축을 통과하지 못함."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "치명적 위반은 없으나, 프롬프트의 핵심인 '손끝 클로즈업'과 '위에서 내려다보는 구도'를 완전히 무시했으며, 꽉 맞잡은 행동 대신 주먹에 손을 얹은 포즈를 묘사하여 지시를 실패했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지시된 구도와 맞잡은 행동을 따랐으나, 금속 손의 손가락이 6개로 묘사되고 사람 손과 융합되는 치명적 위반(해부학적 오류)이 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 수면을 위에서 아래로 직접 내려다보고 있으며, 두 사람이 꽉 맞잡은 손을 정확히 향하고 있음.",
        "built_space": "지시된 클로즈업 구도에 맞게 거친 바닷물만 보이며, 그 외의 배경 구조물은 프레임에 없음.",
        "entities": "찰리의 금속 손과 녹색 소매를 입은 앰버의 손이 등장함. 하지만 금속 손은 왼쪽에 5개의 손가락, 오른쪽에 1개의 엄지가 있어 총 6개의 손가락을 가진 기형적인 모습임.",
        "hard_violations": [
         "physically impossible anatomy (금속 손의 손가락이 6개이며 사람 손과 기형적으로 융합됨)"
        ],
        "physics": "두 손이 서로 꽉 맞잡고 물 위에 떠 있으며 팔에 의해 지탱됨. 포즈 자체는 '빈틈없이 꽉 틀어쥔' 행동과 일치하지만 해부학적 형태가 불가능함."
       },
       {
        "label": "B",
        "direction": "카메라는 수면 높이에서 측면을 향해 두 사람의 팔과 배경의 하수구 배출구, 등대를 바라보고 있음.",
        "built_space": "오른쪽에 하수구 배출구 벽이 있고 멀리 등대가 보이며, 레퍼런스 이미지의 위치 및 구조와 일치함. 인물들은 구조물 앞 물속에 알맞게 위치함.",
        "entities": "레퍼런스와 질감이 일치하는 찰리의 낡은 금속 팔과 앰버의 녹색 재킷, 구명조끼 끈, 금발 머리가 보임.",
        "hard_violations": [],
        "physics": "앰버의 손이 찰리의 꽉 쥔 금속 주먹 위에 가볍게 얹혀 있으며 팔과 물에 의해 지탱됨. 이 포즈는 프롬프트가 지시한 '빈틈없이 꽉 틀어쥔' 행동으로 만들어질 수 없으므로 물리 축을 통과하지 못함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "맞잡은 손과 야간 배수구의 연속성은 좋지만, 수면 높이에서 상체와 수평선까지 보여 주어 위에서 내려다보는 손끝 클로즈업 구도에서 벗어난다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "거친 바닷물로 둘러싸인 두 손을 위에서 밀착 촬영해 지정 구도를 가장 충실히 구현하지만, 벌어진 금속 엄지는 빈틈없이 꽉 쥔 인상을 조금 약화한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 금속 팔은 왼쪽 아래에서 중앙으로, 앰버의 팔은 오른쪽에서 중앙으로 뻗어 서로의 손을 맞잡는다. 금속 손가락은 앰버의 손을 감싸고 앰버의 손가락도 금속 손 위에 걸려 있다. 얼굴과 시선은 보이지 않는다. 카메라는 손 위에서 내려다보기보다 수면 높이에서 접촉점을 바라본다.",
        "built_space": "오른쪽 뒤에 콘크리트 방벽 하나와 직사각형 배수구 하나가 보이며, 배수구에서 바다로 물이 쏟아진다. 왼쪽 배경에는 수평선과 먼 불빛이 있다. 이전 장면의 장소 요소와 부합하지만, 배경 구조물과 앰버의 몸통까지 포함하여 요청한 손 중심 구도보다 넓다. 수면은 직접 보이며 인물의 반사상으로 대체되지 않았다.",
        "entities": "낡고 긁힌 굵은 금속 손 하나와 작고 밝은 피부의 손 하나가 보인다. 금속 팔의 마모와 재질은 찰리 참고 이미지에 부합한다. 앰버의 작은 손, 젖은 회녹색 소매, 화면 오른쪽의 젖은 금발과 붉은 어깨끈은 이전 장면과 대체로 일치한다. 얼굴이 잘려 정확한 나이와 혼혈 정체성은 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 손은 각각 손목과 팔에 연결되어 있으며, 금속 손가락과 사람 손가락의 상호 접촉이 보인다. 앰버의 팔과 몸통 일부, 찰리의 팔 일부는 물에 잠겨 있다. 팔을 들어 맞잡는 자세는 가능하며, 지지 없이 떠 있는 분리된 신체나 물체는 없다. 수영 추진 동작 자체는 프레임 밖이다."
       },
       {
        "label": "B",
        "direction": "찰리의 팔은 화면 위에서 아래쪽 중앙으로, 앰버의 팔은 아래에서 위쪽 중앙으로 뻗어 만난다. 앰버의 손가락은 금속 손 양옆을 감싸고 금속 손가락은 작은 손의 아래쪽으로 굽혀져 있다. 다만 금속 엄지는 오른쪽으로 벌어져 있어 완전히 조여 잡은 모습은 덜 분명하다. 시선은 프레임 밖이며, 카메라는 두 손과 주변 바닷물을 위에서 직접 내려다본다.",
        "built_space": "프레임 전체가 출렁이는 바닷물이며 중앙의 맞잡은 손과 두 팔 일부만 포함한다. 방벽과 배수구 등 고정 구조물은 이 클로즈업 범위 밖이다. 구조물의 중복이나 불가능한 반사는 없고, 물이 접촉점 사방을 둘러싸는 지정 구도를 따른다.",
        "entities": "굵고 마모된 금속 손 하나, 그보다 작은 맨손 하나, 젖은 회녹색 소매가 보인다. 찰리의 금속에는 긁힘과 변색, 관절 분할이 있어 낡은 기계 신체 참고와 부합한다. 앰버의 손과 가는 손목은 어린아이의 체격에 맞고 소매도 이전 장면과 일치한다. 얼굴과 머리카락이 없어 혼혈 정체성이나 정확한 연령은 직접 확인할 수 없다. 추가 인물, 소품, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "금속 손은 위쪽 팔에, 작은 손은 아래쪽 소매에서 나온 손목에 자연스럽게 연결된다. 서로 감싼 손가락이 접촉을 유지하며 두 팔 주위로 물과 포말이 흐른다. 손은 팔에 의해 지지되고 수면 가까이에서 맞잡혀 있어 무지지 부유로 보이지 않는다. 벌어진 엄지는 악력 표현의 약점이지 불가능한 자세는 아니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "맞잡은 손과 야간 배수구의 연속성은 좋지만, 수면 높이에서 상체와 수평선까지 보여 주어 위에서 내려다보는 손끝 클로즈업 구도에서 벗어난다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "거친 바닷물로 둘러싸인 두 손을 위에서 밀착 촬영해 지정 구도를 가장 충실히 구현하지만, 벌어진 금속 엄지는 빈틈없이 꽉 쥔 인상을 조금 약화한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 금속 팔은 왼쪽 아래에서 중앙으로, 앰버의 팔은 오른쪽에서 중앙으로 뻗어 서로의 손을 맞잡는다. 금속 손가락은 앰버의 손을 감싸고 앰버의 손가락도 금속 손 위에 걸려 있다. 얼굴과 시선은 보이지 않는다. 카메라는 손 위에서 내려다보기보다 수면 높이에서 접촉점을 바라본다.",
        "built_space": "오른쪽 뒤에 콘크리트 방벽 하나와 직사각형 배수구 하나가 보이며, 배수구에서 바다로 물이 쏟아진다. 왼쪽 배경에는 수평선과 먼 불빛이 있다. 이전 장면의 장소 요소와 부합하지만, 배경 구조물과 앰버의 몸통까지 포함하여 요청한 손 중심 구도보다 넓다. 수면은 직접 보이며 인물의 반사상으로 대체되지 않았다.",
        "entities": "낡고 긁힌 굵은 금속 손 하나와 작고 밝은 피부의 손 하나가 보인다. 금속 팔의 마모와 재질은 찰리 참고 이미지에 부합한다. 앰버의 작은 손, 젖은 회녹색 소매, 화면 오른쪽의 젖은 금발과 붉은 어깨끈은 이전 장면과 대체로 일치한다. 얼굴이 잘려 정확한 나이와 혼혈 정체성은 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 손은 각각 손목과 팔에 연결되어 있으며, 금속 손가락과 사람 손가락의 상호 접촉이 보인다. 앰버의 팔과 몸통 일부, 찰리의 팔 일부는 물에 잠겨 있다. 팔을 들어 맞잡는 자세는 가능하며, 지지 없이 떠 있는 분리된 신체나 물체는 없다. 수영 추진 동작 자체는 프레임 밖이다."
       },
       {
        "label": "A",
        "direction": "찰리의 팔은 화면 위에서 아래쪽 중앙으로, 앰버의 팔은 아래에서 위쪽 중앙으로 뻗어 만난다. 앰버의 손가락은 금속 손 양옆을 감싸고 금속 손가락은 작은 손의 아래쪽으로 굽혀져 있다. 다만 금속 엄지는 오른쪽으로 벌어져 있어 완전히 조여 잡은 모습은 덜 분명하다. 시선은 프레임 밖이며, 카메라는 두 손과 주변 바닷물을 위에서 직접 내려다본다.",
        "built_space": "프레임 전체가 출렁이는 바닷물이며 중앙의 맞잡은 손과 두 팔 일부만 포함한다. 방벽과 배수구 등 고정 구조물은 이 클로즈업 범위 밖이다. 구조물의 중복이나 불가능한 반사는 없고, 물이 접촉점 사방을 둘러싸는 지정 구도를 따른다.",
        "entities": "굵고 마모된 금속 손 하나, 그보다 작은 맨손 하나, 젖은 회녹색 소매가 보인다. 찰리의 금속에는 긁힘과 변색, 관절 분할이 있어 낡은 기계 신체 참고와 부합한다. 앰버의 손과 가는 손목은 어린아이의 체격에 맞고 소매도 이전 장면과 일치한다. 얼굴과 머리카락이 없어 혼혈 정체성이나 정확한 연령은 직접 확인할 수 없다. 추가 인물, 소품, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "금속 손은 위쪽 팔에, 작은 손은 아래쪽 소매에서 나온 손목에 자연스럽게 연결된다. 서로 감싼 손가락이 접촉을 유지하며 두 팔 주위로 물과 포말이 흐른다. 손은 팔에 의해 지지되고 수면 가까이에서 맞잡혀 있어 무지지 부유로 보이지 않는다. 벌어진 엄지는 악력 표현의 약점이지 불가능한 자세는 아니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.667
   },
   "violations": {
    "A": [
     "[gemini-pro] physically impossible anatomy (금속 손의 손가락이 6개이며 사람 손과 기형적으로 융합됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1667,
   "A": 1417
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "치명적 위반은 없으나, 프롬프트의 핵심인 '손끝 클로즈업'과 '위에서 내려다보는 구도'를 완전히 무시했으며, 꽉 맞잡은 행동 대신 주먹에 손을 얹은 포즈를 묘사하여 지시를 실패했습니다."
   },
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "지시된 구도와 맞잡은 행동을 따랐으나, 금속 손의 손가락이 6개로 묘사되고 사람 손과 융합되는 치명적 위반(해부학적 오류)이 발생했습니다.  ★위반: [gemini-pro] physically impossible anatomy (금속 손의 손가락이 6개이며 사람 손과 기형적으로 융합됨)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S36sh5_sel.png",
    "asset_id": "ee66d578-c1fd-4ceb-8067-381f0837d836",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1597480>",
    "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdbc-0feb-7ead-8f85-5b2b3a648c97",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S36sh5"
  }
 },
 "S36sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:35:45.007673+00:00",
  "fingerprint": "189700210e85911b11995e81888024072e1005736e9cce23a9423d3636163577",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S36sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S36sh6_sel.png",
  "source_sha256": "f5682d8807ac1c0cc3103210501820ba427445cfa5dc3fdede24cef793ee4422",
  "file": "S36sh6_cine.png",
  "staged_sha256": "ff6436ae7732121a41347b5bc4985ee88f7b14da2df65a4844a5f242141d2cb5",
  "latency_ms": 11920
 },
 "S37sh2::signage": {
  "fp": "613c8329452788b2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S37sh2": {
  "input_fingerprint": "4a00b185ee4ac913",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 유리 창문 너머로 칠흑같이 검은 거대한 해일이 관리사무소 건물을 향해 무서운 기세로 솟구쳐 오른 위압적인 전경.\n\nLOCATION (lock): Beside a corridor window inside the lit administration building, looking out into the night at the approaching seawater. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Corridor window (Still intact before the wave strikes) — Viewed obliquely through its glass, with the approaching wave visible beyond and an edge retained at frame left; used as Places a fragile physical boundary between the interior camera and the exterior threat; Approaching tidal wave (Rising and advancing toward the management office) — Visible beyond the window along a slightly upward viewing axis; used as Provides the exterior scale and impending impact without inserting people into the view; Corridor interior (Not yet overtaken by the wave in this moment) — A limited interior edge remains beneath the window view; used as Anchors the camera inside the building.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime contrast separates the intact window edge from the black wave beyond, with no dreamlike treatment or invented exterior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A huge incoming wave is approaching the management building beyond the still-intact corridor windows. The seawall breach continues to flood the settlement.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 유리 창문 너머로 칠흑같이 검은 거대한 해일이 관리사무소 건물을 향해 무서운 기세로 솟구쳐 오른 위압적인 전경.\n\nLOCATION (lock): Beside a corridor window inside the lit administration building, looking out into the night at the approaching seawater. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Corridor window (Still intact before the wave strikes) — Viewed obliquely through its glass, with the approaching wave visible beyond and an edge retained at frame left; used as Places a fragile physical boundary between the interior camera and the exterior threat; Approaching tidal wave (Rising and advancing toward the management office) — Visible beyond the window along a slightly upward viewing axis; used as Provides the exterior scale and impending impact without inserting people into the view; Corridor interior (Not yet overtaken by the wave in this moment) — A limited interior edge remains beneath the window view; used as Anchors the camera inside the building.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime contrast separates the intact window edge from the black wave beyond, with no dreamlike treatment or invented exterior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A huge incoming wave is approaching the management building beyond the still-intact corridor windows. The seawall breach continues to flood the settlement.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 유리 창문 너머로 칠흑같이 검은 거대한 해일이 관리사무소 건물을 향해 무서운 기세로 솟구쳐 오른 위압적인 전경.\n\nLOCATION (lock): Beside a corridor window inside the lit administration building, looking out into the night at the approaching seawater. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Corridor window (Still intact before the wave strikes) — Viewed obliquely through its glass, with the approaching wave visible beyond and an edge retained at frame left; used as Places a fragile physical boundary between the interior camera and the exterior threat; Approaching tidal wave (Rising and advancing toward the management office) — Visible beyond the window along a slightly upward viewing axis; used as Provides the exterior scale and impending impact without inserting people into the view; Corridor interior (Not yet overtaken by the wave in this moment) — A limited interior edge remains beneath the window view; used as Anchors the camera inside the building.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime contrast separates the intact window edge from the black wave beyond, with no dreamlike treatment or invented exterior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A huge incoming wave is approaching the management building beyond the still-intact corridor windows. The seawall breach continues to flood the settlement.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 복도를 향하며, 우측의 개방된 공간을 통해 검은 해일이 다가오는 것을 보여줍니다.",
    "built_space": "좌측에는 프레임이 있는 창문이 위치하나, 우측은 유리가 없는 거대한 구멍 혹은 발코니 형태로 원본 건축 구조물과 다르게 변형되었습니다.",
    "entities": "검고 거대한 해일, 복도의 창틀 및 벽면이 존재하며 인물은 없습니다.",
    "hard_violations": [
     "[gemini-pro] 건축 구조 왜곡 및 발명 (지시된 복도 창문 구조를 무시하고 우측에 유리가 없는 거대한 개방형 구멍을 만듦)",
     "[gemini-pro] 물리적으로 불가능한 반사 (좌측 창문에 반사된 조명과 문의 위치가 카메라 시점상 맞지 않음)",
     "[gpt-high] 참고 장소의 왼쪽 철문 벽에 추가 미닫이창을 만들고 맞은편에도 별도의 큰 개구부를 두어, 고정 창호의 수와 공간 구조를 변경했습니다."
    ],
    "physics": "좌측 창문에 맺힌 실내조명과 문의 반사상이 물리적 공간의 위치와 호환되지 않습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 복도 끝을 정면으로 향하며, 정면에 나타난 해일을 바라봅니다.",
    "built_space": "좌측에 창문이 있으나, 복도 정면의 끝부분 벽과 천장이 해일 이미지로 덮여 공간 구조가 완전히 단절되었습니다.",
    "entities": "평면적인 사각형으로 잘린 해일 이미지, 좌측의 창문과 벽면이 있습니다.",
    "hard_violations": [
     "[gemini-pro] 그래픽 오버레이 및 스티커 사용 (해일 이미지가 천장과 배관을 뚫고 완벽한 직사각형의 평면 스티커처럼 덧붙여짐)",
     "[gemini-pro] 물리적으로 불가능한 공간 구조 (해일이 건물의 원근법과 구조를 무시하고 삽입됨)",
     "[gpt-high] 참고 장소에서 철문들이 놓인 왼쪽 벽에 별도의 두 짝 창을 추가하여, 고정 창호의 수와 공간 구조를 변경했습니다."
    ],
    "physics": "정면의 해일 이미지가 어떠한 건축적 테두리나 물리적 지지 없이 공간 위에 평면으로 떠 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "창문 너머로 해일을 비스듬히 본다는 지시와 달리, 우측에 유리가 없는 거대한 개방형 벽면을 임의로 만들어 공간 구조를 심각하게 훼손했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "해일이 복도 정면 끝에 평면 스티커나 그래픽 오버레이처럼 직사각형으로 합성되어 물리적 공간감과 지시 사항을 완전히 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 복도를 향하며, 우측의 개방된 공간을 통해 검은 해일이 다가오는 것을 보여줍니다.",
        "built_space": "좌측에는 프레임이 있는 창문이 위치하나, 우측은 유리가 없는 거대한 구멍 혹은 발코니 형태로 원본 건축 구조물과 다르게 변형되었습니다.",
        "entities": "검고 거대한 해일, 복도의 창틀 및 벽면이 존재하며 인물은 없습니다.",
        "hard_violations": [
         "건축 구조 왜곡 및 발명 (지시된 복도 창문 구조를 무시하고 우측에 유리가 없는 거대한 개방형 구멍을 만듦)",
         "물리적으로 불가능한 반사 (좌측 창문에 반사된 조명과 문의 위치가 카메라 시점상 맞지 않음)"
        ],
        "physics": "좌측 창문에 맺힌 실내조명과 문의 반사상이 물리적 공간의 위치와 호환되지 않습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 복도 끝을 정면으로 향하며, 정면에 나타난 해일을 바라봅니다.",
        "built_space": "좌측에 창문이 있으나, 복도 정면의 끝부분 벽과 천장이 해일 이미지로 덮여 공간 구조가 완전히 단절되었습니다.",
        "entities": "평면적인 사각형으로 잘린 해일 이미지, 좌측의 창문과 벽면이 있습니다.",
        "hard_violations": [
         "그래픽 오버레이 및 스티커 사용 (해일 이미지가 천장과 배관을 뚫고 완벽한 직사각형의 평면 스티커처럼 덧붙여짐)",
         "물리적으로 불가능한 공간 구조 (해일이 건물의 원근법과 구조를 무시하고 삽입됨)"
        ],
        "physics": "정면의 해일 이미지가 어떠한 건축적 테두리나 물리적 지지 없이 공간 위에 평면으로 떠 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "창문 너머로 해일을 비스듬히 본다는 지시와 달리, 우측에 유리가 없는 거대한 개방형 벽면을 임의로 만들어 공간 구조를 심각하게 훼손했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "해일이 복도 정면 끝에 평면 스티커나 그래픽 오버레이처럼 직사각형으로 합성되어 물리적 공간감과 지시 사항을 완전히 위반했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 복도를 향하며, 우측의 개방된 공간을 통해 검은 해일이 다가오는 것을 보여줍니다.",
        "built_space": "좌측에는 프레임이 있는 창문이 위치하나, 우측은 유리가 없는 거대한 구멍 혹은 발코니 형태로 원본 건축 구조물과 다르게 변형되었습니다.",
        "entities": "검고 거대한 해일, 복도의 창틀 및 벽면이 존재하며 인물은 없습니다.",
        "hard_violations": [
         "건축 구조 왜곡 및 발명 (지시된 복도 창문 구조를 무시하고 우측에 유리가 없는 거대한 개방형 구멍을 만듦)",
         "물리적으로 불가능한 반사 (좌측 창문에 반사된 조명과 문의 위치가 카메라 시점상 맞지 않음)"
        ],
        "physics": "좌측 창문에 맺힌 실내조명과 문의 반사상이 물리적 공간의 위치와 호환되지 않습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 복도 끝을 정면으로 향하며, 정면에 나타난 해일을 바라봅니다.",
        "built_space": "좌측에 창문이 있으나, 복도 정면의 끝부분 벽과 천장이 해일 이미지로 덮여 공간 구조가 완전히 단절되었습니다.",
        "entities": "평면적인 사각형으로 잘린 해일 이미지, 좌측의 창문과 벽면이 있습니다.",
        "hard_violations": [
         "그래픽 오버레이 및 스티커 사용 (해일 이미지가 천장과 배관을 뚫고 완벽한 직사각형의 평면 스티커처럼 덧붙여짐)",
         "물리적으로 불가능한 공간 구조 (해일이 건물의 원근법과 구조를 무시하고 삽입됨)"
        ],
        "physics": "정면의 해일 이미지가 어떠한 건축적 테두리나 물리적 지지 없이 공간 위에 평면으로 떠 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "밤의 검은 해일은 보이지만 창을 추가해 장소 구조를 바꾸었고, 실내 비중이 커서 유리 너머 솟구치는 해일 중심의 구도에 덜 부합합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "추가 창과 불명확한 전면 유리 때문에 장소·경계 조건에는 실패하지만, 비스듬히 올려다보는 거대한 검은 해일의 규모와 임박한 위협은 A보다 가깝습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물이나 지향성 소품은 없습니다. 정면 창 너머 파도 전면이 실내 카메라 쪽을 향하고 마루가 위로 솟아 있지만, 시선 축은 거의 수평이고 파도의 접근 방향은 B보다 덜 강하게 드러납니다.",
        "built_space": "왼쪽에 두 짝 미닫이창 한 세트, 맞은편에 중앙 세로대 없는 큰 창 하나, 왼쪽 끝에 녹색 요철 면이 있습니다. 참고 장소의 왼쪽 철문들이 있는 벽에 창이 추가되었으며, 끝벽의 두 짝 창도 다른 구조가 되었습니다. 낡은 밝은 벽과 짙은 하부 도장, 노출 배관은 유사합니다. 왼쪽 유리에는 형광등과 창 손잡이의 반사가 보이며 이것만으로 광학적 불가능을 단정할 수는 없습니다. 바닥과 창 아래 벽이 넓게 들어와 제한된 실내 가장자리라는 지시에서 벗어납니다.",
        "entities": "인물, 읽을 수 있는 글자, 표식은 없습니다. 밤하늘 아래 검은 바닷물과 흰 포말은 해일로 읽힙니다. 왼쪽 미닫이창 유리는 온전하지만, 주된 해일이 보이는 맞은편 창에서는 유리면의 존재가 뚜렷하지 않습니다. 관리사무소 복도의 마감은 닮았으나 고정 창호 배치는 일치하지 않습니다.",
        "hard_violations": [
         "참고 장소에서 철문들이 놓인 왼쪽 벽에 별도의 두 짝 창을 추가하여, 고정 창호의 수와 공간 구조를 변경했습니다."
        ],
        "physics": "파도는 화면 아래 수면과 이어진 연속적인 물 덩어리이며, 마루의 포말도 물결에서 발생한 것으로 보입니다. 근거 없이 떠 있는 물체나 사람은 없습니다. 창틀은 벽에 고정되어 있고 실내 바닥에는 해일이 들어오지 않았습니다."
       },
       {
        "label": "B",
        "direction": "인물이나 지향성 소품은 없습니다. 거대한 파도 전면이 창과 실내 카메라 쪽을 향하며, 마루가 오른쪽 위까지 치솟습니다. 비스듬한 상향 시선과 건물에 닥쳐오는 물벽의 방향이 A보다 분명합니다.",
        "built_space": "왼쪽에 두 짝 미닫이창 한 세트, 맞은편에 큰 사각 개구부 하나, 오른쪽 가장자리에 철문 일부가 있습니다. 참고의 끝벽 두 짝 창과 왼쪽 철문 배치가 그대로 유지되지 않았습니다. 콘크리트성 벽, 짙은 하부 도장과 배관은 참고와 유사합니다. 왼쪽 유리에는 형광등과 실내 문이 반사되며, 해당 반사만으로 불가능한 광학 배치라고 판단할 근거는 부족합니다. 주된 해일 앞 개구부에는 유리나 중앙 창대가 뚜렷하지 않아 온전한 유리 경계가 불명확합니다.",
        "entities": "인물과 읽을 수 있는 글자는 없습니다. 밤의 검은 해일이 화면 오른쪽을 크게 채우고, 소량의 포말이 높은 마루를 드러냅니다. 왼쪽에는 온전한 유리창이 있으나 해일의 주된 시야를 가로막는 유리는 확실히 식별되지 않습니다. 복도 재질은 유사하지만 참고 장소의 창호 구성은 다릅니다.",
        "hard_violations": [
         "참고 장소의 왼쪽 철문 벽에 추가 미닫이창을 만들고 맞은편에도 별도의 큰 개구부를 두어, 고정 창호의 수와 공간 구조를 변경했습니다."
        ],
        "physics": "파도는 아래쪽으로 이어지는 거대한 수괴이며, 높은 마루와 그 아래 경사진 수면이 연결되어 있습니다. 지지나 운동 원인 없이 떠 있는 별도 물체는 없습니다. 창틀과 문은 벽에 고정되어 있고, 보이는 실내에는 아직 물이 밀려들지 않았습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "밤의 검은 해일은 보이지만 창을 추가해 장소 구조를 바꾸었고, 실내 비중이 커서 유리 너머 솟구치는 해일 중심의 구도에 덜 부합합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "추가 창과 불명확한 전면 유리 때문에 장소·경계 조건에는 실패하지만, 비스듬히 올려다보는 거대한 검은 해일의 규모와 임박한 위협은 A보다 가깝습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "인물이나 지향성 소품은 없습니다. 정면 창 너머 파도 전면이 실내 카메라 쪽을 향하고 마루가 위로 솟아 있지만, 시선 축은 거의 수평이고 파도의 접근 방향은 B보다 덜 강하게 드러납니다.",
        "built_space": "왼쪽에 두 짝 미닫이창 한 세트, 맞은편에 중앙 세로대 없는 큰 창 하나, 왼쪽 끝에 녹색 요철 면이 있습니다. 참고 장소의 왼쪽 철문들이 있는 벽에 창이 추가되었으며, 끝벽의 두 짝 창도 다른 구조가 되었습니다. 낡은 밝은 벽과 짙은 하부 도장, 노출 배관은 유사합니다. 왼쪽 유리에는 형광등과 창 손잡이의 반사가 보이며 이것만으로 광학적 불가능을 단정할 수는 없습니다. 바닥과 창 아래 벽이 넓게 들어와 제한된 실내 가장자리라는 지시에서 벗어납니다.",
        "entities": "인물, 읽을 수 있는 글자, 표식은 없습니다. 밤하늘 아래 검은 바닷물과 흰 포말은 해일로 읽힙니다. 왼쪽 미닫이창 유리는 온전하지만, 주된 해일이 보이는 맞은편 창에서는 유리면의 존재가 뚜렷하지 않습니다. 관리사무소 복도의 마감은 닮았으나 고정 창호 배치는 일치하지 않습니다.",
        "hard_violations": [
         "참고 장소에서 철문들이 놓인 왼쪽 벽에 별도의 두 짝 창을 추가하여, 고정 창호의 수와 공간 구조를 변경했습니다."
        ],
        "physics": "파도는 화면 아래 수면과 이어진 연속적인 물 덩어리이며, 마루의 포말도 물결에서 발생한 것으로 보입니다. 근거 없이 떠 있는 물체나 사람은 없습니다. 창틀은 벽에 고정되어 있고 실내 바닥에는 해일이 들어오지 않았습니다."
       },
       {
        "label": "A",
        "direction": "인물이나 지향성 소품은 없습니다. 거대한 파도 전면이 창과 실내 카메라 쪽을 향하며, 마루가 오른쪽 위까지 치솟습니다. 비스듬한 상향 시선과 건물에 닥쳐오는 물벽의 방향이 A보다 분명합니다.",
        "built_space": "왼쪽에 두 짝 미닫이창 한 세트, 맞은편에 큰 사각 개구부 하나, 오른쪽 가장자리에 철문 일부가 있습니다. 참고의 끝벽 두 짝 창과 왼쪽 철문 배치가 그대로 유지되지 않았습니다. 콘크리트성 벽, 짙은 하부 도장과 배관은 참고와 유사합니다. 왼쪽 유리에는 형광등과 실내 문이 반사되며, 해당 반사만으로 불가능한 광학 배치라고 판단할 근거는 부족합니다. 주된 해일 앞 개구부에는 유리나 중앙 창대가 뚜렷하지 않아 온전한 유리 경계가 불명확합니다.",
        "entities": "인물과 읽을 수 있는 글자는 없습니다. 밤의 검은 해일이 화면 오른쪽을 크게 채우고, 소량의 포말이 높은 마루를 드러냅니다. 왼쪽에는 온전한 유리창이 있으나 해일의 주된 시야를 가로막는 유리는 확실히 식별되지 않습니다. 복도 재질은 유사하지만 참고 장소의 창호 구성은 다릅니다.",
        "hard_violations": [
         "참고 장소의 왼쪽 철문 벽에 추가 미닫이창을 만들고 맞은편에도 별도의 큰 개구부를 두어, 고정 창호의 수와 공간 구조를 변경했습니다."
        ],
        "physics": "파도는 아래쪽으로 이어지는 거대한 수괴이며, 높은 마루와 그 아래 경사진 수면이 연결되어 있습니다. 지지나 운동 원인 없이 떠 있는 별도 물체는 없습니다. 창틀과 문은 벽에 고정되어 있고, 보이는 실내에는 아직 물이 밀려들지 않았습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.333
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.083
   },
   "violations": {
    "A": [
     "[gemini-pro] 건축 구조 왜곡 및 발명 (지시된 복도 창문 구조를 무시하고 우측에 유리가 없는 거대한 개방형 구멍을 만듦)",
     "[gemini-pro] 물리적으로 불가능한 반사 (좌측 창문에 반사된 조명과 문의 위치가 카메라 시점상 맞지 않음)",
     "[gpt-high] 참고 장소의 왼쪽 철문 벽에 추가 미닫이창을 만들고 맞은편에도 별도의 큰 개구부를 두어, 고정 창호의 수와 공간 구조를 변경했습니다."
    ],
    "B": [
     "[gemini-pro] 그래픽 오버레이 및 스티커 사용 (해일 이미지가 천장과 배관을 뚫고 완벽한 직사각형의 평면 스티커처럼 덧붙여짐)",
     "[gemini-pro] 물리적으로 불가능한 공간 구조 (해일이 건물의 원근법과 구조를 무시하고 삽입됨)",
     "[gpt-high] 참고 장소에서 철문들이 놓인 왼쪽 벽에 별도의 두 짝 창을 추가하여, 고정 창호의 수와 공간 구조를 변경했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1083
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "창문 너머로 해일을 비스듬히 본다는 지시와 달리, 우측에 유리가 없는 거대한 개방형 벽면을 임의로 만들어 공간 구조를 심각하게 훼손했습니다.  ★위반: [gemini-pro] 건축 구조 왜곡 및 발명 (지시된 복도 창문 구조를 무시하고 우측에 유리가 없는 거대한 개방형 구멍을 만듦) / [gemini-pro] 물리적으로 불가능한 반사 (좌측 창문에 반사된 조명과 문의 위치가 카메라 시점상 맞지 않음) / [gpt-high] 참고 장소의 왼쪽 철문 벽에 추가 미닫이창을 만들고 맞은편에도 별도의 큰 개구부를 두어, 고정 창호의 수와 공간 구조를 변경했습니다."
   },
   {
    "label": "B",
    "score": 1083,
    "verdict_ko": "해일이 복도 정면 끝에 평면 스티커나 그래픽 오버레이처럼 직사각형으로 합성되어 물리적 공간감과 지시 사항을 완전히 위반했습니다.  ★위반: [gemini-pro] 그래픽 오버레이 및 스티커 사용 (해일 이미지가 천장과 배관을 뚫고 완벽한 직사각형의 평면 스티커처럼 덧붙여짐) / [gemini-pro] 물리적으로 불가능한 공간 구조 (해일이 건물의 원근법과 구조를 무시하고 삽입됨) / [gpt-high] 참고 장소에서 철문들이 놓인 왼쪽 벽에 별도의 두 짝 창을 추가하여, 고정 창호의 수와 공간 구조를 변경했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L44B02.png",
    "asset_id": "6a3add4a-a856-45ed-8b07-5c69e4800eed",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdbc-d419-747c-b2e2-e9f9520fb0a3",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S37sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:35:57.267815+00:00",
  "fingerprint": "59af8ce4d43565551888ed06ef7239eb361f47b371c30f413971d697113bb516",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S37sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S37sh2_sel.png",
  "source_sha256": "766e38d3ef1d2e2a76f77f6d69e621febe94750c14e658bef6ff46be0a06d94a",
  "file": "S37sh2_cine.png",
  "staged_sha256": "c831f65bd48fe67fb9c9bbf2d8ae063493669d946a0c5315b5a6ace1a1e58d93",
  "latency_ms": 10476
 },
 "S37sh6::signage": {
  "fp": "de840a7a1e50852f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S37sh6::bgfirst_bg": {
  "input_fingerprint": "4a7340ce8d500871",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 허공으로 튀어 오르는 날카로운 유리 파편과 거친 바닷물이 서로를 껴안은 미연과 이현우를 완전히 덮치기 직전의 찰나.\n\nLOCATION (lock): At a windowed upper-floor circulation area inside the administration building, as floodwater breaks through the glass into the lit interior.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: breaking window behind the embracing pair in the upper-right of the frame, background; incoming water and scattered glass in the middle-right of the frame, midground, moves toward embracing pair left of center.\n- KEY BACKGROUND ELEMENTS: Breaking corridor window (Shattering as the wave enters) — Seen obliquely behind and to the right of the embracing pair, with water crossing its opening; used as Establishes the origin of the incoming impact; Airborne glass fragments (Scattering inward from the breaking window) — Distributed along the diagonal between the window and the pair; used as Traces the impact route without oversized foreground shards; Incoming seawater (Entering the corridor immediately before covering the pair) — Surges from the window toward the interior foreground; used as Connects the exterior disaster physically to the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the corridor's nighttime tonal continuity, preserving separation among the embracing figures, glass fragments, and seawater without adding an impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 허공으로 튀어 오르는 날카로운 유리 파편과 거친 바닷물이 서로를 껴안은 미연과 이현우를 완전히 덮치기 직전의 찰나.\n\nLOCATION (lock): At a windowed upper-floor circulation area inside the administration building, as floodwater breaks through the glass into the lit interior.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: breaking window behind the embracing pair in the upper-right of the frame, background; incoming water and scattered glass in the middle-right of the frame, midground, moves toward embracing pair left of center.\n- KEY BACKGROUND ELEMENTS: Breaking corridor window (Shattering as the wave enters) — Seen obliquely behind and to the right of the embracing pair, with water crossing its opening; used as Establishes the origin of the incoming impact; Airborne glass fragments (Scattering inward from the breaking window) — Distributed along the diagonal between the window and the pair; used as Traces the impact route without oversized foreground shards; Incoming seawater (Entering the corridor immediately before covering the pair) — Surges from the window toward the interior foreground; used as Connects the exterior disaster physically to the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the corridor's nighttime tonal continuity, preserving separation among the embracing figures, glass fragments, and seawater without adding an impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S37sh6__bgfirst_bg.png",
  "asset_id": "1c3ee55e-e3d2-4e32-9973-3c43e0604a89",
  "input_asset_ids": [
   "a9e9e63f-b94c-4a13-94c0-695c11fe79c9",
   "6a3add4a-a856-45ed-8b07-5c69e4800eed"
  ]
 },
 "S37sh6": {
  "input_fingerprint": "9f6eb3f5de7e5730",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 허공으로 튀어 오르는 날카로운 유리 파편과 거친 바닷물이 서로를 껴안은 미연과 이현우를 완전히 덮치기 직전의 찰나.\n\nLOCATION (lock): At a windowed upper-floor circulation area inside the administration building, as floodwater breaks through the glass into the lit interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: breaking window behind the embracing pair in the upper-right of the frame, background; incoming water and scattered glass in the middle-right of the frame, midground, moves toward embracing pair left of center.\n- KEY BACKGROUND ELEMENTS: Breaking corridor window (Shattering as the wave enters) — Seen obliquely behind and to the right of the embracing pair, with water crossing its opening; used as Establishes the origin of the incoming impact; Airborne glass fragments (Scattering inward from the breaking window) — Distributed along the diagonal between the window and the pair; used as Traces the impact route without oversized foreground shards; Incoming seawater (Entering the corridor immediately before covering the pair) — Surges from the window toward the interior foreground; used as Connects the exterior disaster physically to the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the corridor's nighttime tonal continuity, preserving separation among the embracing figures, glass fragments, and seawater without adding an impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management-building windows are breaking as the wave strikes, sending glass fragments and seawater into the interior. This damage replaces the intact-window state. 이현우: He is on a higher floor with his upper body drawn into a protective huddle. His facial bruises and leg injury remain, and the contact card is still concealed inside his shoe. 미연: Her arms are wrapped forward in a protective embrace immediately before the wave engulfs her. Her face remains swollen from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 허공으로 튀어 오르는 날카로운 유리 파편과 거친 바닷물이 서로를 껴안은 미연과 이현우를 완전히 덮치기 직전의 찰나.\n\nLOCATION (lock): At a windowed upper-floor circulation area inside the administration building, as floodwater breaks through the glass into the lit interior. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: breaking window behind the embracing pair in the upper-right of the frame, background; incoming water and scattered glass in the middle-right of the frame, midground, moves toward embracing pair left of center.\n- KEY BACKGROUND ELEMENTS: Breaking corridor window (Shattering as the wave enters) — Seen obliquely behind and to the right of the embracing pair, with water crossing its opening; used as Establishes the origin of the incoming impact; Airborne glass fragments (Scattering inward from the breaking window) — Distributed along the diagonal between the window and the pair; used as Traces the impact route without oversized foreground shards; Incoming seawater (Entering the corridor immediately before covering the pair) — Surges from the window toward the interior foreground; used as Connects the exterior disaster physically to the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the corridor's nighttime tonal continuity, preserving separation among the embracing figures, glass fragments, and seawater without adding an impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management-building windows are breaking as the wave strikes, sending glass fragments and seawater into the interior. This damage replaces the intact-window state. 이현우: He is on a higher floor with his upper body drawn into a protective huddle. His facial bruises and leg injury remain, and the contact card is still concealed inside his shoe. 미연: Her arms are wrapped forward in a protective embrace immediately before the wave engulfs her. Her face remains swollen from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 허공으로 튀어 오르는 날카로운 유리 파편과 거친 바닷물이 서로를 껴안은 미연과 이현우를 완전히 덮치기 직전의 찰나.\n\nLOCATION (lock): At a windowed upper-floor circulation area inside the administration building, as floodwater breaks through the glass into the lit interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: breaking window behind the embracing pair in the upper-right of the frame, background; incoming water and scattered glass in the middle-right of the frame, midground, moves toward embracing pair left of center.\n- KEY BACKGROUND ELEMENTS: Breaking corridor window (Shattering as the wave enters) — Seen obliquely behind and to the right of the embracing pair, with water crossing its opening; used as Establishes the origin of the incoming impact; Airborne glass fragments (Scattering inward from the breaking window) — Distributed along the diagonal between the window and the pair; used as Traces the impact route without oversized foreground shards; Incoming seawater (Entering the corridor immediately before covering the pair) — Surges from the window toward the interior foreground; used as Connects the exterior disaster physically to the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the corridor's nighttime tonal continuity, preserving separation among the embracing figures, glass fragments, and seawater without adding an impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management-building windows are breaking as the wave strikes, sending glass fragments and seawater into the interior. This damage replaces the intact-window state. 이현우: He is on a higher floor with his upper body drawn into a protective huddle. His facial bruises and leg injury remain, and the contact card is still concealed inside his shoe. 미연: Her arms are wrapped forward in a protective embrace immediately before the wave engulfs her. Her face remains swollen from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S37sh6__bgfirst_bg.png",
     "asset_id": "1c3ee55e-e3d2-4e32-9973-3c43e0604a89",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S37sh6.png",
     "asset_id": "a9e9e63f-b94c-4a13-94c0-695c11fe79c9",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:878964>",
     "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L44B02.png",
     "asset_id": "6a3add4a-a856-45ed-8b07-5c69e4800eed",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:878964>",
     "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "창문이 깨지며 쏟아지는 물살과 유리 파편들이 화면 우측에서 좌측에 안고 있는 두 사람을 향해 정확히 날아가고 있음.",
    "built_space": "지정된 복도 레퍼런스와 완전히 다르게, 우측 벽면 전체가 거대한 유리창으로 대체된 구조로 렌더링됨. 인물들은 좌측에 위치함.",
    "entities": "이현우(후드티, 찢어진 청바지, 멍든 얼굴)와 미연(갈색 재킷, 모자) 모두 레퍼런스의 인상착의와 잘 일치함. 바닷물과 파편도 사실적으로 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 공간 구조 왜곡 (레퍼런스에 없는 거대한 측면 유리창을 임의로 생성함)",
     "[gpt-high] 고정된 장소 참조에서 창이 있어야 하는 복도 끝에 새로운 문을 만들고, 유일한 끝창을 오른쪽 측벽 창으로 재구성했다."
    ],
    "physics": "두 인물은 바닥에 발을 딛고 서로를 지탱하고 있음. 쏟아지는 물살과 유리는 들이닥치는 동력에 의해 공중에 떠서 이동 중임."
   },
   {
    "label": "B",
    "direction": "배경 끝의 창문에서 유리 파편이 화면 앞쪽으로 흩어지고 있으나, 바닷물 형상은 창문 주변에 정지된 듯 뭉쳐 있음.",
    "built_space": "양옆에 문이 있고 끝에 창문이 있는 레퍼런스의 복도 구조를 올바르게 유지함. 인물들은 프레임 좌측에 위치함.",
    "entities": "이현우는 인상착의가 일치하나, 미연은 레퍼런스에 있던 모자를 쓰지 않음. 들이닥치는 바닷물이 액체가 아닌 정체불명의 짙은 고체나 먼지 덩어리처럼 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 물리적 지지가 없는 허공의 부유물 (바닷물이 유체의 흐름이나 중력의 영향 없이 허공에 뜬 암석 같은 고체 덩어리로 렌더링됨)"
    ],
    "physics": "인물들은 바닥에 서서 안고 있음. 유리는 날아가고 있으나, 물 형상은 액체로서의 물리적 특성(쏟아짐, 고임 등) 없이 허공에 덩어리째 부유하고 있어 물리적으로 불가능한 상태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지정된 복도 구조는 유지했으나 미연의 모자가 누락되었고, 가장 중요한 바닷물이 유체의 특성을 상실한 채 허공에 뜬 고체 덩어리로 렌더링되어 재질 묘사 및 물리적 사실성에서 크게 실패했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "액션의 역동성과 프레임 구성은 프롬프트에 부합하나, 잠금 조건인 레퍼런스의 공간 구조를 무시하고 복도 우측 벽 전체를 거대한 창문으로 임의 변형하는 치명적인 규정 위반을 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "창문이 깨지며 쏟아지는 물살과 유리 파편들이 화면 우측에서 좌측에 안고 있는 두 사람을 향해 정확히 날아가고 있음.",
        "built_space": "지정된 복도 레퍼런스와 완전히 다르게, 우측 벽면 전체가 거대한 유리창으로 대체된 구조로 렌더링됨. 인물들은 좌측에 위치함.",
        "entities": "이현우(후드티, 찢어진 청바지, 멍든 얼굴)와 미연(갈색 재킷, 모자) 모두 레퍼런스의 인상착의와 잘 일치함. 바닷물과 파편도 사실적으로 묘사됨.",
        "hard_violations": [
         "공간 구조 왜곡 (레퍼런스에 없는 거대한 측면 유리창을 임의로 생성함)"
        ],
        "physics": "두 인물은 바닥에 발을 딛고 서로를 지탱하고 있음. 쏟아지는 물살과 유리는 들이닥치는 동력에 의해 공중에 떠서 이동 중임."
       },
       {
        "label": "B",
        "direction": "배경 끝의 창문에서 유리 파편이 화면 앞쪽으로 흩어지고 있으나, 바닷물 형상은 창문 주변에 정지된 듯 뭉쳐 있음.",
        "built_space": "양옆에 문이 있고 끝에 창문이 있는 레퍼런스의 복도 구조를 올바르게 유지함. 인물들은 프레임 좌측에 위치함.",
        "entities": "이현우는 인상착의가 일치하나, 미연은 레퍼런스에 있던 모자를 쓰지 않음. 들이닥치는 바닷물이 액체가 아닌 정체불명의 짙은 고체나 먼지 덩어리처럼 묘사됨.",
        "hard_violations": [
         "물리적 지지가 없는 허공의 부유물 (바닷물이 유체의 흐름이나 중력의 영향 없이 허공에 뜬 암석 같은 고체 덩어리로 렌더링됨)"
        ],
        "physics": "인물들은 바닥에 서서 안고 있음. 유리는 날아가고 있으나, 물 형상은 액체로서의 물리적 특성(쏟아짐, 고임 등) 없이 허공에 덩어리째 부유하고 있어 물리적으로 불가능한 상태임."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지정된 복도 구조는 유지했으나 미연의 모자가 누락되었고, 가장 중요한 바닷물이 유체의 특성을 상실한 채 허공에 뜬 고체 덩어리로 렌더링되어 재질 묘사 및 물리적 사실성에서 크게 실패했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "액션의 역동성과 프레임 구성은 프롬프트에 부합하나, 잠금 조건인 레퍼런스의 공간 구조를 무시하고 복도 우측 벽 전체를 거대한 창문으로 임의 변형하는 치명적인 규정 위반을 범했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "창문이 깨지며 쏟아지는 물살과 유리 파편들이 화면 우측에서 좌측에 안고 있는 두 사람을 향해 정확히 날아가고 있음.",
        "built_space": "지정된 복도 레퍼런스와 완전히 다르게, 우측 벽면 전체가 거대한 유리창으로 대체된 구조로 렌더링됨. 인물들은 좌측에 위치함.",
        "entities": "이현우(후드티, 찢어진 청바지, 멍든 얼굴)와 미연(갈색 재킷, 모자) 모두 레퍼런스의 인상착의와 잘 일치함. 바닷물과 파편도 사실적으로 묘사됨.",
        "hard_violations": [
         "공간 구조 왜곡 (레퍼런스에 없는 거대한 측면 유리창을 임의로 생성함)"
        ],
        "physics": "두 인물은 바닥에 발을 딛고 서로를 지탱하고 있음. 쏟아지는 물살과 유리는 들이닥치는 동력에 의해 공중에 떠서 이동 중임."
       },
       {
        "label": "B",
        "direction": "배경 끝의 창문에서 유리 파편이 화면 앞쪽으로 흩어지고 있으나, 바닷물 형상은 창문 주변에 정지된 듯 뭉쳐 있음.",
        "built_space": "양옆에 문이 있고 끝에 창문이 있는 레퍼런스의 복도 구조를 올바르게 유지함. 인물들은 프레임 좌측에 위치함.",
        "entities": "이현우는 인상착의가 일치하나, 미연은 레퍼런스에 있던 모자를 쓰지 않음. 들이닥치는 바닷물이 액체가 아닌 정체불명의 짙은 고체나 먼지 덩어리처럼 묘사됨.",
        "hard_violations": [
         "물리적 지지가 없는 허공의 부유물 (바닷물이 유체의 흐름이나 중력의 영향 없이 허공에 뜬 암석 같은 고체 덩어리로 렌더링됨)"
        ],
        "physics": "인물들은 바닥에 서서 안고 있음. 유리는 날아가고 있으나, 물 형상은 액체로서의 물리적 특성(쏟아짐, 고임 등) 없이 허공에 덩어리째 부유하고 있어 물리적으로 불가능한 상태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "참조 복도의 구조와 오른쪽 뒤 창에서 포옹한 두 사람으로 향하는 충격 경로를 보존하지만, 와이드 구도가 다소 좁고 바닷물이 두 사람을 덮치기까지의 간격이 크며 미연의 모자가 빠졌다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "와이드 구도와 덮치기 직전의 파도는 더 정확하지만, 참조의 복도 끝 창을 오른쪽 벽 창과 새로운 막다른 문으로 바꾸어 고정된 장소 구조를 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 사람은 화면 중앙보다 왼쪽에서 얼굴을 서로 가까이 대고 눈을 내리거나 감고 있다. 오른쪽 위 배경의 깨진 창에서 물이 실내 앞쪽으로 솟구치고, 유리 조각은 창과 두 사람 사이의 왼쪽 아래 대각선을 따라 퍼진다. 충격의 목적지는 포옹한 두 사람으로 읽히지만, 물의 선단은 아직 상당히 뒤에 있다.",
        "built_space": "복도 끝에 세로 중앙틀이 있는 창 한 개, 천장 형광등 한 개, 왼쪽에 전경 철문과 그 뒤 두 출입구, 벽 상단 노출 배관이 보인다. 낡은 회색 벽과 어두운 하단 도장, 콘크리트 바닥도 참조와 대응한다. 두 사람은 왼쪽 전경 문 옆의 통행 공간에 있다. 창을 보는 각도는 요청한 사선보다 정면에 가깝고, 인물은 무릎 아래가 잘려 와이드의 여유가 부족하다.",
        "entities": "등장인물은 젊은 동아시아계 남성 한 명과 중년 동아시아계 여성 한 명뿐이다. 남성의 헝클어진 검은 머리, 어두운 후드와 청바지는 이현우 참조에 가깝고, 여성의 단발과 갈색 작업복도 미연에 대응한다. 미연의 얼굴에는 뚜렷한 타박상이 있지만 참조의 모자는 없다. 이현우의 얼굴 멍은 뚜렷하지 않다. 다리 부상과 신발 속 카드는 프레임에서 확인할 수 없다. 깨진 유리와 유입되는 바닷물이 있으며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람은 상체를 붙이고 서로의 어깨와 허리를 팔로 감싸 지지한다. 발은 프레임 밖이지만 몸통과 다리는 아래로 이어지며 공중에 떠 있다는 징후는 없다. 물은 창 개구부의 외부 파도와 연결된 채 압력으로 분출되고, 유리 파편도 같은 충격으로 날아간 것으로 읽힌다. 물 덩어리가 바닥 위로 들려 있지만 창을 통한 분출이라는 발사 원인이 보인다."
       },
       {
        "label": "B",
        "direction": "이현우는 미연의 머리 쪽으로 얼굴을 숙이고, 미연은 그의 가슴에 기대어 왼쪽 앞을 바라본다. 오른쪽 위의 깨진 창에서 나온 파도와 유리는 왼쪽 아래의 두 사람 쪽으로 향한다. 물의 선단이 두 사람 바로 뒤까지 와 있어 완전히 덮치기 직전이라는 순간이 선명하다.",
        "built_space": "오른쪽 벽에 큰 두 칸 창 한 개, 복도 끝에 닫힌 문 한 개, 왼쪽 벽에 전경 문을 포함한 여러 출입구, 천장 형광등 한 개와 노출 배관이 보인다. 두 사람은 왼쪽 벽 앞 바닥에 서 있고 통행 공간과 충돌하지 않는다. 다만 참조는 복도 끝 벽에 창이 있는 구조인데, 이 이미지는 그 끝에 문을 만들고 창을 오른쪽 측벽으로 옮겼다. 이는 단순한 카메라 각도 변화로 설명되지 않는다.",
        "entities": "젊은 동아시아계 남성과 중년 동아시아계 여성 두 명만 보인다. 이현우의 검은 머리, 어두운 후드, 찢어진 청바지와 운동화, 미연의 모자와 갈색 작업복 및 부츠는 참조에 대체로 맞는다. 이현우의 볼에 멍이 보이며 미연의 얼굴에도 붉은 상처가 보이지만 붓기는 명확하지 않다. 찢어진 무릎은 보이나 다리 부상 자체는 확정하기 어렵다. 신발 속 카드는 보이지 않아 은닉 상태와 모순되지 않는다. 바닷물과 유리 파편이 있고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "고정된 장소 참조에서 창이 있어야 하는 복도 끝에 새로운 문을 만들고, 유일한 끝창을 오른쪽 측벽 창으로 재구성했다."
        ],
        "physics": "두 사람 모두 신발이 젖은 바닥에 닿고, 벌린 다리와 굽힌 무릎으로 체중을 받는다. 미연의 팔은 이현우의 몸통을 감싸고 이현우의 손은 미연의 머리를 보호한다. 파도는 창에서 바닥까지 연속적으로 이어지고, 유리 조각은 창 파손의 충격으로 비산한다. 지지나 발사 원인 없이 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "참조 복도의 구조와 오른쪽 뒤 창에서 포옹한 두 사람으로 향하는 충격 경로를 보존하지만, 와이드 구도가 다소 좁고 바닷물이 두 사람을 덮치기까지의 간격이 크며 미연의 모자가 빠졌다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "와이드 구도와 덮치기 직전의 파도는 더 정확하지만, 참조의 복도 끝 창을 오른쪽 벽 창과 새로운 막다른 문으로 바꾸어 고정된 장소 구조를 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 사람은 화면 중앙보다 왼쪽에서 얼굴을 서로 가까이 대고 눈을 내리거나 감고 있다. 오른쪽 위 배경의 깨진 창에서 물이 실내 앞쪽으로 솟구치고, 유리 조각은 창과 두 사람 사이의 왼쪽 아래 대각선을 따라 퍼진다. 충격의 목적지는 포옹한 두 사람으로 읽히지만, 물의 선단은 아직 상당히 뒤에 있다.",
        "built_space": "복도 끝에 세로 중앙틀이 있는 창 한 개, 천장 형광등 한 개, 왼쪽에 전경 철문과 그 뒤 두 출입구, 벽 상단 노출 배관이 보인다. 낡은 회색 벽과 어두운 하단 도장, 콘크리트 바닥도 참조와 대응한다. 두 사람은 왼쪽 전경 문 옆의 통행 공간에 있다. 창을 보는 각도는 요청한 사선보다 정면에 가깝고, 인물은 무릎 아래가 잘려 와이드의 여유가 부족하다.",
        "entities": "등장인물은 젊은 동아시아계 남성 한 명과 중년 동아시아계 여성 한 명뿐이다. 남성의 헝클어진 검은 머리, 어두운 후드와 청바지는 이현우 참조에 가깝고, 여성의 단발과 갈색 작업복도 미연에 대응한다. 미연의 얼굴에는 뚜렷한 타박상이 있지만 참조의 모자는 없다. 이현우의 얼굴 멍은 뚜렷하지 않다. 다리 부상과 신발 속 카드는 프레임에서 확인할 수 없다. 깨진 유리와 유입되는 바닷물이 있으며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람은 상체를 붙이고 서로의 어깨와 허리를 팔로 감싸 지지한다. 발은 프레임 밖이지만 몸통과 다리는 아래로 이어지며 공중에 떠 있다는 징후는 없다. 물은 창 개구부의 외부 파도와 연결된 채 압력으로 분출되고, 유리 파편도 같은 충격으로 날아간 것으로 읽힌다. 물 덩어리가 바닥 위로 들려 있지만 창을 통한 분출이라는 발사 원인이 보인다."
       },
       {
        "label": "A",
        "direction": "이현우는 미연의 머리 쪽으로 얼굴을 숙이고, 미연은 그의 가슴에 기대어 왼쪽 앞을 바라본다. 오른쪽 위의 깨진 창에서 나온 파도와 유리는 왼쪽 아래의 두 사람 쪽으로 향한다. 물의 선단이 두 사람 바로 뒤까지 와 있어 완전히 덮치기 직전이라는 순간이 선명하다.",
        "built_space": "오른쪽 벽에 큰 두 칸 창 한 개, 복도 끝에 닫힌 문 한 개, 왼쪽 벽에 전경 문을 포함한 여러 출입구, 천장 형광등 한 개와 노출 배관이 보인다. 두 사람은 왼쪽 벽 앞 바닥에 서 있고 통행 공간과 충돌하지 않는다. 다만 참조는 복도 끝 벽에 창이 있는 구조인데, 이 이미지는 그 끝에 문을 만들고 창을 오른쪽 측벽으로 옮겼다. 이는 단순한 카메라 각도 변화로 설명되지 않는다.",
        "entities": "젊은 동아시아계 남성과 중년 동아시아계 여성 두 명만 보인다. 이현우의 검은 머리, 어두운 후드, 찢어진 청바지와 운동화, 미연의 모자와 갈색 작업복 및 부츠는 참조에 대체로 맞는다. 이현우의 볼에 멍이 보이며 미연의 얼굴에도 붉은 상처가 보이지만 붓기는 명확하지 않다. 찢어진 무릎은 보이나 다리 부상 자체는 확정하기 어렵다. 신발 속 카드는 보이지 않아 은닉 상태와 모순되지 않는다. 바닷물과 유리 파편이 있고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "고정된 장소 참조에서 창이 있어야 하는 복도 끝에 새로운 문을 만들고, 유일한 끝창을 오른쪽 측벽 창으로 재구성했다."
        ],
        "physics": "두 사람 모두 신발이 젖은 바닥에 닿고, 벌린 다리와 굽힌 무릎으로 체중을 받는다. 미연의 팔은 이현우의 몸통을 감싸고 이현우의 손은 미연의 머리를 보호한다. 파도는 창에서 바닥까지 연속적으로 이어지고, 유리 조각은 창 파손의 충격으로 비산한다. 지지나 발사 원인 없이 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.095,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.845,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 공간 구조 왜곡 (레퍼런스에 없는 거대한 측면 유리창을 임의로 생성함)",
     "[gpt-high] 고정된 장소 참조에서 창이 있어야 하는 복도 끝에 새로운 문을 만들고, 유일한 끝창을 오른쪽 측벽 창으로 재구성했다."
    ],
    "B": [
     "[gemini-pro] 물리적 지지가 없는 허공의 부유물 (바닷물이 유체의 흐름이나 중력의 영향 없이 허공에 뜬 암석 같은 고체 덩어리로 렌더링됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 845
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지정된 복도 구조는 유지했으나 미연의 모자가 누락되었고, 가장 중요한 바닷물이 유체의 특성을 상실한 채 허공에 뜬 고체 덩어리로 렌더링되어 재질 묘사 및 물리적 사실성에서 크게 실패했습니다.  ★위반: [gemini-pro] 물리적 지지가 없는 허공의 부유물 (바닷물이 유체의 흐름이나 중력의 영향 없이 허공에 뜬 암석 같은 고체 덩어리로 렌더링됨)"
   },
   {
    "label": "A",
    "score": 845,
    "verdict_ko": "액션의 역동성과 프레임 구성은 프롬프트에 부합하나, 잠금 조건인 레퍼런스의 공간 구조를 무시하고 복도 우측 벽 전체를 거대한 창문으로 임의 변형하는 치명적인 규정 위반을 범했습니다.  ★위반: [gemini-pro] 공간 구조 왜곡 (레퍼런스에 없는 거대한 측면 유리창을 임의로 생성함) / [gpt-high] 고정된 장소 참조에서 창이 있어야 하는 복도 끝에 새로운 문을 만들고, 유일한 끝창을 오른쪽 측벽 창으로 재구성했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L44B02.png",
    "asset_id": "6a3add4a-a856-45ed-8b07-5c69e4800eed",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:878964>",
    "asset_id": "cfd8b787-f919-4600-bcb7-ff0db73044dc",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdbd-813a-714c-8919-35af77c631cd",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S37sh6__bgfirst_bg.png",
   "bg_asset_id": "1c3ee55e-e3d2-4e32-9973-3c43e0604a89",
   "bg_record_key": "S37sh6::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S37sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:36:08.163193+00:00",
  "fingerprint": "7317761c1711fbf684c268e2f9245b0531849a68b697e49e8f2f868d4adc44ea",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S37sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S37sh6_sel.png",
  "source_sha256": "f6316de100cc62829a78db8148228f9833cfa0eb1abae528ba0a98e218145bea",
  "file": "S37sh6_cine.png",
  "staged_sha256": "83fb0b2cb49dc43cea5f854dfb04f87ec43674231c9b7c819a9126cc20a57396",
  "latency_ms": 12109
 },
 "S38sh2::signage": {
  "fp": "147157cdc7ddf084",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::c4d912045e439ffe": {
  "subjects": [],
  "subject_text": "바다에 떠 있는 컨테이너 내부\n금속 벽으로 둘러싸인 상자형 컨테이너 실내. 적재용 상자들이 놓여 있고 한쪽 끝에는 외부로 열리는 큰 문이 있다.",
  "identity": "canonical",
  "scope_id": "L48",
  "scope_role": "location_interior",
  "scope_sha": "7b8620ec3576d5e8"
 },
 "S38sh2::bgfirst_bg": {
  "input_fingerprint": "94f1e3234f825087",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 바다 저편에 떠 있는 다른 컨테이너 판자 위에 미연이 피를 흘린 채 미동 없이 쓰러져 있는 구도.\n\nLOCATION (lock): On a detached container panel floating on the open sea at dawn, apart from the other drifting wreckage.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: distant panel supporting 미연 in the upper-right of the frame, background; uninterrupted water separating the camera from 미연 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 미연's container panel (Floating separately with 미연 lying on it) — Its upper surface is seen obliquely from above beneath her full body; used as Supports the distant figure at a natural relative scale; Intervening sea (Separating 미연's panel from 이현우's position) — Extends from the near foreground to the distant panel; used as Makes separation the principal compositional fact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Early-dawn ambient illumination allows restrained separation between 미연, her panel, and the surrounding sea without prescribing a sunrise color or direction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 바다 저편에 떠 있는 다른 컨테이너 판자 위에 미연이 피를 흘린 채 미동 없이 쓰러져 있는 구도.\n\nLOCATION (lock): On a detached container panel floating on the open sea at dawn, apart from the other drifting wreckage.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: distant panel supporting 미연 in the upper-right of the frame, background; uninterrupted water separating the camera from 미연 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 미연's container panel (Floating separately with 미연 lying on it) — Its upper surface is seen obliquely from above beneath her full body; used as Supports the distant figure at a natural relative scale; Intervening sea (Separating 미연's panel from 이현우's position) — Extends from the near foreground to the distant panel; used as Makes separation the principal compositional fact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Early-dawn ambient illumination allows restrained separation between 미연, her panel, and the surrounding sea without prescribing a sunrise color or direction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S38sh2__bgfirst_bg.png",
  "asset_id": "15f2d01b-14c8-48e2-b1cf-219d83251c98",
  "input_asset_ids": [
   "35abcbab-54bd-449c-ab76-ae08e7038daf",
   "392c4eb8-924e-444c-8573-409027d6d48e"
  ]
 },
 "S38sh2": {
  "input_fingerprint": "830886a61569e7aa",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 바다 저편에 떠 있는 다른 컨테이너 판자 위에 미연이 피를 흘린 채 미동 없이 쓰러져 있는 구도.\n\nLOCATION (lock): On a detached container panel floating on the open sea at dawn, apart from the other drifting wreckage. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: distant panel supporting 미연 in the upper-right of the frame, background; uninterrupted water separating the camera from 미연 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 미연's container panel (Floating separately with 미연 lying on it) — Its upper surface is seen obliquely from above beneath her full body; used as Supports the distant figure at a natural relative scale; Intervening sea (Separating 미연's panel from 이현우's position) — Extends from the near foreground to the distant panel; used as Makes separation the principal compositional fact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Early-dawn ambient illumination allows restrained separation between 미연, her panel, and the surrounding sea without prescribing a sunrise color or direction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon is reclining helplessly on a floating container panel, supported by the panel and bleeding from her abdomen. Her precise head angle, arm and leg positions, and face-up or sideways orientation are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At dawn, detached container panels and containers float on the sea. Charlie remains inside a container among storage boxes after a night of impacts and drifting. 미연: She is reclining helplessly on a floating container panel with a bleeding puncture wound in her abdomen. Her earlier facial swelling remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 바다 저편에 떠 있는 다른 컨테이너 판자 위에 미연이 피를 흘린 채 미동 없이 쓰러져 있는 구도.\n\nLOCATION (lock): On a detached container panel floating on the open sea at dawn, apart from the other drifting wreckage. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: distant panel supporting 미연 in the upper-right of the frame, background; uninterrupted water separating the camera from 미연 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 미연's container panel (Floating separately with 미연 lying on it) — Its upper surface is seen obliquely from above beneath her full body; used as Supports the distant figure at a natural relative scale; Intervening sea (Separating 미연's panel from 이현우's position) — Extends from the near foreground to the distant panel; used as Makes separation the principal compositional fact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Early-dawn ambient illumination allows restrained separation between 미연, her panel, and the surrounding sea without prescribing a sunrise color or direction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon is reclining helplessly on a floating container panel, supported by the panel and bleeding from her abdomen. Her precise head angle, arm and leg positions, and face-up or sideways orientation are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At dawn, detached container panels and containers float on the sea. Charlie remains inside a container among storage boxes after a night of impacts and drifting. 미연: She is reclining helplessly on a floating container panel with a bleeding puncture wound in her abdomen. Her earlier facial swelling remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 바다 저편에 떠 있는 다른 컨테이너 판자 위에 미연이 피를 흘린 채 미동 없이 쓰러져 있는 구도.\n\nLOCATION (lock): On a detached container panel floating on the open sea at dawn, apart from the other drifting wreckage. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: distant panel supporting 미연 in the upper-right of the frame, background; uninterrupted water separating the camera from 미연 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 미연's container panel (Floating separately with 미연 lying on it) — Its upper surface is seen obliquely from above beneath her full body; used as Supports the distant figure at a natural relative scale; Intervening sea (Separating 미연's panel from 이현우's position) — Extends from the near foreground to the distant panel; used as Makes separation the principal compositional fact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Early-dawn ambient illumination allows restrained separation between 미연, her panel, and the surrounding sea without prescribing a sunrise color or direction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon is reclining helplessly on a floating container panel, supported by the panel and bleeding from her abdomen. Her precise head angle, arm and leg positions, and face-up or sideways orientation are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At dawn, detached container panels and containers float on the sea. Charlie remains inside a container among storage boxes after a night of impacts and drifting. 미연: She is reclining helplessly on a floating container panel with a bleeding puncture wound in her abdomen. Her earlier facial swelling remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S38sh2__bgfirst_bg.png",
     "asset_id": "15f2d01b-14c8-48e2-b1cf-219d83251c98",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S38sh2.png",
     "asset_id": "35abcbab-54bd-449c-ab76-ae08e7038daf",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1387085>",
     "asset_id": "29884448-7e28-4e7c-8652-3354113e4e44",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L48B01.png",
     "asset_id": "392c4eb8-924e-444c-8573-409027d6d48e",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1387085>",
     "asset_id": "29884448-7e28-4e7c-8652-3354113e4e44",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 새벽 바다를 가로질러 원경을 바라보고 있으며, 시선 끝 프레임 우측 상단에 미연이 누워있는 패널이 위치합니다.",
    "built_space": "전경부터 중경까지 끊김 없는 바다가 펼쳐져 카메라와 피사체 사이의 거리감을 주며, 우측 상단 원경에 레퍼런스와 유사한 평평한 금속 판자가 떠 있습니다.",
    "entities": "미연이 레퍼런스와 일치하는 복장(챙이 있는 야구 모자, 재킷, 셔츠, 바지)을 착용하고 있으며, 복부에 피를 흘린 채 정면을 향해 쓰러져 있습니다.",
    "hard_violations": [],
    "physics": "미연의 몸 전체가 패널 표면에 밀착되어 중력에 순응하며 누워 있고, 금속 패널 역시 바다 수면 위에 자연스럽게 떠 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 수평선을 향하고 있으나, 목표가 되는 피사체의 패널이 프레임의 우측 상단이 아닌 중앙 우측 부근에 치우쳐 있습니다.",
    "built_space": "바다 한가운데 늑골 구조를 가진 컨테이너 판자가 떠 있으며, 샷 텍스트가 요구한 '원경(distant)'의 스케일보다는 피사체가 다소 가깝게 배치되었습니다.",
    "entities": "미연이 복부에 피를 흘리며 누워 있으나, 레퍼런스 이미지에서 명확히 확인되는 챙 달린 모자가 누락되어 헤어스타일이 노출되어 있습니다.",
    "hard_violations": [],
    "physics": "미연의 몸은 패널의 기울기에 맞춰 지지되어 있으나, 판자에서 수면으로 흘러내린 피가 물과 섞이는 액체의 질감이 아니라 물결 위에 덧칠해진 평면적인 붉은 그래픽처럼 묘사되어 물질적 사실성(Material Realism)을 위반합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프레임 우측 상단이라는 정확한 위치 지정, 원경의 스케일감, 캐릭터의 의상(모자 포함) 및 복부 출혈 등 프롬프트의 모든 지시를 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 우측 상단이 아닌 중앙에 가깝게 피사체가 배치되었고, 레퍼런스의 모자가 누락되었으며, 수면 위의 피가 평면적인 그래픽 스티커처럼 렌더링되어 사실성이 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 새벽 바다를 가로질러 원경을 바라보고 있으며, 시선 끝 프레임 우측 상단에 미연이 누워있는 패널이 위치합니다.",
        "built_space": "전경부터 중경까지 끊김 없는 바다가 펼쳐져 카메라와 피사체 사이의 거리감을 주며, 우측 상단 원경에 레퍼런스와 유사한 평평한 금속 판자가 떠 있습니다.",
        "entities": "미연이 레퍼런스와 일치하는 복장(챙이 있는 야구 모자, 재킷, 셔츠, 바지)을 착용하고 있으며, 복부에 피를 흘린 채 정면을 향해 쓰러져 있습니다.",
        "hard_violations": [],
        "physics": "미연의 몸 전체가 패널 표면에 밀착되어 중력에 순응하며 누워 있고, 금속 패널 역시 바다 수면 위에 자연스럽게 떠 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 수평선을 향하고 있으나, 목표가 되는 피사체의 패널이 프레임의 우측 상단이 아닌 중앙 우측 부근에 치우쳐 있습니다.",
        "built_space": "바다 한가운데 늑골 구조를 가진 컨테이너 판자가 떠 있으며, 샷 텍스트가 요구한 '원경(distant)'의 스케일보다는 피사체가 다소 가깝게 배치되었습니다.",
        "entities": "미연이 복부에 피를 흘리며 누워 있으나, 레퍼런스 이미지에서 명확히 확인되는 챙 달린 모자가 누락되어 헤어스타일이 노출되어 있습니다.",
        "hard_violations": [],
        "physics": "미연의 몸은 패널의 기울기에 맞춰 지지되어 있으나, 판자에서 수면으로 흘러내린 피가 물과 섞이는 액체의 질감이 아니라 물결 위에 덧칠해진 평면적인 붉은 그래픽처럼 묘사되어 물질적 사실성(Material Realism)을 위반합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프레임 우측 상단이라는 정확한 위치 지정, 원경의 스케일감, 캐릭터의 의상(모자 포함) 및 복부 출혈 등 프롬프트의 모든 지시를 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 우측 상단이 아닌 중앙에 가깝게 피사체가 배치되었고, 레퍼런스의 모자가 누락되었으며, 수면 위의 피가 평면적인 그래픽 스티커처럼 렌더링되어 사실성이 떨어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 새벽 바다를 가로질러 원경을 바라보고 있으며, 시선 끝 프레임 우측 상단에 미연이 누워있는 패널이 위치합니다.",
        "built_space": "전경부터 중경까지 끊김 없는 바다가 펼쳐져 카메라와 피사체 사이의 거리감을 주며, 우측 상단 원경에 레퍼런스와 유사한 평평한 금속 판자가 떠 있습니다.",
        "entities": "미연이 레퍼런스와 일치하는 복장(챙이 있는 야구 모자, 재킷, 셔츠, 바지)을 착용하고 있으며, 복부에 피를 흘린 채 정면을 향해 쓰러져 있습니다.",
        "hard_violations": [],
        "physics": "미연의 몸 전체가 패널 표면에 밀착되어 중력에 순응하며 누워 있고, 금속 패널 역시 바다 수면 위에 자연스럽게 떠 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 수평선을 향하고 있으나, 목표가 되는 피사체의 패널이 프레임의 우측 상단이 아닌 중앙 우측 부근에 치우쳐 있습니다.",
        "built_space": "바다 한가운데 늑골 구조를 가진 컨테이너 판자가 떠 있으며, 샷 텍스트가 요구한 '원경(distant)'의 스케일보다는 피사체가 다소 가깝게 배치되었습니다.",
        "entities": "미연이 복부에 피를 흘리며 누워 있으나, 레퍼런스 이미지에서 명확히 확인되는 챙 달린 모자가 누락되어 헤어스타일이 노출되어 있습니다.",
        "hard_violations": [],
        "physics": "미연의 몸은 패널의 기울기에 맞춰 지지되어 있으나, 판자에서 수면으로 흘러내린 피가 물과 섞이는 액체의 질감이 아니라 물결 위에 덧칠해진 평면적인 붉은 그래픽처럼 묘사되어 물질적 사실성(Material Realism)을 위반합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "새벽 바다와 복부 출혈, 판자에 지지된 부동 자세는 맞지만, 판자와 인물이 더 크게 보여 우측 상단 먼 배경이라는 거리감이 B보다 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "미연과 판자를 우측 상단에 작게 두고 중앙·전경을 끊김 없는 바다로 채워, 분리된 먼 판자를 보는 와이드 구도를 더 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 머리를 화면 오른쪽 위, 발을 왼쪽 아래로 두고 바로 누워 있다. 얼굴은 하늘 쪽이며 특정 대상을 바라보는 시선은 식별되지 않는다. 조준물이나 이동 동작은 없다.",
        "built_space": "바다에 분리된 직사각형 금속 패널 한 장과 그 위의 인물 한 명이 보인다. 패널의 골진 윗면, 테두리, 네 모서리 고정부와 측면 고정부가 드러나며 참조의 낡은 컨테이너 재질과 부합한다. 전신 아래의 윗면을 비스듬히 내려다본다. 중앙과 전경은 바다지만 패널이 화면 너비의 약 3분의 1을 차지해 먼 배경치고 존재감이 크다.",
        "entities": "미연에 해당하는 중년 동아시아계 여성 한 명이며 다른 사람은 없다. 짙은 머리, 갈색 작업 재킷과 바지, 회색 계열 속옷이 보인다. 정확한 얼굴 동일성, 40대 후반의 세부 인상과 기존 얼굴 부종은 이 거리에서 확정하기 어렵다. 복부의 붉은 상처 부위와 판자 위 혈흔, 가장자리 아래 물에 번진 붉은 흔적이 보인다. 관통상 형태 자체는 판별하기 어렵다. 새벽 하늘과 열린 바다가 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 몸통, 팔과 손, 뻗은 다리가 패널에 놓여 있다. 발끝은 위를 향하지만 뒤꿈치와 다리가 받쳐져 있어 공중에 떠 있는 자세는 아니다. 몸을 적극적으로 들어 올리는 동작은 보이지 않는다. 패널은 수면에 걸쳐 떠 있고 혈흔은 패널 표면과 인접한 물에 이어진다."
       },
       {
        "label": "B",
        "direction": "미연은 머리를 화면 왼쪽, 발을 오른쪽으로 두고 바로 누워 있다. 얼굴은 위를 향하며 특정 대상에 대한 응시는 보이지 않는다. 팔은 몸 옆에 내려놓았고 이동이나 조준 동작은 없다.",
        "built_space": "분리된 금속 패널 한 장이 우측 상단에 작게 놓여 있고 그 위에 인물 한 명만 있다. 테두리와 모서리 고정부가 보이며, 먼 쪽 일부는 몸에 가려 고정부 전체 개수는 확정하기 어렵다. 패널 윗면은 전신 아래에서 비스듬히 보인다. 녹슨 금속과 골진 표면은 참조 장소의 재질에 부합한다. 화면 중앙부터 가까운 전경까지 다른 구조물 없이 바다가 이어져 거리와 분리가 뚜렷하다.",
        "entities": "미연에 해당하는 여성 한 명이 검은 모자, 갈색 작업 재킷, 회색 속옷과 어두운 바지를 착용해 참조 의상과 대체로 맞는다. 인물이 작아 정확한 얼굴 동일성, 한국인 중년 여성의 세부 인상과 얼굴 부종까지 검증할 수는 없다. 복부에 붉게 젖은 부위와 몸 옆 판자 위 혈흔이 보이나 관통상 구멍은 식별되지 않는다. 새벽 바다 외에 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 등, 골반과 다리가 패널에 지지되고 보이는 손도 몸 옆 패널에 내려앉아 있다. 발끝이 들린 형태는 뒤꿈치가 놓인 누운 자세와 양립하며, 지지 없이 몸을 들어 올린 부분은 보이지 않는다. 패널 가장자리에 파도가 닿아 수면에 떠 있는 상태가 읽히고 혈흔도 금속 표면에 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "새벽 바다와 복부 출혈, 판자에 지지된 부동 자세는 맞지만, 판자와 인물이 더 크게 보여 우측 상단 먼 배경이라는 거리감이 B보다 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "미연과 판자를 우측 상단에 작게 두고 중앙·전경을 끊김 없는 바다로 채워, 분리된 먼 판자를 보는 와이드 구도를 더 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "미연은 머리를 화면 오른쪽 위, 발을 왼쪽 아래로 두고 바로 누워 있다. 얼굴은 하늘 쪽이며 특정 대상을 바라보는 시선은 식별되지 않는다. 조준물이나 이동 동작은 없다.",
        "built_space": "바다에 분리된 직사각형 금속 패널 한 장과 그 위의 인물 한 명이 보인다. 패널의 골진 윗면, 테두리, 네 모서리 고정부와 측면 고정부가 드러나며 참조의 낡은 컨테이너 재질과 부합한다. 전신 아래의 윗면을 비스듬히 내려다본다. 중앙과 전경은 바다지만 패널이 화면 너비의 약 3분의 1을 차지해 먼 배경치고 존재감이 크다.",
        "entities": "미연에 해당하는 중년 동아시아계 여성 한 명이며 다른 사람은 없다. 짙은 머리, 갈색 작업 재킷과 바지, 회색 계열 속옷이 보인다. 정확한 얼굴 동일성, 40대 후반의 세부 인상과 기존 얼굴 부종은 이 거리에서 확정하기 어렵다. 복부의 붉은 상처 부위와 판자 위 혈흔, 가장자리 아래 물에 번진 붉은 흔적이 보인다. 관통상 형태 자체는 판별하기 어렵다. 새벽 하늘과 열린 바다가 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 몸통, 팔과 손, 뻗은 다리가 패널에 놓여 있다. 발끝은 위를 향하지만 뒤꿈치와 다리가 받쳐져 있어 공중에 떠 있는 자세는 아니다. 몸을 적극적으로 들어 올리는 동작은 보이지 않는다. 패널은 수면에 걸쳐 떠 있고 혈흔은 패널 표면과 인접한 물에 이어진다."
       },
       {
        "label": "A",
        "direction": "미연은 머리를 화면 왼쪽, 발을 오른쪽으로 두고 바로 누워 있다. 얼굴은 위를 향하며 특정 대상에 대한 응시는 보이지 않는다. 팔은 몸 옆에 내려놓았고 이동이나 조준 동작은 없다.",
        "built_space": "분리된 금속 패널 한 장이 우측 상단에 작게 놓여 있고 그 위에 인물 한 명만 있다. 테두리와 모서리 고정부가 보이며, 먼 쪽 일부는 몸에 가려 고정부 전체 개수는 확정하기 어렵다. 패널 윗면은 전신 아래에서 비스듬히 보인다. 녹슨 금속과 골진 표면은 참조 장소의 재질에 부합한다. 화면 중앙부터 가까운 전경까지 다른 구조물 없이 바다가 이어져 거리와 분리가 뚜렷하다.",
        "entities": "미연에 해당하는 여성 한 명이 검은 모자, 갈색 작업 재킷, 회색 속옷과 어두운 바지를 착용해 참조 의상과 대체로 맞는다. 인물이 작아 정확한 얼굴 동일성, 한국인 중년 여성의 세부 인상과 얼굴 부종까지 검증할 수는 없다. 복부에 붉게 젖은 부위와 몸 옆 판자 위 혈흔이 보이나 관통상 구멍은 식별되지 않는다. 새벽 바다 외에 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 등, 골반과 다리가 패널에 지지되고 보이는 손도 몸 옆 패널에 내려앉아 있다. 발끝이 들린 형태는 뒤꿈치가 놓인 누운 자세와 양립하며, 지지 없이 몸을 들어 올린 부분은 보이지 않는다. 패널 가장자리에 파도가 닿아 수면에 떠 있는 상태가 읽히고 혈흔도 금속 표면에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.489
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.489
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1489
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프레임 우측 상단이라는 정확한 위치 지정, 원경의 스케일감, 캐릭터의 의상(모자 포함) 및 복부 출혈 등 프롬프트의 모든 지시를 완벽하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1489,
    "verdict_ko": "지정된 우측 상단이 아닌 중앙에 가깝게 피사체가 배치되었고, 레퍼런스의 모자가 누락되었으며, 수면 위의 피가 평면적인 그래픽 스티커처럼 렌더링되어 사실성이 떨어집니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L48B01.png",
    "asset_id": "392c4eb8-924e-444c-8573-409027d6d48e",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1387085>",
    "asset_id": "29884448-7e28-4e7c-8652-3354113e4e44",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdbe-49ab-7ead-bb5c-af7ea998b38d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S38sh2__bgfirst_bg.png",
   "bg_asset_id": "15f2d01b-14c8-48e2-b1cf-219d83251c98",
   "bg_record_key": "S38sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S38sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:36:20.698641+00:00",
  "fingerprint": "056372cfc3a7d784e5d98a23831dbace3ba9eebd0ba5502f48acd226e2d67db7",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S38sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S38sh2_sel.png",
  "source_sha256": "4617edf3f5c1a203a968891f072942081daa1f3fcbcfbdcc34a5026888fff83c",
  "file": "S38sh2_cine.png",
  "staged_sha256": "ab873a430dba88eea31c9a4b5c883e94de5d9d28a697cf7104dcf384750ed5df",
  "latency_ms": 12147
 },
 "S38sh6::signage": {
  "fp": "69ff6f71c3ffaf4a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S38sh6": {
  "input_fingerprint": "2e8ae75478976ebf",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 눈을 감은 미연의 상체를 꽉 끌어안은 이현우가 하늘을 향해 목에 핏대를 세우며 포효하는 광기 어린 표정.\n\nLOCATION (lock): On the floating container panel supporting the injured mother amid the flooded settlement's open waters at dawn. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container panel beneath the pair (Floating and supporting their embrace) — Only the near edge and a limited area beneath their bodies remain visible; used as Grounds the embrace without competing with the faces; Surrounding sea (Visible beyond the floating panel) — Seen directly through the open space beside the pair; used as Maintains the exposed setting behind the intimate composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained early-dawn ambient illumination, giving the faces readable tonal weight without a new color cue or heightened fantasy effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's lifeless upper body is held tightly against Hyunwoo in his arms on the floating container panel, with her eyes closed. Her head's exact support, the positions of her arms and legs, and the extent to which her lower body rests on the panel are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container panels and containers remain afloat on the dawn sea. Charlie is still inside the drifting container among storage boxes before opening its door. 미연: She is dead with her eyes closed, still on the floating panel. The bleeding abdominal puncture wound and swollen face remain. 이현우: He is on the floating panel, bent into a tight embrace and sobbing. His facial bruises and leg injury persist, and the contact card remains inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 눈을 감은 미연의 상체를 꽉 끌어안은 이현우가 하늘을 향해 목에 핏대를 세우며 포효하는 광기 어린 표정.\n\nLOCATION (lock): On the floating container panel supporting the injured mother amid the flooded settlement's open waters at dawn. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container panel beneath the pair (Floating and supporting their embrace) — Only the near edge and a limited area beneath their bodies remain visible; used as Grounds the embrace without competing with the faces; Surrounding sea (Visible beyond the floating panel) — Seen directly through the open space beside the pair; used as Maintains the exposed setting behind the intimate composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained early-dawn ambient illumination, giving the faces readable tonal weight without a new color cue or heightened fantasy effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's lifeless upper body is held tightly against Hyunwoo in his arms on the floating container panel, with her eyes closed. Her head's exact support, the positions of her arms and legs, and the extent to which her lower body rests on the panel are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container panels and containers remain afloat on the dawn sea. Charlie is still inside the drifting container among storage boxes before opening its door. 미연: She is dead with her eyes closed, still on the floating panel. The bleeding abdominal puncture wound and swollen face remain. 이현우: He is on the floating panel, bent into a tight embrace and sobbing. His facial bruises and leg injury persist, and the contact card remains inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 눈을 감은 미연의 상체를 꽉 끌어안은 이현우가 하늘을 향해 목에 핏대를 세우며 포효하는 광기 어린 표정.\n\nLOCATION (lock): On the floating container panel supporting the injured mother amid the flooded settlement's open waters at dawn. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container panel beneath the pair (Floating and supporting their embrace) — Only the near edge and a limited area beneath their bodies remain visible; used as Grounds the embrace without competing with the faces; Surrounding sea (Visible beyond the floating panel) — Seen directly through the open space beside the pair; used as Maintains the exposed setting behind the intimate composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained early-dawn ambient illumination, giving the faces readable tonal weight without a new color cue or heightened fantasy effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's lifeless upper body is held tightly against Hyunwoo in his arms on the floating container panel, with her eyes closed. Her head's exact support, the positions of her arms and legs, and the extent to which her lower body rests on the panel are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container panels and containers remain afloat on the dawn sea. Charlie is still inside the drifting container among storage boxes before opening its door. 미연: She is dead with her eyes closed, still on the floating panel. The bleeding abdominal puncture wound and swollen face remain. 이현우: He is on the floating panel, bent into a tight embrace and sobbing. His facial bruises and leg injury persist, and the contact card remains inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우가 하늘을 향해 고개를 젖히고 포효하고 있음.",
    "built_space": "바다 위 컨테이너 패널 모서리 쪽에 두 사람이 위치함.",
    "entities": "이현우(어두운 후드티), 미연(갈색 재킷, 모자 착용).",
    "hard_violations": [
     "[gemini-pro] 의식이 없는 캐릭터(미연)가 팔을 들어 이현우의 어깨를 꽉 잡고 있는 자세 (중력 및 근육 미사용 조건 위반)",
     "[gemini-pro] 미연의 허리를 감싼 팔이 갈색 소매를 입고 있어 누구의 팔인지 알 수 없는 해부학적 구조 및 의상 오류",
     "[gpt-high] 죽은 미연의 팔이 가슴 높이로 들리고 팔꿈치가 굽혀진 채 손이 현우의 목·어깨를 감싼다. 손끝의 어깨 접촉 외에는 이 들어 올린 자세를 수동적으로 유지할 받침이 드러나지 않아, 무근력 시신이 아니라 스스로 매달려 포옹하는 동작이 된다."
    ],
    "physics": "미연이 의식을 잃은 상태임에도 손에 힘을 주어 이현우를 잡고 있으며, 뒤엉킨 팔의 소매와 신체 연결이 물리적으로 불가능함."
   },
   {
    "label": "B",
    "direction": "이현우가 고개를 들고 하늘을 향해 입을 벌려 포효하고 있음.",
    "built_space": "바다에 떠 있는 컨테이너 패널 위이며, 바닥에 핏물이 고여 있음.",
    "entities": "이현우(얼굴 멍 자국, 어두운 후드티), 미연(갈색 재킷, 모자 없음).",
    "hard_violations": [],
    "physics": "미연은 근육의 힘이 완전히 빠진 채 이현우의 품에 축 늘어져 안겨 있으며, 이현우의 두 팔이 그녀를 온전히 지탱하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "사망한 캐릭터가 능동적으로 상대를 붙잡는 중력 위반 및 팔의 소매 색상이 뒤섞인 해부학적 오류가 발생하여 실격입니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "미연의 모자가 누락된 점은 아쉬우나, 핏대를 세우며 포효하는 이현우와 축 늘어진 미연의 물리적 상태를 지시문대로 훌륭하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우가 하늘을 향해 고개를 젖히고 포효하고 있음.",
        "built_space": "바다 위 컨테이너 패널 모서리 쪽에 두 사람이 위치함.",
        "entities": "이현우(어두운 후드티), 미연(갈색 재킷, 모자 착용).",
        "hard_violations": [
         "의식이 없는 캐릭터(미연)가 팔을 들어 이현우의 어깨를 꽉 잡고 있는 자세 (중력 및 근육 미사용 조건 위반)",
         "미연의 허리를 감싼 팔이 갈색 소매를 입고 있어 누구의 팔인지 알 수 없는 해부학적 구조 및 의상 오류"
        ],
        "physics": "미연이 의식을 잃은 상태임에도 손에 힘을 주어 이현우를 잡고 있으며, 뒤엉킨 팔의 소매와 신체 연결이 물리적으로 불가능함."
       },
       {
        "label": "B",
        "direction": "이현우가 고개를 들고 하늘을 향해 입을 벌려 포효하고 있음.",
        "built_space": "바다에 떠 있는 컨테이너 패널 위이며, 바닥에 핏물이 고여 있음.",
        "entities": "이현우(얼굴 멍 자국, 어두운 후드티), 미연(갈색 재킷, 모자 없음).",
        "hard_violations": [],
        "physics": "미연은 근육의 힘이 완전히 빠진 채 이현우의 품에 축 늘어져 안겨 있으며, 이현우의 두 팔이 그녀를 온전히 지탱하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "사망한 캐릭터가 능동적으로 상대를 붙잡는 중력 위반 및 팔의 소매 색상이 뒤섞인 해부학적 오류가 발생하여 실격입니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "미연의 모자가 누락된 점은 아쉬우나, 핏대를 세우며 포효하는 이현우와 축 늘어진 미연의 물리적 상태를 지시문대로 훌륭하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우가 하늘을 향해 고개를 젖히고 포효하고 있음.",
        "built_space": "바다 위 컨테이너 패널 모서리 쪽에 두 사람이 위치함.",
        "entities": "이현우(어두운 후드티), 미연(갈색 재킷, 모자 착용).",
        "hard_violations": [
         "의식이 없는 캐릭터(미연)가 팔을 들어 이현우의 어깨를 꽉 잡고 있는 자세 (중력 및 근육 미사용 조건 위반)",
         "미연의 허리를 감싼 팔이 갈색 소매를 입고 있어 누구의 팔인지 알 수 없는 해부학적 구조 및 의상 오류"
        ],
        "physics": "미연이 의식을 잃은 상태임에도 손에 힘을 주어 이현우를 잡고 있으며, 뒤엉킨 팔의 소매와 신체 연결이 물리적으로 불가능함."
       },
       {
        "label": "B",
        "direction": "이현우가 고개를 들고 하늘을 향해 입을 벌려 포효하고 있음.",
        "built_space": "바다에 떠 있는 컨테이너 패널 위이며, 바닥에 핏물이 고여 있음.",
        "entities": "이현우(얼굴 멍 자국, 어두운 후드티), 미연(갈색 재킷, 모자 없음).",
        "hard_violations": [],
        "physics": "미연은 근육의 힘이 완전히 빠진 채 이현우의 품에 축 늘어져 안겨 있으며, 이현우의 두 팔이 그녀를 온전히 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "하늘을 향한 포효와 미연의 머리를 가슴에 받친 밀착 포옹은 충실하지만, 미연의 모자가 사라졌고 패널 노출 면적이 지나치게 넓다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "모자와 새벽 바다는 잘 유지했지만, 죽은 미연이 팔을 들어 현우의 목을 감싸는 자세가 무근력·중력 준수 지시를 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 고개와 벌린 입을 화면 왼쪽 위 하늘로 향하고 목의 힘줄을 드러낸다. 포효의 방향은 지시와 맞는다. 미연은 눈을 감고 얼굴을 아래로 기울여 현우의 가슴에 댄다. 겨누는 도구나 이동하는 물체는 없다.",
        "built_space": "두 사람은 녹슨 컨테이너 패널 하나 위에 붙어 앉아 있다. 패널 뒤쪽 양끝에 돌출 고정부 두 개, 왼쪽 가장자리에 추가 돌출 부품들이 보인다. 주변은 열린 바다이며 별도 건물이나 인물은 없다. 다만 패널이 화면 하단 대부분과 인물 뒤까지 넓게 차지하여, 가까운 가장자리와 몸 아래의 제한된 면적만 보이라는 구도에서 벗어난다.",
        "entities": "젊은 한국계 남성으로 보이는 현우와 중년 한국인 여성으로 보이는 미연 두 명만 있다. 현우의 헝클어진 검은 머리, 해진 짙은 후드, 얼굴의 멍은 참조와 부합한다. 미연의 검은 머리와 낡은 갈색 겉옷, 감긴 눈은 맞지만 참조의 모자가 없다. 얼굴 부종은 뚜렷하지 않다. 복부 상처는 포옹에 가려 확인할 수 없고, 패널에는 검붉은 액체가 보인다. 신발 속 카드는 프레임 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "현우의 앉은 하체는 패널에 지지되고, 양팔은 미연의 어깨와 몸통을 감싼다. 미연의 머리는 현우의 가슴에 기대어 아래로 처져 있어 지지가 명확하다. 보이는 팔도 낮은 위치에서 두 몸 사이와 현우의 팔 아래에 놓여 있다. 손이 허리를 감싸는 모양은 다소 능동적으로 보이지만, 공중에 떠 있는 팔다리로 단정할 근거는 없다. 패널은 수면에 떠서 두 사람을 받친다."
       },
       {
        "label": "B",
        "direction": "현우는 눈을 질끈 감고 얼굴과 벌린 입을 왼쪽 위 하늘로 향한다. 목의 힘줄과 포효 방향은 지시대로다. 미연은 눈을 감은 채 얼굴이 위로 젖혀져 있으며, 현우를 바라보는 시선은 없다. 무기나 방향성 소품은 없다.",
        "built_space": "녹슨 패널 하나가 두 사람을 받치며 왼쪽 뒤 모서리에 구멍 난 고정부 하나가 선명하게 보인다. 두 사람 옆과 뒤로 바다가 직접 이어져 장소 관계는 맞는다. 그러나 현우의 허벅지와 무릎, 미연의 하체 일부, 패널의 넓은 면적까지 드러나 요청한 밀착 미디엄 숏보다 하체와 바닥의 비중이 크다.",
        "entities": "지정된 두 인물만 보인다. 현우의 젊은 한국계 남성 외형, 검은 머리, 짙은 후드와 찢어진 청바지는 참조에 가깝다. 얼굴에는 상처 흔적이 있으나 멍은 약하다. 미연은 중년 한국인 여성 외형이며 참조의 어두운 모자, 검은 머리와 갈색 작업 겉옷을 유지한다. 눈은 감겨 있지만 얼굴 부종은 분명하지 않다. 복부 상처는 가려져 확인되지 않는다. 카드와 신발은 보이지 않으며 읽을 수 있는 글자도 없다.",
        "hard_violations": [
         "죽은 미연의 팔이 가슴 높이로 들리고 팔꿈치가 굽혀진 채 손이 현우의 목·어깨를 감싼다. 손끝의 어깨 접촉 외에는 이 들어 올린 자세를 수동적으로 유지할 받침이 드러나지 않아, 무근력 시신이 아니라 스스로 매달려 포옹하는 동작이 된다."
        ],
        "physics": "현우의 골반과 굽힌 다리는 패널에 놓이고, 그의 팔과 손은 미연의 등과 몸통을 받친다. 미연의 하체도 무릎과 패널 쪽에 놓여 있어 몸 전체가 떠 있지는 않다. 뒤로 젖혀진 머리는 상체의 기울기에 따라 처지는 것으로 읽을 수 있다. 그러나 목을 감싼 미연의 팔은 위로 들어 굽힌 상태이며, 손을 어깨에 걸친 것만으로 그 팔 전체가 힘없이 놓인 자세라고 보기 어렵다. 패널 자체는 수면의 부력으로 지지된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "하늘을 향한 포효와 미연의 머리를 가슴에 받친 밀착 포옹은 충실하지만, 미연의 모자가 사라졌고 패널 노출 면적이 지나치게 넓다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "모자와 새벽 바다는 잘 유지했지만, 죽은 미연이 팔을 들어 현우의 목을 감싸는 자세가 무근력·중력 준수 지시를 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 고개와 벌린 입을 화면 왼쪽 위 하늘로 향하고 목의 힘줄을 드러낸다. 포효의 방향은 지시와 맞는다. 미연은 눈을 감고 얼굴을 아래로 기울여 현우의 가슴에 댄다. 겨누는 도구나 이동하는 물체는 없다.",
        "built_space": "두 사람은 녹슨 컨테이너 패널 하나 위에 붙어 앉아 있다. 패널 뒤쪽 양끝에 돌출 고정부 두 개, 왼쪽 가장자리에 추가 돌출 부품들이 보인다. 주변은 열린 바다이며 별도 건물이나 인물은 없다. 다만 패널이 화면 하단 대부분과 인물 뒤까지 넓게 차지하여, 가까운 가장자리와 몸 아래의 제한된 면적만 보이라는 구도에서 벗어난다.",
        "entities": "젊은 한국계 남성으로 보이는 현우와 중년 한국인 여성으로 보이는 미연 두 명만 있다. 현우의 헝클어진 검은 머리, 해진 짙은 후드, 얼굴의 멍은 참조와 부합한다. 미연의 검은 머리와 낡은 갈색 겉옷, 감긴 눈은 맞지만 참조의 모자가 없다. 얼굴 부종은 뚜렷하지 않다. 복부 상처는 포옹에 가려 확인할 수 없고, 패널에는 검붉은 액체가 보인다. 신발 속 카드는 프레임 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "현우의 앉은 하체는 패널에 지지되고, 양팔은 미연의 어깨와 몸통을 감싼다. 미연의 머리는 현우의 가슴에 기대어 아래로 처져 있어 지지가 명확하다. 보이는 팔도 낮은 위치에서 두 몸 사이와 현우의 팔 아래에 놓여 있다. 손이 허리를 감싸는 모양은 다소 능동적으로 보이지만, 공중에 떠 있는 팔다리로 단정할 근거는 없다. 패널은 수면에 떠서 두 사람을 받친다."
       },
       {
        "label": "A",
        "direction": "현우는 눈을 질끈 감고 얼굴과 벌린 입을 왼쪽 위 하늘로 향한다. 목의 힘줄과 포효 방향은 지시대로다. 미연은 눈을 감은 채 얼굴이 위로 젖혀져 있으며, 현우를 바라보는 시선은 없다. 무기나 방향성 소품은 없다.",
        "built_space": "녹슨 패널 하나가 두 사람을 받치며 왼쪽 뒤 모서리에 구멍 난 고정부 하나가 선명하게 보인다. 두 사람 옆과 뒤로 바다가 직접 이어져 장소 관계는 맞는다. 그러나 현우의 허벅지와 무릎, 미연의 하체 일부, 패널의 넓은 면적까지 드러나 요청한 밀착 미디엄 숏보다 하체와 바닥의 비중이 크다.",
        "entities": "지정된 두 인물만 보인다. 현우의 젊은 한국계 남성 외형, 검은 머리, 짙은 후드와 찢어진 청바지는 참조에 가깝다. 얼굴에는 상처 흔적이 있으나 멍은 약하다. 미연은 중년 한국인 여성 외형이며 참조의 어두운 모자, 검은 머리와 갈색 작업 겉옷을 유지한다. 눈은 감겨 있지만 얼굴 부종은 분명하지 않다. 복부 상처는 가려져 확인되지 않는다. 카드와 신발은 보이지 않으며 읽을 수 있는 글자도 없다.",
        "hard_violations": [
         "죽은 미연의 팔이 가슴 높이로 들리고 팔꿈치가 굽혀진 채 손이 현우의 목·어깨를 감싼다. 손끝의 어깨 접촉 외에는 이 들어 올린 자세를 수동적으로 유지할 받침이 드러나지 않아, 무근력 시신이 아니라 스스로 매달려 포옹하는 동작이 된다."
        ],
        "physics": "현우의 골반과 굽힌 다리는 패널에 놓이고, 그의 팔과 손은 미연의 등과 몸통을 받친다. 미연의 하체도 무릎과 패널 쪽에 놓여 있어 몸 전체가 떠 있지는 않다. 뒤로 젖혀진 머리는 상체의 기울기에 따라 처지는 것으로 읽을 수 있다. 그러나 목을 감싼 미연의 팔은 위로 들어 굽힌 상태이며, 손을 어깨에 걸친 것만으로 그 팔 전체가 힘없이 놓인 자세라고 보기 어렵다. 패널 자체는 수면의 부력으로 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.75,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 의식이 없는 캐릭터(미연)가 팔을 들어 이현우의 어깨를 꽉 잡고 있는 자세 (중력 및 근육 미사용 조건 위반)",
     "[gemini-pro] 미연의 허리를 감싼 팔이 갈색 소매를 입고 있어 누구의 팔인지 알 수 없는 해부학적 구조 및 의상 오류",
     "[gpt-high] 죽은 미연의 팔이 가슴 높이로 들리고 팔꿈치가 굽혀진 채 손이 현우의 목·어깨를 감싼다. 손끝의 어깨 접촉 외에는 이 들어 올린 자세를 수동적으로 유지할 받침이 드러나지 않아, 무근력 시신이 아니라 스스로 매달려 포옹하는 동작이 된다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 750,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "사망한 캐릭터가 능동적으로 상대를 붙잡는 중력 위반 및 팔의 소매 색상이 뒤섞인 해부학적 오류가 발생하여 실격입니다.  ★위반: [gemini-pro] 의식이 없는 캐릭터(미연)가 팔을 들어 이현우의 어깨를 꽉 잡고 있는 자세 (중력 및 근육 미사용 조건 위반) / [gemini-pro] 미연의 허리를 감싼 팔이 갈색 소매를 입고 있어 누구의 팔인지 알 수 없는 해부학적 구조 및 의상 오류 / [gpt-high] 죽은 미연의 팔이 가슴 높이로 들리고 팔꿈치가 굽혀진 채 손이 현우의 목·어깨를 감싼다. 손끝의 어깨 접촉 외에는 이 들어 올린 자세를 수동적으로 유지할 받침이 드러나지 않아, 무근력 시신이 아니라 스스로 매달려 포옹하는 동작이 된다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "미연의 모자가 누락된 점은 아쉬우나, 핏대를 세우며 포효하는 이현우와 축 늘어진 미연의 물리적 상태를 지시문대로 훌륭하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S38sh2_sel.png",
    "asset_id": "79ceb555-4796-43d5-a2ab-c6964f690a1b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1421993>",
    "asset_id": "f4820086-7e89-49e3-b8cf-fe160a7afdcd",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdbf-12d3-7ef6-b3ed-556d05d10853",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S38sh2"
  }
 },
 "S38sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:36:33.189080+00:00",
  "fingerprint": "eb2598d856e67173538203b36803ea628c0e0770be8361e0842786536f368594",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S38sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S38sh6_sel.png",
  "source_sha256": "314da6a50c4cdde579ca4c7f7d513e091668c5cf3d4824746756548d4c46f3f5",
  "file": "S38sh6_cine.png",
  "staged_sha256": "c059da6c72d95bd32b0f6201ca815b662680d822b8dcecb27b0ede9a937b4eb8",
  "latency_ms": 12411
 },
 "S38sh12::signage": {
  "fp": "88ceccd4b06a1276",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S38sh12": {
  "input_fingerprint": "d2d8ee99174014cd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 이현우 옆에 선 채 피투성이인 미연의 얼굴을 바라보며 두 눈이 크게 확장된 앰버의 충격받은 표정 클로즈업.\n\nLOCATION (lock): On the floating wreckage beside the mother's body and her son, surrounded by open water at dawn. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sea beyond the group (Visible as limited background around 앰버) — Seen directly beyond her shoulder, without a reflected image; used as Provides restrained setting continuity while remaining secondary to her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established early-dawn ambient illumination and restrained contrast, letting 앰버's widened eyes convey shock without an artificial lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's body is reclining motionlessly on the floating container panel beside the crouching Hyunwoo, with her eyes closed as though asleep. Her exact head angle, arm and leg positions, and face-up or sideways orientation are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Floating containers and detached panels remain clustered on the dawn sea, and the container door opened for escape remains open. Charlie is now outside the container, with his old metal body and peeling chest logo unchanged. 앰버: She has climbed out of the container and reached the floating wreckage. Her expression has changed from relief to extreme shock and grief.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 이현우 옆에 선 채 피투성이인 미연의 얼굴을 바라보며 두 눈이 크게 확장된 앰버의 충격받은 표정 클로즈업.\n\nLOCATION (lock): On the floating wreckage beside the mother's body and her son, surrounded by open water at dawn. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sea beyond the group (Visible as limited background around 앰버) — Seen directly beyond her shoulder, without a reflected image; used as Provides restrained setting continuity while remaining secondary to her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established early-dawn ambient illumination and restrained contrast, letting 앰버's widened eyes convey shock without an artificial lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's body is reclining motionlessly on the floating container panel beside the crouching Hyunwoo, with her eyes closed as though asleep. Her exact head angle, arm and leg positions, and face-up or sideways orientation are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Floating containers and detached panels remain clustered on the dawn sea, and the container door opened for escape remains open. Charlie is now outside the container, with his old metal body and peeling chest logo unchanged. 앰버: She has climbed out of the container and reached the floating wreckage. Her expression has changed from relief to extreme shock and grief.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 이현우 옆에 선 채 피투성이인 미연의 얼굴을 바라보며 두 눈이 크게 확장된 앰버의 충격받은 표정 클로즈업.\n\nLOCATION (lock): On the floating wreckage beside the mother's body and her son, surrounded by open water at dawn. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sea beyond the group (Visible as limited background around 앰버) — Seen directly beyond her shoulder, without a reflected image; used as Provides restrained setting continuity while remaining secondary to her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established early-dawn ambient illumination and restrained contrast, letting 앰버's widened eyes convey shock without an artificial lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's body is reclining motionlessly on the floating container panel beside the crouching Hyunwoo, with her eyes closed as though asleep. Her exact head angle, arm and leg positions, and face-up or sideways orientation are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Floating containers and detached panels remain clustered on the dawn sea, and the container door opened for escape remains open. Charlie is now outside the container, with his old metal body and peeling chest logo unchanged. 앰버: She has climbed out of the container and reached the floating wreckage. Her expression has changed from relief to extreme shock and grief.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "앰버의 두 눈은 화면 왼쪽의 검은 옷 인물 쪽, 거의 수평보다 약간 위를 향한다. 아래에 누워 있는 미연의 피투성이 얼굴을 내려다보는 시선은 아니다. 미연의 얼굴은 화면 밖이라 실제 시선 도착점은 확인되지 않는다.",
    "built_space": "얼굴과 어깨를 크게 잡았으며, 뒤로 녹슨 금속 패널 한 면과 그 가장자리, 오른쪽의 파란 컨테이너 일부가 보인다. 바다가 어깨 뒤로 직접 보이고 반사 영상은 없다. 참조의 부식된 해상 잔해와 대체로 연결되지만, 배경 잔해가 차지하는 면적은 다소 크다. 왼쪽 인물은 몸통 일부만 보여 쪼그려 앉은 현우의 자세인지 확인할 수 없다. 탈출문은 프레임에 없어 개폐 상태를 판단할 수 없다.",
    "entities": "금발의 약 10세 여자아이 한 명이 중심이며, 얼굴·체격과 회녹색 재킷 및 회색 안쪽 옷은 앰버 참조에 대체로 가깝다. 한국계·백인 혼혈이라는 정체성은 외모만으로 확정할 수 없다. 참조의 흰 마스크는 없다. 왼쪽에는 현우로 해석되는 검은 옷의 몸통 일부만 있고, 미연과 찰리는 보이지 않는다. 눈은 정상적인 해부학적 형태지만 크게 확장된 충격보다는 걱정스러운 표정에 가깝다. 읽을 수 있는 글자는 없다.",
    "hard_violations": [],
    "physics": "앰버의 머리는 목과 몸통에 자연스럽게 이어지고 머리카락과 옷은 중력에 맞게 내려온다. 발과 바닥 접촉은 클로즈업 밖이므로 서 있는 지지 상태를 직접 확인할 수 없지만, 공중에 떠 있다는 증거는 없다. 왼쪽 인물도 하체가 잘려 지지점을 확인할 수 없다. 미연은 보이지 않아 신체의 지지와 무동작 자세는 평가할 수 없다."
   },
   {
    "label": "A",
    "direction": "앰버는 눈을 크게 뜨고 화면 왼쪽의 프레임 밖을 거의 수평으로 바라본다. 누운 미연의 얼굴을 향하는 하향 시선은 보이지 않는다. 왼쪽에서 들어온 손은 앰버의 어깨에 닿아 있으며, 미연을 가리키거나 받치는 동작은 아니다.",
    "built_space": "앰버의 얼굴과 어깨 중심 클로즈업이며, 어깨 뒤로 새벽 바다가 직접 보인다. 아래쪽에는 흐린 녹슨 패널 한 면, 오른쪽 뒤에는 작은 모서리 부속 한 개가 보인다. 반사 영상이나 중복된 고정 설비는 없다. 현우로 해석되는 인물의 팔만 왼쪽에서 들어와 그의 앉은 위치와 자세는 확인되지 않는다. 탈출문과 다른 컨테이너는 프레임 밖이다.",
    "entities": "금발의 약 10세 여자아이로 얼굴과 체격은 앰버 참조에 대체로 가깝고, 자연스러운 눈을 크게 뜬 충격이 분명하다. 혼혈 정체성 자체는 외모만으로 확정할 수 없다. 회녹색 재킷 위에 참조와 지문에 없는 주황색·회색 구명조끼가 추가되었고 흰 마스크는 없다. 왼쪽에는 현우로 해석되는 성인 크기의 손과 검은 소매가 보이지만 얼굴은 없다. 미연과 찰리는 보이지 않으며 읽을 수 있는 글자는 없다.",
    "hard_violations": [
     "참조 의상과 지문에 없는 주황색·회색 구명조끼를 앰버의 상체에 큰 소품으로 추가했다."
    ],
    "physics": "왼쪽 손은 소매에서 이어져 앰버의 어깨 위에 자연스럽게 얹혀 있고, 구명조끼는 어깨와 몸통에 걸쳐 지지된다. 머리와 목의 연결 및 머리카락의 처짐도 자연스럽다. 하체는 화면 밖이므로 발의 접촉은 확인할 수 없지만, 떠 있거나 지지 없이 들린 신체는 보이지 않는다. 미연의 무동작 신체는 화면 밖이라 평가할 수 없다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "얼굴 클로즈업과 기존 의상·새벽의 녹슨 잔해는 더 충실하지만, 누운 미연이 아니라 왼쪽 위를 바라보고 눈을 크게 뜬 충격도 약하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "크게 뜬 눈의 충격은 더 분명하지만, 설정에 없는 구명조끼를 추가했고 시선도 아래에 누운 미연을 향하지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 두 눈은 화면 왼쪽의 검은 옷 인물 쪽, 거의 수평보다 약간 위를 향한다. 아래에 누워 있는 미연의 피투성이 얼굴을 내려다보는 시선은 아니다. 미연의 얼굴은 화면 밖이라 실제 시선 도착점은 확인되지 않는다.",
        "built_space": "얼굴과 어깨를 크게 잡았으며, 뒤로 녹슨 금속 패널 한 면과 그 가장자리, 오른쪽의 파란 컨테이너 일부가 보인다. 바다가 어깨 뒤로 직접 보이고 반사 영상은 없다. 참조의 부식된 해상 잔해와 대체로 연결되지만, 배경 잔해가 차지하는 면적은 다소 크다. 왼쪽 인물은 몸통 일부만 보여 쪼그려 앉은 현우의 자세인지 확인할 수 없다. 탈출문은 프레임에 없어 개폐 상태를 판단할 수 없다.",
        "entities": "금발의 약 10세 여자아이 한 명이 중심이며, 얼굴·체격과 회녹색 재킷 및 회색 안쪽 옷은 앰버 참조에 대체로 가깝다. 한국계·백인 혼혈이라는 정체성은 외모만으로 확정할 수 없다. 참조의 흰 마스크는 없다. 왼쪽에는 현우로 해석되는 검은 옷의 몸통 일부만 있고, 미연과 찰리는 보이지 않는다. 눈은 정상적인 해부학적 형태지만 크게 확장된 충격보다는 걱정스러운 표정에 가깝다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앰버의 머리는 목과 몸통에 자연스럽게 이어지고 머리카락과 옷은 중력에 맞게 내려온다. 발과 바닥 접촉은 클로즈업 밖이므로 서 있는 지지 상태를 직접 확인할 수 없지만, 공중에 떠 있다는 증거는 없다. 왼쪽 인물도 하체가 잘려 지지점을 확인할 수 없다. 미연은 보이지 않아 신체의 지지와 무동작 자세는 평가할 수 없다."
       },
       {
        "label": "B",
        "direction": "앰버는 눈을 크게 뜨고 화면 왼쪽의 프레임 밖을 거의 수평으로 바라본다. 누운 미연의 얼굴을 향하는 하향 시선은 보이지 않는다. 왼쪽에서 들어온 손은 앰버의 어깨에 닿아 있으며, 미연을 가리키거나 받치는 동작은 아니다.",
        "built_space": "앰버의 얼굴과 어깨 중심 클로즈업이며, 어깨 뒤로 새벽 바다가 직접 보인다. 아래쪽에는 흐린 녹슨 패널 한 면, 오른쪽 뒤에는 작은 모서리 부속 한 개가 보인다. 반사 영상이나 중복된 고정 설비는 없다. 현우로 해석되는 인물의 팔만 왼쪽에서 들어와 그의 앉은 위치와 자세는 확인되지 않는다. 탈출문과 다른 컨테이너는 프레임 밖이다.",
        "entities": "금발의 약 10세 여자아이로 얼굴과 체격은 앰버 참조에 대체로 가깝고, 자연스러운 눈을 크게 뜬 충격이 분명하다. 혼혈 정체성 자체는 외모만으로 확정할 수 없다. 회녹색 재킷 위에 참조와 지문에 없는 주황색·회색 구명조끼가 추가되었고 흰 마스크는 없다. 왼쪽에는 현우로 해석되는 성인 크기의 손과 검은 소매가 보이지만 얼굴은 없다. 미연과 찰리는 보이지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참조 의상과 지문에 없는 주황색·회색 구명조끼를 앰버의 상체에 큰 소품으로 추가했다."
        ],
        "physics": "왼쪽 손은 소매에서 이어져 앰버의 어깨 위에 자연스럽게 얹혀 있고, 구명조끼는 어깨와 몸통에 걸쳐 지지된다. 머리와 목의 연결 및 머리카락의 처짐도 자연스럽다. 하체는 화면 밖이므로 발의 접촉은 확인할 수 없지만, 떠 있거나 지지 없이 들린 신체는 보이지 않는다. 미연의 무동작 신체는 화면 밖이라 평가할 수 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "얼굴 클로즈업과 기존 의상·새벽의 녹슨 잔해는 더 충실하지만, 누운 미연이 아니라 왼쪽 위를 바라보고 눈을 크게 뜬 충격도 약하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "크게 뜬 눈의 충격은 더 분명하지만, 설정에 없는 구명조끼를 추가했고 시선도 아래에 누운 미연을 향하지 않는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 두 눈은 화면 왼쪽의 검은 옷 인물 쪽, 거의 수평보다 약간 위를 향한다. 아래에 누워 있는 미연의 피투성이 얼굴을 내려다보는 시선은 아니다. 미연의 얼굴은 화면 밖이라 실제 시선 도착점은 확인되지 않는다.",
        "built_space": "얼굴과 어깨를 크게 잡았으며, 뒤로 녹슨 금속 패널 한 면과 그 가장자리, 오른쪽의 파란 컨테이너 일부가 보인다. 바다가 어깨 뒤로 직접 보이고 반사 영상은 없다. 참조의 부식된 해상 잔해와 대체로 연결되지만, 배경 잔해가 차지하는 면적은 다소 크다. 왼쪽 인물은 몸통 일부만 보여 쪼그려 앉은 현우의 자세인지 확인할 수 없다. 탈출문은 프레임에 없어 개폐 상태를 판단할 수 없다.",
        "entities": "금발의 약 10세 여자아이 한 명이 중심이며, 얼굴·체격과 회녹색 재킷 및 회색 안쪽 옷은 앰버 참조에 대체로 가깝다. 한국계·백인 혼혈이라는 정체성은 외모만으로 확정할 수 없다. 참조의 흰 마스크는 없다. 왼쪽에는 현우로 해석되는 검은 옷의 몸통 일부만 있고, 미연과 찰리는 보이지 않는다. 눈은 정상적인 해부학적 형태지만 크게 확장된 충격보다는 걱정스러운 표정에 가깝다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앰버의 머리는 목과 몸통에 자연스럽게 이어지고 머리카락과 옷은 중력에 맞게 내려온다. 발과 바닥 접촉은 클로즈업 밖이므로 서 있는 지지 상태를 직접 확인할 수 없지만, 공중에 떠 있다는 증거는 없다. 왼쪽 인물도 하체가 잘려 지지점을 확인할 수 없다. 미연은 보이지 않아 신체의 지지와 무동작 자세는 평가할 수 없다."
       },
       {
        "label": "A",
        "direction": "앰버는 눈을 크게 뜨고 화면 왼쪽의 프레임 밖을 거의 수평으로 바라본다. 누운 미연의 얼굴을 향하는 하향 시선은 보이지 않는다. 왼쪽에서 들어온 손은 앰버의 어깨에 닿아 있으며, 미연을 가리키거나 받치는 동작은 아니다.",
        "built_space": "앰버의 얼굴과 어깨 중심 클로즈업이며, 어깨 뒤로 새벽 바다가 직접 보인다. 아래쪽에는 흐린 녹슨 패널 한 면, 오른쪽 뒤에는 작은 모서리 부속 한 개가 보인다. 반사 영상이나 중복된 고정 설비는 없다. 현우로 해석되는 인물의 팔만 왼쪽에서 들어와 그의 앉은 위치와 자세는 확인되지 않는다. 탈출문과 다른 컨테이너는 프레임 밖이다.",
        "entities": "금발의 약 10세 여자아이로 얼굴과 체격은 앰버 참조에 대체로 가깝고, 자연스러운 눈을 크게 뜬 충격이 분명하다. 혼혈 정체성 자체는 외모만으로 확정할 수 없다. 회녹색 재킷 위에 참조와 지문에 없는 주황색·회색 구명조끼가 추가되었고 흰 마스크는 없다. 왼쪽에는 현우로 해석되는 성인 크기의 손과 검은 소매가 보이지만 얼굴은 없다. 미연과 찰리는 보이지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참조 의상과 지문에 없는 주황색·회색 구명조끼를 앰버의 상체에 큰 소품으로 추가했다."
        ],
        "physics": "왼쪽 손은 소매에서 이어져 앰버의 어깨 위에 자연스럽게 얹혀 있고, 구명조끼는 어깨와 몸통에 걸쳐 지지된다. 머리와 목의 연결 및 머리카락의 처짐도 자연스럽다. 하체는 화면 밖이므로 발의 접촉은 확인할 수 없지만, 떠 있거나 지지 없이 들린 신체는 보이지 않는다. 미연의 무동작 신체는 화면 밖이라 평가할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "totals": {
   "B": 5,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "얼굴 클로즈업과 기존 의상·새벽의 녹슨 잔해는 더 충실하지만, 누운 미연이 아니라 왼쪽 위를 바라보고 눈을 크게 뜬 충격도 약하다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "크게 뜬 눈의 충격은 더 분명하지만, 설정에 없는 구명조끼를 추가했고 시선도 아래에 누운 미연을 향하지 않는다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S38sh6_sel.png",
    "asset_id": "f051c526-60ed-46e6-b08d-034d2e664e3f",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdbf-dedf-7559-8b13-e0ac141ca135",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S38sh6"
  }
 },
 "S38sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:36:45.938995+00:00",
  "fingerprint": "f1c224680d259ec80aaefcb77a00c7e05c77a450690d6ad51b597d2f33214089",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S38sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S38sh12_sel.png",
  "source_sha256": "34d2f6d78daa7a7b68f933af42d70a15a0c3c0f3131e88717d97959337ed1b4e",
  "file": "S38sh12_cine.png",
  "staged_sha256": "baa7ebfe3dcb248735560445368309b16d90e653daefbdd1d0250a41d2bb7a83",
  "latency_ms": 15494
 },
 "S39sh1::signage": {
  "fp": "f7b4b53d881f888d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S39sh1": {
  "input_fingerprint": "23de3882665b903b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 바닷물 아래로 완전히 수몰되어 지붕과 쓰레기 더미만 곳곳에 떠 있는 난민촌의 참혹한 아침 전경.\n\nLOCATION (lock): Over the open waters covering the former refugee settlement, with only rooftops and floating rubbish visible in morning daylight. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Submerged refugee settlement (Completely inundated) — Its remaining layout is viewed steeply from above; used as Establishes the scale of destruction at the start of the aerial passage; Roof portions (Remaining visible amid the flooded settlement) — Their upper surfaces appear in scattered, small patches beneath the camera; used as Supply settlement-scale references without an oversized foreground structure; Floating rubbish (Scattered across the seawater) — Distributed irregularly in the overhead view; used as Breaks up the expanses of water and conveys the aftermath without geometric repetition; Seawater covering the settlement (Rough water covering the former inhabited area) — Viewed directly downward through the drone camera; used as Connects the scattered remnants into one continuous flooded landscape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination gives the drone image sober, readable contrast across the submerged settlement without importing the later helicopter searchlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sector 7 remains completely submerged in daylight, with floating containers, wreckage, and bodies on the water. Rescue and patrol activity takes place across the inundated settlement.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상); 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 바닷물 아래로 완전히 수몰되어 지붕과 쓰레기 더미만 곳곳에 떠 있는 난민촌의 참혹한 아침 전경.\n\nLOCATION (lock): Over the open waters covering the former refugee settlement, with only rooftops and floating rubbish visible in morning daylight. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Submerged refugee settlement (Completely inundated) — Its remaining layout is viewed steeply from above; used as Establishes the scale of destruction at the start of the aerial passage; Roof portions (Remaining visible amid the flooded settlement) — Their upper surfaces appear in scattered, small patches beneath the camera; used as Supply settlement-scale references without an oversized foreground structure; Floating rubbish (Scattered across the seawater) — Distributed irregularly in the overhead view; used as Breaks up the expanses of water and conveys the aftermath without geometric repetition; Seawater covering the settlement (Rough water covering the former inhabited area) — Viewed directly downward through the drone camera; used as Connects the scattered remnants into one continuous flooded landscape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination gives the drone image sober, readable contrast across the submerged settlement without importing the later helicopter searchlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sector 7 remains completely submerged in daylight, with floating containers, wreckage, and bodies on the water. Rescue and patrol activity takes place across the inundated settlement.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상); 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 바닷물 아래로 완전히 수몰되어 지붕과 쓰레기 더미만 곳곳에 떠 있는 난민촌의 참혹한 아침 전경.\n\nLOCATION (lock): Over the open waters covering the former refugee settlement, with only rooftops and floating rubbish visible in morning daylight. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Submerged refugee settlement (Completely inundated) — Its remaining layout is viewed steeply from above; used as Establishes the scale of destruction at the start of the aerial passage; Roof portions (Remaining visible amid the flooded settlement) — Their upper surfaces appear in scattered, small patches beneath the camera; used as Supply settlement-scale references without an oversized foreground structure; Floating rubbish (Scattered across the seawater) — Distributed irregularly in the overhead view; used as Breaks up the expanses of water and conveys the aftermath without geometric repetition; Seawater covering the settlement (Rough water covering the former inhabited area) — Viewed directly downward through the drone camera; used as Connects the scattered remnants into one continuous flooded landscape.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination gives the drone image sober, readable contrast across the submerged settlement without importing the later helicopter searchlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sector 7 remains completely submerged in daylight, with floating containers, wreckage, and bodies on the water. Rescue and patrol activity takes place across the inundated settlement.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상); 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 수면을 향해 아래로 비스듬히 내려다봄.",
    "built_space": "수몰된 지붕들이 떠 있음. 우측에 레퍼런스와 달리 양쪽에 물이 있는 방파제와 뼈대만 남은 건물 구조물이 세워져 있음.",
    "entities": "수몰된 지붕, 떠다니는 쓰레기 및 컨테이너. 구조용 보트나 인물은 보이지 않음.",
    "hard_violations": [
     "[gemini-pro] 지정된 로케이션 구조물 형태 왜곡 (해안 벽이 아닌 양면이 물에 닿은 부두 형태로 변형)"
    ],
    "physics": "잔해와 쓰레기들이 수면에 떠 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 높은 고도에서 수몰 지역을 넓게 내려다봄.",
    "built_space": "넓은 수몰 지역에 지붕들이 흩어져 있으며, 좌측 상단 배경에 레퍼런스와 유사한 해안 콘크리트 장벽과 육지 구조물이 보임.",
    "entities": "수몰된 판잣집 지붕, 쓰레기 더미, 원경에 다수의 소형 구조/순찰 보트들.",
    "hard_violations": [
     "[gpt-high] 지붕과 쓰레기만 보이는 숏에 여러 보트 탑승자를 추가했습니다. 장면 단위의 구조·순찰 활동 설정은 이 숏에 인물을 추가할 근거가 되지 않습니다."
    ],
    "physics": "보트와 잔해들이 수면에 안정적으로 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 로케이션 구조물의 형태를 잘 유지하였고, 구조 및 순찰 활동(소형 보트)을 화면에 성공적으로 포함시켰습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "레퍼런스의 해안 방파제 구조를 양쪽에 물이 있는 뼈대 형태의 부두로 왜곡했으며, 지시된 구조/순찰 활동이 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 수면을 향해 아래로 비스듬히 내려다봄.",
        "built_space": "수몰된 지붕들이 떠 있음. 우측에 레퍼런스와 달리 양쪽에 물이 있는 방파제와 뼈대만 남은 건물 구조물이 세워져 있음.",
        "entities": "수몰된 지붕, 떠다니는 쓰레기 및 컨테이너. 구조용 보트나 인물은 보이지 않음.",
        "hard_violations": [
         "지정된 로케이션 구조물 형태 왜곡 (해안 벽이 아닌 양면이 물에 닿은 부두 형태로 변형)"
        ],
        "physics": "잔해와 쓰레기들이 수면에 떠 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 높은 고도에서 수몰 지역을 넓게 내려다봄.",
        "built_space": "넓은 수몰 지역에 지붕들이 흩어져 있으며, 좌측 상단 배경에 레퍼런스와 유사한 해안 콘크리트 장벽과 육지 구조물이 보임.",
        "entities": "수몰된 판잣집 지붕, 쓰레기 더미, 원경에 다수의 소형 구조/순찰 보트들.",
        "hard_violations": [],
        "physics": "보트와 잔해들이 수면에 안정적으로 떠 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 로케이션 구조물의 형태를 잘 유지하였고, 구조 및 순찰 활동(소형 보트)을 화면에 성공적으로 포함시켰습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "레퍼런스의 해안 방파제 구조를 양쪽에 물이 있는 뼈대 형태의 부두로 왜곡했으며, 지시된 구조/순찰 활동이 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 수면을 향해 아래로 비스듬히 내려다봄.",
        "built_space": "수몰된 지붕들이 떠 있음. 우측에 레퍼런스와 달리 양쪽에 물이 있는 방파제와 뼈대만 남은 건물 구조물이 세워져 있음.",
        "entities": "수몰된 지붕, 떠다니는 쓰레기 및 컨테이너. 구조용 보트나 인물은 보이지 않음.",
        "hard_violations": [
         "지정된 로케이션 구조물 형태 왜곡 (해안 벽이 아닌 양면이 물에 닿은 부두 형태로 변형)"
        ],
        "physics": "잔해와 쓰레기들이 수면에 떠 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 높은 고도에서 수몰 지역을 넓게 내려다봄.",
        "built_space": "넓은 수몰 지역에 지붕들이 흩어져 있으며, 좌측 상단 배경에 레퍼런스와 유사한 해안 콘크리트 장벽과 육지 구조물이 보임.",
        "entities": "수몰된 판잣집 지붕, 쓰레기 더미, 원경에 다수의 소형 구조/순찰 보트들.",
        "hard_violations": [],
        "physics": "보트와 잔해들이 수면에 안정적으로 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "수몰 규모와 흩어진 지붕은 잘 드러나지만, 숏에 없는 보트 탑승자들을 추가했으며 수평선과 큰 전경 지붕도 수직 하향 구도에 어긋납니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "명확한 인물 추가 없이 참조의 콘크리트 방벽과 수몰 잔해를 보여주지만, 큰 전경 지붕과 노출된 건물군 중심의 사선 조망이라 요청한 하향 항공 전경에는 미달합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 수면을 비스듬히 내려다보며 상단의 수평선까지 향합니다. 요구된 수직 하향 시선은 아닙니다. 먼 보트 탑승자들의 시선은 판별할 수 없고, 무기나 특정 대상을 겨누는 물체는 없습니다. 상단 중앙의 작은 보트 뒤로 옆 방향 항적이 보입니다.",
        "built_space": "화면 전역에 수십 개의 금속판·방수포 지붕과 일부 기둥이 흩어져 있습니다. 왼쪽 상단에는 긴 콘크리트 방벽과 폐건물 줄이 보이며, 참조의 해안 구조물과 재질은 유사합니다. 하단에는 크게 잘린 지붕이 있어 지붕들을 작은 조각으로만 배치하라는 조건과 다릅니다. 대부분의 벽체와 지면은 잠겨 있습니다.",
        "entities": "바닷물, 수몰 지붕, 목재와 플라스틱 쓰레기 더미가 보입니다. 상단 수면에는 약 다섯 척의 작은 보트와 여러 탑승자가 추가되어 있습니다. 사람들의 나이·성별·민족적 정체성은 이 크기에서 확인할 수 없으며, 숏 텍스트는 이들을 등장시키지 않습니다. 낮의 자연광이며 판독 가능한 글자는 없습니다. 물결은 있으나 거친 바다의 표현은 비교적 약합니다.",
        "hard_violations": [
         "지붕과 쓰레기만 보이는 숏에 여러 보트 탑승자를 추가했습니다. 장면 단위의 구조·순찰 활동 설정은 이 숏에 인물을 추가할 근거가 되지 않습니다."
        ],
        "physics": "보트 탑승자들은 선체 안에 있고 보트는 물에 떠 있어 지지 관계가 성립합니다. 쓰레기는 수면에 접촉하고 부력으로 떠 있으며, 지붕은 잠긴 구조물의 상부로 읽힙니다. 공중에 지지 없이 떠 있는 사람이나 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 가까운 지붕 너머 방벽과 먼 도시를 향해 비스듬히 내려다봅니다. 상단에 수평선이 보여 직접 아래를 보는 드론 시점과 다릅니다. 식별 가능한 인물의 시선이나 무기 조준은 없으며, 먼 선박의 진행 방향도 확정하기 어렵습니다.",
        "built_space": "약 여덟 군데의 지붕 잔해가 보이고, 오른쪽에서 상단 중앙으로 긴 콘크리트 방벽 하나가 이어집니다. 방벽에는 원형 구멍과 계단이 있으며, 뒤쪽에는 여러 층이 노출된 폐건물들이 서 있습니다. 이는 참조의 방벽·폐건물 재질과 형태를 이어가지만, 지붕과 쓰레기만 남은 수면보다 수직 건축물이 두드러집니다. 좌하단과 우하단의 큰 지붕도 요구된 작은 지붕 패치 배치와 맞지 않습니다.",
        "entities": "골판 지붕, 파란 방수포, 거친 바닷물, 목재 상자, 병과 플라스틱류, 떠 있는 컨테이너와 선박 잔해가 보입니다. 컨테이너는 지속 상태 설정에 부합합니다. 명확하게 식별되는 사람은 없으며, 구도환이나 페드로를 억지로 넣지 않았습니다. 낮의 자연광과 물결의 포말이 보이고 읽을 수 있는 문구는 없습니다.",
        "hard_violations": [],
        "physics": "쓰레기와 상자들은 수면에 걸쳐 떠 있고, 선박과 컨테이너도 일부가 잠겨 있어 부력에 의한 지지가 표현됩니다. 지붕은 잠긴 건물의 상부로, 방벽과 폐건물은 수중으로 이어지는 고정 구조물로 읽힙니다. 지지 없이 공중에 매달린 물체나 불가능한 인체 자세는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "수몰 규모와 흩어진 지붕은 잘 드러나지만, 숏에 없는 보트 탑승자들을 추가했으며 수평선과 큰 전경 지붕도 수직 하향 구도에 어긋납니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "명확한 인물 추가 없이 참조의 콘크리트 방벽과 수몰 잔해를 보여주지만, 큰 전경 지붕과 노출된 건물군 중심의 사선 조망이라 요청한 하향 항공 전경에는 미달합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 수면을 비스듬히 내려다보며 상단의 수평선까지 향합니다. 요구된 수직 하향 시선은 아닙니다. 먼 보트 탑승자들의 시선은 판별할 수 없고, 무기나 특정 대상을 겨누는 물체는 없습니다. 상단 중앙의 작은 보트 뒤로 옆 방향 항적이 보입니다.",
        "built_space": "화면 전역에 수십 개의 금속판·방수포 지붕과 일부 기둥이 흩어져 있습니다. 왼쪽 상단에는 긴 콘크리트 방벽과 폐건물 줄이 보이며, 참조의 해안 구조물과 재질은 유사합니다. 하단에는 크게 잘린 지붕이 있어 지붕들을 작은 조각으로만 배치하라는 조건과 다릅니다. 대부분의 벽체와 지면은 잠겨 있습니다.",
        "entities": "바닷물, 수몰 지붕, 목재와 플라스틱 쓰레기 더미가 보입니다. 상단 수면에는 약 다섯 척의 작은 보트와 여러 탑승자가 추가되어 있습니다. 사람들의 나이·성별·민족적 정체성은 이 크기에서 확인할 수 없으며, 숏 텍스트는 이들을 등장시키지 않습니다. 낮의 자연광이며 판독 가능한 글자는 없습니다. 물결은 있으나 거친 바다의 표현은 비교적 약합니다.",
        "hard_violations": [
         "지붕과 쓰레기만 보이는 숏에 여러 보트 탑승자를 추가했습니다. 장면 단위의 구조·순찰 활동 설정은 이 숏에 인물을 추가할 근거가 되지 않습니다."
        ],
        "physics": "보트 탑승자들은 선체 안에 있고 보트는 물에 떠 있어 지지 관계가 성립합니다. 쓰레기는 수면에 접촉하고 부력으로 떠 있으며, 지붕은 잠긴 구조물의 상부로 읽힙니다. 공중에 지지 없이 떠 있는 사람이나 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "카메라는 가까운 지붕 너머 방벽과 먼 도시를 향해 비스듬히 내려다봅니다. 상단에 수평선이 보여 직접 아래를 보는 드론 시점과 다릅니다. 식별 가능한 인물의 시선이나 무기 조준은 없으며, 먼 선박의 진행 방향도 확정하기 어렵습니다.",
        "built_space": "약 여덟 군데의 지붕 잔해가 보이고, 오른쪽에서 상단 중앙으로 긴 콘크리트 방벽 하나가 이어집니다. 방벽에는 원형 구멍과 계단이 있으며, 뒤쪽에는 여러 층이 노출된 폐건물들이 서 있습니다. 이는 참조의 방벽·폐건물 재질과 형태를 이어가지만, 지붕과 쓰레기만 남은 수면보다 수직 건축물이 두드러집니다. 좌하단과 우하단의 큰 지붕도 요구된 작은 지붕 패치 배치와 맞지 않습니다.",
        "entities": "골판 지붕, 파란 방수포, 거친 바닷물, 목재 상자, 병과 플라스틱류, 떠 있는 컨테이너와 선박 잔해가 보입니다. 컨테이너는 지속 상태 설정에 부합합니다. 명확하게 식별되는 사람은 없으며, 구도환이나 페드로를 억지로 넣지 않았습니다. 낮의 자연광과 물결의 포말이 보이고 읽을 수 있는 문구는 없습니다.",
        "hard_violations": [],
        "physics": "쓰레기와 상자들은 수면에 걸쳐 떠 있고, 선박과 컨테이너도 일부가 잠겨 있어 부력에 의한 지지가 표현됩니다. 지붕은 잠긴 건물의 상부로, 방벽과 폐건물은 수중으로 이어지는 고정 구조물로 읽힙니다. 지지 없이 공중에 매달린 물체나 불가능한 인체 자세는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 지정된 로케이션 구조물 형태 왜곡 (해안 벽이 아닌 양면이 물에 닿은 부두 형태로 변형)"
    ],
    "B": [
     "[gpt-high] 지붕과 쓰레기만 보이는 숏에 여러 보트 탑승자를 추가했습니다. 장면 단위의 구조·순찰 활동 설정은 이 숏에 인물을 추가할 근거가 되지 않습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1500,
   "A": 1179
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "지정된 로케이션 구조물의 형태를 잘 유지하였고, 구조 및 순찰 활동(소형 보트)을 화면에 성공적으로 포함시켰습니다.  ★위반: [gpt-high] 지붕과 쓰레기만 보이는 숏에 여러 보트 탑승자를 추가했습니다. 장면 단위의 구조·순찰 활동 설정은 이 숏에 인물을 추가할 근거가 되지 않습니다."
   },
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "레퍼런스의 해안 방파제 구조를 양쪽에 물이 있는 뼈대 형태의 부두로 왜곡했으며, 지시된 구조/순찰 활동이 누락되었습니다.  ★위반: [gemini-pro] 지정된 로케이션 구조물 형태 왜곡 (해안 벽이 아닌 양면이 물에 닿은 부두 형태로 변형)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_flooded_settlement_surface_sel.png",
    "asset_id": "6b1f4dda-6ed9-43b9-8740-3d4044f03bc4",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc0-dc59-7dba-afe9-c90a49c65711",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S39sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:37:01.782928+00:00",
  "fingerprint": "72c385dcb032c4581753ef81bd12d1b6a801f6e26f035f34cf232e39ee99a0db",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S39sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S39sh1_sel.png",
  "source_sha256": "b93a17ac3d9ddd523bef40a3da594cc937e02a4172ab49d83424c432e72341df",
  "file": "S39sh1_cine.png",
  "staged_sha256": "9e63a048a638d9dae36240079ded812b1756fb0474fa27042d05e821413843db",
  "latency_ms": 12620
 },
 "S39sh2::signage": {
  "fp": "7488cb73c96a94d6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::1e17550c6588005e": {
  "subjects": [],
  "subject_text": "수몰된 인천 난민촌 해상과 떠다니는 잔해 위\n넓은 해수면 위에 컨테이너 판자와 각종 잔해가 떠 있는 공간. 수면 위로 담장 일부가 드러나며 새벽빛이 물결에 비친다.",
  "identity": "canonical",
  "scope_id": "L47",
  "scope_role": "location_exterior",
  "scope_sha": "aa1023f853618d1a"
 },
 "groupbg::submerged_camp_panorama": {
  "input_fingerprint": "a74c8d95f1411065",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "submerged_camp_panorama",
    "tags": [
     "S39sh1",
     "S39sh2"
    ]
   },
   "context_sig": "3e399175e9d1127d"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the open water surface above the drowned refugee settlement, beside small drifting wooden fragments.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n수몰된 인천 난민촌 해상과 떠다니는 잔해 위: 동틀 무렵, 건물 꼭대기만 남고 완전히 물에 잠긴 도시 위에 컨테이너와 판자 조각들이 둥둥 떠다니는 수면. (특징: 수면에 떠 있는 나무 판자와 잔해들; 물 위를 비추는 헬기의 서치라이트; 잔해 사이를 수색하는 고무보트)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 중계용 드론이 부감으로 바다를 훑는다.\n- 수몰된 난민촌의 풍경. 페드로와 구도환 등도 둥둥 떠다니고....\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the open water surface above the drowned refugee settlement, beside small drifting wooden fragments.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n수몰된 인천 난민촌 해상과 떠다니는 잔해 위: 동틀 무렵, 건물 꼭대기만 남고 완전히 물에 잠긴 도시 위에 컨테이너와 판자 조각들이 둥둥 떠다니는 수면. (특징: 수면에 떠 있는 나무 판자와 잔해들; 물 위를 비추는 헬기의 서치라이트; 잔해 사이를 수색하는 고무보트)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 중계용 드론이 부감으로 바다를 훑는다.\n- 수몰된 난민촌의 풍경. 페드로와 구도환 등도 둥둥 떠다니고....\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_submerged_camp_panorama_63b629.png",
  "asset_id": "a17edad1-a6bc-48ed-8f40-9c438ee35b8b",
  "input_asset_ids": [
   "07af7eae-d00a-466b-bf36-56e945935fc1"
  ],
  "origin_tag": "S39sh2",
  "place_text": "At the open water surface above the drowned refugee settlement, beside small drifting wooden fragments.",
  "origin_inputs": {
   "place_text": "At the open water surface above the drowned refugee settlement, beside small drifting wooden fragments.",
   "time_of_day_en": "day",
   "conti_asset_id": "07af7eae-d00a-466b-bf36-56e945935fc1"
  }
 },
 "S39sh2::bgfirst_bg": {
  "input_fingerprint": "9c0d076540930a7c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 물에 잠긴 난민촌 바다 위, 페드로와 구도환이 낡은 널빤지 조각에 간신히 매달려 탈진한 채 눈을 감은 위태로운 구도.\n\nLOCATION (lock): At the open water surface above the drowned refugee settlement, beside small drifting wooden fragments.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Sea over the submerged refugee camp (The refugee camp is inundated; both people float in the water); used as Open water surrounds the pair and establishes their lack of secure footing; Old board fragment (Floating, with both people barely holding on) — Its upper face and near edge are visible diagonally between their hands; used as A small shared support connects the two otherwise drifting bodies.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps both closed faces legible within a restrained, somber broadcast image.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 물에 잠긴 난민촌 바다 위, 페드로와 구도환이 낡은 널빤지 조각에 간신히 매달려 탈진한 채 눈을 감은 위태로운 구도.\n\nLOCATION (lock): At the open water surface above the drowned refugee settlement, beside small drifting wooden fragments.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Sea over the submerged refugee camp (The refugee camp is inundated; both people float in the water); used as Open water surrounds the pair and establishes their lack of secure footing; Old board fragment (Floating, with both people barely holding on) — Its upper face and near edge are visible diagonally between their hands; used as A small shared support connects the two otherwise drifting bodies.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps both closed faces legible within a restrained, somber broadcast image.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S39sh2__bgfirst_bg.png",
  "asset_id": "b3c53239-1cd3-4403-80a5-e3d2bd448cbd",
  "input_asset_ids": [
   "07af7eae-d00a-466b-bf36-56e945935fc1",
   "a17edad1-a6bc-48ed-8f40-9c438ee35b8b"
  ]
 },
 "S39sh2": {
  "input_fingerprint": "2fc7141da5266b58",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물에 잠긴 난민촌 바다 위, 페드로와 구도환이 낡은 널빤지 조각에 간신히 매달려 탈진한 채 눈을 감은 위태로운 구도.\n\nLOCATION (lock): At the open water surface above the drowned refugee settlement, beside small drifting wooden fragments. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Sea over the submerged refugee camp (The refugee camp is inundated; both people float in the water); used as Open water surrounds the pair and establishes their lack of secure footing; Old board fragment (Floating, with both people barely holding on) — Its upper face and near edge are visible diagonally between their hands; used as A small shared support connects the two otherwise drifting bodies.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps both closed faces legible within a restrained, somber broadcast image.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Pedro floats exhausted in the floodwater with his eyes closed, clinging to a fragment of wooden planking that supports him. His exact grip, head angle, torso orientation, and submerged leg positions are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Dohwan floats exhausted in the floodwater with his eyes closed, clinging to a fragment of wooden planking that supports him. His exact grip, head angle, torso orientation, and submerged leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement remains submerged beneath seawater, with containers, debris, and bodies floating across it. 페드로: He is still floating in the flooded settlement. 구도환: He is still floating in the flooded settlement; his exact grip, head angle, torso orientation, and submerged leg positions are unspecified.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 페드로와 구도환 right now, so 페드로와 구도환's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 페드로와 구도환: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴); 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물에 잠긴 난민촌 바다 위, 페드로와 구도환이 낡은 널빤지 조각에 간신히 매달려 탈진한 채 눈을 감은 위태로운 구도.\n\nLOCATION (lock): At the open water surface above the drowned refugee settlement, beside small drifting wooden fragments. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Sea over the submerged refugee camp (The refugee camp is inundated; both people float in the water); used as Open water surrounds the pair and establishes their lack of secure footing; Old board fragment (Floating, with both people barely holding on) — Its upper face and near edge are visible diagonally between their hands; used as A small shared support connects the two otherwise drifting bodies.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps both closed faces legible within a restrained, somber broadcast image.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Pedro floats exhausted in the floodwater with his eyes closed, clinging to a fragment of wooden planking that supports him. His exact grip, head angle, torso orientation, and submerged leg positions are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Dohwan floats exhausted in the floodwater with his eyes closed, clinging to a fragment of wooden planking that supports him. His exact grip, head angle, torso orientation, and submerged leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement remains submerged beneath seawater, with containers, debris, and bodies floating across it. 페드로: He is still floating in the flooded settlement. 구도환: He is still floating in the flooded settlement; his exact grip, head angle, torso orientation, and submerged leg positions are unspecified.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 페드로와 구도환 right now, so 페드로와 구도환's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 페드로와 구도환: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴); 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물에 잠긴 난민촌 바다 위, 페드로와 구도환이 낡은 널빤지 조각에 간신히 매달려 탈진한 채 눈을 감은 위태로운 구도.\n\nLOCATION (lock): At the open water surface above the drowned refugee settlement, beside small drifting wooden fragments. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Sea over the submerged refugee camp (The refugee camp is inundated; both people float in the water); used as Open water surrounds the pair and establishes their lack of secure footing; Old board fragment (Floating, with both people barely holding on) — Its upper face and near edge are visible diagonally between their hands; used as A small shared support connects the two otherwise drifting bodies.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps both closed faces legible within a restrained, somber broadcast image.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Pedro floats exhausted in the floodwater with his eyes closed, clinging to a fragment of wooden planking that supports him. His exact grip, head angle, torso orientation, and submerged leg positions are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Dohwan floats exhausted in the floodwater with his eyes closed, clinging to a fragment of wooden planking that supports him. His exact grip, head angle, torso orientation, and submerged leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement remains submerged beneath seawater, with containers, debris, and bodies floating across it. 페드로: He is still floating in the flooded settlement. 구도환: He is still floating in the flooded settlement; his exact grip, head angle, torso orientation, and submerged leg positions are unspecified.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 페드로와 구도환 right now, so 페드로와 구도환's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 페드로와 구도환: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 얼굴, 18세 남성, 청소년의 젊은 얼굴); 구도환 (50대 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S39sh2__bgfirst_bg.png",
     "asset_id": "b3c53239-1cd3-4403-80a5-e3d2bd448cbd",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S39sh2.png",
     "asset_id": "07af7eae-d00a-466b-bf36-56e945935fc1",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:826925>",
     "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929233>",
     "asset_id": "1297d7df-4792-4e39-bdf8-1c0f5b59585c",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_submerged_camp_panorama_63b629.png",
     "asset_id": "a17edad1-a6bc-48ed-8f40-9c438ee35b8b",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:826925>",
     "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929233>",
     "asset_id": "1297d7df-4792-4e39-bdf8-1c0f5b59585c",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 눈을 감은 상태로, 얼굴과 시선 방향이 아래쪽의 널빤지를 향하고 있음.",
    "built_space": "수몰된 난민촌 바다 위 배경. 원경의 도시 스카이라인과 물에 잠긴 건물 지붕 등 제시된 위치 레퍼런스의 공간적 특성을 정확히 반영함.",
    "entities": "널빤지와 주변 부유물은 잘 묘사되었으나, 페드로와 구도환 모두 제공된 캐릭터 레퍼런스의 외모와 의상(비니, 가죽 재킷, 페도라, 트렌치코트 등)을 완전히 누락하고 임의의 모습과 복장으로 등장함.",
    "hard_violations": [],
    "physics": "두 인물은 물과 널빤지에 의지해 떠 있으며, 손으로 널빤지를 단단히 쥐고 팔에 머리를 기댄 채 중력에 맞게 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 눈을 감고 있음. 구도환은 하늘을 향해 위를 보고, 페드로는 널빤지 쪽으로 얼굴을 숙임.",
    "built_space": "수몰된 난민촌 바다 위 배경. 물에 잠긴 구조물과 드럼통 등의 부유물들이 위치 레퍼런스와 정확히 일치하며 올바르게 배치됨.",
    "entities": "페드로(비니, 가죽재킷)와 구도환(페도라, 트렌치코트, 수염) 모두 캐릭터 레퍼런스의 외모와 의상 요소를 완벽히 재현함.",
    "hard_violations": [],
    "physics": "페드로는 상체를 널빤지에 기댄 채 떠 있고, 구도환은 널빤지를 잡지 않고 물 위에 등을 대고 떠 있음. 두 몸 모두 지지체(물, 널빤지)에 기대어 중력 지침을 준수함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "캐릭터 레퍼런스(외모 및 의상)를 전혀 반영하지 못했으나, 두 인물이 널빤지에 매달려 있고 손이 프레임에 노출되어야 하는 최우선 액션 및 프레이밍 지침을 충실히 따랐습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "캐릭터의 외모와 지정된 의상은 완벽하게 구현했으나, 구도환이 널빤지를 붙잡지 않고 떠 있으며 두 사람의 손이 화면에 보이지 않아 주요 액션 및 프레이밍 지침을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 눈을 감은 상태로, 얼굴과 시선 방향이 아래쪽의 널빤지를 향하고 있음.",
        "built_space": "수몰된 난민촌 바다 위 배경. 원경의 도시 스카이라인과 물에 잠긴 건물 지붕 등 제시된 위치 레퍼런스의 공간적 특성을 정확히 반영함.",
        "entities": "널빤지와 주변 부유물은 잘 묘사되었으나, 페드로와 구도환 모두 제공된 캐릭터 레퍼런스의 외모와 의상(비니, 가죽 재킷, 페도라, 트렌치코트 등)을 완전히 누락하고 임의의 모습과 복장으로 등장함.",
        "hard_violations": [],
        "physics": "두 인물은 물과 널빤지에 의지해 떠 있으며, 손으로 널빤지를 단단히 쥐고 팔에 머리를 기댄 채 중력에 맞게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 눈을 감고 있음. 구도환은 하늘을 향해 위를 보고, 페드로는 널빤지 쪽으로 얼굴을 숙임.",
        "built_space": "수몰된 난민촌 바다 위 배경. 물에 잠긴 구조물과 드럼통 등의 부유물들이 위치 레퍼런스와 정확히 일치하며 올바르게 배치됨.",
        "entities": "페드로(비니, 가죽재킷)와 구도환(페도라, 트렌치코트, 수염) 모두 캐릭터 레퍼런스의 외모와 의상 요소를 완벽히 재현함.",
        "hard_violations": [],
        "physics": "페드로는 상체를 널빤지에 기댄 채 떠 있고, 구도환은 널빤지를 잡지 않고 물 위에 등을 대고 떠 있음. 두 몸 모두 지지체(물, 널빤지)에 기대어 중력 지침을 준수함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "캐릭터 레퍼런스(외모 및 의상)를 전혀 반영하지 못했으나, 두 인물이 널빤지에 매달려 있고 손이 프레임에 노출되어야 하는 최우선 액션 및 프레이밍 지침을 충실히 따랐습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "캐릭터의 외모와 지정된 의상은 완벽하게 구현했으나, 구도환이 널빤지를 붙잡지 않고 떠 있으며 두 사람의 손이 화면에 보이지 않아 주요 액션 및 프레이밍 지침을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 눈을 감은 상태로, 얼굴과 시선 방향이 아래쪽의 널빤지를 향하고 있음.",
        "built_space": "수몰된 난민촌 바다 위 배경. 원경의 도시 스카이라인과 물에 잠긴 건물 지붕 등 제시된 위치 레퍼런스의 공간적 특성을 정확히 반영함.",
        "entities": "널빤지와 주변 부유물은 잘 묘사되었으나, 페드로와 구도환 모두 제공된 캐릭터 레퍼런스의 외모와 의상(비니, 가죽 재킷, 페도라, 트렌치코트 등)을 완전히 누락하고 임의의 모습과 복장으로 등장함.",
        "hard_violations": [],
        "physics": "두 인물은 물과 널빤지에 의지해 떠 있으며, 손으로 널빤지를 단단히 쥐고 팔에 머리를 기댄 채 중력에 맞게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 눈을 감고 있음. 구도환은 하늘을 향해 위를 보고, 페드로는 널빤지 쪽으로 얼굴을 숙임.",
        "built_space": "수몰된 난민촌 바다 위 배경. 물에 잠긴 구조물과 드럼통 등의 부유물들이 위치 레퍼런스와 정확히 일치하며 올바르게 배치됨.",
        "entities": "페드로(비니, 가죽재킷)와 구도환(페도라, 트렌치코트, 수염) 모두 캐릭터 레퍼런스의 외모와 의상 요소를 완벽히 재현함.",
        "hard_violations": [],
        "physics": "페드로는 상체를 널빤지에 기댄 채 떠 있고, 구도환은 널빤지를 잡지 않고 물 위에 등을 대고 떠 있음. 두 몸 모두 지지체(물, 널빤지)에 기대어 중력 지침을 준수함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "넓은 수면과 참조 의상은 잘 재현했지만, 두 사람이 작은 널빤지를 간신히 붙잡기보다 기대어 누워 있어 핵심 행동과 손의 접촉이 불명확하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "인물의 얼굴·머리·의상 일치는 떨어지지만, 물에 잠긴 두 사람이 눈을 감고 같은 널빤지에 손을 얹어 매달리는 핵심 순간을 더 충실히 보여준다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 구도환은 얼굴을 위로 젖히고, 오른쪽 페드로는 얼굴을 왼쪽 아래 널빤지 쪽으로 기울인다. 둘 다 눈을 감아 시선의 대상은 없다. 몸은 공유 널빤지에서 각각 왼쪽과 오른쪽으로 뻗어 있으며, 판자의 윗면과 가까운 가장자리는 화면 오른쪽 아래로 비스듬히 보인다.",
        "built_space": "수면을 넓게 담은 와이드 숏이다. 왼쪽 전경의 작은 삼각형 구조물 하나, 먼 왼쪽과 오른쪽의 큰 지붕 두 개, 오른쪽의 녹슨 드럼통 하나, 여러 상자형 잔해와 먼 도시 윤곽이 장소 참조에 대응한다. 두 사람은 중앙 아래의 같은 목재 잔해 양쪽에 있고 안전한 육지는 없다. 공유 지지물은 작은 단일 널빤지보다는 여러 판자가 붙은 비교적 넓은 파편으로 보인다.",
        "entities": "주요 인물은 젊은 남성과 중년 남성 두 명이다. 페드로의 젊은 얼굴, 검은 비니, 낡은 검은 가죽 재킷과 찢어진 바지는 참조와 가깝다. 구도환의 중년 얼굴, 회색 수염, 챙 있는 모자와 베이지색 외투도 참조를 따른다. 페드로의 얼굴은 작고 기울어져 있어 지정된 라틴계 혼혈 인상의 세부 일치는 확인하기 어렵다. 낡은 목재와 해수, 떠다니는 잔해가 보이며 읽을 수 있는 글자는 없다. 두 사람의 손이 판자를 붙잡는 모습은 명확히 드러나지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 하체는 물에 잠겨 부력을 받고, 구도환의 등과 어깨 및 페드로의 머리와 상체는 목재 가장자리에 기대어 있다. 공중에 지지 없이 떠 있는 몸은 아니다. 다만 판자에 매달리는 손의 지지가 불명확하고, 구도환의 뒤로 젖힌 머리는 받쳐진 접점이 잘 보이지 않는다. 전체 자세는 위태롭게 붙잡는 동작보다 떠 있는 잔해에 기대어 쉬는 모습에 가깝다."
       },
       {
        "label": "B",
        "direction": "왼쪽 중년 남성과 오른쪽 젊은 남성 모두 눈을 감고 얼굴을 손과 널빤지 쪽으로 숙인다. 상대나 카메라를 바라보지 않는다. 두 사람의 팔은 동일한 판자로 향하며 손이 윗면과 가장자리에 닿는다. 판자의 윗면과 가까운 모서리는 왼쪽 위에서 오른쪽 아래로 대각선을 이룬다.",
        "built_space": "두 사람이 A보다 크게 나오지만 주변 수면, 먼 지붕과 도시를 충분히 포함하는 환경 중심의 와이드 구도다. 왼쪽 전경의 삼각형 구조물 하나, 먼 큰 지붕 두 개, 오른쪽 드럼통 하나와 상자형 잔해들이 장소 참조에 대응한다. 두 사람은 널빤지 뒤쪽 수면에 나란히 잠겨 있고, 앞에는 다른 목재 파편이 맞닿아 있다. 서 있을 바닥이나 안전한 발판은 보이지 않는다.",
        "entities": "젊은 남성과 중년 남성 두 명만 주요 인물로 보이며 둘 다 눈을 감았다. 젊은 남성은 짙은 곱슬머리와 검은 반소매 차림으로, 페드로 참조의 비니와 가죽 재킷이 없고 얼굴도 다르다. 중년 남성은 짧은 회색 섞인 머리와 베이지색 겉옷을 입었으나 구도환 참조의 모자, 수염 및 얼굴형과 차이가 있다. 연령대와 남성 외형은 대체로 맞지만 정확한 인물 정체성 재현은 약하다. 두 사람의 손과 손목이 공유 목재에 닿아 있으며 해수와 낡은 나무의 질감은 실물처럼 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몸통과 하체는 수중 부력을 받고, 드러난 팔뚝과 손은 떠 있는 판자에 놓이거나 가장자리를 감싼다. 목재 자체도 수면에 잠긴 채 떠 있어 지지 관계가 성립한다. 머리는 앞으로 숙였지만 얼굴 전체가 판자에 완전히 얹혀 있지는 않아 완전한 무긴장 자세의 표현은 다소 약하다. 그래도 지지 없이 공중에 뜬 신체나 물체는 없고, 탈진한 상태로 판자에 매달리는 행동은 분명하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "넓은 수면과 참조 의상은 잘 재현했지만, 두 사람이 작은 널빤지를 간신히 붙잡기보다 기대어 누워 있어 핵심 행동과 손의 접촉이 불명확하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "인물의 얼굴·머리·의상 일치는 떨어지지만, 물에 잠긴 두 사람이 눈을 감고 같은 널빤지에 손을 얹어 매달리는 핵심 순간을 더 충실히 보여준다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 구도환은 얼굴을 위로 젖히고, 오른쪽 페드로는 얼굴을 왼쪽 아래 널빤지 쪽으로 기울인다. 둘 다 눈을 감아 시선의 대상은 없다. 몸은 공유 널빤지에서 각각 왼쪽과 오른쪽으로 뻗어 있으며, 판자의 윗면과 가까운 가장자리는 화면 오른쪽 아래로 비스듬히 보인다.",
        "built_space": "수면을 넓게 담은 와이드 숏이다. 왼쪽 전경의 작은 삼각형 구조물 하나, 먼 왼쪽과 오른쪽의 큰 지붕 두 개, 오른쪽의 녹슨 드럼통 하나, 여러 상자형 잔해와 먼 도시 윤곽이 장소 참조에 대응한다. 두 사람은 중앙 아래의 같은 목재 잔해 양쪽에 있고 안전한 육지는 없다. 공유 지지물은 작은 단일 널빤지보다는 여러 판자가 붙은 비교적 넓은 파편으로 보인다.",
        "entities": "주요 인물은 젊은 남성과 중년 남성 두 명이다. 페드로의 젊은 얼굴, 검은 비니, 낡은 검은 가죽 재킷과 찢어진 바지는 참조와 가깝다. 구도환의 중년 얼굴, 회색 수염, 챙 있는 모자와 베이지색 외투도 참조를 따른다. 페드로의 얼굴은 작고 기울어져 있어 지정된 라틴계 혼혈 인상의 세부 일치는 확인하기 어렵다. 낡은 목재와 해수, 떠다니는 잔해가 보이며 읽을 수 있는 글자는 없다. 두 사람의 손이 판자를 붙잡는 모습은 명확히 드러나지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 하체는 물에 잠겨 부력을 받고, 구도환의 등과 어깨 및 페드로의 머리와 상체는 목재 가장자리에 기대어 있다. 공중에 지지 없이 떠 있는 몸은 아니다. 다만 판자에 매달리는 손의 지지가 불명확하고, 구도환의 뒤로 젖힌 머리는 받쳐진 접점이 잘 보이지 않는다. 전체 자세는 위태롭게 붙잡는 동작보다 떠 있는 잔해에 기대어 쉬는 모습에 가깝다."
       },
       {
        "label": "A",
        "direction": "왼쪽 중년 남성과 오른쪽 젊은 남성 모두 눈을 감고 얼굴을 손과 널빤지 쪽으로 숙인다. 상대나 카메라를 바라보지 않는다. 두 사람의 팔은 동일한 판자로 향하며 손이 윗면과 가장자리에 닿는다. 판자의 윗면과 가까운 모서리는 왼쪽 위에서 오른쪽 아래로 대각선을 이룬다.",
        "built_space": "두 사람이 A보다 크게 나오지만 주변 수면, 먼 지붕과 도시를 충분히 포함하는 환경 중심의 와이드 구도다. 왼쪽 전경의 삼각형 구조물 하나, 먼 큰 지붕 두 개, 오른쪽 드럼통 하나와 상자형 잔해들이 장소 참조에 대응한다. 두 사람은 널빤지 뒤쪽 수면에 나란히 잠겨 있고, 앞에는 다른 목재 파편이 맞닿아 있다. 서 있을 바닥이나 안전한 발판은 보이지 않는다.",
        "entities": "젊은 남성과 중년 남성 두 명만 주요 인물로 보이며 둘 다 눈을 감았다. 젊은 남성은 짙은 곱슬머리와 검은 반소매 차림으로, 페드로 참조의 비니와 가죽 재킷이 없고 얼굴도 다르다. 중년 남성은 짧은 회색 섞인 머리와 베이지색 겉옷을 입었으나 구도환 참조의 모자, 수염 및 얼굴형과 차이가 있다. 연령대와 남성 외형은 대체로 맞지만 정확한 인물 정체성 재현은 약하다. 두 사람의 손과 손목이 공유 목재에 닿아 있으며 해수와 낡은 나무의 질감은 실물처럼 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몸통과 하체는 수중 부력을 받고, 드러난 팔뚝과 손은 떠 있는 판자에 놓이거나 가장자리를 감싼다. 목재 자체도 수면에 잠긴 채 떠 있어 지지 관계가 성립한다. 머리는 앞으로 숙였지만 얼굴 전체가 판자에 완전히 얹혀 있지는 않아 완전한 무긴장 자세의 표현은 다소 약하다. 그래도 지지 없이 공중에 뜬 신체나 물체는 없고, 탈진한 상태로 판자에 매달리는 행동은 분명하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.675
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.675
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1675
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "캐릭터 레퍼런스(외모 및 의상)를 전혀 반영하지 못했으나, 두 인물이 널빤지에 매달려 있고 손이 프레임에 노출되어야 하는 최우선 액션 및 프레이밍 지침을 충실히 따랐습니다."
   },
   {
    "label": "B",
    "score": 1675,
    "verdict_ko": "캐릭터의 외모와 지정된 의상은 완벽하게 구현했으나, 구도환이 널빤지를 붙잡지 않고 떠 있으며 두 사람의 손이 화면에 보이지 않아 주요 액션 및 프레이밍 지침을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_submerged_camp_panorama_63b629.png",
    "asset_id": "a17edad1-a6bc-48ed-8f40-9c438ee35b8b",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:826925>",
    "asset_id": "cb4ef899-82ca-4ae6-9120-f23ed0f62f4e",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929233>",
    "asset_id": "1297d7df-4792-4e39-bdf8-1c0f5b59585c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc1-abc9-7a02-8550-d81e424e9cd4",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S39sh2__bgfirst_bg.png",
   "bg_asset_id": "b3c53239-1cd3-4403-80a5-e3d2bd448cbd",
   "bg_record_key": "S39sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "submerged_camp_panorama",
   "groupbg_asset_id": "a17edad1-a6bc-48ed-8f40-9c438ee35b8b"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S39sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:37:14.906976+00:00",
  "fingerprint": "a3fd03811fef0461251c0c2463e820bdfb85cddcd4934da193f606887e87152d",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S39sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S39sh2_sel.png",
  "source_sha256": "f67aeba3af85ced71377642f68ed2ec70553149fb2a458d7216629d2eef8f7be",
  "file": "S39sh2_cine.png",
  "staged_sha256": "7bccd509604b89bfc29d222b4fa16c824210410dc86baaadb4a81d1a319e76c9",
  "latency_ms": 11783
 },
 "S40sh2::signage": {
  "fp": "1f2cd279321af47c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::357769839556c95c": {
  "subjects": [],
  "subject_text": "인천 난민촌 외곽 도로와 무인점포 앞\n컨테이너 주거지 외곽에서 도심 방향으로 이어지는 도로. 길 건너에 무인점포 출입구와 매장 전면이 보이는 낮의 거리다.",
  "identity": "canonical",
  "scope_id": "L50",
  "scope_role": "location_exterior",
  "scope_sha": "b3644454255ef758"
 },
 "S40sh2::bgfirst_bg": {
  "input_fingerprint": "d7e05b56c120c44e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 제자리에 우뚝 선 앰버가 눈물범벅이 된 얼굴로 난민촌 쪽을 바라보는 애절한 시선.\n\nLOCATION (lock): On the exposed road at the refugee settlement's outskirts, looking back toward the flooded area in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Route from the camp toward the city (The route on which the group has been running) — Only a soft, partial stretch remains behind her turned shoulders; used as Maintains the movement context without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light preserves the tears and facial tension without a separate dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 제자리에 우뚝 선 앰버가 눈물범벅이 된 얼굴로 난민촌 쪽을 바라보는 애절한 시선.\n\nLOCATION (lock): On the exposed road at the refugee settlement's outskirts, looking back toward the flooded area in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Route from the camp toward the city (The route on which the group has been running) — Only a soft, partial stretch remains behind her turned shoulders; used as Maintains the movement context without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light preserves the tears and facial tension without a separate dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S40sh2__bgfirst_bg.png",
  "asset_id": "c4ab2b7a-28f3-417f-a7c5-72377ed6f9dc",
  "input_asset_ids": [
   "45544ccf-d081-428c-95f0-721b0bf45331",
   "af367c24-34b8-4f48-bf44-df9c7c54d6d7"
  ]
 },
 "S40sh2": {
  "input_fingerprint": "4214e622ce939573",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 제자리에 우뚝 선 앰버가 눈물범벅이 된 얼굴로 난민촌 쪽을 바라보는 애절한 시선.\n\nLOCATION (lock): On the exposed road at the refugee settlement's outskirts, looking back toward the flooded area in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Route from the camp toward the city (The route on which the group has been running) — Only a soft, partial stretch remains behind her turned shoulders; used as Maintains the movement context without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light preserves the tears and facial tension without a separate dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement behind the escape route remains flooded. Charlie accompanies the escape on foot, with his old metal exterior and peeling UBIC chest logo unrepaired. 앰버: She has stopped from exhaustion and keeps looking back. Her face is covered in tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 제자리에 우뚝 선 앰버가 눈물범벅이 된 얼굴로 난민촌 쪽을 바라보는 애절한 시선.\n\nLOCATION (lock): On the exposed road at the refugee settlement's outskirts, looking back toward the flooded area in daylight. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Route from the camp toward the city (The route on which the group has been running) — Only a soft, partial stretch remains behind her turned shoulders; used as Maintains the movement context without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light preserves the tears and facial tension without a separate dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement behind the escape route remains flooded. Charlie accompanies the escape on foot, with his old metal exterior and peeling UBIC chest logo unrepaired. 앰버: She has stopped from exhaustion and keeps looking back. Her face is covered in tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 제자리에 우뚝 선 앰버가 눈물범벅이 된 얼굴로 난민촌 쪽을 바라보는 애절한 시선.\n\nLOCATION (lock): On the exposed road at the refugee settlement's outskirts, looking back toward the flooded area in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Route from the camp toward the city (The route on which the group has been running) — Only a soft, partial stretch remains behind her turned shoulders; used as Maintains the movement context without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light preserves the tears and facial tension without a separate dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement behind the escape route remains flooded. Charlie accompanies the escape on foot, with his old metal exterior and peeling UBIC chest logo unrepaired. 앰버: She has stopped from exhaustion and keeps looking back. Her face is covered in tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S40sh2__bgfirst_bg.png",
     "asset_id": "c4ab2b7a-28f3-417f-a7c5-72377ed6f9dc",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S40sh2.png",
     "asset_id": "45544ccf-d081-428c-95f0-721b0bf45331",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L50B01.png",
     "asset_id": "af367c24-34b8-4f48-bf44-df9c7c54d6d7",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "앰버는 화면 왼쪽 바깥을 바라보고 있으며, 뒤쪽의 로봇은 화면 앞쪽으로 시선과 방향을 향함.",
    "built_space": "도로와 왼쪽의 판잣집은 보이나, 로케이션 레퍼런스의 메인 콘크리트 상점 건물이 없고 다른 초소 건물이 우측에 배치됨.",
    "entities": "앰버는 레퍼런스의 초록 재킷을 입지 않았으며, 샷 텍스트에 언급되지 않은 로봇(찰리)이 프레임에 추가로 등장함. 로봇 가슴에 명확한 'UBIC' 텍스트가 있음.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 없는 인물(로봇 찰리) 추가",
     "[gemini-pro] 읽을 수 있는 텍스트('UBIC') 노출",
     "[gpt-high] 이 숏의 가시적 출연 대상을 앰버로 한정했는데도 찰리로 보이는 인간형 로봇을 뒤에 추가했습니다.",
     "[gpt-high] 로봇 가슴에 'UBIC'가 선명하게 읽혀, 읽을 수 있는 글자와 로고를 금지한 지시를 위반합니다."
    ],
    "physics": "앰버는 서 있고 로봇은 걷는 자세로, 둘 다 지면에 지탱되어 정상적인 물리 법칙을 따름."
   },
   {
    "label": "B",
    "direction": "앰버가 오른쪽 어깨 너머로 카메라 렌즈 쪽을 향해 시선을 던지고 있음.",
    "built_space": "아스팔트 도로와 뒤편의 콘크리트 상점 건물, 좌측의 난민촌 등 로케이션 레퍼런스의 구조와 배치를 정확히 재현함.",
    "entities": "눈물범벅인 얼굴의 앰버가 단독으로 등장하며, 레퍼런스와 일치하는 초록색 재킷을 입고 있음.",
    "hard_violations": [],
    "physics": "도로 위에 안정적으로 두 발을 디디고 서 있는 자연스러운 무게 중심을 보여줌."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷 텍스트에 없는 인물이 등장하고 위치 레퍼런스 및 의상을 무시했으며, 금지된 텍스트 노출까지 포함되어 조건 위반이 큽니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "샷 텍스트가 요구한 인물(앰버)만 단독으로 프레이밍하고, 로케이션 건물과 의상 레퍼런스를 정확히 반영한 우수한 결과물입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 화면 왼쪽 바깥을 바라보고 있으며, 뒤쪽의 로봇은 화면 앞쪽으로 시선과 방향을 향함.",
        "built_space": "도로와 왼쪽의 판잣집은 보이나, 로케이션 레퍼런스의 메인 콘크리트 상점 건물이 없고 다른 초소 건물이 우측에 배치됨.",
        "entities": "앰버는 레퍼런스의 초록 재킷을 입지 않았으며, 샷 텍스트에 언급되지 않은 로봇(찰리)이 프레임에 추가로 등장함. 로봇 가슴에 명확한 'UBIC' 텍스트가 있음.",
        "hard_violations": [
         "샷 텍스트에 없는 인물(로봇 찰리) 추가",
         "읽을 수 있는 텍스트('UBIC') 노출"
        ],
        "physics": "앰버는 서 있고 로봇은 걷는 자세로, 둘 다 지면에 지탱되어 정상적인 물리 법칙을 따름."
       },
       {
        "label": "B",
        "direction": "앰버가 오른쪽 어깨 너머로 카메라 렌즈 쪽을 향해 시선을 던지고 있음.",
        "built_space": "아스팔트 도로와 뒤편의 콘크리트 상점 건물, 좌측의 난민촌 등 로케이션 레퍼런스의 구조와 배치를 정확히 재현함.",
        "entities": "눈물범벅인 얼굴의 앰버가 단독으로 등장하며, 레퍼런스와 일치하는 초록색 재킷을 입고 있음.",
        "hard_violations": [],
        "physics": "도로 위에 안정적으로 두 발을 디디고 서 있는 자연스러운 무게 중심을 보여줌."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷 텍스트에 없는 인물이 등장하고 위치 레퍼런스 및 의상을 무시했으며, 금지된 텍스트 노출까지 포함되어 조건 위반이 큽니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "샷 텍스트가 요구한 인물(앰버)만 단독으로 프레이밍하고, 로케이션 건물과 의상 레퍼런스를 정확히 반영한 우수한 결과물입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 화면 왼쪽 바깥을 바라보고 있으며, 뒤쪽의 로봇은 화면 앞쪽으로 시선과 방향을 향함.",
        "built_space": "도로와 왼쪽의 판잣집은 보이나, 로케이션 레퍼런스의 메인 콘크리트 상점 건물이 없고 다른 초소 건물이 우측에 배치됨.",
        "entities": "앰버는 레퍼런스의 초록 재킷을 입지 않았으며, 샷 텍스트에 언급되지 않은 로봇(찰리)이 프레임에 추가로 등장함. 로봇 가슴에 명확한 'UBIC' 텍스트가 있음.",
        "hard_violations": [
         "샷 텍스트에 없는 인물(로봇 찰리) 추가",
         "읽을 수 있는 텍스트('UBIC') 노출"
        ],
        "physics": "앰버는 서 있고 로봇은 걷는 자세로, 둘 다 지면에 지탱되어 정상적인 물리 법칙을 따름."
       },
       {
        "label": "B",
        "direction": "앰버가 오른쪽 어깨 너머로 카메라 렌즈 쪽을 향해 시선을 던지고 있음.",
        "built_space": "아스팔트 도로와 뒤편의 콘크리트 상점 건물, 좌측의 난민촌 등 로케이션 레퍼런스의 구조와 배치를 정확히 재현함.",
        "entities": "눈물범벅인 얼굴의 앰버가 단독으로 등장하며, 레퍼런스와 일치하는 초록색 재킷을 입고 있음.",
        "hard_violations": [],
        "physics": "도로 위에 안정적으로 두 발을 디디고 서 있는 자연스러운 무게 중심을 보여줌."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "장소와 앰버 단독 출연은 더 충실하지만, 난민촌 대신 카메라 쪽을 보는 시선과 얼굴보다 넓은 상반신 구도가 핵심 지시에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "눈물범벅 얼굴의 클로즈업과 돌아보는 동작은 더 가깝지만, 불필요한 찰리의 등장과 읽히는 가슴 로고가 실격 사유이며 배경도 지나치게 선명합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 몸통과 어깨를 비스듬히 둔 채 얼굴을 카메라 쪽으로 돌리고, 눈도 거의 렌즈를 향합니다. 보이는 난민촌 천막과 진입로는 왼쪽 뒤에 있어 그곳을 바라보는 시선으로 연결되지 않습니다. 무기나 방향을 확인할 휴대 물체는 없습니다.",
        "built_space": "뒤쪽에 단층 콘크리트 상점 한 동, 왼쪽 유리 출입구 한 곳, 가려진 전면 유리창 구획, 옥상 실외기 한 대가 보입니다. 왼쪽 천막길과 오른쪽 도로, 앞쪽 차도·연석·배수구·점자블록의 관계가 장소 사진과 대체로 일치합니다. 앰버는 카메라 쪽 보도 가장자리에 선 것으로 보입니다. 다만 상점 전면과 차도가 화면 상당 부분을 차지해, 어깨 뒤로 부드럽게 남는 짧은 도로 조각이라는 지시보다 배경 범위가 넓습니다. 보이는 길은 건조하며 침수 상태는 확인되지 않습니다.",
        "entities": "인물은 금발에 큰 갈색 눈을 가진 약 10세 여자아이 한 명뿐입니다. 참고 인물과 대체로 유사한 얼굴과 체격이지만 혼혈 정체성 자체는 외형만으로 확정할 수 없습니다. 낡은 연녹색 재킷은 참고와 맞고, 참고의 흰 마스크는 없습니다. 양 볼의 눈물 자국과 방울, 붉어진 눈가와 슬픈 표정이 보입니다. 찰리나 추가 인물, 읽을 수 있는 글자는 보이지 않습니다. 낮의 자연광입니다.",
        "hard_violations": [],
        "physics": "목과 어깨, 몸통의 연결과 뒤돌아본 자세는 자연스럽습니다. 발과 지면 접촉은 프레임 밖이지만 서 있는 상반신으로 읽히며, 공중에 뜬 신체나 물체는 없습니다. 눈물은 볼 표면을 따라 아래로 흐르고 재킷과 머리카락도 중력에 맞게 처집니다. 달리는 동작보다 멈춰 있는 순간에 가깝습니다."
       },
       {
        "label": "B",
        "direction": "앰버는 어깨 너머 화면 왼쪽을 바라봅니다. 그쪽에 천막촌과 침수 지역이 있어 난민촌을 돌아보는 방향은 A보다 잘 연결됩니다. 뒤의 로봇은 머리를 왼쪽 아래로 약간 돌렸지만 구체적인 시선 대상은 확인되지 않습니다.",
        "built_space": "왼쪽에 철망 울타리와 천막 열, 가로등이 달린 전주 한 개 및 뒤로 이어지는 전주들이 있고, 도로는 화면 안쪽으로 길게 뻗습니다. 오른쪽에는 상점 일부와 옥상 실외기 한 대가 보이며, 중앙 뒤에는 넓은 침수 수면이 있습니다. 참고 장소의 요소는 있으나 상점·천막길·차도의 공간 관계가 그대로 유지되는지는 불명확합니다. 앰버는 도로 앞쪽, 로봇은 바로 뒤 오른쪽에 있습니다. 도로와 천막촌이 넓고 선명해 얼굴 뒤에 흐릿한 도로 일부만 남기라는 지시와 맞지 않습니다.",
        "entities": "금발과 큰 갈색 눈을 가진 약 10세 여자아이의 얼굴이 크게 보이고, 볼과 턱까지 이어진 여러 눈물 줄기가 눈물범벅 상태를 분명히 표현합니다. 참고와 유사한 아동 얼굴이지만 혼혈 정체성은 영상만으로 확정할 수 없습니다. 회색 상의는 보이나 참고의 연녹색 재킷과 흰 마스크는 없습니다. 오른쪽에 낡고 도장이 벗겨진 인간형 로봇 한 대가 추가로 등장하며 가슴의 'UBIC' 문자가 읽힙니다. 낮의 자연광과 침수 배경은 구현되어 있습니다.",
        "hard_violations": [
         "이 숏의 가시적 출연 대상을 앰버로 한정했는데도 찰리로 보이는 인간형 로봇을 뒤에 추가했습니다.",
         "로봇 가슴에 'UBIC'가 선명하게 읽혀, 읽을 수 있는 글자와 로고를 금지한 지시를 위반합니다."
        ],
        "physics": "앰버의 목과 몸통은 자연스럽게 이어지고, 몸을 멈춘 채 고개를 돌리는 자세가 가능합니다. 발은 프레임 밖이어서 접지는 직접 확인되지 않지만 부유를 나타내는 단서는 없습니다. 로봇도 다리가 아래 프레임으로 이어지는 직립 자세이며 지지 없는 공중 자세는 아닙니다. 눈물은 피부를 따라 아래로 흐르고 턱에 맺힙니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "장소와 앰버 단독 출연은 더 충실하지만, 난민촌 대신 카메라 쪽을 보는 시선과 얼굴보다 넓은 상반신 구도가 핵심 지시에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "눈물범벅 얼굴의 클로즈업과 돌아보는 동작은 더 가깝지만, 불필요한 찰리의 등장과 읽히는 가슴 로고가 실격 사유이며 배경도 지나치게 선명합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 몸통과 어깨를 비스듬히 둔 채 얼굴을 카메라 쪽으로 돌리고, 눈도 거의 렌즈를 향합니다. 보이는 난민촌 천막과 진입로는 왼쪽 뒤에 있어 그곳을 바라보는 시선으로 연결되지 않습니다. 무기나 방향을 확인할 휴대 물체는 없습니다.",
        "built_space": "뒤쪽에 단층 콘크리트 상점 한 동, 왼쪽 유리 출입구 한 곳, 가려진 전면 유리창 구획, 옥상 실외기 한 대가 보입니다. 왼쪽 천막길과 오른쪽 도로, 앞쪽 차도·연석·배수구·점자블록의 관계가 장소 사진과 대체로 일치합니다. 앰버는 카메라 쪽 보도 가장자리에 선 것으로 보입니다. 다만 상점 전면과 차도가 화면 상당 부분을 차지해, 어깨 뒤로 부드럽게 남는 짧은 도로 조각이라는 지시보다 배경 범위가 넓습니다. 보이는 길은 건조하며 침수 상태는 확인되지 않습니다.",
        "entities": "인물은 금발에 큰 갈색 눈을 가진 약 10세 여자아이 한 명뿐입니다. 참고 인물과 대체로 유사한 얼굴과 체격이지만 혼혈 정체성 자체는 외형만으로 확정할 수 없습니다. 낡은 연녹색 재킷은 참고와 맞고, 참고의 흰 마스크는 없습니다. 양 볼의 눈물 자국과 방울, 붉어진 눈가와 슬픈 표정이 보입니다. 찰리나 추가 인물, 읽을 수 있는 글자는 보이지 않습니다. 낮의 자연광입니다.",
        "hard_violations": [],
        "physics": "목과 어깨, 몸통의 연결과 뒤돌아본 자세는 자연스럽습니다. 발과 지면 접촉은 프레임 밖이지만 서 있는 상반신으로 읽히며, 공중에 뜬 신체나 물체는 없습니다. 눈물은 볼 표면을 따라 아래로 흐르고 재킷과 머리카락도 중력에 맞게 처집니다. 달리는 동작보다 멈춰 있는 순간에 가깝습니다."
       },
       {
        "label": "A",
        "direction": "앰버는 어깨 너머 화면 왼쪽을 바라봅니다. 그쪽에 천막촌과 침수 지역이 있어 난민촌을 돌아보는 방향은 A보다 잘 연결됩니다. 뒤의 로봇은 머리를 왼쪽 아래로 약간 돌렸지만 구체적인 시선 대상은 확인되지 않습니다.",
        "built_space": "왼쪽에 철망 울타리와 천막 열, 가로등이 달린 전주 한 개 및 뒤로 이어지는 전주들이 있고, 도로는 화면 안쪽으로 길게 뻗습니다. 오른쪽에는 상점 일부와 옥상 실외기 한 대가 보이며, 중앙 뒤에는 넓은 침수 수면이 있습니다. 참고 장소의 요소는 있으나 상점·천막길·차도의 공간 관계가 그대로 유지되는지는 불명확합니다. 앰버는 도로 앞쪽, 로봇은 바로 뒤 오른쪽에 있습니다. 도로와 천막촌이 넓고 선명해 얼굴 뒤에 흐릿한 도로 일부만 남기라는 지시와 맞지 않습니다.",
        "entities": "금발과 큰 갈색 눈을 가진 약 10세 여자아이의 얼굴이 크게 보이고, 볼과 턱까지 이어진 여러 눈물 줄기가 눈물범벅 상태를 분명히 표현합니다. 참고와 유사한 아동 얼굴이지만 혼혈 정체성은 영상만으로 확정할 수 없습니다. 회색 상의는 보이나 참고의 연녹색 재킷과 흰 마스크는 없습니다. 오른쪽에 낡고 도장이 벗겨진 인간형 로봇 한 대가 추가로 등장하며 가슴의 'UBIC' 문자가 읽힙니다. 낮의 자연광과 침수 배경은 구현되어 있습니다.",
        "hard_violations": [
         "이 숏의 가시적 출연 대상을 앰버로 한정했는데도 찰리로 보이는 인간형 로봇을 뒤에 추가했습니다.",
         "로봇 가슴에 'UBIC'가 선명하게 읽혀, 읽을 수 있는 글자와 로고를 금지한 지시를 위반합니다."
        ],
        "physics": "앰버의 목과 몸통은 자연스럽게 이어지고, 몸을 멈춘 채 고개를 돌리는 자세가 가능합니다. 발은 프레임 밖이어서 접지는 직접 확인되지 않지만 부유를 나타내는 단서는 없습니다. 로봇도 다리가 아래 프레임으로 이어지는 직립 자세이며 지지 없는 공중 자세는 아닙니다. 눈물은 피부를 따라 아래로 흐르고 턱에 맺힙니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.829,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.579,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 샷 텍스트에 없는 인물(로봇 찰리) 추가",
     "[gemini-pro] 읽을 수 있는 텍스트('UBIC') 노출",
     "[gpt-high] 이 숏의 가시적 출연 대상을 앰버로 한정했는데도 찰리로 보이는 인간형 로봇을 뒤에 추가했습니다.",
     "[gpt-high] 로봇 가슴에 'UBIC'가 선명하게 읽혀, 읽을 수 있는 글자와 로고를 금지한 지시를 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 579,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 579,
    "verdict_ko": "샷 텍스트에 없는 인물이 등장하고 위치 레퍼런스 및 의상을 무시했으며, 금지된 텍스트 노출까지 포함되어 조건 위반이 큽니다.  ★위반: [gemini-pro] 샷 텍스트에 없는 인물(로봇 찰리) 추가 / [gemini-pro] 읽을 수 있는 텍스트('UBIC') 노출 / [gpt-high] 이 숏의 가시적 출연 대상을 앰버로 한정했는데도 찰리로 보이는 인간형 로봇을 뒤에 추가했습니다. / [gpt-high] 로봇 가슴에 'UBIC'가 선명하게 읽혀, 읽을 수 있는 글자와 로고를 금지한 지시를 위반합니다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "샷 텍스트가 요구한 인물(앰버)만 단독으로 프레이밍하고, 로케이션 건물과 의상 레퍼런스를 정확히 반영한 우수한 결과물입니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L50B01.png",
    "asset_id": "af367c24-34b8-4f48-bf44-df9c7c54d6d7",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc2-7067-7653-8267-987dc3936b15",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S40sh2__bgfirst_bg.png",
   "bg_asset_id": "c4ab2b7a-28f3-417f-a7c5-72377ed6f9dc",
   "bg_record_key": "S40sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S40sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:37:27.126908+00:00",
  "fingerprint": "c6cef988c052982ad44734ed3036c3ca3fd8dbc1c1468e00a9c73c9068fbf93e",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S40sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S40sh2_sel.png",
  "source_sha256": "493eab6b8abeb650b8f1487b70664ae66300d9680d98715e450112796a863533",
  "file": "S40sh2_cine.png",
  "staged_sha256": "114392b39243c22a674956a89c8450e7f370b9d85688e653d80a8e4cf259fdcd",
  "latency_ms": 11300
 },
 "S40sh3::signage": {
  "fp": "e0112994982dd542",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S40sh3": {
  "input_fingerprint": "02b61786857cc729",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 미간에 굵은 주름을 잡은 채 앰버의 얇은 팔목을 거칠게 쥐고, 앞을 향해 체중을 실어 당긴 상태로 힘이 들어간 자세.\n\nLOCATION (lock): Along the cityward roadside outside the flooded refugee settlement, before reaching the unattended shop. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Route toward the city (The escape route continues beyond the pair) — The route continues toward screen right; used as Keeps the force of the pull aligned with the already established destination.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daytime ambient treatment, allowing facial tension and the force in their joined arms to carry the urgency.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The flooded settlement remains behind the escape route, and an unmanned shop is visible across a street without police. Charlie continues along the route with his unrepaired old exterior. 앰버: She remains exhausted and tear-streaked as she resumes moving toward the city. 이현우: He is moving forward urgently while avoiding public attention. His facial bruising, leg injury, and shoe-concealed contact card persist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 미간에 굵은 주름을 잡은 채 앰버의 얇은 팔목을 거칠게 쥐고, 앞을 향해 체중을 실어 당긴 상태로 힘이 들어간 자세.\n\nLOCATION (lock): Along the cityward roadside outside the flooded refugee settlement, before reaching the unattended shop. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Route toward the city (The escape route continues beyond the pair) — The route continues toward screen right; used as Keeps the force of the pull aligned with the already established destination.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daytime ambient treatment, allowing facial tension and the force in their joined arms to carry the urgency.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The flooded settlement remains behind the escape route, and an unmanned shop is visible across a street without police. Charlie continues along the route with his unrepaired old exterior. 앰버: She remains exhausted and tear-streaked as she resumes moving toward the city. 이현우: He is moving forward urgently while avoiding public attention. His facial bruising, leg injury, and shoe-concealed contact card persist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 미간에 굵은 주름을 잡은 채 앰버의 얇은 팔목을 거칠게 쥐고, 앞을 향해 체중을 실어 당긴 상태로 힘이 들어간 자세.\n\nLOCATION (lock): Along the cityward roadside outside the flooded refugee settlement, before reaching the unattended shop. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Route toward the city (The escape route continues beyond the pair) — The route continues toward screen right; used as Keeps the force of the pull aligned with the already established destination.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daytime ambient treatment, allowing facial tension and the force in their joined arms to carry the urgency.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The flooded settlement remains behind the escape route, and an unmanned shop is visible across a street without police. Charlie continues along the route with his unrepaired old exterior. 앰버: She remains exhausted and tear-streaked as she resumes moving toward the city. 이현우: He is moving forward urgently while avoiding public attention. His facial bruising, leg injury, and shoe-concealed contact card persist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선과 몸이 화면 우측(도시 방향)을 향하고 있으며, 앰버를 우측으로 강하게 당기고 있음.",
    "built_space": "참조와 동일한 무인 상점이 화면 중앙 뒤편에 위치하며, 우측으로 이어지는 도로와 공간적 스케일이 올바름.",
    "entities": "이현우(미간 주름, 멍든 얼굴, 의상 일치), 앰버(금발, 눈물 자국, 의상 일치) 모두 완벽히 구현됨. 배경 우측에 낡은 외형의 찰리(로봇)가 존재함.",
    "hard_violations": [
     "[gpt-high] 숏 본문이 지정한 이현우와 앰버 외에 완전한 인간형 로봇 한 체를 배경에 추가했다. 이는 이 숏의 등장 대상 제한을 벗어난 제3의 형상이다."
    ],
    "physics": "이현우가 앞발에 체중을 싣고 오른손으로 앰버의 왼쪽 팔목을 단단히 쥐고 당기는 물리적 자세와 지면 접지력이 매우 자연스러움."
   },
   {
    "label": "B",
    "direction": "이현우가 화면 좌측을 돌아보고 있어, 우측으로 향해야 할 당김의 방향과 힘의 축이 모순됨.",
    "built_space": "참조와 동일한 상점과 도로가 배경에 올바르게 배치됨.",
    "entities": "이현우와 앰버의 외형과 복장은 참조와 일치하나, 프롬프트가 요구한 찰리가 화면에서 완전히 누락됨.",
    "hard_violations": [],
    "physics": "이현우가 왼손으로 앰버의 오른팔을 잡고 있으나, 몸통이 꼬이고 체중이 분산되어 강하게 당기는 물리적 힘이 느껴지지 않음. 지면 접촉은 정상적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "화면 우측으로 향하는 체중 이동과 당기는 힘의 방향성을 완벽하게 연출했으며, 명시된 찰리(로봇)의 배경 배치도 충실함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이현우의 시선과 자세가 목적지(우측)가 아닌 뒤쪽을 향해 당기는 방향성이 어긋나며, 찰리가 누락됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 몸이 화면 우측(도시 방향)을 향하고 있으며, 앰버를 우측으로 강하게 당기고 있음.",
        "built_space": "참조와 동일한 무인 상점이 화면 중앙 뒤편에 위치하며, 우측으로 이어지는 도로와 공간적 스케일이 올바름.",
        "entities": "이현우(미간 주름, 멍든 얼굴, 의상 일치), 앰버(금발, 눈물 자국, 의상 일치) 모두 완벽히 구현됨. 배경 우측에 낡은 외형의 찰리(로봇)가 존재함.",
        "hard_violations": [],
        "physics": "이현우가 앞발에 체중을 싣고 오른손으로 앰버의 왼쪽 팔목을 단단히 쥐고 당기는 물리적 자세와 지면 접지력이 매우 자연스러움."
       },
       {
        "label": "B",
        "direction": "이현우가 화면 좌측을 돌아보고 있어, 우측으로 향해야 할 당김의 방향과 힘의 축이 모순됨.",
        "built_space": "참조와 동일한 상점과 도로가 배경에 올바르게 배치됨.",
        "entities": "이현우와 앰버의 외형과 복장은 참조와 일치하나, 프롬프트가 요구한 찰리가 화면에서 완전히 누락됨.",
        "hard_violations": [],
        "physics": "이현우가 왼손으로 앰버의 오른팔을 잡고 있으나, 몸통이 꼬이고 체중이 분산되어 강하게 당기는 물리적 힘이 느껴지지 않음. 지면 접촉은 정상적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "화면 우측으로 향하는 체중 이동과 당기는 힘의 방향성을 완벽하게 연출했으며, 명시된 찰리(로봇)의 배경 배치도 충실함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이현우의 시선과 자세가 목적지(우측)가 아닌 뒤쪽을 향해 당기는 방향성이 어긋나며, 찰리가 누락됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 몸이 화면 우측(도시 방향)을 향하고 있으며, 앰버를 우측으로 강하게 당기고 있음.",
        "built_space": "참조와 동일한 무인 상점이 화면 중앙 뒤편에 위치하며, 우측으로 이어지는 도로와 공간적 스케일이 올바름.",
        "entities": "이현우(미간 주름, 멍든 얼굴, 의상 일치), 앰버(금발, 눈물 자국, 의상 일치) 모두 완벽히 구현됨. 배경 우측에 낡은 외형의 찰리(로봇)가 존재함.",
        "hard_violations": [],
        "physics": "이현우가 앞발에 체중을 싣고 오른손으로 앰버의 왼쪽 팔목을 단단히 쥐고 당기는 물리적 자세와 지면 접지력이 매우 자연스러움."
       },
       {
        "label": "B",
        "direction": "이현우가 화면 좌측을 돌아보고 있어, 우측으로 향해야 할 당김의 방향과 힘의 축이 모순됨.",
        "built_space": "참조와 동일한 상점과 도로가 배경에 올바르게 배치됨.",
        "entities": "이현우와 앰버의 외형과 복장은 참조와 일치하나, 프롬프트가 요구한 찰리가 화면에서 완전히 누락됨.",
        "hard_violations": [],
        "physics": "이현우가 왼손으로 앰버의 오른팔을 잡고 있으나, 몸통이 꼬이고 체중이 분산되어 강하게 당기는 물리적 힘이 느껴지지 않음. 지면 접촉은 정상적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "두 인물만 담은 미디엄 숏과 손목을 움켜쥔 긴장은 충실하지만, 이현우가 오른쪽 도시 방향이 아니라 왼쪽으로 당기는 것이 핵심 오류다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "오른쪽으로 체중을 실어 앰버를 당기는 동작은 맞지만, 이 숏에 허용되지 않은 제3의 로봇 형상을 추가해 탈락한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 화면 왼쪽에 앞서 있고 앰버는 오른쪽에 뒤처져 있으며, 잡힌 팔의 인장 방향은 앰버에게서 왼쪽 이현우 쪽으로 향한다. 이현우의 시선도 화면 왼쪽 바깥을 향하고 앰버는 그를 바라본다. 배경 도로는 오른쪽으로 계속되지만 실제 당김은 지정된 도시행 방향과 반대다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "길 건너 단층 콘크리트 상점 한 채, 전면의 큰 유리창 구획 두 곳, 옥상 상자형 설비 한 개가 보인다. 왼쪽 출입구 쪽은 이현우에게 상당 부분 가려진다. 두 사람은 상점 맞은편의 경계석과 노란 점자블록이 있는 보도에 있다. 오른쪽 도로, 왼쪽 낮은 건물과 전봇대, 낮의 조명은 이전 장소와 대체로 이어진다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "이현우와 앰버로 읽히는 두 사람만 있다. 이현우는 젊은 한국계 남성 외형, 헝클어진 검은 머리, 해진 짙은 후드와 청바지, 얼굴의 멍을 갖췄으며 참조와 대체로 일치한다. 앰버는 약 10세의 금발 소녀로, 큰 눈과 참조의 얼굴 인상, 연녹색 재킷과 회색 원피스, 지치고 눈물 어린 표정을 유지한다. 마스크가 없는 것은 이전 숏과 일치한다. 다리 부상과 신발 속 카드는 프레임 밖이므로 확인할 수 없다. 읽을 수 있는 글씨는 없다.",
        "hard_violations": [],
        "physics": "이현우의 손이 앰버의 손목을 실제로 감싸고 있으며 앰버의 손은 그 아래로 처진다. 연결된 팔과 굽힌 무릎, 낮춘 골반은 당기는 긴장을 만들 수 있는 자세다. 발의 접지는 화면 밖이라 직접 확인되지 않지만, 보이는 하체는 아래로 이어져 있고 공중에 뜬 자세가 아니다. 앰버도 몸통이 세워진 채 팔이 당겨지며, 불가능한 관절이나 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우가 화면 오른쪽에서 앞서 몸을 기울이고 왼쪽의 앰버를 오른쪽으로 끌어당긴다. 이현우의 시선은 오른쪽 앞쪽을, 앰버의 시선은 이현우를 향한다. 당김과 배경 도로의 오른쪽 연장이 지정된 목적지 방향에 부합한다. 뒤쪽 로봇은 정면을 향해 서 있어 도시 방향으로 계속 이동하는지는 분명하지 않다.",
        "built_space": "길 건너 상점 한 채에 왼쪽 유리 출입구 한 곳, 큰 전면 창 구획 두 곳, 일부 가려진 옥상 설비 한 개가 보인다. 두 주인공은 카메라 쪽 길가에 있고 로봇은 그 뒤 차도에 서 있다. 오른쪽 도로와 왼쪽 낮은 건물들은 참조 장소와 대체로 맞지만, 전경은 참조의 포장 보도보다 거친 흙·자갈 면이 크게 나타난다. 중복된 고정 설비나 광학적으로 불가능한 반사는 없다.",
        "entities": "참조에 가까운 검은 머리의 젊은 한국계 남성과 금발 소녀가 보이며, 후드·청바지와 연녹색 재킷·회색 원피스도 유지된다. 이현우의 얼굴 멍과 강하게 찌푸린 미간, 앰버의 지친 표정과 눈물 자국이 보인다. 그러나 뒤에 낡은 금속제 인간형 로봇 한 체가 추가된다. 상태 유지 문구에 찰리가 언급되더라도, 이 숏의 등장 대상을 정하는 본문은 두 주인공만 보여 주며 찰리를 로봇으로 명시하지 않는다. 읽을 수 있는 글씨는 없다.",
        "hard_violations": [
         "숏 본문이 지정한 이현우와 앰버 외에 완전한 인간형 로봇 한 체를 배경에 추가했다. 이는 이 숏의 등장 대상 제한을 벗어난 제3의 형상이다."
        ],
        "physics": "이현우의 두 손이 앰버의 손목과 소매 끝을 잡고 있고, 앰버의 팔은 오른쪽으로 당겨져 있다. 이현우의 벌어진 다리와 오른쪽으로 기울인 상체는 체중을 실은 당김으로 성립한다. 두 사람의 발은 프레임 밖이라 접지를 직접 확인할 수 없으나 공중에 떠 있다는 징후는 없다. 배경 로봇의 두 발은 차도에 닿아 있고 그림자도 바닥에 이어져 있어, 추가 등장 문제와 별개로 지지 자체는 가능하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "두 인물만 담은 미디엄 숏과 손목을 움켜쥔 긴장은 충실하지만, 이현우가 오른쪽 도시 방향이 아니라 왼쪽으로 당기는 것이 핵심 오류다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "오른쪽으로 체중을 실어 앰버를 당기는 동작은 맞지만, 이 숏에 허용되지 않은 제3의 로봇 형상을 추가해 탈락한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 화면 왼쪽에 앞서 있고 앰버는 오른쪽에 뒤처져 있으며, 잡힌 팔의 인장 방향은 앰버에게서 왼쪽 이현우 쪽으로 향한다. 이현우의 시선도 화면 왼쪽 바깥을 향하고 앰버는 그를 바라본다. 배경 도로는 오른쪽으로 계속되지만 실제 당김은 지정된 도시행 방향과 반대다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "길 건너 단층 콘크리트 상점 한 채, 전면의 큰 유리창 구획 두 곳, 옥상 상자형 설비 한 개가 보인다. 왼쪽 출입구 쪽은 이현우에게 상당 부분 가려진다. 두 사람은 상점 맞은편의 경계석과 노란 점자블록이 있는 보도에 있다. 오른쪽 도로, 왼쪽 낮은 건물과 전봇대, 낮의 조명은 이전 장소와 대체로 이어진다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "이현우와 앰버로 읽히는 두 사람만 있다. 이현우는 젊은 한국계 남성 외형, 헝클어진 검은 머리, 해진 짙은 후드와 청바지, 얼굴의 멍을 갖췄으며 참조와 대체로 일치한다. 앰버는 약 10세의 금발 소녀로, 큰 눈과 참조의 얼굴 인상, 연녹색 재킷과 회색 원피스, 지치고 눈물 어린 표정을 유지한다. 마스크가 없는 것은 이전 숏과 일치한다. 다리 부상과 신발 속 카드는 프레임 밖이므로 확인할 수 없다. 읽을 수 있는 글씨는 없다.",
        "hard_violations": [],
        "physics": "이현우의 손이 앰버의 손목을 실제로 감싸고 있으며 앰버의 손은 그 아래로 처진다. 연결된 팔과 굽힌 무릎, 낮춘 골반은 당기는 긴장을 만들 수 있는 자세다. 발의 접지는 화면 밖이라 직접 확인되지 않지만, 보이는 하체는 아래로 이어져 있고 공중에 뜬 자세가 아니다. 앰버도 몸통이 세워진 채 팔이 당겨지며, 불가능한 관절이나 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우가 화면 오른쪽에서 앞서 몸을 기울이고 왼쪽의 앰버를 오른쪽으로 끌어당긴다. 이현우의 시선은 오른쪽 앞쪽을, 앰버의 시선은 이현우를 향한다. 당김과 배경 도로의 오른쪽 연장이 지정된 목적지 방향에 부합한다. 뒤쪽 로봇은 정면을 향해 서 있어 도시 방향으로 계속 이동하는지는 분명하지 않다.",
        "built_space": "길 건너 상점 한 채에 왼쪽 유리 출입구 한 곳, 큰 전면 창 구획 두 곳, 일부 가려진 옥상 설비 한 개가 보인다. 두 주인공은 카메라 쪽 길가에 있고 로봇은 그 뒤 차도에 서 있다. 오른쪽 도로와 왼쪽 낮은 건물들은 참조 장소와 대체로 맞지만, 전경은 참조의 포장 보도보다 거친 흙·자갈 면이 크게 나타난다. 중복된 고정 설비나 광학적으로 불가능한 반사는 없다.",
        "entities": "참조에 가까운 검은 머리의 젊은 한국계 남성과 금발 소녀가 보이며, 후드·청바지와 연녹색 재킷·회색 원피스도 유지된다. 이현우의 얼굴 멍과 강하게 찌푸린 미간, 앰버의 지친 표정과 눈물 자국이 보인다. 그러나 뒤에 낡은 금속제 인간형 로봇 한 체가 추가된다. 상태 유지 문구에 찰리가 언급되더라도, 이 숏의 등장 대상을 정하는 본문은 두 주인공만 보여 주며 찰리를 로봇으로 명시하지 않는다. 읽을 수 있는 글씨는 없다.",
        "hard_violations": [
         "숏 본문이 지정한 이현우와 앰버 외에 완전한 인간형 로봇 한 체를 배경에 추가했다. 이는 이 숏의 등장 대상 제한을 벗어난 제3의 형상이다."
        ],
        "physics": "이현우의 두 손이 앰버의 손목과 소매 끝을 잡고 있고, 앰버의 팔은 오른쪽으로 당겨져 있다. 이현우의 벌어진 다리와 오른쪽으로 기울인 상체는 체중을 실은 당김으로 성립한다. 두 사람의 발은 프레임 밖이라 접지를 직접 확인할 수 없으나 공중에 떠 있다는 징후는 없다. 배경 로봇의 두 발은 차도에 닿아 있고 그림자도 바닥에 이어져 있어, 추가 등장 문제와 별개로 지지 자체는 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.4,
    "B": 1.444
   },
   "adjusted": {
    "A": 1.15,
    "B": 1.444
   },
   "violations": {
    "A": [
     "[gpt-high] 숏 본문이 지정한 이현우와 앰버 외에 완전한 인간형 로봇 한 체를 배경에 추가했다. 이는 이 숏의 등장 대상 제한을 벗어난 제3의 형상이다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1150,
   "B": 1444
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1150,
    "verdict_ko": "화면 우측으로 향하는 체중 이동과 당기는 힘의 방향성을 완벽하게 연출했으며, 명시된 찰리(로봇)의 배경 배치도 충실함.  ★위반: [gpt-high] 숏 본문이 지정한 이현우와 앰버 외에 완전한 인간형 로봇 한 체를 배경에 추가했다. 이는 이 숏의 등장 대상 제한을 벗어난 제3의 형상이다."
   },
   {
    "label": "B",
    "score": 1444,
    "verdict_ko": "이현우의 시선과 자세가 목적지(우측)가 아닌 뒤쪽을 향해 당기는 방향성이 어긋나며, 찰리가 누락됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S40sh2_sel.png",
    "asset_id": "b611158e-a4c8-4110-8908-fcb2b04bfed7",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc3-2c2b-78b6-83f4-e7b63c42e97c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S40sh2"
  }
 },
 "S40sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:37:38.774631+00:00",
  "fingerprint": "2932ed17610be392b59cf8df841062447ac62a5a21b2136f3ead5daac430050d",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S40sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S40sh3_sel.png",
  "source_sha256": "0e68c01ef2802cacdea756ed36660b52f134dacc0c4a579d9af72b63b206a6cb",
  "file": "S40sh3_cine.png",
  "staged_sha256": "e4bd83e77f7e058378e0f05edef001980baed56b1f0a2a87fa8a984dc3fe587d",
  "latency_ms": 12099
 },
 "S41sh11::signage": {
  "fp": "076220fb48a2a275",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S41sh11": {
  "input_fingerprint": "4c8b71f005133850",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 주먹이 현금인출기 전면 패널을 맹렬히 뚫고 들어간 파괴적인 타격의 순간.\n\nLOCATION (lock): Directly in front of the cash machine inside the unattended shop, in daytime retail ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cash machine front panel (Being penetrated by 찰리's fist at the instant of impact) — The front face is seen obliquely, with the entry point unobstructed; used as Provides the impact surface across the left third while retaining spatial context; Shop display shelving (Present behind the action) — A partial aisle-facing section remains visible beyond the arm; used as Supplies an ordinary shop-scale reference without exaggerated foreground perspective.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral interior ambient light gives the metal fist and breached panel clear tactile definition without invented sparks or colored effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A duffel bag and a backpack have been taken from the shelves and loaded with disposable phones, a map, canned food, drinks, instant noodles, and COPD medication. Charlie's fist has penetrated the ATM casing, creating a hole that remains after the punch.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 주먹이 현금인출기 전면 패널을 맹렬히 뚫고 들어간 파괴적인 타격의 순간.\n\nLOCATION (lock): Directly in front of the cash machine inside the unattended shop, in daytime retail ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cash machine front panel (Being penetrated by 찰리's fist at the instant of impact) — The front face is seen obliquely, with the entry point unobstructed; used as Provides the impact surface across the left third while retaining spatial context; Shop display shelving (Present behind the action) — A partial aisle-facing section remains visible beyond the arm; used as Supplies an ordinary shop-scale reference without exaggerated foreground perspective.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral interior ambient light gives the metal fist and breached panel clear tactile definition without invented sparks or colored effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A duffel bag and a backpack have been taken from the shelves and loaded with disposable phones, a map, canned food, drinks, instant noodles, and COPD medication. Charlie's fist has penetrated the ATM casing, creating a hole that remains after the punch.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 주먹이 현금인출기 전면 패널을 맹렬히 뚫고 들어간 파괴적인 타격의 순간.\n\nLOCATION (lock): Directly in front of the cash machine inside the unattended shop, in daytime retail ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Cash machine front panel (Being penetrated by 찰리's fist at the instant of impact) — The front face is seen obliquely, with the entry point unobstructed; used as Provides the impact surface across the left third while retaining spatial context; Shop display shelving (Present behind the action) — A partial aisle-facing section remains visible beyond the arm; used as Supplies an ordinary shop-scale reference without exaggerated foreground perspective.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral interior ambient light gives the metal fist and breached panel clear tactile definition without invented sparks or colored effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A duffel bag and a backpack have been taken from the shelves and loaded with disposable phones, a map, canned food, drinks, instant noodles, and COPD medication. Charlie's fist has penetrated the ATM casing, creating a hole that remains after the punch.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "금속 주먹이 ATM이 아닌 그 옆의 회색 금속 캐비닛 패널을 향해 뻗어 관통하고 있음.",
    "built_space": "왼쪽에 ATM이 존재하고 그 옆에 참조 이미지의 공간에는 없는 회색 캐비닛이 추가로 배치됨. 배경에 진열대가 보임.",
    "entities": "찰리의 기계 팔 외형은 참조와 일치함. 지시된 더플백과 배낭이 팔 주위에 나타남.",
    "hard_violations": [
     "[gemini-pro] 발명된 구조물 (위치 참조에 없는 회색 캐비닛을 추가하여 타격함)",
     "[gemini-pro] 지지대 없이 떠 있는 물체 (배낭과 더플백의 끈이 팔 근처 허공에 떠 있음)"
    ],
    "physics": "배낭과 더플백의 끈이 팔이나 손에 쥐어지지 않은 채 허공에 뜬 상태로 팔뚝 주위에 매달려 있음."
   },
   {
    "label": "B",
    "direction": "금속 주먹이 ATM의 전면 패널을 정확히 겨냥하여 뚫고 들어가고 있음.",
    "built_space": "왼쪽에 단일 ATM이 있고 배경에 상점 진열대가 배치되어 있어 참조 사진의 공간 구성과 일치함.",
    "entities": "찰리의 낡은 금속 기계 팔과 ATM의 형태가 참조와 잘 맞음. 클로즈업 프레임으로 인해 가방은 보이지 않음.",
    "hard_violations": [],
    "physics": "주먹은 프레임 밖의 팔과 몸체에 의해 지지되며 타격 동작을 수행함. 파편들은 바닥에 자연스럽게 떨어져 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "참조 이미지에 없는 별도의 캐비닛을 생성하여 타격하고 있으며, 가방이 물리적 지지 없이 허공에 떠 있는 치명적인 오류가 있습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 클로즈업 샷과 ATM 패널을 관통하는 금속 주먹의 타격 순간을 참조 공간에 맞게 정확히 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "금속 주먹이 ATM이 아닌 그 옆의 회색 금속 캐비닛 패널을 향해 뻗어 관통하고 있음.",
        "built_space": "왼쪽에 ATM이 존재하고 그 옆에 참조 이미지의 공간에는 없는 회색 캐비닛이 추가로 배치됨. 배경에 진열대가 보임.",
        "entities": "찰리의 기계 팔 외형은 참조와 일치함. 지시된 더플백과 배낭이 팔 주위에 나타남.",
        "hard_violations": [
         "발명된 구조물 (위치 참조에 없는 회색 캐비닛을 추가하여 타격함)",
         "지지대 없이 떠 있는 물체 (배낭과 더플백의 끈이 팔 근처 허공에 떠 있음)"
        ],
        "physics": "배낭과 더플백의 끈이 팔이나 손에 쥐어지지 않은 채 허공에 뜬 상태로 팔뚝 주위에 매달려 있음."
       },
       {
        "label": "B",
        "direction": "금속 주먹이 ATM의 전면 패널을 정확히 겨냥하여 뚫고 들어가고 있음.",
        "built_space": "왼쪽에 단일 ATM이 있고 배경에 상점 진열대가 배치되어 있어 참조 사진의 공간 구성과 일치함.",
        "entities": "찰리의 낡은 금속 기계 팔과 ATM의 형태가 참조와 잘 맞음. 클로즈업 프레임으로 인해 가방은 보이지 않음.",
        "hard_violations": [],
        "physics": "주먹은 프레임 밖의 팔과 몸체에 의해 지지되며 타격 동작을 수행함. 파편들은 바닥에 자연스럽게 떨어져 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "참조 이미지에 없는 별도의 캐비닛을 생성하여 타격하고 있으며, 가방이 물리적 지지 없이 허공에 떠 있는 치명적인 오류가 있습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 클로즈업 샷과 ATM 패널을 관통하는 금속 주먹의 타격 순간을 참조 공간에 맞게 정확히 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "금속 주먹이 ATM이 아닌 그 옆의 회색 금속 캐비닛 패널을 향해 뻗어 관통하고 있음.",
        "built_space": "왼쪽에 ATM이 존재하고 그 옆에 참조 이미지의 공간에는 없는 회색 캐비닛이 추가로 배치됨. 배경에 진열대가 보임.",
        "entities": "찰리의 기계 팔 외형은 참조와 일치함. 지시된 더플백과 배낭이 팔 주위에 나타남.",
        "hard_violations": [
         "발명된 구조물 (위치 참조에 없는 회색 캐비닛을 추가하여 타격함)",
         "지지대 없이 떠 있는 물체 (배낭과 더플백의 끈이 팔 근처 허공에 떠 있음)"
        ],
        "physics": "배낭과 더플백의 끈이 팔이나 손에 쥐어지지 않은 채 허공에 뜬 상태로 팔뚝 주위에 매달려 있음."
       },
       {
        "label": "B",
        "direction": "금속 주먹이 ATM의 전면 패널을 정확히 겨냥하여 뚫고 들어가고 있음.",
        "built_space": "왼쪽에 단일 ATM이 있고 배경에 상점 진열대가 배치되어 있어 참조 사진의 공간 구성과 일치함.",
        "entities": "찰리의 낡은 금속 기계 팔과 ATM의 형태가 참조와 잘 맞음. 클로즈업 프레임으로 인해 가방은 보이지 않음.",
        "hard_violations": [],
        "physics": "주먹은 프레임 밖의 팔과 몸체에 의해 지지되며 타격 동작을 수행함. 파편들은 바닥에 자연스럽게 떨어져 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "금속 주먹과 왼쪽 현금인출기 전면, 뒤쪽 매대의 관계는 더 가깝지만, 주먹이 구멍 안을 관통하기보다 이미 떨어져 나온 패널 앞에 있어 결정적인 타격 순간이 약하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "주먹의 진입 방향이 조작부가 있는 전면이 아니라 오른쪽 측판을 향하며, 어깨와 가방까지 포함해 요구된 주먹·전완 중심의 클로즈업을 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "팔은 오른쪽 위에서 왼쪽 아래로 뻗고, 주먹은 현금인출기 전면에서 뜯겨 내려온 패널을 향한다. 큰 관통 구멍은 주먹보다 위쪽에 있으며, 주먹 자체는 그 구멍 안에 들어가 있지 않다. 시선이나 무기는 보이지 않는다.",
        "built_space": "왼쪽에 현금인출기 한 대, 바로 뒤에 침구가 놓인 좁은 선반 한 구획, 오른쪽 뒤에 식품 매대가 보인다. 회색 기기와 금속 선반, 바닥 및 상품 구성은 장소 참조와 비교적 잘 맞는다. 기기 전면은 비스듬히 보이지만 왼쪽 약 절반을 차지하고, 바닥까지 넓게 포함되어 요구된 왼쪽 삼분의 일 배치와 밀착된 구도에서는 벗어난다. 사람 전신이나 반사는 없다.",
        "entities": "보이는 신체는 찰리의 금속 주먹과 팔 하나뿐이다. 마모된 은회색 장갑판, 분절된 손가락과 기계 관절은 손 소유자 참조에 가깝다. 현금인출기, 찢어진 전면, 배경 진열대가 존재한다. 바닥의 지폐와 잔해는 보이지만 지폐의 국가·액면은 판별하기 어렵다. 가방과 적재물은 프레임 밖이므로 누락으로 판단하지 않는다. 불꽃이나 색채 효과는 없고 뚜렷하게 읽히는 문구도 확인되지 않는다.",
        "hard_violations": [],
        "physics": "주먹은 손목과 전완에 연결되어 화면 밖으로 이어지므로 떠 있는 물체가 아니다. 뜯긴 패널은 남은 연결부와 케이블에 걸쳐 내려온 모습이며, 바닥의 지폐와 큰 잔해는 바닥이 받친다. 작은 공중 파편은 파손 충격으로 설명 가능하다. 다만 주먹과 큰 구멍의 위치가 어긋나 있어 전면을 맹렬히 뚫고 들어가는 바로 그 순간보다는 파손 이후의 자세로 읽힌다."
       },
       {
        "label": "B",
        "direction": "금속 팔은 오른쪽에서 왼쪽으로 뻗으며 주먹은 중앙의 찢어진 금속판 구멍을 향한다. 화면과 키패드가 있는 현금인출기 전면은 그보다 왼쪽에 별도로 보이므로, 실제 타격 대상은 전면이 아니라 오른쪽 측판으로 읽힌다. 시선이나 무기는 보이지 않는다.",
        "built_space": "왼쪽에 화면과 키패드를 갖춘 현금인출기 한 대가 있고, 그 오른쪽 넓은 측판에 관통 구멍이 있다. 뒤에는 식품 선반과 오른쪽의 작은 포장상품 진열 구역이 보인다. 금속 매대와 실내광은 참조와 유사하지만 현금인출기 외함 구성과 타격면의 배치가 다르다. 진입점이 중앙에 놓이고 어깨 일부와 가방 두 개까지 들어와, 전면을 왼쪽 삼분의 일에 놓는 주먹 중심 구도에서 벗어난다.",
        "entities": "찰리의 주먹·전완·상완과 어깨 일부가 보이며, 낡은 금속 장갑판과 관절은 참조의 기계 팔에 가깝다. 다른 인물은 없다. 현금인출기와 매대 외에 더플백 하나와 배낭 하나가 보이고, 내부에는 병과 식품처럼 보이는 물품이 있다. 일회용 전화기·지도·특정 약품인지는 식별할 수 없다. 가방을 더 보여 준 것은 요구된 프레이밍의 장점이 아니다. 불꽃이나 색채 효과는 없다.",
        "hard_violations": [],
        "physics": "주먹은 연결된 기계 팔이 지지하며, 찢긴 판 가장자리 사이에 일부 들어가 있는 모습이다. 구멍 아래 금속 조각은 외함에 붙어 접혀 있어 지지 없는 부유물로 보이지 않는다. 두 가방의 끈은 팔 쪽으로 올라가 걸린 형태여서 공중에 아무 지지 없이 떠 있다고 단정할 근거는 없다. 물리적인 타격은 가능하지만 그 타격이 요구된 전면이 아닌 측판에 이루어진다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "금속 주먹과 왼쪽 현금인출기 전면, 뒤쪽 매대의 관계는 더 가깝지만, 주먹이 구멍 안을 관통하기보다 이미 떨어져 나온 패널 앞에 있어 결정적인 타격 순간이 약하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "주먹의 진입 방향이 조작부가 있는 전면이 아니라 오른쪽 측판을 향하며, 어깨와 가방까지 포함해 요구된 주먹·전완 중심의 클로즈업을 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "팔은 오른쪽 위에서 왼쪽 아래로 뻗고, 주먹은 현금인출기 전면에서 뜯겨 내려온 패널을 향한다. 큰 관통 구멍은 주먹보다 위쪽에 있으며, 주먹 자체는 그 구멍 안에 들어가 있지 않다. 시선이나 무기는 보이지 않는다.",
        "built_space": "왼쪽에 현금인출기 한 대, 바로 뒤에 침구가 놓인 좁은 선반 한 구획, 오른쪽 뒤에 식품 매대가 보인다. 회색 기기와 금속 선반, 바닥 및 상품 구성은 장소 참조와 비교적 잘 맞는다. 기기 전면은 비스듬히 보이지만 왼쪽 약 절반을 차지하고, 바닥까지 넓게 포함되어 요구된 왼쪽 삼분의 일 배치와 밀착된 구도에서는 벗어난다. 사람 전신이나 반사는 없다.",
        "entities": "보이는 신체는 찰리의 금속 주먹과 팔 하나뿐이다. 마모된 은회색 장갑판, 분절된 손가락과 기계 관절은 손 소유자 참조에 가깝다. 현금인출기, 찢어진 전면, 배경 진열대가 존재한다. 바닥의 지폐와 잔해는 보이지만 지폐의 국가·액면은 판별하기 어렵다. 가방과 적재물은 프레임 밖이므로 누락으로 판단하지 않는다. 불꽃이나 색채 효과는 없고 뚜렷하게 읽히는 문구도 확인되지 않는다.",
        "hard_violations": [],
        "physics": "주먹은 손목과 전완에 연결되어 화면 밖으로 이어지므로 떠 있는 물체가 아니다. 뜯긴 패널은 남은 연결부와 케이블에 걸쳐 내려온 모습이며, 바닥의 지폐와 큰 잔해는 바닥이 받친다. 작은 공중 파편은 파손 충격으로 설명 가능하다. 다만 주먹과 큰 구멍의 위치가 어긋나 있어 전면을 맹렬히 뚫고 들어가는 바로 그 순간보다는 파손 이후의 자세로 읽힌다."
       },
       {
        "label": "A",
        "direction": "금속 팔은 오른쪽에서 왼쪽으로 뻗으며 주먹은 중앙의 찢어진 금속판 구멍을 향한다. 화면과 키패드가 있는 현금인출기 전면은 그보다 왼쪽에 별도로 보이므로, 실제 타격 대상은 전면이 아니라 오른쪽 측판으로 읽힌다. 시선이나 무기는 보이지 않는다.",
        "built_space": "왼쪽에 화면과 키패드를 갖춘 현금인출기 한 대가 있고, 그 오른쪽 넓은 측판에 관통 구멍이 있다. 뒤에는 식품 선반과 오른쪽의 작은 포장상품 진열 구역이 보인다. 금속 매대와 실내광은 참조와 유사하지만 현금인출기 외함 구성과 타격면의 배치가 다르다. 진입점이 중앙에 놓이고 어깨 일부와 가방 두 개까지 들어와, 전면을 왼쪽 삼분의 일에 놓는 주먹 중심 구도에서 벗어난다.",
        "entities": "찰리의 주먹·전완·상완과 어깨 일부가 보이며, 낡은 금속 장갑판과 관절은 참조의 기계 팔에 가깝다. 다른 인물은 없다. 현금인출기와 매대 외에 더플백 하나와 배낭 하나가 보이고, 내부에는 병과 식품처럼 보이는 물품이 있다. 일회용 전화기·지도·특정 약품인지는 식별할 수 없다. 가방을 더 보여 준 것은 요구된 프레이밍의 장점이 아니다. 불꽃이나 색채 효과는 없다.",
        "hard_violations": [],
        "physics": "주먹은 연결된 기계 팔이 지지하며, 찢긴 판 가장자리 사이에 일부 들어가 있는 모습이다. 구멍 아래 금속 조각은 외함에 붙어 접혀 있어 지지 없는 부유물로 보이지 않는다. 두 가방의 끈은 팔 쪽으로 올라가 걸린 형태여서 공중에 아무 지지 없이 떠 있다고 단정할 근거는 없다. 물리적인 타격은 가능하지만 그 타격이 요구된 전면이 아닌 측판에 이루어진다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.029,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.779,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 발명된 구조물 (위치 참조에 없는 회색 캐비닛을 추가하여 타격함)",
     "[gemini-pro] 지지대 없이 떠 있는 물체 (배낭과 더플백의 끈이 팔 근처 허공에 떠 있음)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 779,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 779,
    "verdict_ko": "참조 이미지에 없는 별도의 캐비닛을 생성하여 타격하고 있으며, 가방이 물리적 지지 없이 허공에 떠 있는 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 발명된 구조물 (위치 참조에 없는 회색 캐비닛을 추가하여 타격함) / [gemini-pro] 지지대 없이 떠 있는 물체 (배낭과 더플백의 끈이 팔 근처 허공에 떠 있음)"
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요구된 클로즈업 샷과 ATM 패널을 관통하는 금속 주먹의 타격 순간을 참조 공간에 맞게 정확히 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L51B01.png",
    "asset_id": "ebd17e28-2ffe-4520-8a2d-6f7928470bd2",
    "role": "location_plate"
   },
   {
    "label": "HAND OWNER REFERENCE — 찰리: the person whose hand is on the object in this shot. Use this photograph ONLY to get that hand right — its size, build, skin tone and texture, apparent age, grooming, nails, sleeve and anything worn on the wrist or fingers. Do NOT bring their face, body or clothing into the frame; the framing stays on the object and reaches no further than the wrist and forearm.",
    "path": "<bytes:1597480>",
    "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc3-f341-782e-8393-27f1a3d452fa",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S41sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:37:51.220302+00:00",
  "fingerprint": "af6f81a64a54157aad51c54a1e05b9e241251ee417a9d2638fecbc63bf51888b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S41sh11_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S41sh11_sel.png",
  "source_sha256": "6bd5f3e11a4b7d627c4d32e02a32eb22a4da8bb2fbfbd3b29e12234c5fa13168",
  "file": "S41sh11_cine.png",
  "staged_sha256": "cfd2d169b36b12103ed1b3ff8eb3d32f024b09aa82da38ff51481103c678d31c",
  "latency_ms": 12705
 },
 "S41sh22::signage": {
  "fp": "c857c132b528cb9b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S41sh22": {
  "input_fingerprint": "9bf0efacf2e039f3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 달려들어 라울의 품에 와락 안기며 부딪히는 역동적인 찰나.\n\nLOCATION (lock): In the unattended shop's aisle near the entrance and the shelves used as cover, with daylight entering through the open doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Display shelving along the aisle (The aisle between the displays is accessible) — Receding shelf faces flank the reunion rather than overlapping either face; used as Preserves the path 앰버 has crossed and contains the movement within the shop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent neutral interior light lets the relief in their faces supply the warmth without a new light source or color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM remains punctured, with cash and coins spilled out and money collected in a bag; the gathered food, map, disposable phones, and medication remain packed. Charlie has the newly supplied blanket for concealment, and the shop entrance has been opened. 앰버: She has emerged from hiding behind the shelves and rushed forward in relief. She retains the loaded bag. 라울: He has entered the shop and stopped in delighted recognition.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 달려들어 라울의 품에 와락 안기며 부딪히는 역동적인 찰나.\n\nLOCATION (lock): In the unattended shop's aisle near the entrance and the shelves used as cover, with daylight entering through the open doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Display shelving along the aisle (The aisle between the displays is accessible) — Receding shelf faces flank the reunion rather than overlapping either face; used as Preserves the path 앰버 has crossed and contains the movement within the shop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent neutral interior light lets the relief in their faces supply the warmth without a new light source or color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM remains punctured, with cash and coins spilled out and money collected in a bag; the gathered food, map, disposable phones, and medication remain packed. Charlie has the newly supplied blanket for concealment, and the shop entrance has been opened. 앰버: She has emerged from hiding behind the shelves and rushed forward in relief. She retains the loaded bag. 라울: He has entered the shop and stopped in delighted recognition.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 달려들어 라울의 품에 와락 안기며 부딪히는 역동적인 찰나.\n\nLOCATION (lock): In the unattended shop's aisle near the entrance and the shelves used as cover, with daylight entering through the open doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Display shelving along the aisle (The aisle between the displays is accessible) — Receding shelf faces flank the reunion rather than overlapping either face; used as Preserves the path 앰버 has crossed and contains the movement within the shop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent neutral interior light lets the relief in their faces supply the warmth without a new light source or color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM remains punctured, with cash and coins spilled out and money collected in a bag; the gathered food, map, disposable phones, and medication remain packed. Charlie has the newly supplied blanket for concealment, and the shop entrance has been opened. 앰버: She has emerged from hiding behind the shelves and rushed forward in relief. She retains the loaded bag. 라울: He has entered the shop and stopped in delighted recognition.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "앰버는 화면 왼쪽에서 라울 쪽으로 몸을 기울이고 양팔로 그의 상체를 감싼다. 눈은 거의 감겨 있고 얼굴은 라울의 얼굴 가까이에 있다. 라울 역시 앰버 쪽으로 얼굴을 돌리고 그녀를 받아 안는다. 두 몸의 접촉 방향은 달려와 품에 안기는 행동과 맞는다.",
    "built_space": "좌우에 금속 진열대 열이 하나씩 있고 중앙 통로는 열려 있다. 뒤쪽에는 열린 유리 출입구 하나, 오른쪽에는 파손된 현금자동입출금기 한 대가 보인다. 진열대는 얼굴을 가리지 않고 뒤로 이어진다. 회색 금속, 포장식품, 쌀 포대, 파편이 있는 바닥은 장소 참고와 대체로 이어진다. 다만 두 사람의 신발부터 머리까지와 넓은 바닥을 모두 담은 전신 구도로, 지정된 미디엄 숏은 아니다.",
    "entities": "등장인물은 금발 여자아이와 짙은 피부의 남자아이 두 명뿐이며, 둘 다 약 10세로 보인다. 앰버의 큰 눈과 얼굴은 표정 때문에 일부만 확인되며, 금발·회녹색 재킷·회색 원피스·갈색 부츠는 참고와 맞는다. 참고의 흰 마스크는 없다. 라울의 어린 얼굴, 뒤로 묶인 곱슬머리, 낡은 녹색 티셔츠와 청반바지는 대체로 맞는다. 두 아이의 외형은 명시된 혼혈 설정과 모순되지 않지만 사진만으로 혈통을 확정할 수는 없다. 앰버의 옆구리에 불룩한 가방이 있고 오른쪽 바닥에도 짐가방이 있다. 파손 기계와 흩어진 지폐는 보이나 화폐의 국가·액면, 동전과 가방 속 물품은 확인되지 않는다. 여러 상품 포장에는 읽을 수 있는 글자가 남아 있다.",
    "hard_violations": [
     "진열대의 과자·컵라면 포장에 식별 가능한 상품명과 글자가 보여, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반한다."
    ],
    "physics": "라울은 두 발을 바닥에 대고 무릎을 굽혀 앰버의 충격과 체중을 받는다. 앰버는 다리를 뒤로 굽힌 채 발이 바닥에서 떨어져 있지만 라울의 팔이 등과 몸통을 감싸 지지하고, 자신의 팔도 그의 어깨를 잡고 있다. 도약 직후 받아 안긴 자세로 성립하며 근거 없는 부유는 아니다. 앰버의 가방은 몸에 걸린 끈으로, 바닥 짐가방과 파편은 바닥으로 지지된다."
   },
   {
    "label": "A",
    "direction": "앰버는 왼쪽에서 라울의 가슴 쪽으로 몸을 밀착시키고 팔을 그의 어깨와 등 뒤로 두른다. 라울은 앰버의 얼굴을 올려다보며 웃는다. 앰버의 얼굴과 시선은 라울보다 카메라 쪽을 향해 있어 상호 재회의 집중은 약하지만, 몸의 이동과 포옹 대상은 분명히 라울이다.",
    "built_space": "왼쪽에 열린 유리 출입구 하나, 그 안쪽에 좁은 진열대 하나, 중앙 뒤에 현금자동입출금기 한 대가 있다. 뒤쪽에는 포대가 놓인 선반과 상품 선반들이 있고 오른쪽 전경의 긴 진열대가 통로를 한정한다. 두 아이는 출입구 가까운 빈 통로에 있으며 선반이 얼굴을 가리지 않는다. 회색 금속 선반, 포대, 기계 아래 파편과 낮의 유입광은 참고 장소와 대체로 맞는다. 라울의 하체를 화면 아래로 잘라 A보다 인물 중심이지만, 무릎과 앰버의 부츠까지 포함해 엄밀한 미디엄 숏보다는 넓다.",
    "entities": "두 명의 어린이만 보인다. 앰버는 참고와 유사한 금발, 어린 얼굴, 회녹색 재킷과 회색 원피스를 갖췄고 옆구리에 불룩한 가방을 유지한다. 흰 마스크는 없다. 라울의 짙은 피부, 어린 얼굴, 낡은 녹색 티셔츠와 청반바지는 맞지만 머리는 참고의 느슨한 곱슬 묶음보다 촘촘하게 땋은 형태다. 명시된 혼혈 정체성과 외형상 명백한 모순은 없으나 혈통 자체는 판별할 수 없다. 현금자동입출금기의 하부 손상과 바닥 파편·지폐가 일부 보이며, 동전과 포장된 지도·전화기·약품은 확인되지 않는다. 찰리나 이전 장면의 인물은 추가되지 않았다. 과자 포장의 로고와 한글 상품명이 읽힌다.",
    "hard_violations": [
     "오른쪽 진열대의 과자 포장에서 로고와 상품명이 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반한다."
    ],
    "physics": "라울의 팔이 앰버의 허리와 등을 감싸 체중을 받치며, 앰버의 팔도 그의 어깨를 붙잡는다. 앰버는 무릎을 굽히고 발을 뒤로 들어 달려와 안긴 직후의 자세를 만든다. 라울의 발은 프레임 밖이지만 보이는 하체는 아래로 이어지고, 공중에 뜬 것으로 볼 근거는 없다. 앰버는 실제 팔 접촉으로 지지되므로 무지지 부유가 아니다. 가방은 어깨끈으로 몸에 매달리고 바닥의 지폐와 파편은 바닥에 놓여 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "달려와 부딪히는 포옹과 통로 배치는 잘 보이지만, 전신과 넓은 바닥을 담아 지정된 미디엄 숏을 벗어나며 상품의 읽을 수 있는 글자가 금지 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 아이의 상체와 포옹을 더 크게 담아 지정 구도에 상대적으로 가깝지만, 여전히 프레임이 넓고 상품 로고와 글자가 읽혀 최종 사용 조건에는 실패한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 화면 왼쪽에서 라울 쪽으로 몸을 기울이고 양팔로 그의 상체를 감싼다. 눈은 거의 감겨 있고 얼굴은 라울의 얼굴 가까이에 있다. 라울 역시 앰버 쪽으로 얼굴을 돌리고 그녀를 받아 안는다. 두 몸의 접촉 방향은 달려와 품에 안기는 행동과 맞는다.",
        "built_space": "좌우에 금속 진열대 열이 하나씩 있고 중앙 통로는 열려 있다. 뒤쪽에는 열린 유리 출입구 하나, 오른쪽에는 파손된 현금자동입출금기 한 대가 보인다. 진열대는 얼굴을 가리지 않고 뒤로 이어진다. 회색 금속, 포장식품, 쌀 포대, 파편이 있는 바닥은 장소 참고와 대체로 이어진다. 다만 두 사람의 신발부터 머리까지와 넓은 바닥을 모두 담은 전신 구도로, 지정된 미디엄 숏은 아니다.",
        "entities": "등장인물은 금발 여자아이와 짙은 피부의 남자아이 두 명뿐이며, 둘 다 약 10세로 보인다. 앰버의 큰 눈과 얼굴은 표정 때문에 일부만 확인되며, 금발·회녹색 재킷·회색 원피스·갈색 부츠는 참고와 맞는다. 참고의 흰 마스크는 없다. 라울의 어린 얼굴, 뒤로 묶인 곱슬머리, 낡은 녹색 티셔츠와 청반바지는 대체로 맞는다. 두 아이의 외형은 명시된 혼혈 설정과 모순되지 않지만 사진만으로 혈통을 확정할 수는 없다. 앰버의 옆구리에 불룩한 가방이 있고 오른쪽 바닥에도 짐가방이 있다. 파손 기계와 흩어진 지폐는 보이나 화폐의 국가·액면, 동전과 가방 속 물품은 확인되지 않는다. 여러 상품 포장에는 읽을 수 있는 글자가 남아 있다.",
        "hard_violations": [
         "진열대의 과자·컵라면 포장에 식별 가능한 상품명과 글자가 보여, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반한다."
        ],
        "physics": "라울은 두 발을 바닥에 대고 무릎을 굽혀 앰버의 충격과 체중을 받는다. 앰버는 다리를 뒤로 굽힌 채 발이 바닥에서 떨어져 있지만 라울의 팔이 등과 몸통을 감싸 지지하고, 자신의 팔도 그의 어깨를 잡고 있다. 도약 직후 받아 안긴 자세로 성립하며 근거 없는 부유는 아니다. 앰버의 가방은 몸에 걸린 끈으로, 바닥 짐가방과 파편은 바닥으로 지지된다."
       },
       {
        "label": "B",
        "direction": "앰버는 왼쪽에서 라울의 가슴 쪽으로 몸을 밀착시키고 팔을 그의 어깨와 등 뒤로 두른다. 라울은 앰버의 얼굴을 올려다보며 웃는다. 앰버의 얼굴과 시선은 라울보다 카메라 쪽을 향해 있어 상호 재회의 집중은 약하지만, 몸의 이동과 포옹 대상은 분명히 라울이다.",
        "built_space": "왼쪽에 열린 유리 출입구 하나, 그 안쪽에 좁은 진열대 하나, 중앙 뒤에 현금자동입출금기 한 대가 있다. 뒤쪽에는 포대가 놓인 선반과 상품 선반들이 있고 오른쪽 전경의 긴 진열대가 통로를 한정한다. 두 아이는 출입구 가까운 빈 통로에 있으며 선반이 얼굴을 가리지 않는다. 회색 금속 선반, 포대, 기계 아래 파편과 낮의 유입광은 참고 장소와 대체로 맞는다. 라울의 하체를 화면 아래로 잘라 A보다 인물 중심이지만, 무릎과 앰버의 부츠까지 포함해 엄밀한 미디엄 숏보다는 넓다.",
        "entities": "두 명의 어린이만 보인다. 앰버는 참고와 유사한 금발, 어린 얼굴, 회녹색 재킷과 회색 원피스를 갖췄고 옆구리에 불룩한 가방을 유지한다. 흰 마스크는 없다. 라울의 짙은 피부, 어린 얼굴, 낡은 녹색 티셔츠와 청반바지는 맞지만 머리는 참고의 느슨한 곱슬 묶음보다 촘촘하게 땋은 형태다. 명시된 혼혈 정체성과 외형상 명백한 모순은 없으나 혈통 자체는 판별할 수 없다. 현금자동입출금기의 하부 손상과 바닥 파편·지폐가 일부 보이며, 동전과 포장된 지도·전화기·약품은 확인되지 않는다. 찰리나 이전 장면의 인물은 추가되지 않았다. 과자 포장의 로고와 한글 상품명이 읽힌다.",
        "hard_violations": [
         "오른쪽 진열대의 과자 포장에서 로고와 상품명이 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반한다."
        ],
        "physics": "라울의 팔이 앰버의 허리와 등을 감싸 체중을 받치며, 앰버의 팔도 그의 어깨를 붙잡는다. 앰버는 무릎을 굽히고 발을 뒤로 들어 달려와 안긴 직후의 자세를 만든다. 라울의 발은 프레임 밖이지만 보이는 하체는 아래로 이어지고, 공중에 뜬 것으로 볼 근거는 없다. 앰버는 실제 팔 접촉으로 지지되므로 무지지 부유가 아니다. 가방은 어깨끈으로 몸에 매달리고 바닥의 지폐와 파편은 바닥에 놓여 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "달려와 부딪히는 포옹과 통로 배치는 잘 보이지만, 전신과 넓은 바닥을 담아 지정된 미디엄 숏을 벗어나며 상품의 읽을 수 있는 글자가 금지 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 아이의 상체와 포옹을 더 크게 담아 지정 구도에 상대적으로 가깝지만, 여전히 프레임이 넓고 상품 로고와 글자가 읽혀 최종 사용 조건에는 실패한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 화면 왼쪽에서 라울 쪽으로 몸을 기울이고 양팔로 그의 상체를 감싼다. 눈은 거의 감겨 있고 얼굴은 라울의 얼굴 가까이에 있다. 라울 역시 앰버 쪽으로 얼굴을 돌리고 그녀를 받아 안는다. 두 몸의 접촉 방향은 달려와 품에 안기는 행동과 맞는다.",
        "built_space": "좌우에 금속 진열대 열이 하나씩 있고 중앙 통로는 열려 있다. 뒤쪽에는 열린 유리 출입구 하나, 오른쪽에는 파손된 현금자동입출금기 한 대가 보인다. 진열대는 얼굴을 가리지 않고 뒤로 이어진다. 회색 금속, 포장식품, 쌀 포대, 파편이 있는 바닥은 장소 참고와 대체로 이어진다. 다만 두 사람의 신발부터 머리까지와 넓은 바닥을 모두 담은 전신 구도로, 지정된 미디엄 숏은 아니다.",
        "entities": "등장인물은 금발 여자아이와 짙은 피부의 남자아이 두 명뿐이며, 둘 다 약 10세로 보인다. 앰버의 큰 눈과 얼굴은 표정 때문에 일부만 확인되며, 금발·회녹색 재킷·회색 원피스·갈색 부츠는 참고와 맞는다. 참고의 흰 마스크는 없다. 라울의 어린 얼굴, 뒤로 묶인 곱슬머리, 낡은 녹색 티셔츠와 청반바지는 대체로 맞는다. 두 아이의 외형은 명시된 혼혈 설정과 모순되지 않지만 사진만으로 혈통을 확정할 수는 없다. 앰버의 옆구리에 불룩한 가방이 있고 오른쪽 바닥에도 짐가방이 있다. 파손 기계와 흩어진 지폐는 보이나 화폐의 국가·액면, 동전과 가방 속 물품은 확인되지 않는다. 여러 상품 포장에는 읽을 수 있는 글자가 남아 있다.",
        "hard_violations": [
         "진열대의 과자·컵라면 포장에 식별 가능한 상품명과 글자가 보여, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반한다."
        ],
        "physics": "라울은 두 발을 바닥에 대고 무릎을 굽혀 앰버의 충격과 체중을 받는다. 앰버는 다리를 뒤로 굽힌 채 발이 바닥에서 떨어져 있지만 라울의 팔이 등과 몸통을 감싸 지지하고, 자신의 팔도 그의 어깨를 잡고 있다. 도약 직후 받아 안긴 자세로 성립하며 근거 없는 부유는 아니다. 앰버의 가방은 몸에 걸린 끈으로, 바닥 짐가방과 파편은 바닥으로 지지된다."
       },
       {
        "label": "A",
        "direction": "앰버는 왼쪽에서 라울의 가슴 쪽으로 몸을 밀착시키고 팔을 그의 어깨와 등 뒤로 두른다. 라울은 앰버의 얼굴을 올려다보며 웃는다. 앰버의 얼굴과 시선은 라울보다 카메라 쪽을 향해 있어 상호 재회의 집중은 약하지만, 몸의 이동과 포옹 대상은 분명히 라울이다.",
        "built_space": "왼쪽에 열린 유리 출입구 하나, 그 안쪽에 좁은 진열대 하나, 중앙 뒤에 현금자동입출금기 한 대가 있다. 뒤쪽에는 포대가 놓인 선반과 상품 선반들이 있고 오른쪽 전경의 긴 진열대가 통로를 한정한다. 두 아이는 출입구 가까운 빈 통로에 있으며 선반이 얼굴을 가리지 않는다. 회색 금속 선반, 포대, 기계 아래 파편과 낮의 유입광은 참고 장소와 대체로 맞는다. 라울의 하체를 화면 아래로 잘라 A보다 인물 중심이지만, 무릎과 앰버의 부츠까지 포함해 엄밀한 미디엄 숏보다는 넓다.",
        "entities": "두 명의 어린이만 보인다. 앰버는 참고와 유사한 금발, 어린 얼굴, 회녹색 재킷과 회색 원피스를 갖췄고 옆구리에 불룩한 가방을 유지한다. 흰 마스크는 없다. 라울의 짙은 피부, 어린 얼굴, 낡은 녹색 티셔츠와 청반바지는 맞지만 머리는 참고의 느슨한 곱슬 묶음보다 촘촘하게 땋은 형태다. 명시된 혼혈 정체성과 외형상 명백한 모순은 없으나 혈통 자체는 판별할 수 없다. 현금자동입출금기의 하부 손상과 바닥 파편·지폐가 일부 보이며, 동전과 포장된 지도·전화기·약품은 확인되지 않는다. 찰리나 이전 장면의 인물은 추가되지 않았다. 과자 포장의 로고와 한글 상품명이 읽힌다.",
        "hard_violations": [
         "오른쪽 진열대의 과자 포장에서 로고와 상품명이 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반한다."
        ],
        "physics": "라울의 팔이 앰버의 허리와 등을 감싸 체중을 받치며, 앰버의 팔도 그의 어깨를 붙잡는다. 앰버는 무릎을 굽히고 발을 뒤로 들어 달려와 안긴 직후의 자세를 만든다. 라울의 발은 프레임 밖이지만 보이는 하체는 아래로 이어지고, 공중에 뜬 것으로 볼 근거는 없다. 앰버는 실제 팔 접촉으로 지지되므로 무지지 부유가 아니다. 가방은 어깨끈으로 몸에 매달리고 바닥의 지폐와 파편은 바닥에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "totals": {
   "B": 3,
   "A": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "달려와 부딪히는 포옹과 통로 배치는 잘 보이지만, 전신과 넓은 바닥을 담아 지정된 미디엄 숏을 벗어나며 상품의 읽을 수 있는 글자가 금지 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "두 아이의 상체와 포옹을 더 크게 담아 지정 구도에 상대적으로 가깝지만, 여전히 프레임이 넓고 상품 로고와 글자가 읽혀 최종 사용 조건에는 실패한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S41sh11_sel.png",
    "asset_id": "e8cb4c5c-a851-4d5f-b6ea-be7da4ba823e",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:943979>",
    "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc4-c402-7752-bec6-42861709b346",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S41sh11"
  }
 },
 "S41sh22::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:38:04.265328+00:00",
  "fingerprint": "9f38c7bd9e37ab60ba7029a2ba018df0ccdddf05dfd99aa67741badb6a715523",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S41sh22_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S41sh22_sel.png",
  "source_sha256": "f59d22dbb59fae6775b54a8ba7cb8d32969b14d4992124842c9528f0dbc09c4e",
  "file": "S41sh22_cine.png",
  "staged_sha256": "8745448598fb0931aa3e8f8a719da0feced175ed4ebaa913ef933fe7a82a1785",
  "latency_ms": 11372
 },
 "S41sh27::signage": {
  "fp": "2451a88731b25dc0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S41sh27": {
  "input_fingerprint": "56fa55871ff90a91",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 라울을 향해 턱을 살짝 내린 채 단호한 표정으로 가방 끈을 꽉 움켜쥔 찰나.\n\nLOCATION (lock): In the merchandise aisle inside the unattended shop, near the group's packed bags in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Bag and strap held by 이현우 (The strap is gripped firmly) — The gripped strap runs down the visible side of his torso; only part of the bag enters below; used as Makes readiness to leave legible within the same frame as his expression; Shop display shelving (Still present behind the conversation) — A subdued section of the shelf fronts appears behind 이현우; used as Maintains the shared shop setting while keeping the shoulder-to-face sightline clear.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the shop's neutral ambient light consistent, using restrained facial contrast to support 이현우's resolve.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged ATM and spilled currency remain in the shop, while the gathered cash, supplies, disposable phones, map, and medication are packed for departure. Charlie retains the supplied blanket. 이현우: He is ready to leave with the loaded luggage and stolen cash. His facial bruising and injured leg remain, and the contact card is still hidden inside his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 라울을 향해 턱을 살짝 내린 채 단호한 표정으로 가방 끈을 꽉 움켜쥔 찰나.\n\nLOCATION (lock): In the merchandise aisle inside the unattended shop, near the group's packed bags in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Bag and strap held by 이현우 (The strap is gripped firmly) — The gripped strap runs down the visible side of his torso; only part of the bag enters below; used as Makes readiness to leave legible within the same frame as his expression; Shop display shelving (Still present behind the conversation) — A subdued section of the shelf fronts appears behind 이현우; used as Maintains the shared shop setting while keeping the shoulder-to-face sightline clear.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the shop's neutral ambient light consistent, using restrained facial contrast to support 이현우's resolve.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged ATM and spilled currency remain in the shop, while the gathered cash, supplies, disposable phones, map, and medication are packed for departure. Charlie retains the supplied blanket. 이현우: He is ready to leave with the loaded luggage and stolen cash. His facial bruising and injured leg remain, and the contact card is still hidden inside his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 라울을 향해 턱을 살짝 내린 채 단호한 표정으로 가방 끈을 꽉 움켜쥔 찰나.\n\nLOCATION (lock): In the merchandise aisle inside the unattended shop, near the group's packed bags in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Bag and strap held by 이현우 (The strap is gripped firmly) — The gripped strap runs down the visible side of his torso; only part of the bag enters below; used as Makes readiness to leave legible within the same frame as his expression; Shop display shelving (Still present behind the conversation) — A subdued section of the shelf fronts appears behind 이현우; used as Maintains the shared shop setting while keeping the shoulder-to-face sightline clear.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the shop's neutral ambient light consistent, using restrained facial contrast to support 이현우's resolve.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged ATM and spilled currency remain in the shop, while the gathered cash, supplies, disposable phones, map, and medication are packed for departure. Charlie retains the supplied blanket. 이현우: He is ready to leave with the loaded luggage and stolen cash. His facial bruising and injured leg remain, and the contact card is still hidden inside his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선이 프레임 밖의 인물(라울)이 아닌 카메라 렌즈를 정면으로 응시하고 있습니다.",
    "built_space": "레퍼런스 이미지에 나타난 장소의 특징(왼쪽의 유리문, 중앙의 ATM 기기, 진열대의 정확한 배치 등)을 완벽하게 재현하고 유지했습니다. 인물 뒤편으로 바닥에 놓인 짐가방들도 보입니다.",
    "entities": "이현우의 얼굴, 헝클어진 머리, 의상이 캐릭터 레퍼런스와 정확히 일치하며 얼굴의 멍 자국도 표현되었습니다. 진열대의 기물들 역시 레퍼런스의 형태를 따릅니다.",
    "hard_violations": [
     "[gpt-high] 왼쪽 진열 상자와 오른쪽 과자 포장 등에 읽을 수 있는 브랜드 글자와 로고가 남아 있어 무문자·무로고 조건을 위반한다."
    ],
    "physics": "가방 끈을 움켜쥔 손의 형태가 자연스러우며, 손의 위치와 가방의 무게감이 사실적으로 연결되어 있습니다."
   },
   {
    "label": "B",
    "direction": "인물이 턱을 살짝 내린 채 화면 우측 밖을 바라보고 있어, 오프스크린의 대상을 향하는 텍스트의 지시를 잘 따르고 있습니다.",
    "built_space": "이전 샷의 장소를 유지해야 한다는 지침안에도 불구하고 ATM, 문 등 주요 기물이 사라지고 진열대 배치가 완전히 임의로 변형된 전혀 다른 실내 공간이 생성되었습니다.",
    "entities": "이현우의 인상과 의상은 레퍼런스에 잘 부합합니다. 하지만 우측 진열대의 과자 봉지들에 명확히 읽을 수 있는 텍스트('체카칩' 등)가 생성되었습니다.",
    "hard_violations": [
     "[gemini-pro] 지정된 장소(Location lock)의 고정된 구조와 기물을 무시하고 완전히 다른 공간을 생성함",
     "[gemini-pro] 이미지 내 읽을 수 있는 텍스트(진열대 과자의 한글 이름) 생성 금지 지침 위반",
     "[gpt-high] 오른쪽 상품 포장에 읽을 수 있는 한글과 브랜드 로고가 선명하게 남아 있어, 글자와 로고를 일절 읽을 수 없게 하라는 명시적 조건을 위반한다."
    ],
    "physics": "가방 끈을 단호하게 쥐고 있는 손의 관절과 주름이 물리적으로 자연스럽게 표현되었으며 어깨의 자세도 안정적입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 샷의 배경(공간 구조 및 기물)과 인물의 외형을 레퍼런스와 완벽하게 일치시켰으나, 시선이 오프스크린의 인물이 아닌 카메라 정면을 향한 점이 유일한 아쉬움입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시된 시선 방향과 표정 연출은 훌륭하지만, 엄격하게 고정된 이전 샷의 장소 구조를 완전히 무시하고 읽을 수 있는 텍스트를 생성하여 핵심 지침을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선이 프레임 밖의 인물(라울)이 아닌 카메라 렌즈를 정면으로 응시하고 있습니다.",
        "built_space": "레퍼런스 이미지에 나타난 장소의 특징(왼쪽의 유리문, 중앙의 ATM 기기, 진열대의 정확한 배치 등)을 완벽하게 재현하고 유지했습니다. 인물 뒤편으로 바닥에 놓인 짐가방들도 보입니다.",
        "entities": "이현우의 얼굴, 헝클어진 머리, 의상이 캐릭터 레퍼런스와 정확히 일치하며 얼굴의 멍 자국도 표현되었습니다. 진열대의 기물들 역시 레퍼런스의 형태를 따릅니다.",
        "hard_violations": [],
        "physics": "가방 끈을 움켜쥔 손의 형태가 자연스러우며, 손의 위치와 가방의 무게감이 사실적으로 연결되어 있습니다."
       },
       {
        "label": "B",
        "direction": "인물이 턱을 살짝 내린 채 화면 우측 밖을 바라보고 있어, 오프스크린의 대상을 향하는 텍스트의 지시를 잘 따르고 있습니다.",
        "built_space": "이전 샷의 장소를 유지해야 한다는 지침안에도 불구하고 ATM, 문 등 주요 기물이 사라지고 진열대 배치가 완전히 임의로 변형된 전혀 다른 실내 공간이 생성되었습니다.",
        "entities": "이현우의 인상과 의상은 레퍼런스에 잘 부합합니다. 하지만 우측 진열대의 과자 봉지들에 명확히 읽을 수 있는 텍스트('체카칩' 등)가 생성되었습니다.",
        "hard_violations": [
         "지정된 장소(Location lock)의 고정된 구조와 기물을 무시하고 완전히 다른 공간을 생성함",
         "이미지 내 읽을 수 있는 텍스트(진열대 과자의 한글 이름) 생성 금지 지침 위반"
        ],
        "physics": "가방 끈을 단호하게 쥐고 있는 손의 관절과 주름이 물리적으로 자연스럽게 표현되었으며 어깨의 자세도 안정적입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 샷의 배경(공간 구조 및 기물)과 인물의 외형을 레퍼런스와 완벽하게 일치시켰으나, 시선이 오프스크린의 인물이 아닌 카메라 정면을 향한 점이 유일한 아쉬움입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시된 시선 방향과 표정 연출은 훌륭하지만, 엄격하게 고정된 이전 샷의 장소 구조를 완전히 무시하고 읽을 수 있는 텍스트를 생성하여 핵심 지침을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선이 프레임 밖의 인물(라울)이 아닌 카메라 렌즈를 정면으로 응시하고 있습니다.",
        "built_space": "레퍼런스 이미지에 나타난 장소의 특징(왼쪽의 유리문, 중앙의 ATM 기기, 진열대의 정확한 배치 등)을 완벽하게 재현하고 유지했습니다. 인물 뒤편으로 바닥에 놓인 짐가방들도 보입니다.",
        "entities": "이현우의 얼굴, 헝클어진 머리, 의상이 캐릭터 레퍼런스와 정확히 일치하며 얼굴의 멍 자국도 표현되었습니다. 진열대의 기물들 역시 레퍼런스의 형태를 따릅니다.",
        "hard_violations": [],
        "physics": "가방 끈을 움켜쥔 손의 형태가 자연스러우며, 손의 위치와 가방의 무게감이 사실적으로 연결되어 있습니다."
       },
       {
        "label": "B",
        "direction": "인물이 턱을 살짝 내린 채 화면 우측 밖을 바라보고 있어, 오프스크린의 대상을 향하는 텍스트의 지시를 잘 따르고 있습니다.",
        "built_space": "이전 샷의 장소를 유지해야 한다는 지침안에도 불구하고 ATM, 문 등 주요 기물이 사라지고 진열대 배치가 완전히 임의로 변형된 전혀 다른 실내 공간이 생성되었습니다.",
        "entities": "이현우의 인상과 의상은 레퍼런스에 잘 부합합니다. 하지만 우측 진열대의 과자 봉지들에 명확히 읽을 수 있는 텍스트('체카칩' 등)가 생성되었습니다.",
        "hard_violations": [
         "지정된 장소(Location lock)의 고정된 구조와 기물을 무시하고 완전히 다른 공간을 생성함",
         "이미지 내 읽을 수 있는 텍스트(진열대 과자의 한글 이름) 생성 금지 지침 위반"
        ],
        "physics": "가방 끈을 단호하게 쥐고 있는 손의 관절과 주름이 물리적으로 자연스럽게 표현되었으며 어깨의 자세도 안정적입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "턱을 내리고 화면 밖 상대를 향해 끈을 꽉 쥔 미디엄 숏은 더 충실하지만, 선명한 상품 글자와 로고가 무문자 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "장소와 인물 외형은 잘 이어지지만, 라울을 향한 순간보다 렌즈를 정면 응시하는 자세이며 상품 글자와 로고도 노출된다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 눈이 화면 오른쪽 아래의 화면 밖 지점을 향하고 턱도 조금 내려가 있다. 라울은 보이지 않아 실제 시선 도착점은 확인할 수 없지만, 화면 밖 상대를 보는 대화 방향으로 읽힌다. 무기나 이동 중인 물체는 없다.",
        "built_space": "왼쪽 끝의 일부 진열대, 그 옆 좁은 진열대, 뒤쪽 넓은 진열대, 몸 오른쪽 뒤의 부분 진열대, 오른쪽 가까운 진열대가 보인다. 인물은 그 사이 통로에 서 있다. 회색 금속 선반과 밝은 벽은 장소 참고와 맞으며, 카메라 각도가 달라 출입문과 현금인출기는 보이지 않는다. 오른쪽 상품 면은 요구된 차분한 배경보다 선명하고 두드러진다. 반사는 없다.",
        "entities": "한 명의 젊은 한국계 남성으로 보이며 헝클어진 검은 머리, 낡은 짙은 회색 후드, 마른 체격과 얼굴의 상처가 참고 및 지시와 대체로 맞는다. 얼굴 윤곽은 참고보다 조금 길고 성숙해 보인다. 본인의 손과 소매가 가방 끈을 감싸고, 몸 옆으로 내려오는 끈과 하단에 잘린 천 가방이 보인다. 다리 부상, 신발 속 카드와 포장된 내용물은 이 구도에서 확인할 수 없다. 오른쪽 과자 포장의 글자와 로고는 읽을 수 있다.",
        "hard_violations": [
         "오른쪽 상품 포장에 읽을 수 있는 한글과 브랜드 로고가 선명하게 남아 있어, 글자와 로고를 일절 읽을 수 없게 하라는 명시적 조건을 위반한다."
        ],
        "physics": "가방은 어깨에 걸친 끈으로 지지되고, 손가락이 끈을 감싸 하중을 붙잡는 형태다. 가방 천은 아래로 처지며 손목과 팔꿈치 연결도 자연스럽다. 하체는 프레임 밖이지만 상체에 부유나 불가능한 자세의 징후는 없다. 상품은 선반 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "얼굴이 거의 정면이고 눈은 렌즈 쪽을 향한다. 턱을 약간 당겼지만 라울을 향한 시선이라는 단서는 없으며, 직접 카메라를 응시하는 인상이 강하다. 끈을 쥔 손은 가슴 앞에 있고 방향성 있는 무기나 이동 물체는 없다.",
        "built_space": "왼쪽 유리 출입문 하나와 뒤쪽 현금인출기 하나가 보인다. 왼쪽 좁은 선반, 중앙 뒤의 부분 선반, 오른쪽 중간의 좁은 선반과 넓은 후면 선반, 오른쪽 전경 선반, 왼쪽 가장자리의 일부 진열대가 참고 장소 배치와 잘 이어진다. 인물은 통로 중앙에 서고 포장된 가방들은 왼쪽 뒤 바닥에 놓여 있다. 중복된 고정 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "한 명의 젊은 한국계 남성으로 보이며 검은 헝클어진 머리, 얼굴 형태, 낡은 회색 후드와 체격이 인물 참고에 가깝다. 양 볼의 멍과 상처가 보이고 본인의 손이 어깨 가방 끈을 움켜쥔다. 끈은 몸 앞쪽으로 내려오며 가방 일부가 하단에 걸친다. 뒤에는 출발용 짐이 보인다. 다리, 신발 속 카드와 가방 내용물은 확인 범위 밖이다. 여러 상품의 브랜드 글자와 로고가 식별된다.",
        "hard_violations": [
         "왼쪽 진열 상자와 오른쪽 과자 포장 등에 읽을 수 있는 브랜드 글자와 로고가 남아 있어 무문자·무로고 조건을 위반한다."
        ],
        "physics": "가방 끈이 어깨에 걸려 있고 주먹이 끈을 직접 잡아 가방을 지지한다. 천 가방의 처짐과 팔의 굽힘은 가능한 자세다. 뒤쪽 짐은 바닥과 서로의 가방 위에 안정적으로 놓여 있다. 발은 보이지 않지만 인물이 떠 있다는 징후는 없고, 선반 상품에도 지지 없는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "턱을 내리고 화면 밖 상대를 향해 끈을 꽉 쥔 미디엄 숏은 더 충실하지만, 선명한 상품 글자와 로고가 무문자 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "장소와 인물 외형은 잘 이어지지만, 라울을 향한 순간보다 렌즈를 정면 응시하는 자세이며 상품 글자와 로고도 노출된다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 눈이 화면 오른쪽 아래의 화면 밖 지점을 향하고 턱도 조금 내려가 있다. 라울은 보이지 않아 실제 시선 도착점은 확인할 수 없지만, 화면 밖 상대를 보는 대화 방향으로 읽힌다. 무기나 이동 중인 물체는 없다.",
        "built_space": "왼쪽 끝의 일부 진열대, 그 옆 좁은 진열대, 뒤쪽 넓은 진열대, 몸 오른쪽 뒤의 부분 진열대, 오른쪽 가까운 진열대가 보인다. 인물은 그 사이 통로에 서 있다. 회색 금속 선반과 밝은 벽은 장소 참고와 맞으며, 카메라 각도가 달라 출입문과 현금인출기는 보이지 않는다. 오른쪽 상품 면은 요구된 차분한 배경보다 선명하고 두드러진다. 반사는 없다.",
        "entities": "한 명의 젊은 한국계 남성으로 보이며 헝클어진 검은 머리, 낡은 짙은 회색 후드, 마른 체격과 얼굴의 상처가 참고 및 지시와 대체로 맞는다. 얼굴 윤곽은 참고보다 조금 길고 성숙해 보인다. 본인의 손과 소매가 가방 끈을 감싸고, 몸 옆으로 내려오는 끈과 하단에 잘린 천 가방이 보인다. 다리 부상, 신발 속 카드와 포장된 내용물은 이 구도에서 확인할 수 없다. 오른쪽 과자 포장의 글자와 로고는 읽을 수 있다.",
        "hard_violations": [
         "오른쪽 상품 포장에 읽을 수 있는 한글과 브랜드 로고가 선명하게 남아 있어, 글자와 로고를 일절 읽을 수 없게 하라는 명시적 조건을 위반한다."
        ],
        "physics": "가방은 어깨에 걸친 끈으로 지지되고, 손가락이 끈을 감싸 하중을 붙잡는 형태다. 가방 천은 아래로 처지며 손목과 팔꿈치 연결도 자연스럽다. 하체는 프레임 밖이지만 상체에 부유나 불가능한 자세의 징후는 없다. 상품은 선반 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "얼굴이 거의 정면이고 눈은 렌즈 쪽을 향한다. 턱을 약간 당겼지만 라울을 향한 시선이라는 단서는 없으며, 직접 카메라를 응시하는 인상이 강하다. 끈을 쥔 손은 가슴 앞에 있고 방향성 있는 무기나 이동 물체는 없다.",
        "built_space": "왼쪽 유리 출입문 하나와 뒤쪽 현금인출기 하나가 보인다. 왼쪽 좁은 선반, 중앙 뒤의 부분 선반, 오른쪽 중간의 좁은 선반과 넓은 후면 선반, 오른쪽 전경 선반, 왼쪽 가장자리의 일부 진열대가 참고 장소 배치와 잘 이어진다. 인물은 통로 중앙에 서고 포장된 가방들은 왼쪽 뒤 바닥에 놓여 있다. 중복된 고정 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "한 명의 젊은 한국계 남성으로 보이며 검은 헝클어진 머리, 얼굴 형태, 낡은 회색 후드와 체격이 인물 참고에 가깝다. 양 볼의 멍과 상처가 보이고 본인의 손이 어깨 가방 끈을 움켜쥔다. 끈은 몸 앞쪽으로 내려오며 가방 일부가 하단에 걸친다. 뒤에는 출발용 짐이 보인다. 다리, 신발 속 카드와 가방 내용물은 확인 범위 밖이다. 여러 상품의 브랜드 글자와 로고가 식별된다.",
        "hard_violations": [
         "왼쪽 진열 상자와 오른쪽 과자 포장 등에 읽을 수 있는 브랜드 글자와 로고가 남아 있어 무문자·무로고 조건을 위반한다."
        ],
        "physics": "가방 끈이 어깨에 걸려 있고 주먹이 끈을 직접 잡아 가방을 지지한다. 천 가방의 처짐과 팔의 굽힘은 가능한 자세다. 뒤쪽 짐은 바닥과 서로의 가방 위에 안정적으로 놓여 있다. 발은 보이지 않지만 인물이 떠 있다는 징후는 없고, 선반 상품에도 지지 없는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gemini-pro] 지정된 장소(Location lock)의 고정된 구조와 기물을 무시하고 완전히 다른 공간을 생성함",
     "[gemini-pro] 이미지 내 읽을 수 있는 텍스트(진열대 과자의 한글 이름) 생성 금지 지침 위반",
     "[gpt-high] 오른쪽 상품 포장에 읽을 수 있는 한글과 브랜드 로고가 선명하게 남아 있어, 글자와 로고를 일절 읽을 수 없게 하라는 명시적 조건을 위반한다."
    ],
    "A": [
     "[gpt-high] 왼쪽 진열 상자와 오른쪽 과자 포장 등에 읽을 수 있는 브랜드 글자와 로고가 남아 있어 무문자·무로고 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1500,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "지정된 샷의 배경(공간 구조 및 기물)과 인물의 외형을 레퍼런스와 완벽하게 일치시켰으나, 시선이 오프스크린의 인물이 아닌 카메라 정면을 향한 점이 유일한 아쉬움입니다.  ★위반: [gpt-high] 왼쪽 진열 상자와 오른쪽 과자 포장 등에 읽을 수 있는 브랜드 글자와 로고가 남아 있어 무문자·무로고 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "지시된 시선 방향과 표정 연출은 훌륭하지만, 엄격하게 고정된 이전 샷의 장소 구조를 완전히 무시하고 읽을 수 있는 텍스트를 생성하여 핵심 지침을 위반했습니다.  ★위반: [gemini-pro] 지정된 장소(Location lock)의 고정된 구조와 기물을 무시하고 완전히 다른 공간을 생성함 / [gemini-pro] 이미지 내 읽을 수 있는 텍스트(진열대 과자의 한글 이름) 생성 금지 지침 위반 / [gpt-high] 오른쪽 상품 포장에 읽을 수 있는 한글과 브랜드 로고가 선명하게 남아 있어, 글자와 로고를 일절 읽을 수 없게 하라는 명시적 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S41sh22_sel.png",
    "asset_id": "6f92d0cc-a612-4909-a076-b8c0f8976c4e",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc5-7fb5-7bf6-9c18-dd90c25fd6b9",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S41sh22"
  }
 },
 "S41sh27::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:38:15.994467+00:00",
  "fingerprint": "7451db3118058d588697f46b91d72d16141573ed7ab5cfb779946237ae5a188f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S41sh27_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S41sh27_sel.png",
  "source_sha256": "836cc86be8d61bea8060bb01f6ac6e8d877b68600341e0df8e8b8fbbcb4263c7",
  "file": "S41sh27_cine.png",
  "staged_sha256": "8b6c6b5c226b37d0ac9cf6479c5f73743019a178bb8211200c275af453e59643",
  "latency_ms": 11541
 },
 "S42sh2::signage": {
  "fp": "9c4f3fffd489befd",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::camp_gate": {
  "input_fingerprint": "6f5a88bc77ce5046",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "camp_gate",
    "tags": [
     "S11sh5",
     "S42sh10",
     "S42sh18",
     "S42sh2"
    ]
   },
   "context_sig": "913f05321a69543a"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the muddy entrance road of the refugee settlement at night, beside an opened body bag.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민 수용소 출입문과 입구 도로: 통금 시간을 알리는 거대한 스피커와 철조망, 닫히기 직전의 육중한 철제 출입문이 있는 통제 구역. (특징: 높은 철조망과 육중한 철제 출입문; 곳곳에 설치된 통제용 확성기 스피커)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 출입문 닫기 전에 헐레벌떡 뛰어오는 난민들, 그 틈에 현우도 끼여서 간신히 들어온다.\n- 42. 난민촌 입구 – N\n- 진입로마다 검문소 설치했습니다. 감시 드론도 띄웠구요.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the muddy entrance road of the refugee settlement at night, beside an opened body bag.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민 수용소 출입문과 입구 도로: 통금 시간을 알리는 거대한 스피커와 철조망, 닫히기 직전의 육중한 철제 출입문이 있는 통제 구역. (특징: 높은 철조망과 육중한 철제 출입문; 곳곳에 설치된 통제용 확성기 스피커)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 출입문 닫기 전에 헐레벌떡 뛰어오는 난민들, 그 틈에 현우도 끼여서 간신히 들어온다.\n- 42. 난민촌 입구 – N\n- 진입로마다 검문소 설치했습니다. 감시 드론도 띄웠구요.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_camp_gate_b39179.png",
  "asset_id": "0dcec927-ae46-47a9-b35a-9da3b801a1fa",
  "input_asset_ids": [
   "82c158c8-bc9e-4527-ab9e-386ee7694053"
  ],
  "origin_tag": "S42sh2",
  "place_text": "At the muddy entrance road of the refugee settlement at night, beside an opened body bag.",
  "origin_inputs": {
   "place_text": "At the muddy entrance road of the refugee settlement at night, beside an opened body bag.",
   "time_of_day_en": "night",
   "conti_asset_id": "82c158c8-bc9e-4527-ab9e-386ee7694053"
  }
 },
 "S42sh2::bgfirst_bg": {
  "input_fingerprint": "7b97900962444085",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박철진이 보디백 안의 창백한 구도환의 시체를 향해 턱을 치켜들고 무심하게 내려다보는 시선.\n\nLOCATION (lock): At the muddy entrance road of the refugee settlement at night, beside an opened body bag.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Opened body bag around the exposed face in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Body bag (Unzipped sufficiently to expose 구도환's face) — The opening faces upward and is visible obliquely from beside the head; used as Anchors the low foreground and connects the corpse to 박철진's downward attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime ambient illumination distinguishes 구도환's pale face and 박철진's expression without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박철진이 보디백 안의 창백한 구도환의 시체를 향해 턱을 치켜들고 무심하게 내려다보는 시선.\n\nLOCATION (lock): At the muddy entrance road of the refugee settlement at night, beside an opened body bag.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Opened body bag around the exposed face in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Body bag (Unzipped sufficiently to expose 구도환's face) — The opening faces upward and is visible obliquely from beside the head; used as Anchors the low foreground and connects the corpse to 박철진's downward attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime ambient illumination distinguishes 구도환's pale face and 박철진's expression without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S42sh2__bgfirst_bg.png",
  "asset_id": "74676c16-4bd5-425e-b8ed-a3cf6c9ebedf",
  "input_asset_ids": [
   "82c158c8-bc9e-4527-ab9e-386ee7694053",
   "0dcec927-ae46-47a9-b35a-9da3b801a1fa"
  ]
 },
 "S42sh2": {
  "input_fingerprint": "eb5826a4658bf142",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 보디백 안의 창백한 구도환의 시체를 향해 턱을 치켜들고 무심하게 내려다보는 시선.\n\nLOCATION (lock): At the muddy entrance road of the refugee settlement at night, beside an opened body bag. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Opened body bag around the exposed face in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Body bag (Unzipped sufficiently to expose 구도환's face) — The opening faces upward and is visible obliquely from beside the head; used as Anchors the low foreground and connects the corpse to 박철진's downward attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime ambient illumination distinguishes 구도환's pale face and 박철진's expression without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Dohwan's corpse is enclosed in a body bag with the opened zipper exposing his pale face for inspection from above. His exact head angle, torso orientation, limb positions, and the surface supporting the bag are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At the settlement entrance at night, a body bag lies with its zipper opened to expose the face inside. 박철진: He is standing over the opened body bag.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 보디백 안의 창백한 구도환의 시체를 향해 턱을 치켜들고 무심하게 내려다보는 시선.\n\nLOCATION (lock): At the muddy entrance road of the refugee settlement at night, beside an opened body bag. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Opened body bag around the exposed face in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Body bag (Unzipped sufficiently to expose 구도환's face) — The opening faces upward and is visible obliquely from beside the head; used as Anchors the low foreground and connects the corpse to 박철진's downward attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime ambient illumination distinguishes 구도환's pale face and 박철진's expression without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Dohwan's corpse is enclosed in a body bag with the opened zipper exposing his pale face for inspection from above. His exact head angle, torso orientation, limb positions, and the surface supporting the bag are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At the settlement entrance at night, a body bag lies with its zipper opened to expose the face inside. 박철진: He is standing over the opened body bag.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 보디백 안의 창백한 구도환의 시체를 향해 턱을 치켜들고 무심하게 내려다보는 시선.\n\nLOCATION (lock): At the muddy entrance road of the refugee settlement at night, beside an opened body bag. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Opened body bag around the exposed face in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Body bag (Unzipped sufficiently to expose 구도환's face) — The opening faces upward and is visible obliquely from beside the head; used as Anchors the low foreground and connects the corpse to 박철진's downward attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime ambient illumination distinguishes 구도환's pale face and 박철진's expression without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Dohwan's corpse is enclosed in a body bag with the opened zipper exposing his pale face for inspection from above. His exact head angle, torso orientation, limb positions, and the surface supporting the bag are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At the settlement entrance at night, a body bag lies with its zipper opened to expose the face inside. 박철진: He is standing over the opened body bag.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S42sh2__bgfirst_bg.png",
     "asset_id": "74676c16-4bd5-425e-b8ed-a3cf6c9ebedf",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S42sh2.png",
     "asset_id": "82c158c8-bc9e-4527-ab9e-386ee7694053",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822047>",
     "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_camp_gate_b39179.png",
     "asset_id": "0dcec927-ae46-47a9-b35a-9da3b801a1fa",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822047>",
     "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "박철진은 턱을 위로 치켜든 채 화면 좌측 하단에 위치한 보디백 안의 얼굴을 시선으로 향하고 있음.",
    "built_space": "야간의 진흙탕 길, 철조망, 초소, 가로등, 배경의 텐트 등 참조 사진의 난민촌 입구 환경이 정확히 구성됨.",
    "entities": "박철진(참조 이미지의 얼굴 및 헌병 완장이 포함된 군복 착용 일치), 얼굴이 노출된 채 보디백 안에 누워있는 시체.",
    "hard_violations": [
     "[gpt-high] 박철진의 완장에 영문 ‘MP’와 한글 일부가 읽히므로, 이미지 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반한다."
    ],
    "physics": "박철진은 두 발로 진흙 바닥에 안정적으로 서 있으며, 시체는 중력에 순응하여 바닥에 뉘여 있음."
   },
   {
    "label": "B",
    "direction": "박철진은 턱을 들지 않고 고개를 아래로 숙여 바닥에 놓인 보디백 전체를 내려다봄.",
    "built_space": "야간 진흙탕 길과 초소, 철조망, 조명 등 지정된 장소의 배경 요소들이 적절히 배치됨.",
    "entities": "박철진(얼굴 및 군복은 유사하나 헌병 완장 누락됨), 전신이 드러나게 바닥에 놓인 보디백과 시체.",
    "hard_violations": [],
    "physics": "박철진은 지면에 두 발을 딛고 서 있으며, 보디백은 바닥에 완전히 밀착되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 구도(좌측 하단 전경에 열린 보디백 배치)와 '턱을 치켜들고' 있는 구체적인 포즈, 캐릭터의 복장(완장 포함)을 가장 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시된 미디엄 샷과 전경 프레이밍을 무시하고 넓은 풀 샷으로 렌더링했으며, 턱을 치켜든 포즈와 참조 이미지의 완장 디테일이 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진은 턱을 위로 치켜든 채 화면 좌측 하단에 위치한 보디백 안의 얼굴을 시선으로 향하고 있음.",
        "built_space": "야간의 진흙탕 길, 철조망, 초소, 가로등, 배경의 텐트 등 참조 사진의 난민촌 입구 환경이 정확히 구성됨.",
        "entities": "박철진(참조 이미지의 얼굴 및 헌병 완장이 포함된 군복 착용 일치), 얼굴이 노출된 채 보디백 안에 누워있는 시체.",
        "hard_violations": [],
        "physics": "박철진은 두 발로 진흙 바닥에 안정적으로 서 있으며, 시체는 중력에 순응하여 바닥에 뉘여 있음."
       },
       {
        "label": "B",
        "direction": "박철진은 턱을 들지 않고 고개를 아래로 숙여 바닥에 놓인 보디백 전체를 내려다봄.",
        "built_space": "야간 진흙탕 길과 초소, 철조망, 조명 등 지정된 장소의 배경 요소들이 적절히 배치됨.",
        "entities": "박철진(얼굴 및 군복은 유사하나 헌병 완장 누락됨), 전신이 드러나게 바닥에 놓인 보디백과 시체.",
        "hard_violations": [],
        "physics": "박철진은 지면에 두 발을 딛고 서 있으며, 보디백은 바닥에 완전히 밀착되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 구도(좌측 하단 전경에 열린 보디백 배치)와 '턱을 치켜들고' 있는 구체적인 포즈, 캐릭터의 복장(완장 포함)을 가장 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시된 미디엄 샷과 전경 프레이밍을 무시하고 넓은 풀 샷으로 렌더링했으며, 턱을 치켜든 포즈와 참조 이미지의 완장 디테일이 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박철진은 턱을 위로 치켜든 채 화면 좌측 하단에 위치한 보디백 안의 얼굴을 시선으로 향하고 있음.",
        "built_space": "야간의 진흙탕 길, 철조망, 초소, 가로등, 배경의 텐트 등 참조 사진의 난민촌 입구 환경이 정확히 구성됨.",
        "entities": "박철진(참조 이미지의 얼굴 및 헌병 완장이 포함된 군복 착용 일치), 얼굴이 노출된 채 보디백 안에 누워있는 시체.",
        "hard_violations": [],
        "physics": "박철진은 두 발로 진흙 바닥에 안정적으로 서 있으며, 시체는 중력에 순응하여 바닥에 뉘여 있음."
       },
       {
        "label": "B",
        "direction": "박철진은 턱을 들지 않고 고개를 아래로 숙여 바닥에 놓인 보디백 전체를 내려다봄.",
        "built_space": "야간 진흙탕 길과 초소, 철조망, 조명 등 지정된 장소의 배경 요소들이 적절히 배치됨.",
        "entities": "박철진(얼굴 및 군복은 유사하나 헌병 완장 누락됨), 전신이 드러나게 바닥에 놓인 보디백과 시체.",
        "hard_violations": [],
        "physics": "박철진은 지면에 두 발을 딛고 서 있으며, 보디백은 바닥에 완전히 밀착되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "시체로 향하는 하향 시선과 장소·지지 관계는 맞지만, 미디엄 숏 대신 전신 와이드 숏이며 턱을 치켜드는 연기와 기준 모자가 빠졌다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "왼쪽 아래 얼굴과 치켜든 턱은 구현했지만 시체를 내려다보는 시선이 불명확하고 전신 구도이며, 완장의 읽히는 글자가 명시적 금지 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 얼굴과 눈은 왼쪽 아래 보디백 속 구도환의 얼굴 쪽으로 내려가 있다. 시선 대상은 맞지만 턱까지 내려가 있어 ‘턱을 치켜들고’라는 연기는 약하다. 구도환은 눈을 감고 얼굴을 위로 향한다. 무기나 이동하는 물체는 없다.",
        "built_space": "왼쪽 큰 경비 건물 한 채, 중앙의 작은 초소 한 채, 그 사이 출입문 한 곳이 보인다. 왼쪽 철조망 울타리와 콘크리트 기단, 오른쪽 천막 열과 울타리, 확성기·감시카메라가 달린 기둥들과 가로등이 장소 사진의 배치를 따른다. 박철진은 진흙길 위 보디백의 오른쪽 뒤에 서 있다. 물웅덩이의 가로등 반사는 자연스럽다. 다만 인물 전신과 보디백 전체를 넣은 넓은 구도로, 요구된 미디엄 숏이 아니다.",
        "entities": "살아 있는 중년 남성 한 명과 보디백에 누운 창백한 남성 시체 한 명만 보인다. 박철진의 짧은 짙은 머리, 회색 섞인 수염, 얼굴과 체격은 기준 인물에 대체로 가깝고 올리브색 군복과 검은 부츠도 맞는다. 기준의 모자와 검은 완장은 보이지 않는다. 구도환의 별도 얼굴 기준은 없으며, 창백한 얼굴이 지퍼가 열린 검은 보디백 안에 있다는 조건은 충족한다. 읽히는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "박철진은 진흙길에 선 두 다리로 몸을 지지하며 팔은 자연스럽게 내려져 있다. 부츠 일부가 보디백에 가리지만 공중에 뜬 모습은 아니다. 보디백은 지면에 놓여 있고 구도환의 머리와 몸은 그 안에서 바닥의 지지를 받는다. 시체가 스스로 신체를 들어 올리는 부분이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "박철진은 턱을 들고 머리를 왼쪽으로 기울였다. 눈은 화면 왼쪽을 향하지만, 왼쪽 아래 가까운 시체 얼굴에 시선이 꽂힌 모습은 명확하지 않으며 그보다 높은 곳을 보는 듯하다. 구도환은 눈을 감은 채 얼굴을 위로 향한다. 보디백의 열린 면도 위를 향한다.",
        "built_space": "왼쪽 큰 경비 건물 한 채와 중앙 작은 초소 한 채, 출입문 한 곳, 왼쪽 철조망과 오른쪽 천막 열이 보이며 기준 장소의 구조와 재료를 잘 유지한다. 확성기·감시카메라 기둥 및 가로등도 대응하는 위치에 있다. 카메라는 시체 머리 옆 지면 가까이에 있고 박철진은 보디백 너머 오른쪽 진흙길에 서 있다. 얼굴의 큰 전경 크기는 가까운 거리로 설명되지만, 박철진의 부츠까지 포함한 전신 구도라 미디엄 숏 조건은 충족하지 않는다.",
        "entities": "중년 남성 박철진 한 명과 창백한 남성 시체 구도환 한 명이 보인다. 박철진의 수염, 얼굴 윤곽과 올리브색 군복은 기준에 대체로 가깝지만 모자는 없다. 검은 완장은 있으나 흰색 한글 일부와 영문 ‘MP’가 읽힌다. 구도환의 얼굴은 검은 보디백의 열린 지퍼 사이로 노출되어 있고 추가 인물은 없다.",
        "hard_violations": [
         "박철진의 완장에 영문 ‘MP’와 한글 일부가 읽히므로, 이미지 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반한다."
        ],
        "physics": "박철진의 두 부츠는 진흙길에 닿고, 약간 기울어진 몸은 벌린 다리로 지지된다. 양손은 주머니에 넣은 자세로 보인다. 구도환의 뒤통수와 목은 지면에 놓인 보디백 내부에 기대어 있으며, 몸을 능동적으로 들어 올리는 모습은 없다. 보디백의 접힌 천도 지면과 내부 신체에 받쳐져 있어 지지 없는 부유는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "시체로 향하는 하향 시선과 장소·지지 관계는 맞지만, 미디엄 숏 대신 전신 와이드 숏이며 턱을 치켜드는 연기와 기준 모자가 빠졌다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "왼쪽 아래 얼굴과 치켜든 턱은 구현했지만 시체를 내려다보는 시선이 불명확하고 전신 구도이며, 완장의 읽히는 글자가 명시적 금지 조건을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박철진의 얼굴과 눈은 왼쪽 아래 보디백 속 구도환의 얼굴 쪽으로 내려가 있다. 시선 대상은 맞지만 턱까지 내려가 있어 ‘턱을 치켜들고’라는 연기는 약하다. 구도환은 눈을 감고 얼굴을 위로 향한다. 무기나 이동하는 물체는 없다.",
        "built_space": "왼쪽 큰 경비 건물 한 채, 중앙의 작은 초소 한 채, 그 사이 출입문 한 곳이 보인다. 왼쪽 철조망 울타리와 콘크리트 기단, 오른쪽 천막 열과 울타리, 확성기·감시카메라가 달린 기둥들과 가로등이 장소 사진의 배치를 따른다. 박철진은 진흙길 위 보디백의 오른쪽 뒤에 서 있다. 물웅덩이의 가로등 반사는 자연스럽다. 다만 인물 전신과 보디백 전체를 넣은 넓은 구도로, 요구된 미디엄 숏이 아니다.",
        "entities": "살아 있는 중년 남성 한 명과 보디백에 누운 창백한 남성 시체 한 명만 보인다. 박철진의 짧은 짙은 머리, 회색 섞인 수염, 얼굴과 체격은 기준 인물에 대체로 가깝고 올리브색 군복과 검은 부츠도 맞는다. 기준의 모자와 검은 완장은 보이지 않는다. 구도환의 별도 얼굴 기준은 없으며, 창백한 얼굴이 지퍼가 열린 검은 보디백 안에 있다는 조건은 충족한다. 읽히는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "박철진은 진흙길에 선 두 다리로 몸을 지지하며 팔은 자연스럽게 내려져 있다. 부츠 일부가 보디백에 가리지만 공중에 뜬 모습은 아니다. 보디백은 지면에 놓여 있고 구도환의 머리와 몸은 그 안에서 바닥의 지지를 받는다. 시체가 스스로 신체를 들어 올리는 부분이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "박철진은 턱을 들고 머리를 왼쪽으로 기울였다. 눈은 화면 왼쪽을 향하지만, 왼쪽 아래 가까운 시체 얼굴에 시선이 꽂힌 모습은 명확하지 않으며 그보다 높은 곳을 보는 듯하다. 구도환은 눈을 감은 채 얼굴을 위로 향한다. 보디백의 열린 면도 위를 향한다.",
        "built_space": "왼쪽 큰 경비 건물 한 채와 중앙 작은 초소 한 채, 출입문 한 곳, 왼쪽 철조망과 오른쪽 천막 열이 보이며 기준 장소의 구조와 재료를 잘 유지한다. 확성기·감시카메라 기둥 및 가로등도 대응하는 위치에 있다. 카메라는 시체 머리 옆 지면 가까이에 있고 박철진은 보디백 너머 오른쪽 진흙길에 서 있다. 얼굴의 큰 전경 크기는 가까운 거리로 설명되지만, 박철진의 부츠까지 포함한 전신 구도라 미디엄 숏 조건은 충족하지 않는다.",
        "entities": "중년 남성 박철진 한 명과 창백한 남성 시체 구도환 한 명이 보인다. 박철진의 수염, 얼굴 윤곽과 올리브색 군복은 기준에 대체로 가깝지만 모자는 없다. 검은 완장은 있으나 흰색 한글 일부와 영문 ‘MP’가 읽힌다. 구도환의 얼굴은 검은 보디백의 열린 지퍼 사이로 노출되어 있고 추가 인물은 없다.",
        "hard_violations": [
         "박철진의 완장에 영문 ‘MP’와 한글 일부가 읽히므로, 이미지 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반한다."
        ],
        "physics": "박철진의 두 부츠는 진흙길에 닿고, 약간 기울어진 몸은 벌린 다리로 지지된다. 양손은 주머니에 넣은 자세로 보인다. 구도환의 뒤통수와 목은 지면에 놓인 보디백 내부에 기대어 있으며, 몸을 능동적으로 들어 올리는 모습은 없다. 보디백의 접힌 천도 지면과 내부 신체에 받쳐져 있어 지지 없는 부유는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.429
   },
   "violations": {
    "A": [
     "[gpt-high] 박철진의 완장에 영문 ‘MP’와 한글 일부가 읽히므로, 이미지 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1429
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "요구된 구도(좌측 하단 전경에 열린 보디백 배치)와 '턱을 치켜들고' 있는 구체적인 포즈, 캐릭터의 복장(완장 포함)을 가장 정확하게 구현했습니다.  ★위반: [gpt-high] 박철진의 완장에 영문 ‘MP’와 한글 일부가 읽히므로, 이미지 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1429,
    "verdict_ko": "지시된 미디엄 샷과 전경 프레이밍을 무시하고 넓은 풀 샷으로 렌더링했으며, 턱을 치켜든 포즈와 참조 이미지의 완장 디테일이 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_camp_gate_b39179.png",
    "asset_id": "0dcec927-ae46-47a9-b35a-9da3b801a1fa",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc6-3e09-7ca1-b876-52710aad84e6",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S42sh2__bgfirst_bg.png",
   "bg_asset_id": "74676c16-4bd5-425e-b8ed-a3cf6c9ebedf",
   "bg_record_key": "S42sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "camp_gate",
   "groupbg_asset_id": "0dcec927-ae46-47a9-b35a-9da3b801a1fa"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S42sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:38:28.044484+00:00",
  "fingerprint": "686c6d4be7d7482017d9d249ad35cccd9b40ac59a38de8aea6f5294fbea661a3",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S42sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S42sh2_sel.png",
  "source_sha256": "872fea9c167785afb3c59aee5328000276afb0fe32ba45b2f532fe01de82ef99",
  "file": "S42sh2_cine.png",
  "staged_sha256": "2b834cabcf6eb0e6524f26224d23a34460d39d26b303410e88914db8c3eb31e1",
  "latency_ms": 11667
 },
 "S42sh10::signage": {
  "fp": "8842f42ef8d29ba6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S42sh10": {
  "input_fingerprint": "8d3e7bd2d5f56946",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 열린 자동차 문을 한 손으로 꽉 틀어쥔 채 차에 타려는 윤성찬을 막아선 찰나.\n\nLOCATION (lock): Beside the open passenger door of a luxury sedan stopped on the muddy refugee-settlement entrance road at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open door edge gripped between the men in the lower-center of the frame, foreground; Passenger opening behind the interrupted entrant in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open sedan passenger door (Held open by 박철진 during 윤성찬's attempted entry) — Its outer edge is nearest the camera, with the passenger opening visible beyond it; used as Creates the physical barrier and places the gripping hand between the two men; Sedan passenger opening (Accessible through the open door, with 윤성찬's entry interrupted) — Seen diagonally from outside rather than from within the vehicle; used as Shows the destination of 윤성찬's bent posture without allowing the car to overwhelm the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination with enough tonal separation to read the hand, door edge, and interrupted entry.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The body bag remains open at the settlement entrance. The luxury sedan is stopped nearby with a door open. 박철진: He is beside the sedan, holding its open door. 윤성찬: He is getting into the sedan with his cane and handkerchief still in his possession. Mud remains on the cane tip.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상); 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 열린 자동차 문을 한 손으로 꽉 틀어쥔 채 차에 타려는 윤성찬을 막아선 찰나.\n\nLOCATION (lock): Beside the open passenger door of a luxury sedan stopped on the muddy refugee-settlement entrance road at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open door edge gripped between the men in the lower-center of the frame, foreground; Passenger opening behind the interrupted entrant in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open sedan passenger door (Held open by 박철진 during 윤성찬's attempted entry) — Its outer edge is nearest the camera, with the passenger opening visible beyond it; used as Creates the physical barrier and places the gripping hand between the two men; Sedan passenger opening (Accessible through the open door, with 윤성찬's entry interrupted) — Seen diagonally from outside rather than from within the vehicle; used as Shows the destination of 윤성찬's bent posture without allowing the car to overwhelm the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination with enough tonal separation to read the hand, door edge, and interrupted entry.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The body bag remains open at the settlement entrance. The luxury sedan is stopped nearby with a door open. 박철진: He is beside the sedan, holding its open door. 윤성찬: He is getting into the sedan with his cane and handkerchief still in his possession. Mud remains on the cane tip.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상); 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 열린 자동차 문을 한 손으로 꽉 틀어쥔 채 차에 타려는 윤성찬을 막아선 찰나.\n\nLOCATION (lock): Beside the open passenger door of a luxury sedan stopped on the muddy refugee-settlement entrance road at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Open door edge gripped between the men in the lower-center of the frame, foreground; Passenger opening behind the interrupted entrant in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open sedan passenger door (Held open by 박철진 during 윤성찬's attempted entry) — Its outer edge is nearest the camera, with the passenger opening visible beyond it; used as Creates the physical barrier and places the gripping hand between the two men; Sedan passenger opening (Accessible through the open door, with 윤성찬's entry interrupted) — Seen diagonally from outside rather than from within the vehicle; used as Shows the destination of 윤성찬's bent posture without allowing the car to overwhelm the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination with enough tonal separation to read the hand, door edge, and interrupted entry.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The body bag remains open at the settlement entrance. The luxury sedan is stopped nearby with a door open. 박철진: He is beside the sedan, holding its open door. 윤성찬: He is getting into the sedan with his cane and handkerchief still in his possession. Mud remains on the cane tip.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상); 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "윤성찬은 카메라 렌즈를 정면으로 바라보고 있다. 박철진의 팔은 화면 우측에서 뻗어 나와 차 문 근처 허공을 향하고 있다.",
    "built_space": "야간의 진흙탕 길. 배경에 위병소와 철조망이 이전 샷과 동일한 위치에 배치되어 있다. 세단의 차 문이 열려 있다.",
    "entities": "윤성찬은 참조와 일치하는 정장 차림이다. 박철진은 팔 부분만 프레임에 들어왔다. 시신낭과 세단이 포함되어 있다.",
    "hard_violations": [
     "[gemini-pro] 지시된 카메라 프레이밍 및 피사체 배치(화면 하단 중앙에 위치해야 할 문, 중간 우측에 위치해야 할 승객석 입구)를 완전히 위반함",
     "[gemini-pro] 박철진의 손이 차 문을 쥐지 않고 허공에 떠 있어 물리적 지지 및 액션 지시를 위반함"
    ],
    "physics": "박철진의 손은 문을 잡지 않은 채 허공에 떠 있어 아무런 물리적 지지를 받지 못한다. 윤성찬은 지팡이와 손수건을 양손에 쥐고 서 있다."
   },
   {
    "label": "B",
    "direction": "박철진은 차를 타려는 윤성찬과 차 문 쪽을 내려다보고, 윤성찬은 차 내부를 향해 고개를 숙이고 있다.",
    "built_space": "야간의 진흙탕 길. 위병소, 텐트, 철조망 등 배경 구조물이 올바르게 위치해 있다. 우측 승객석 차 문이 열려 있으며, 문 외부가 카메라를 향하고 있어 기하학적 구조가 올바르다.",
    "entities": "두 인물 모두 참조 이미지의 얼굴과 체형에 일치하나, 박철진의 군복 깃에 임의로 별 계급장이 덧그려졌다. 세단이 존재하며, 시신낭은 프레임 밖으로 벗어나 보이지 않는다.",
    "hard_violations": [],
    "physics": "박철진의 왼손이 열린 차 문의 상단 틀을 단단히 쥐고 물리적으로 지탱하고 있다. 윤성찬은 오른손으로 진흙 묻은 지팡이를 짚고 차에 오르는 자세를 취하고 있으며, 왼손에 손수건을 쥐고 있으나 손가락의 묘사가 다소 뭉개져 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "요구된 화면 구도와 차 문을 쥐고 탑승을 막아서는 핵심 액션을 매우 충실히 구현했으나, 박철진의 군복 깃에 임의의 계급장이 추가되고 이전 샷의 시신낭이 앵글 밖으로 누락된 점이 감점 요인입니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지시된 카메라 프레이밍을 완전히 무시하였으며, 손이 문을 잡지 않고 허공에 떠 있고 인물이 카메라를 빤히 응시하여 주요 지침들을 다수 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬은 카메라 렌즈를 정면으로 바라보고 있다. 박철진의 팔은 화면 우측에서 뻗어 나와 차 문 근처 허공을 향하고 있다.",
        "built_space": "야간의 진흙탕 길. 배경에 위병소와 철조망이 이전 샷과 동일한 위치에 배치되어 있다. 세단의 차 문이 열려 있다.",
        "entities": "윤성찬은 참조와 일치하는 정장 차림이다. 박철진은 팔 부분만 프레임에 들어왔다. 시신낭과 세단이 포함되어 있다.",
        "hard_violations": [
         "지시된 카메라 프레이밍 및 피사체 배치(화면 하단 중앙에 위치해야 할 문, 중간 우측에 위치해야 할 승객석 입구)를 완전히 위반함",
         "박철진의 손이 차 문을 쥐지 않고 허공에 떠 있어 물리적 지지 및 액션 지시를 위반함"
        ],
        "physics": "박철진의 손은 문을 잡지 않은 채 허공에 떠 있어 아무런 물리적 지지를 받지 못한다. 윤성찬은 지팡이와 손수건을 양손에 쥐고 서 있다."
       },
       {
        "label": "B",
        "direction": "박철진은 차를 타려는 윤성찬과 차 문 쪽을 내려다보고, 윤성찬은 차 내부를 향해 고개를 숙이고 있다.",
        "built_space": "야간의 진흙탕 길. 위병소, 텐트, 철조망 등 배경 구조물이 올바르게 위치해 있다. 우측 승객석 차 문이 열려 있으며, 문 외부가 카메라를 향하고 있어 기하학적 구조가 올바르다.",
        "entities": "두 인물 모두 참조 이미지의 얼굴과 체형에 일치하나, 박철진의 군복 깃에 임의로 별 계급장이 덧그려졌다. 세단이 존재하며, 시신낭은 프레임 밖으로 벗어나 보이지 않는다.",
        "hard_violations": [],
        "physics": "박철진의 왼손이 열린 차 문의 상단 틀을 단단히 쥐고 물리적으로 지탱하고 있다. 윤성찬은 오른손으로 진흙 묻은 지팡이를 짚고 차에 오르는 자세를 취하고 있으며, 왼손에 손수건을 쥐고 있으나 손가락의 묘사가 다소 뭉개져 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "요구된 화면 구도와 차 문을 쥐고 탑승을 막아서는 핵심 액션을 매우 충실히 구현했으나, 박철진의 군복 깃에 임의의 계급장이 추가되고 이전 샷의 시신낭이 앵글 밖으로 누락된 점이 감점 요인입니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지시된 카메라 프레이밍을 완전히 무시하였으며, 손이 문을 잡지 않고 허공에 떠 있고 인물이 카메라를 빤히 응시하여 주요 지침들을 다수 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬은 카메라 렌즈를 정면으로 바라보고 있다. 박철진의 팔은 화면 우측에서 뻗어 나와 차 문 근처 허공을 향하고 있다.",
        "built_space": "야간의 진흙탕 길. 배경에 위병소와 철조망이 이전 샷과 동일한 위치에 배치되어 있다. 세단의 차 문이 열려 있다.",
        "entities": "윤성찬은 참조와 일치하는 정장 차림이다. 박철진은 팔 부분만 프레임에 들어왔다. 시신낭과 세단이 포함되어 있다.",
        "hard_violations": [
         "지시된 카메라 프레이밍 및 피사체 배치(화면 하단 중앙에 위치해야 할 문, 중간 우측에 위치해야 할 승객석 입구)를 완전히 위반함",
         "박철진의 손이 차 문을 쥐지 않고 허공에 떠 있어 물리적 지지 및 액션 지시를 위반함"
        ],
        "physics": "박철진의 손은 문을 잡지 않은 채 허공에 떠 있어 아무런 물리적 지지를 받지 못한다. 윤성찬은 지팡이와 손수건을 양손에 쥐고 서 있다."
       },
       {
        "label": "B",
        "direction": "박철진은 차를 타려는 윤성찬과 차 문 쪽을 내려다보고, 윤성찬은 차 내부를 향해 고개를 숙이고 있다.",
        "built_space": "야간의 진흙탕 길. 위병소, 텐트, 철조망 등 배경 구조물이 올바르게 위치해 있다. 우측 승객석 차 문이 열려 있으며, 문 외부가 카메라를 향하고 있어 기하학적 구조가 올바르다.",
        "entities": "두 인물 모두 참조 이미지의 얼굴과 체형에 일치하나, 박철진의 군복 깃에 임의로 별 계급장이 덧그려졌다. 세단이 존재하며, 시신낭은 프레임 밖으로 벗어나 보이지 않는다.",
        "hard_violations": [],
        "physics": "박철진의 왼손이 열린 차 문의 상단 틀을 단단히 쥐고 물리적으로 지탱하고 있다. 윤성찬은 오른손으로 진흙 묻은 지팡이를 짚고 차에 오르는 자세를 취하고 있으며, 왼손에 손수건을 쥐고 있으나 손가락의 묘사가 다소 뭉개져 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "두 남성을 담은 미디엄 숏에서 박철진의 문 가장자리 파지와 오른쪽 승차구로 몸을 굽힌 윤성찬을 명확히 연결해, 승차를 막는 찰나와 지정 배치를 가장 충실하게 구현했다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "문과 승차구는 보이지만 박철진이 팔만 남도록 잘리고 윤성찬이 중앙 왼쪽에 배치되며, 시선도 박철진이 아닌 카메라 쪽을 향해 요구된 대치 구도가 약해졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 박철진은 오른쪽 윤성찬과 문을 쥔 손 부근을 내려다본다. 윤성찬은 아래를 보면서 오른쪽 뒤의 열린 승차구 쪽으로 상체와 다리를 굽힌다. 시선과 몸의 방향이 승차 중단이라는 행동에 부합한다. 지팡이는 손에서 아래로 향한다.",
        "built_space": "전경에 열린 승객용 문 하나, 그 뒤 오른쪽에 승차구 하나와 밝은 뒷좌석이 보인다. 박철진은 문 바깥 왼쪽, 윤성찬은 문과 승차구 사이에 있다. 문 가장자리와 파지한 손이 두 사람 사이의 중앙 하단에 놓이며, 카메라는 차량 외부에서 승차구를 비스듬히 본다. 배경에는 왼쪽 철조망 울타리와 출입구 구조물, 중앙 경비초소 하나, 오른쪽 천막열 및 여러 가로등이 있어 이전 장소의 특징을 유지한다. 불가능한 반사나 중복 문은 보이지 않는다.",
        "entities": "보이는 인물은 중년 남성 박철진과 고령 남성 윤성찬 두 명이다. 박철진의 짧은 검은색·회색 머리, 수염, 체격과 올리브색 작업복은 이전 장면에 가깝고 모자도 없다. 윤성찬의 흰머리와 노년 얼굴, 짙은 남색 정장, 흰 셔츠와 넥타이는 참고와 대체로 일치한다. 윤성찬은 한 손에 지팡이, 다른 손에 구겨진 손수건을 들고 있다. 지팡이 끝과 시신 가방은 프레임 밖이므로 진흙 및 가방 상태는 확인할 수 없다. 고급 세단, 젖은 진흙길과 야간 조명이 보이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "박철진의 손가락이 문 상단 가장자리를 실제로 감싸고 있으며 손목과 올리브색 소매가 자연스럽게 이어진다. 문은 차량의 경첩으로 지지된다. 윤성찬의 허리와 무릎 굽힘은 차에 들어가다 멈춘 자세로 가능하고, 지팡이와 손수건 모두 손에 잡혀 있다. 두 사람의 발과 지팡이 접지점은 잘렸지만 공중에 떠 있는 자세나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "윤성찬은 몸을 차량 쪽으로 약간 굽혔으나 얼굴과 시선은 화면 오른쪽 박철진의 팔이 오는 방향이 아니라 카메라 가까운 왼쪽 전경을 향한다. 박철진의 얼굴은 보이지 않아 시선을 확인할 수 없다. 지팡이는 아래 진흙길을 향하지만, 윤성찬의 몸은 문 왼쪽 바깥에 머물러 승차구로 들어가는 동작이 A보다 덜 분명하다.",
        "built_space": "열린 승객용 문 하나가 화면 중앙에서 크게 솟고, 승차구 하나와 좌석이 오른쪽에 보인다. 윤성찬은 중앙 왼쪽에 있고 박철진은 오른쪽에서 들어오는 팔과 몸통 일부만 보인다. 손은 중앙 하단에 있지만, 두 남성을 배치한 미디엄 숏보다는 문과 잘린 팔을 강조한 구도다. 왼쪽 배경에는 철조망 울타리, 경비초소 하나, 출입구 가로 구조물과 가로등이 보인다. 열린 시신 가방은 왼쪽 진흙 바닥에 놓여 있다. 중복 설비나 명백히 불가능한 반사는 보이지 않는다.",
        "entities": "윤성찬의 흰머리, 고령 남성 얼굴, 남색 정장과 흰 셔츠는 참고에 가깝다. 지팡이와 손수건을 각각 손에 갖고 있다. 박철진은 성인 남성의 손과 올리브색 소매만 확인되므로 얼굴·머리의 일치는 판단할 수 없다. 손목에는 이전 장면에서 확인되지 않는 큰 시계가 있다. 배경의 열린 가방에는 이전 장면의 시신에 해당하는 얼굴과 몸이 보이지만, 작게 보여 세부 자세의 정확한 지속 여부는 확인하기 어렵다. 세단과 진흙길, 야간 환경은 맞고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "윤성찬은 지팡이 손잡이를 잡고 있으며 지팡이는 아래 진흙 지면까지 이어진다. 손수건도 반대 손에 잡혀 있다. 박철진의 손은 문 잠금장치 부근 가장자리에 닿지만 손가락이 넓게 펴져 있어 꽉 틀어쥔 파지보다 손바닥으로 버티는 모습에 가깝다. 문은 경첩으로, 배경의 시신 가방은 지면으로 지지된다. 발은 프레임 밖이며 명백한 부유나 불가능한 관절 자세는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "두 남성을 담은 미디엄 숏에서 박철진의 문 가장자리 파지와 오른쪽 승차구로 몸을 굽힌 윤성찬을 명확히 연결해, 승차를 막는 찰나와 지정 배치를 가장 충실하게 구현했다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "문과 승차구는 보이지만 박철진이 팔만 남도록 잘리고 윤성찬이 중앙 왼쪽에 배치되며, 시선도 박철진이 아닌 카메라 쪽을 향해 요구된 대치 구도가 약해졌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 박철진은 오른쪽 윤성찬과 문을 쥔 손 부근을 내려다본다. 윤성찬은 아래를 보면서 오른쪽 뒤의 열린 승차구 쪽으로 상체와 다리를 굽힌다. 시선과 몸의 방향이 승차 중단이라는 행동에 부합한다. 지팡이는 손에서 아래로 향한다.",
        "built_space": "전경에 열린 승객용 문 하나, 그 뒤 오른쪽에 승차구 하나와 밝은 뒷좌석이 보인다. 박철진은 문 바깥 왼쪽, 윤성찬은 문과 승차구 사이에 있다. 문 가장자리와 파지한 손이 두 사람 사이의 중앙 하단에 놓이며, 카메라는 차량 외부에서 승차구를 비스듬히 본다. 배경에는 왼쪽 철조망 울타리와 출입구 구조물, 중앙 경비초소 하나, 오른쪽 천막열 및 여러 가로등이 있어 이전 장소의 특징을 유지한다. 불가능한 반사나 중복 문은 보이지 않는다.",
        "entities": "보이는 인물은 중년 남성 박철진과 고령 남성 윤성찬 두 명이다. 박철진의 짧은 검은색·회색 머리, 수염, 체격과 올리브색 작업복은 이전 장면에 가깝고 모자도 없다. 윤성찬의 흰머리와 노년 얼굴, 짙은 남색 정장, 흰 셔츠와 넥타이는 참고와 대체로 일치한다. 윤성찬은 한 손에 지팡이, 다른 손에 구겨진 손수건을 들고 있다. 지팡이 끝과 시신 가방은 프레임 밖이므로 진흙 및 가방 상태는 확인할 수 없다. 고급 세단, 젖은 진흙길과 야간 조명이 보이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "박철진의 손가락이 문 상단 가장자리를 실제로 감싸고 있으며 손목과 올리브색 소매가 자연스럽게 이어진다. 문은 차량의 경첩으로 지지된다. 윤성찬의 허리와 무릎 굽힘은 차에 들어가다 멈춘 자세로 가능하고, 지팡이와 손수건 모두 손에 잡혀 있다. 두 사람의 발과 지팡이 접지점은 잘렸지만 공중에 떠 있는 자세나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "윤성찬은 몸을 차량 쪽으로 약간 굽혔으나 얼굴과 시선은 화면 오른쪽 박철진의 팔이 오는 방향이 아니라 카메라 가까운 왼쪽 전경을 향한다. 박철진의 얼굴은 보이지 않아 시선을 확인할 수 없다. 지팡이는 아래 진흙길을 향하지만, 윤성찬의 몸은 문 왼쪽 바깥에 머물러 승차구로 들어가는 동작이 A보다 덜 분명하다.",
        "built_space": "열린 승객용 문 하나가 화면 중앙에서 크게 솟고, 승차구 하나와 좌석이 오른쪽에 보인다. 윤성찬은 중앙 왼쪽에 있고 박철진은 오른쪽에서 들어오는 팔과 몸통 일부만 보인다. 손은 중앙 하단에 있지만, 두 남성을 배치한 미디엄 숏보다는 문과 잘린 팔을 강조한 구도다. 왼쪽 배경에는 철조망 울타리, 경비초소 하나, 출입구 가로 구조물과 가로등이 보인다. 열린 시신 가방은 왼쪽 진흙 바닥에 놓여 있다. 중복 설비나 명백히 불가능한 반사는 보이지 않는다.",
        "entities": "윤성찬의 흰머리, 고령 남성 얼굴, 남색 정장과 흰 셔츠는 참고에 가깝다. 지팡이와 손수건을 각각 손에 갖고 있다. 박철진은 성인 남성의 손과 올리브색 소매만 확인되므로 얼굴·머리의 일치는 판단할 수 없다. 손목에는 이전 장면에서 확인되지 않는 큰 시계가 있다. 배경의 열린 가방에는 이전 장면의 시신에 해당하는 얼굴과 몸이 보이지만, 작게 보여 세부 자세의 정확한 지속 여부는 확인하기 어렵다. 세단과 진흙길, 야간 환경은 맞고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "윤성찬은 지팡이 손잡이를 잡고 있으며 지팡이는 아래 진흙 지면까지 이어진다. 손수건도 반대 손에 잡혀 있다. 박철진의 손은 문 잠금장치 부근 가장자리에 닿지만 손가락이 넓게 펴져 있어 꽉 틀어쥔 파지보다 손바닥으로 버티는 모습에 가깝다. 문은 경첩으로, 배경의 시신 가방은 지면으로 지지된다. 발은 프레임 밖이며 명백한 부유나 불가능한 관절 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.806,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.556,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지시된 카메라 프레이밍 및 피사체 배치(화면 하단 중앙에 위치해야 할 문, 중간 우측에 위치해야 할 승객석 입구)를 완전히 위반함",
     "[gemini-pro] 박철진의 손이 차 문을 쥐지 않고 허공에 떠 있어 물리적 지지 및 액션 지시를 위반함"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 556
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요구된 화면 구도와 차 문을 쥐고 탑승을 막아서는 핵심 액션을 매우 충실히 구현했으나, 박철진의 군복 깃에 임의의 계급장이 추가되고 이전 샷의 시신낭이 앵글 밖으로 누락된 점이 감점 요인입니다."
   },
   {
    "label": "A",
    "score": 556,
    "verdict_ko": "지시된 카메라 프레이밍을 완전히 무시하였으며, 손이 문을 잡지 않고 허공에 떠 있고 인물이 카메라를 빤히 응시하여 주요 지침들을 다수 위반했습니다.  ★위반: [gemini-pro] 지시된 카메라 프레이밍 및 피사체 배치(화면 하단 중앙에 위치해야 할 문, 중간 우측에 위치해야 할 승객석 입구)를 완전히 위반함 / [gemini-pro] 박철진의 손이 차 문을 쥐지 않고 허공에 떠 있어 물리적 지지 및 액션 지시를 위반함"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S42sh2_sel.png",
    "asset_id": "d3b210f2-1401-4dfa-8e54-95cdbcbd2582",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:781789>",
    "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc7-00c8-7333-96aa-2a95f39c7104",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S42sh2"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S42sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:38:40.062480+00:00",
  "fingerprint": "23f043221b7c8616fd2fd33255db269918964c6a2c9a305a05e9006da86bec13",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S42sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S42sh10_sel.png",
  "source_sha256": "93bddfaf2f6a6e4dd91c2fe3a5ff11debcbb9b0ccb83b1d23f184891f9ca8079",
  "file": "S42sh10_cine.png",
  "staged_sha256": "79043f4fd5c8aa1c1ed8ac5080f5cc957f6abba25e6d4709145aebd0ff8f2b27",
  "latency_ms": 13122
 },
 "S42sh18::signage": {
  "fp": "8fcb3f7802e2f83f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S42sh18": {
  "input_fingerprint": "fe81838a70cf8da9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 뻣뻣하게 선 채 동공이 크게 확장되고 안색이 하얗게 질린 당황한 표정 클로즈업.\n\nLOCATION (lock): On the muddy entrance road outside the refugee settlement after the sedan has departed at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged nighttime ambient light reveals the blanched complexion and widened eyes while preserving restrained background contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The luxury sedan has departed with its door closed. The opened body bag remains at the settlement entrance. 박철진: He remains at the entrance, his earlier smile gone and his face newly tense.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 뻣뻣하게 선 채 동공이 크게 확장되고 안색이 하얗게 질린 당황한 표정 클로즈업.\n\nLOCATION (lock): On the muddy entrance road outside the refugee settlement after the sedan has departed at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged nighttime ambient light reveals the blanched complexion and widened eyes while preserving restrained background contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The luxury sedan has departed with its door closed. The opened body bag remains at the settlement entrance. 박철진: He remains at the entrance, his earlier smile gone and his face newly tense.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 뻣뻣하게 선 채 동공이 크게 확장되고 안색이 하얗게 질린 당황한 표정 클로즈업.\n\nLOCATION (lock): On the muddy entrance road outside the refugee settlement after the sedan has departed at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged nighttime ambient light reveals the blanched complexion and widened eyes while preserving restrained background contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The luxury sedan has departed with its door closed. The opened body bag remains at the settlement entrance. 박철진: He remains at the entrance, his earlier smile gone and his face newly tense.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 정면(카메라 렌즈 방향)을 향해 고정되어 있다.",
    "built_space": "이전 샷의 배경을 완벽히 유지하여, 왼쪽의 철조망과 경비 초소, 오른쪽 배경의 난민촌 천막, 바닥의 진흙탕 길이 정확한 위치와 원근감으로 배치되어 있다.",
    "entities": "박철진 역의 인물이 클로즈업으로 잡혀 있으며, 지문의 요구대로 안색이 창백해지고 양쪽 눈의 동공이 비정상적으로 크게 확장된 당황한 표정을 정확히 보여준다. 군복 의상과 수염, 머리 모양은 레퍼런스와 일치한다.",
    "hard_violations": [],
    "physics": "화면에 하반신이 보이지 않지만, 바닥에 발을 딛고 서서 몸이 뻣뻣하게 굳은 자세를 안정적으로 유지하고 있다."
   },
   {
    "label": "B",
    "direction": "인물의 시선은 화면 우측 하단을 향하고 있다.",
    "built_space": "이전 샷과 동일하게 왼쪽의 철조망과 초소, 오른쪽의 천막과 진흙 길 등 공간적 요소들이 올바르게 배치되어 있다.",
    "entities": "박철진의 외모와 군복 의상은 레퍼런스와 일치하나, 요구된 '동공 확장'이 잘 나타나지 않고 안색도 상대적으로 덜 창백하다. 화면 우측 가장자리에 은색 세단의 후면부 일부가 잡혀 있다.",
    "hard_violations": [
     "[gemini-pro] 지문에서 세단이 이미 떠났다고 명시(the sedan has departed)했으나, 화면 우측 가장자리에 은색 차량의 일부가 버젓이 남아 있음.",
     "[gpt-high] 떠났어야 할 세단의 후미가 오른쪽 전경에 남아 있어, 세단 출발 이후라는 명시적 배치 상태를 위반한다."
    ],
    "physics": "하반신은 프레임 밖이나, 지면에 서서 굳어있는 자세가 무리 없이 표현되어 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지문에 명시된 '크게 확장된 동공'과 '하얗게 질린 안색', '뻣뻣하게 선 자세'를 매우 정확하게 묘사했으며, 세단이 떠난 후의 텅 빈 배경을 올바르게 반영했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "표정 연출에서 동공 확장이 뚜렷하지 않고, 지문에서 이미 떠났다고 명시된 세단 차량의 일부가 화면 우측에 남아 있어 서사적 설정과 충돌합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 정면(카메라 렌즈 방향)을 향해 고정되어 있다.",
        "built_space": "이전 샷의 배경을 완벽히 유지하여, 왼쪽의 철조망과 경비 초소, 오른쪽 배경의 난민촌 천막, 바닥의 진흙탕 길이 정확한 위치와 원근감으로 배치되어 있다.",
        "entities": "박철진 역의 인물이 클로즈업으로 잡혀 있으며, 지문의 요구대로 안색이 창백해지고 양쪽 눈의 동공이 비정상적으로 크게 확장된 당황한 표정을 정확히 보여준다. 군복 의상과 수염, 머리 모양은 레퍼런스와 일치한다.",
        "hard_violations": [],
        "physics": "화면에 하반신이 보이지 않지만, 바닥에 발을 딛고 서서 몸이 뻣뻣하게 굳은 자세를 안정적으로 유지하고 있다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 화면 우측 하단을 향하고 있다.",
        "built_space": "이전 샷과 동일하게 왼쪽의 철조망과 초소, 오른쪽의 천막과 진흙 길 등 공간적 요소들이 올바르게 배치되어 있다.",
        "entities": "박철진의 외모와 군복 의상은 레퍼런스와 일치하나, 요구된 '동공 확장'이 잘 나타나지 않고 안색도 상대적으로 덜 창백하다. 화면 우측 가장자리에 은색 세단의 후면부 일부가 잡혀 있다.",
        "hard_violations": [
         "지문에서 세단이 이미 떠났다고 명시(the sedan has departed)했으나, 화면 우측 가장자리에 은색 차량의 일부가 버젓이 남아 있음."
        ],
        "physics": "하반신은 프레임 밖이나, 지면에 서서 굳어있는 자세가 무리 없이 표현되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지문에 명시된 '크게 확장된 동공'과 '하얗게 질린 안색', '뻣뻣하게 선 자세'를 매우 정확하게 묘사했으며, 세단이 떠난 후의 텅 빈 배경을 올바르게 반영했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "표정 연출에서 동공 확장이 뚜렷하지 않고, 지문에서 이미 떠났다고 명시된 세단 차량의 일부가 화면 우측에 남아 있어 서사적 설정과 충돌합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 정면(카메라 렌즈 방향)을 향해 고정되어 있다.",
        "built_space": "이전 샷의 배경을 완벽히 유지하여, 왼쪽의 철조망과 경비 초소, 오른쪽 배경의 난민촌 천막, 바닥의 진흙탕 길이 정확한 위치와 원근감으로 배치되어 있다.",
        "entities": "박철진 역의 인물이 클로즈업으로 잡혀 있으며, 지문의 요구대로 안색이 창백해지고 양쪽 눈의 동공이 비정상적으로 크게 확장된 당황한 표정을 정확히 보여준다. 군복 의상과 수염, 머리 모양은 레퍼런스와 일치한다.",
        "hard_violations": [],
        "physics": "화면에 하반신이 보이지 않지만, 바닥에 발을 딛고 서서 몸이 뻣뻣하게 굳은 자세를 안정적으로 유지하고 있다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 화면 우측 하단을 향하고 있다.",
        "built_space": "이전 샷과 동일하게 왼쪽의 철조망과 초소, 오른쪽의 천막과 진흙 길 등 공간적 요소들이 올바르게 배치되어 있다.",
        "entities": "박철진의 외모와 군복 의상은 레퍼런스와 일치하나, 요구된 '동공 확장'이 잘 나타나지 않고 안색도 상대적으로 덜 창백하다. 화면 우측 가장자리에 은색 세단의 후면부 일부가 잡혀 있다.",
        "hard_violations": [
         "지문에서 세단이 이미 떠났다고 명시(the sedan has departed)했으나, 화면 우측 가장자리에 은색 차량의 일부가 버젓이 남아 있음."
        ],
        "physics": "하반신은 프레임 밖이나, 지면에 서서 굳어있는 자세가 무리 없이 표현되어 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "크게 뜬 눈과 긴장한 얼굴의 클로즈업은 맞지만, 오른쪽 전경에 세단의 후미가 남아 있어 차량이 떠난 뒤라는 장면 상태를 위반한다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "세단이 떠난 밤의 입구에서 박철진의 굳은 상체, 창백한 얼굴과 커진 동공을 클로즈업으로 보여 주며 인물과 장소의 연속성도 유지한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 거의 정면이고 시선은 렌즈 바로 왼쪽의 화면 밖을 향한다. 응시 대상은 보이지 않지만 지문도 특정 대상을 지정하지 않았다. 조준하거나 이동하는 물체는 없다.",
        "built_space": "왼쪽에는 콘크리트 기단 위 철망과 상단 철조망이 이어지고, 상부 출입구 구조물 한 개와 작은 경비 시설이 보인다. 오른쪽에는 천막들이 줄지어 있으며 여러 가로등이 진흙길을 비춘다. 박철진은 이 길의 전경에 서 있다. 물웅덩이의 조명 반사는 자연스럽지만, 오른쪽 가장자리 가까이에 은색 차체와 붉은 후미등 일부가 남아 있다.",
        "entities": "중년 동아시아계 남성 한 명만 보이며, 검은 머리에 섞인 흰머리, 짧은 회색 수염, 올리브색 제복은 이전 장면의 박철진과 대체로 일치한다. 모자가 없는 것도 이전 장면과 맞는다. 눈은 자연스러운 홍채와 동공을 유지하면서 크게 떠 있고 입이 조금 벌어져 당황함을 보인다. 얼굴의 창백함은 상대적으로 약하다. 열린 시신 가방은 보이지 않지만 클로즈업 밖일 수 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "떠났어야 할 세단의 후미가 오른쪽 전경에 남아 있어, 세단 출발 이후라는 명시적 배치 상태를 위반한다."
        ],
        "physics": "머리는 목과 상체에 정상적으로 연결되고 어깨가 굳어 있다. 발은 프레임 밖이라 접지는 직접 확인할 수 없지만 서 있는 상체에 부유나 불가능한 자세의 징후는 없다. 손에 든 물체나 공중에 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "박철진의 얼굴과 두 눈은 거의 렌즈 방향을 향하고 시선이 정면에 고정돼 있다. 화면 안에 별도의 응시 대상은 없으며 지문이 요구한 얼어붙은 당황함으로 읽힌다. 무기나 방향성 있는 소품, 이동 중인 인물은 없다.",
        "built_space": "왼쪽 철망·철조망과 콘크리트 기단, 출입구 상부 구조물 한 개, 창문이 켜진 경비실 한 개가 보인다. 오른쪽에는 여러 천막이 길을 따라 이어지고 가로등들이 원근에 맞춰 작아진다. 박철진은 진흙 입구 도로 전경에 서 있다. 먼 배경의 어두운 차량 한 대는 이전 장면의 배경 차량과 대응하며, 전경의 고급 세단은 사라졌다. 젖은 노면의 반사도 광원과 양립한다.",
        "entities": "보이는 사람은 박철진에 해당하는 중년 동아시아계 남성 한 명뿐이다. 얼굴 윤곽, 뒤로 올린 짧은 흑백 머리, 회색 수염과 올리브색 제복이 참고 이미지에 부합한다. 미소 없이 눈을 크게 뜨고 동공이 두드러지며, 창백해진 얼굴과 굳은 입 주변이 당황한 상태를 표현한다. 눈의 해부학적 형태는 정상이다. 시신 가방은 클로즈업 밖으로 확인할 수 없으며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 곧게 굳은 정지 자세이며 몸의 연결과 옷의 처짐이 자연스럽다. 하체와 발은 클로즈업 밖이므로 지지점을 직접 볼 수 없지만 떠 있다는 증거는 없다. 별도로 지지가 필요한 휴대 물체나 공중 동작도 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "크게 뜬 눈과 긴장한 얼굴의 클로즈업은 맞지만, 오른쪽 전경에 세단의 후미가 남아 있어 차량이 떠난 뒤라는 장면 상태를 위반한다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "세단이 떠난 밤의 입구에서 박철진의 굳은 상체, 창백한 얼굴과 커진 동공을 클로즈업으로 보여 주며 인물과 장소의 연속성도 유지한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 거의 정면이고 시선은 렌즈 바로 왼쪽의 화면 밖을 향한다. 응시 대상은 보이지 않지만 지문도 특정 대상을 지정하지 않았다. 조준하거나 이동하는 물체는 없다.",
        "built_space": "왼쪽에는 콘크리트 기단 위 철망과 상단 철조망이 이어지고, 상부 출입구 구조물 한 개와 작은 경비 시설이 보인다. 오른쪽에는 천막들이 줄지어 있으며 여러 가로등이 진흙길을 비춘다. 박철진은 이 길의 전경에 서 있다. 물웅덩이의 조명 반사는 자연스럽지만, 오른쪽 가장자리 가까이에 은색 차체와 붉은 후미등 일부가 남아 있다.",
        "entities": "중년 동아시아계 남성 한 명만 보이며, 검은 머리에 섞인 흰머리, 짧은 회색 수염, 올리브색 제복은 이전 장면의 박철진과 대체로 일치한다. 모자가 없는 것도 이전 장면과 맞는다. 눈은 자연스러운 홍채와 동공을 유지하면서 크게 떠 있고 입이 조금 벌어져 당황함을 보인다. 얼굴의 창백함은 상대적으로 약하다. 열린 시신 가방은 보이지 않지만 클로즈업 밖일 수 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "떠났어야 할 세단의 후미가 오른쪽 전경에 남아 있어, 세단 출발 이후라는 명시적 배치 상태를 위반한다."
        ],
        "physics": "머리는 목과 상체에 정상적으로 연결되고 어깨가 굳어 있다. 발은 프레임 밖이라 접지는 직접 확인할 수 없지만 서 있는 상체에 부유나 불가능한 자세의 징후는 없다. 손에 든 물체나 공중에 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "박철진의 얼굴과 두 눈은 거의 렌즈 방향을 향하고 시선이 정면에 고정돼 있다. 화면 안에 별도의 응시 대상은 없으며 지문이 요구한 얼어붙은 당황함으로 읽힌다. 무기나 방향성 있는 소품, 이동 중인 인물은 없다.",
        "built_space": "왼쪽 철망·철조망과 콘크리트 기단, 출입구 상부 구조물 한 개, 창문이 켜진 경비실 한 개가 보인다. 오른쪽에는 여러 천막이 길을 따라 이어지고 가로등들이 원근에 맞춰 작아진다. 박철진은 진흙 입구 도로 전경에 서 있다. 먼 배경의 어두운 차량 한 대는 이전 장면의 배경 차량과 대응하며, 전경의 고급 세단은 사라졌다. 젖은 노면의 반사도 광원과 양립한다.",
        "entities": "보이는 사람은 박철진에 해당하는 중년 동아시아계 남성 한 명뿐이다. 얼굴 윤곽, 뒤로 올린 짧은 흑백 머리, 회색 수염과 올리브색 제복이 참고 이미지에 부합한다. 미소 없이 눈을 크게 뜨고 동공이 두드러지며, 창백해진 얼굴과 굳은 입 주변이 당황한 상태를 표현한다. 눈의 해부학적 형태는 정상이다. 시신 가방은 클로즈업 밖으로 확인할 수 없으며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 곧게 굳은 정지 자세이며 몸의 연결과 옷의 처짐이 자연스럽다. 하체와 발은 클로즈업 밖이므로 지지점을 직접 볼 수 없지만 떠 있다는 증거는 없다. 별도로 지지가 필요한 휴대 물체나 공중 동작도 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.844
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.594
   },
   "violations": {
    "B": [
     "[gemini-pro] 지문에서 세단이 이미 떠났다고 명시(the sedan has departed)했으나, 화면 우측 가장자리에 은색 차량의 일부가 버젓이 남아 있음.",
     "[gpt-high] 떠났어야 할 세단의 후미가 오른쪽 전경에 남아 있어, 세단 출발 이후라는 명시적 배치 상태를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 594
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지문에 명시된 '크게 확장된 동공'과 '하얗게 질린 안색', '뻣뻣하게 선 자세'를 매우 정확하게 묘사했으며, 세단이 떠난 후의 텅 빈 배경을 올바르게 반영했습니다."
   },
   {
    "label": "B",
    "score": 594,
    "verdict_ko": "표정 연출에서 동공 확장이 뚜렷하지 않고, 지문에서 이미 떠났다고 명시된 세단 차량의 일부가 화면 우측에 남아 있어 서사적 설정과 충돌합니다.  ★위반: [gemini-pro] 지문에서 세단이 이미 떠났다고 명시(the sedan has departed)했으나, 화면 우측 가장자리에 은색 차량의 일부가 버젓이 남아 있음. / [gpt-high] 떠났어야 할 세단의 후미가 오른쪽 전경에 남아 있어, 세단 출발 이후라는 명시적 배치 상태를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S42sh10_sel.png",
    "asset_id": "1425fc2b-7e0a-4917-a548-053237b3bc12",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc7-d87e-74ff-b774-0cc8fda617d4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S42sh10"
  }
 },
 "S42sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:38:53.544990+00:00",
  "fingerprint": "4d5cd88527fb5194eb9679d54d719fb27ca0bc50a78b41ff73ab0a208de9010f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S42sh18_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S42sh18_sel.png",
  "source_sha256": "ad95ca9fe240ba2ca5e01948e8c67f01e2088d25be46d602e9515ebc591cf498",
  "file": "S42sh18_cine.png",
  "staged_sha256": "6f51e1eca12620893b6364a7a3206d1deac060ea1b054c65bb4054cb7eb2e1e9",
  "latency_ms": 13905
 },
 "S43sh1::signage": {
  "fp": "3638ad9fe2727e04",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::9fb11e6b69055a70": {
  "subjects": [],
  "subject_text": "국방장관 집무실\n집무용 책상과 좌석이 마련된 독립된 사무실. 책상 주변에 전화기가 놓여 있고 실내에 대면용 좌석 공간이 있다.",
  "identity": "canonical",
  "scope_id": "L52",
  "scope_role": "location_interior",
  "scope_sha": "64d917c1cee1f467"
 },
 "S43sh1::bgfirst_bg": {
  "input_fingerprint": "d91d59d2a62955ed",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 화려한 집무실 안, 국방장관이 한 손에 전화기를 든 채 거만하게 턱을 괸 삐딱한 구도.\n\nLOCATION (lock): At the desk inside an imposing defense minister's private office, under ordinary office lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Telephone (Held in one hand during the call) — Seen obliquely beside his face, without presenting a display to the camera; used as Provides the visible connection to the off-site conversation without becoming a device-view image; Ornate office interior (The minister occupies the office during the call); used as A limited background area establishes status without adding unsupported decorative objects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient light appropriate to the office keeps the supported chin and telephone readable, with restrained contrast and no specified source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 화려한 집무실 안, 국방장관이 한 손에 전화기를 든 채 거만하게 턱을 괸 삐딱한 구도.\n\nLOCATION (lock): At the desk inside an imposing defense minister's private office, under ordinary office lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Telephone (Held in one hand during the call) — Seen obliquely beside his face, without presenting a display to the camera; used as Provides the visible connection to the off-site conversation without becoming a device-view image; Ornate office interior (The minister occupies the office during the call); used as A limited background area establishes status without adding unsupported decorative objects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient light appropriate to the office keeps the supported chin and telephone readable, with restrained contrast and no specified source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S43sh1__bgfirst_bg.png",
  "asset_id": "6e926105-a01e-40f7-a99e-78633d79fc3f",
  "input_asset_ids": [
   "2ad5d662-783d-4274-ba3b-ad66e8b584da",
   "fd0ce497-27ed-4369-8849-bcf02b5baad0"
  ]
 },
 "S43sh1": {
  "input_fingerprint": "f4c57b4b784a469a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 화려한 집무실 안, 국방장관이 한 손에 전화기를 든 채 거만하게 턱을 괸 삐딱한 구도.\n\nLOCATION (lock): At the desk inside an imposing defense minister's private office, under ordinary office lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Telephone (Held in one hand during the call) — Seen obliquely beside his face, without presenting a display to the camera; used as Provides the visible connection to the off-site conversation without becoming a device-view image; Ornate office interior (The minister occupies the office during the call); used as A limited background area establishes status without adding unsupported decorative objects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient light appropriate to the office keeps the supported chin and telephone readable, with restrained contrast and no specified source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 국방장관: He is in his office conducting a telephone call.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 국방장관 right now, so 국방장관's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 국방장관: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 화려한 집무실 안, 국방장관이 한 손에 전화기를 든 채 거만하게 턱을 괸 삐딱한 구도.\n\nLOCATION (lock): At the desk inside an imposing defense minister's private office, under ordinary office lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Telephone (Held in one hand during the call) — Seen obliquely beside his face, without presenting a display to the camera; used as Provides the visible connection to the off-site conversation without becoming a device-view image; Ornate office interior (The minister occupies the office during the call); used as A limited background area establishes status without adding unsupported decorative objects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient light appropriate to the office keeps the supported chin and telephone readable, with restrained contrast and no specified source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 국방장관: He is in his office conducting a telephone call.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 국방장관 right now, so 국방장관's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 국방장관: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 화려한 집무실 안, 국방장관이 한 손에 전화기를 든 채 거만하게 턱을 괸 삐딱한 구도.\n\nLOCATION (lock): At the desk inside an imposing defense minister's private office, under ordinary office lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Telephone (Held in one hand during the call) — Seen obliquely beside his face, without presenting a display to the camera; used as Provides the visible connection to the off-site conversation without becoming a device-view image; Ornate office interior (The minister occupies the office during the call); used as A limited background area establishes status without adding unsupported decorative objects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient light appropriate to the office keeps the supported chin and telephone readable, with restrained contrast and no specified source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 국방장관: He is in his office conducting a telephone call.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 국방장관 right now, so 국방장관's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 국방장관: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S43sh1__bgfirst_bg.png",
     "asset_id": "6e926105-a01e-40f7-a99e-78633d79fc3f",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S43sh1.png",
     "asset_id": "2ad5d662-783d-4274-ba3b-ad66e8b584da",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:684284>",
     "asset_id": "b8636b60-6a4d-4eb8-bb28-08231e30fcd5",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 유선 전화기: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:930067>",
     "asset_id": "bfdc7020-0567-4234-b1bc-a71b7757c95f",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L52B01.png",
     "asset_id": "fd0ce497-27ed-4369-8849-bcf02b5baad0",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:684284>",
     "asset_id": "b8636b60-6a4d-4eb8-bb28-08231e30fcd5",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 유선 전화기: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:930067>",
     "asset_id": "bfdc7020-0567-4234-b1bc-a71b7757c95f",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 왼쪽을 향하며 카메라 구도는 수평임.",
    "built_space": "야간 집무실 배경과 기물 배치가 참조 이미지와 일치함.",
    "entities": "인물 외형은 참조와 일치함. 전화기 선 구조에 심각한 오류가 있음.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 사물 구조(수화기 선이 가슴으로 사라지고 본체 선이 허공을 가로지름)"
    ],
    "physics": "왼손으로 턱을 괴고 오른손으로 수화기를 들고 있으나, 선의 물리적 연결이 불가능함."
   },
   {
    "label": "B",
    "direction": "시선은 오른쪽을 향하며 카메라 구도는 수평임.",
    "built_space": "집무실 배경은 일치하나 책상 위에 참조에 없는 노트북이 추가됨.",
    "entities": "인물 외형은 일치함. 전화기 수화기가 두 개로 중복 생성됨.",
    "hard_violations": [
     "[gemini-pro] 중복된 사물(우측 본체에 수화기가 있는데 손에 또 들고 있음)",
     "[gemini-pro] 물리적으로 불가능한 사물(수화기 선이 책상 안으로 사라짐)",
     "[gemini-pro] 임의로 추가된 사물(노트북)",
     "[gpt-high] 통화에 사용하는 전화기 외에 화면 오른쪽 보조 상판에 두 번째 전화기를 추가했다."
    ],
    "physics": "수화기를 귀에 대고 있으나 턱을 괴지 않았으며, 수화기 선이 바닥으로 단절됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "삐딱한 구도가 누락되고 전화기 선이 몸으로 사라지는 물리적 오류가 있으나, 턱을 괸 자세는 부분적으로 표현됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "턱을 괸 자세와 삐딱한 구도 모두 누락되었으며, 수화기가 중복 생성되는 등 치명적인 구조적 오류가 있음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 왼쪽을 향하며 카메라 구도는 수평임.",
        "built_space": "야간 집무실 배경과 기물 배치가 참조 이미지와 일치함.",
        "entities": "인물 외형은 참조와 일치함. 전화기 선 구조에 심각한 오류가 있음.",
        "hard_violations": [
         "물리적으로 불가능한 사물 구조(수화기 선이 가슴으로 사라지고 본체 선이 허공을 가로지름)"
        ],
        "physics": "왼손으로 턱을 괴고 오른손으로 수화기를 들고 있으나, 선의 물리적 연결이 불가능함."
       },
       {
        "label": "B",
        "direction": "시선은 오른쪽을 향하며 카메라 구도는 수평임.",
        "built_space": "집무실 배경은 일치하나 책상 위에 참조에 없는 노트북이 추가됨.",
        "entities": "인물 외형은 일치함. 전화기 수화기가 두 개로 중복 생성됨.",
        "hard_violations": [
         "중복된 사물(우측 본체에 수화기가 있는데 손에 또 들고 있음)",
         "물리적으로 불가능한 사물(수화기 선이 책상 안으로 사라짐)",
         "임의로 추가된 사물(노트북)"
        ],
        "physics": "수화기를 귀에 대고 있으나 턱을 괴지 않았으며, 수화기 선이 바닥으로 단절됨."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "삐딱한 구도가 누락되고 전화기 선이 몸으로 사라지는 물리적 오류가 있으나, 턱을 괸 자세는 부분적으로 표현됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "턱을 괸 자세와 삐딱한 구도 모두 누락되었으며, 수화기가 중복 생성되는 등 치명적인 구조적 오류가 있음."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 왼쪽을 향하며 카메라 구도는 수평임.",
        "built_space": "야간 집무실 배경과 기물 배치가 참조 이미지와 일치함.",
        "entities": "인물 외형은 참조와 일치함. 전화기 선 구조에 심각한 오류가 있음.",
        "hard_violations": [
         "물리적으로 불가능한 사물 구조(수화기 선이 가슴으로 사라지고 본체 선이 허공을 가로지름)"
        ],
        "physics": "왼손으로 턱을 괴고 오른손으로 수화기를 들고 있으나, 선의 물리적 연결이 불가능함."
       },
       {
        "label": "B",
        "direction": "시선은 오른쪽을 향하며 카메라 구도는 수평임.",
        "built_space": "집무실 배경은 일치하나 책상 위에 참조에 없는 노트북이 추가됨.",
        "entities": "인물 외형은 일치함. 전화기 수화기가 두 개로 중복 생성됨.",
        "hard_violations": [
         "중복된 사물(우측 본체에 수화기가 있는데 손에 또 들고 있음)",
         "물리적으로 불가능한 사물(수화기 선이 책상 안으로 사라짐)",
         "임의로 추가된 사물(노트북)"
        ],
        "physics": "수화기를 귀에 대고 있으나 턱을 괴지 않았으며, 수화기 선이 바닥으로 단절됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴 옆 수화기로 통화하는 방향은 맞지만 턱을 괴지 않았고, 불필요한 두 번째 전화기까지 보여 핵심 연출과 소품 수를 어겼다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "비스듬히 앉아 거만하게 턱을 괸 동작과 집무실은 더 충실하지만, 수화기를 얼굴에서 멀리 들고 바라보아 지정된 통화 순간을 완성하지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남성은 화면 오른쪽을 비껴 바라보며, 오른손으로 수화기를 오른쪽 귀와 입 가까이에 대고 있다. 수화기의 사용 방향은 통화에 적합하고 화면을 카메라에 제시하지 않는다. 다만 손은 수화기를 잡고 뺨 옆에 있을 뿐 턱 아래를 받치지 않는다.",
        "built_space": "뒤쪽에 큰 창과 목재 하부 수납장, 왼쪽 태극기 하나와 오른쪽 기관기 하나, 화분 두 개가 보인다. 책상 하나 뒤 의자 하나에 남성이 앉아 있어 기본 배치는 참조와 부합한다. 책상 왼쪽 전경의 전화기 본체 외에 오른쪽 보조 상판에도 수화기가 얹힌 전화기 한 대가 추가로 보인다. 넓은 책상 전경과 양쪽 깃발까지 포함하여 제한된 배경의 미디엄 숏보다 공간의 비중이 크다. 창문에 비친 실내 조명은 가능한 반사다.",
        "entities": "등장인물은 성인 남성 한 명이다. 갈색 피부, 짧은 검은 머리, 다듬은 수염, 짙은 더블브레스트 정장과 넥타이 및 포켓스퀘어는 인물 참조와 대체로 부합하지만 참조의 검은 중절모는 없다. 민족적 정체성은 외관만으로 단정할 수 없다. 검은 유선 수화기와 나선형 줄은 보이나 본체의 정확한 버튼 구성과 마모는 이 각도에서 확인할 수 없다. 창밖은 밤이다.",
        "hard_violations": [
         "통화에 사용하는 전화기 외에 화면 오른쪽 보조 상판에 두 번째 전화기를 추가했다."
        ],
        "physics": "몸은 책상 뒤 의자에 앉아 지지되고, 기울어진 상체와 팔의 자세도 가능한 범위다. 오른손이 수화기를 실제로 잡으며 전화선은 수화기에서 책상 쪽으로 늘어진다. 책상 소품은 상판에 놓여 있다. 지지 없이 떠 있는 물체나 인체는 없지만, 요구된 턱의 손 지지는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "남성의 시선은 화면 왼쪽으로 뻗은 오른손의 수화기 쪽을 향한다. 왼손 주먹은 턱 아래를 받친다. 수화기는 귀와 입에서 상당히 떨어져 있어 얼굴 옆에 비스듬히 든 통화 자세가 아니라 수화기를 살펴보거나 잠시 귀에서 뗀 순간으로 보인다. 전화기 표시창을 관객에게 보여주지는 않는다.",
        "built_space": "책상 하나와 그 뒤 집무용 의자 하나, 뒤쪽 대형 창, 왼쪽 태극기 하나와 오른쪽 기관기 하나가 보인다. 창 아래 목재 수납장에는 비행기 모형과 원형 기념패가 있고, 오른쪽에는 큰 화분 하나와 선박 모형 진열장, 산수 액자가 있다. 왼쪽 벽에도 액자 두 개가 보여 참조 장소의 재료와 주요 배치를 잘 유지한다. 전화기 본체는 책상 오른쪽에 하나 보인다. 인물을 허벅지 일부까지 담고 사무실 양쪽을 넓게 보여 제한된 배경의 미디엄 숏 요구에는 다소 넓다.",
        "entities": "성인 남성 한 명만 등장한다. 피부색, 짧은 검은 머리, 수염과 얼굴 인상은 인물 참조에 대체로 가깝고 짙은 더블브레스트 정장과 넥타이를 착용한다. 참조의 중절모는 없다. 민족적 정체성은 외관만으로 확정할 수 없다. 검은 유선 수화기, 나선형 전화선, 책상용 본체는 요청된 소품 종류와 맞지만 본체 뒷면이 보여 참조의 세부 버튼과 표시창 일치는 확인할 수 없다. 야경과 일반적인 실내 조명이 보이며 읽을 수 있는 문구는 없다.",
        "hard_violations": [],
        "physics": "엉덩이와 등은 의자 좌판 및 등받이에 놓이고, 비스듬한 상체와 포갠 다리는 착석 상태에서 가능한 자세다. 턱은 왼손 주먹이 받치며 해당 팔은 의자 팔걸이 쪽에 지지된다. 오른손은 수화기를 확실히 쥐고 있고 전화선은 몸 앞을 가로질러 책상 위 본체 쪽으로 이어진다. 본체와 다른 책상 소품은 상판에 놓여 있어 지지 없는 부유는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "얼굴 옆 수화기로 통화하는 방향은 맞지만 턱을 괴지 않았고, 불필요한 두 번째 전화기까지 보여 핵심 연출과 소품 수를 어겼다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "비스듬히 앉아 거만하게 턱을 괸 동작과 집무실은 더 충실하지만, 수화기를 얼굴에서 멀리 들고 바라보아 지정된 통화 순간을 완성하지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "남성은 화면 오른쪽을 비껴 바라보며, 오른손으로 수화기를 오른쪽 귀와 입 가까이에 대고 있다. 수화기의 사용 방향은 통화에 적합하고 화면을 카메라에 제시하지 않는다. 다만 손은 수화기를 잡고 뺨 옆에 있을 뿐 턱 아래를 받치지 않는다.",
        "built_space": "뒤쪽에 큰 창과 목재 하부 수납장, 왼쪽 태극기 하나와 오른쪽 기관기 하나, 화분 두 개가 보인다. 책상 하나 뒤 의자 하나에 남성이 앉아 있어 기본 배치는 참조와 부합한다. 책상 왼쪽 전경의 전화기 본체 외에 오른쪽 보조 상판에도 수화기가 얹힌 전화기 한 대가 추가로 보인다. 넓은 책상 전경과 양쪽 깃발까지 포함하여 제한된 배경의 미디엄 숏보다 공간의 비중이 크다. 창문에 비친 실내 조명은 가능한 반사다.",
        "entities": "등장인물은 성인 남성 한 명이다. 갈색 피부, 짧은 검은 머리, 다듬은 수염, 짙은 더블브레스트 정장과 넥타이 및 포켓스퀘어는 인물 참조와 대체로 부합하지만 참조의 검은 중절모는 없다. 민족적 정체성은 외관만으로 단정할 수 없다. 검은 유선 수화기와 나선형 줄은 보이나 본체의 정확한 버튼 구성과 마모는 이 각도에서 확인할 수 없다. 창밖은 밤이다.",
        "hard_violations": [
         "통화에 사용하는 전화기 외에 화면 오른쪽 보조 상판에 두 번째 전화기를 추가했다."
        ],
        "physics": "몸은 책상 뒤 의자에 앉아 지지되고, 기울어진 상체와 팔의 자세도 가능한 범위다. 오른손이 수화기를 실제로 잡으며 전화선은 수화기에서 책상 쪽으로 늘어진다. 책상 소품은 상판에 놓여 있다. 지지 없이 떠 있는 물체나 인체는 없지만, 요구된 턱의 손 지지는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "남성의 시선은 화면 왼쪽으로 뻗은 오른손의 수화기 쪽을 향한다. 왼손 주먹은 턱 아래를 받친다. 수화기는 귀와 입에서 상당히 떨어져 있어 얼굴 옆에 비스듬히 든 통화 자세가 아니라 수화기를 살펴보거나 잠시 귀에서 뗀 순간으로 보인다. 전화기 표시창을 관객에게 보여주지는 않는다.",
        "built_space": "책상 하나와 그 뒤 집무용 의자 하나, 뒤쪽 대형 창, 왼쪽 태극기 하나와 오른쪽 기관기 하나가 보인다. 창 아래 목재 수납장에는 비행기 모형과 원형 기념패가 있고, 오른쪽에는 큰 화분 하나와 선박 모형 진열장, 산수 액자가 있다. 왼쪽 벽에도 액자 두 개가 보여 참조 장소의 재료와 주요 배치를 잘 유지한다. 전화기 본체는 책상 오른쪽에 하나 보인다. 인물을 허벅지 일부까지 담고 사무실 양쪽을 넓게 보여 제한된 배경의 미디엄 숏 요구에는 다소 넓다.",
        "entities": "성인 남성 한 명만 등장한다. 피부색, 짧은 검은 머리, 수염과 얼굴 인상은 인물 참조에 대체로 가깝고 짙은 더블브레스트 정장과 넥타이를 착용한다. 참조의 중절모는 없다. 민족적 정체성은 외관만으로 확정할 수 없다. 검은 유선 수화기, 나선형 전화선, 책상용 본체는 요청된 소품 종류와 맞지만 본체 뒷면이 보여 참조의 세부 버튼과 표시창 일치는 확인할 수 없다. 야경과 일반적인 실내 조명이 보이며 읽을 수 있는 문구는 없다.",
        "hard_violations": [],
        "physics": "엉덩이와 등은 의자 좌판 및 등받이에 놓이고, 비스듬한 상체와 포갠 다리는 착석 상태에서 가능한 자세다. 턱은 왼손 주먹이 받치며 해당 팔은 의자 팔걸이 쪽에 지지된다. 오른손은 수화기를 확실히 쥐고 있고 전화선은 몸 앞을 가로질러 책상 위 본체 쪽으로 이어진다. 본체와 다른 책상 소품은 상판에 놓여 있어 지지 없는 부유는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.25
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 사물 구조(수화기 선이 가슴으로 사라지고 본체 선이 허공을 가로지름)"
    ],
    "B": [
     "[gemini-pro] 중복된 사물(우측 본체에 수화기가 있는데 손에 또 들고 있음)",
     "[gemini-pro] 물리적으로 불가능한 사물(수화기 선이 책상 안으로 사라짐)",
     "[gemini-pro] 임의로 추가된 사물(노트북)",
     "[gpt-high] 통화에 사용하는 전화기 외에 화면 오른쪽 보조 상판에 두 번째 전화기를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "삐딱한 구도가 누락되고 전화기 선이 몸으로 사라지는 물리적 오류가 있으나, 턱을 괸 자세는 부분적으로 표현됨.  ★위반: [gemini-pro] 물리적으로 불가능한 사물 구조(수화기 선이 가슴으로 사라지고 본체 선이 허공을 가로지름)"
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "턱을 괸 자세와 삐딱한 구도 모두 누락되었으며, 수화기가 중복 생성되는 등 치명적인 구조적 오류가 있음.  ★위반: [gemini-pro] 중복된 사물(우측 본체에 수화기가 있는데 손에 또 들고 있음) / [gemini-pro] 물리적으로 불가능한 사물(수화기 선이 책상 안으로 사라짐) / [gemini-pro] 임의로 추가된 사물(노트북) / [gpt-high] 통화에 사용하는 전화기 외에 화면 오른쪽 보조 상판에 두 번째 전화기를 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L52B01.png",
    "asset_id": "fd0ce497-27ed-4369-8849-bcf02b5baad0",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:684284>",
    "asset_id": "b8636b60-6a4d-4eb8-bb28-08231e30fcd5",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 유선 전화기: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:930067>",
    "asset_id": "bfdc7020-0567-4234-b1bc-a71b7757c95f",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc8-bcb8-7466-971f-3c2562e13af1",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S43sh1__bgfirst_bg.png",
   "bg_asset_id": "6e926105-a01e-40f7-a99e-78633d79fc3f",
   "bg_record_key": "S43sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S43sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:39:07.891758+00:00",
  "fingerprint": "073f3dc6301ad71bb4c967589583cd487d87f71f7004a01e056cb9cd43a372d0",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S43sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S43sh1_sel.png",
  "source_sha256": "52e4f7a848e6d50583bd9e14dc27d85e19a397bd179c233c5163fde6542ed4c6",
  "file": "S43sh1_cine.png",
  "staged_sha256": "bd08c71b5a6a564121c4a2779be95bd8a30ccb5cec063fd5112076118a8229eb",
  "latency_ms": 12477
 },
 "S43sh5::signage": {
  "fp": "65f9f7c22142a066",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S43sh5": {
  "input_fingerprint": "257826ef511afb20",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 텅 빈 사무실에서 수화기를 든 채 허리를 90도로 푹 숙인 굴욕적인 자세.\n\nLOCATION (lock): Inside the militia commander's sparsely occupied office, beside his telephone under ordinary office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Office floor (Visible around the deeply bowed figure) — Seen obliquely from above; used as Provides negative space around his folded body; Telephone receiver (Held during the call in his trembling hand) — Seen beside his lowered head without concealing the hand; used as Connects the humiliating posture to the distant speaker.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the office preserves the bowed silhouette and trembling hand without introducing a new source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 박철진: He remains on the telephone with a trembling hand and a submissively bowed posture.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 텅 빈 사무실에서 수화기를 든 채 허리를 90도로 푹 숙인 굴욕적인 자세.\n\nLOCATION (lock): Inside the militia commander's sparsely occupied office, beside his telephone under ordinary office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Office floor (Visible around the deeply bowed figure) — Seen obliquely from above; used as Provides negative space around his folded body; Telephone receiver (Held during the call in his trembling hand) — Seen beside his lowered head without concealing the hand; used as Connects the humiliating posture to the distant speaker.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the office preserves the bowed silhouette and trembling hand without introducing a new source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 박철진: He remains on the telephone with a trembling hand and a submissively bowed posture.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 텅 빈 사무실에서 수화기를 든 채 허리를 90도로 푹 숙인 굴욕적인 자세.\n\nLOCATION (lock): Inside the militia commander's sparsely occupied office, beside his telephone under ordinary office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Office floor (Visible around the deeply bowed figure) — Seen obliquely from above; used as Provides negative space around his folded body; Telephone receiver (Held during the call in his trembling hand) — Seen beside his lowered head without concealing the hand; used as Connects the humiliating posture to the distant speaker.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the office preserves the bowed silhouette and trembling hand without introducing a new source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 박철진: He remains on the telephone with a trembling hand and a submissively bowed posture.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물이 측면 방향으로 바닥을 향해 90도로 허리를 굽히고 수화기를 귀에 대고 있음.",
    "built_space": "나무 패널 벽과 책상이 있는 사무실 구조이나, 배경 좌측 책상에 지시되지 않은 인물이 앉아 있음.",
    "entities": "박철진의 군복과 완장이 참조와 유사하며, 수화기를 들고 있으나 전화기 본체의 디테일이 약간 다름.",
    "hard_violations": [
     "[gemini-pro] 텍스트에 명시되지 않은 추가 인물이 배경에 등장함 (invented people)",
     "[gpt-high] 뒤쪽 책상에 지시되지 않은 인물이 한 명 추가되어, 텅 빈 사무실에 박철진만 등장해야 한다는 조건을 위반한다.",
     "[gpt-high] 통화 중인 수화기 외에 같은 전화기 받침에도 수화기가 남아 있어 소품이 중복되었다.",
     "[gpt-high] 완장에 읽을 수 있는 '헌병'과 'MP'가 있어 문자 금지 조건을 위반한다."
    ],
    "physics": "두 발로 바닥을 딛고 서서 허리를 굽힌 상태를 지탱하며, 손으로 수화기를 자연스럽게 쥐고 있음."
   },
   {
    "label": "B",
    "direction": "인물이 정면 카메라를 향해 허리를 90도로 깊이 숙인 채 수화기를 귀에 대고 있음.",
    "built_space": "참조 이미지와 일치하는 나무 벽면과 책상, 의자들이 있는 빈 사무실 구조가 올바르게 배치됨.",
    "entities": "박철진의 군복과 헌병 완장이 일치하며, 데스크 위에 놓인 전화기 본체와 손에 든 수화기가 존재함.",
    "hard_violations": [
     "[gpt-high] 완장에 '헌병'과 'MP'가 뚜렷하게 읽혀, 읽을 수 있는 문자를 전면 금지한 조건을 위반한다."
    ],
    "physics": "굽힌 자세로 서서 한 손으로 수화기를 쥐고 있으며, 늘어진 전화선이 중력에 맞게 표현됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "텅 빈 사무실에서 수화기를 들고 허리를 깊게 숙인 굴욕적인 자세를 프롬프트의 지시대로 충실히 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "자세는 잘 표현되었으나, 프롬프트에 명시되지 않은 추가 인물이 배경에 등장하여 핵심 규칙을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물이 측면 방향으로 바닥을 향해 90도로 허리를 굽히고 수화기를 귀에 대고 있음.",
        "built_space": "나무 패널 벽과 책상이 있는 사무실 구조이나, 배경 좌측 책상에 지시되지 않은 인물이 앉아 있음.",
        "entities": "박철진의 군복과 완장이 참조와 유사하며, 수화기를 들고 있으나 전화기 본체의 디테일이 약간 다름.",
        "hard_violations": [
         "텍스트에 명시되지 않은 추가 인물이 배경에 등장함 (invented people)"
        ],
        "physics": "두 발로 바닥을 딛고 서서 허리를 굽힌 상태를 지탱하며, 손으로 수화기를 자연스럽게 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "인물이 정면 카메라를 향해 허리를 90도로 깊이 숙인 채 수화기를 귀에 대고 있음.",
        "built_space": "참조 이미지와 일치하는 나무 벽면과 책상, 의자들이 있는 빈 사무실 구조가 올바르게 배치됨.",
        "entities": "박철진의 군복과 헌병 완장이 일치하며, 데스크 위에 놓인 전화기 본체와 손에 든 수화기가 존재함.",
        "hard_violations": [],
        "physics": "굽힌 자세로 서서 한 손으로 수화기를 쥐고 있으며, 늘어진 전화선이 중력에 맞게 표현됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "텅 빈 사무실에서 수화기를 들고 허리를 깊게 숙인 굴욕적인 자세를 프롬프트의 지시대로 충실히 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "자세는 잘 표현되었으나, 프롬프트에 명시되지 않은 추가 인물이 배경에 등장하여 핵심 규칙을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물이 측면 방향으로 바닥을 향해 90도로 허리를 굽히고 수화기를 귀에 대고 있음.",
        "built_space": "나무 패널 벽과 책상이 있는 사무실 구조이나, 배경 좌측 책상에 지시되지 않은 인물이 앉아 있음.",
        "entities": "박철진의 군복과 완장이 참조와 유사하며, 수화기를 들고 있으나 전화기 본체의 디테일이 약간 다름.",
        "hard_violations": [
         "텍스트에 명시되지 않은 추가 인물이 배경에 등장함 (invented people)"
        ],
        "physics": "두 발로 바닥을 딛고 서서 허리를 굽힌 상태를 지탱하며, 손으로 수화기를 자연스럽게 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "인물이 정면 카메라를 향해 허리를 90도로 깊이 숙인 채 수화기를 귀에 대고 있음.",
        "built_space": "참조 이미지와 일치하는 나무 벽면과 책상, 의자들이 있는 빈 사무실 구조가 올바르게 배치됨.",
        "entities": "박철진의 군복과 헌병 완장이 일치하며, 데스크 위에 놓인 전화기 본체와 손에 든 수화기가 존재함.",
        "hard_violations": [],
        "physics": "굽힌 자세로 서서 한 손으로 수화기를 쥐고 있으며, 늘어진 전화선이 중력에 맞게 표현됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "혼자 허리를 숙여 통화하는 핵심 행동은 더 충실하지만, 읽히는 완장 문자와 바닥 여백이 부족한 타이트한 구도로 요구를 충족하지 못한다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "넓은 바닥과 깊이 숙인 전신은 요구에 가깝지만, 배경의 추가 인물과 중복 수화기, 읽히는 완장 문자 때문에 사용할 수 없다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 얼굴과 시선은 아래쪽 바닥을 향한다. 한 손으로 검은 수화기를 귀와 입 옆에 붙이고 있어 통화 방향은 자연스럽다. 손가락과 손등이 드러나며 수화기가 손을 가리지 않는다. 조준하거나 이동하는 대상은 없다.",
        "built_space": "왼쪽 가장자리, 뒤쪽, 오른쪽 전경에 책상 세 개가 보이며 녹색 의자는 세 개가 보인다. 뒤에는 큰 창 하나와 왼쪽에 일부 잘린 창, 열린 출입구가 있고 오른쪽 벽은 목재 패널이다. 창밖은 어둡다. 참고의 목재 벽·금속 창틀·녹색 의자·회의용 전화기는 이어지지만 책상과 서류가 늘어난 배치이고 벽에 박힌 칼은 확인되지 않는다. 인물이 화면 대부분의 높이를 차지하고 발이 잘려, 비스듬히 내려다본 넓은 바닥 여백이 충분하지 않다.",
        "entities": "인물은 한 명이며 짧은 회색 머리, 중년 남성의 체격, 올리브색 군복과 검은 완장이 참고와 대체로 맞는다. 숙인 얼굴 때문에 정확한 얼굴 일치와 민족적 외양은 제한적으로만 확인된다. 손에 든 검은 유선 수화기와 오른쪽 책상 위 본체는 소품의 종류에 맞지만 본체의 정확한 형태는 참고와 다르다. 뒤쪽에는 이전 장면의 회의용 전화기도 보인다. 완장의 '헌병'과 'MP'는 읽을 수 있다.",
        "hard_violations": [
         "완장에 '헌병'과 'MP'가 뚜렷하게 읽혀, 읽을 수 있는 문자를 전면 금지한 조건을 위반한다."
        ],
        "physics": "다리는 아래로 곧게 내려가고 상체는 허리에서 깊이 접혀 있어 실제로 가능한 절 자세다. 발의 접지점은 화면 밖이지만 공중에 떠 있다는 증거는 없다. 수화기는 손가락으로 확실히 잡고 있으며 본체는 책상에 놓여 있다. 전화선은 손에 든 수화기와 본체 사이에 늘어져 지지 관계가 자연스럽다. 손의 떨림 자체는 정지 화면에서 명확히 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "박철진은 왼쪽 아래 바닥을 향해 머리를 낮추고 수화기를 귀 옆에 대고 있다. 수화기를 잡은 손이 보이며 통화 자세의 방향은 맞는다. 뒤쪽에 추가된 인물은 자기 책상 위를 내려다본다.",
        "built_space": "책상은 왼쪽 전경부터 뒤쪽과 오른쪽까지 일곱 개가 보인다. 오른쪽 큰 창 아래에는 대기 의자 세 개가 나란히 있고 각 책상 주변에도 의자들이 있다. 목재 패널 벽, 여러 창과 오른쪽 출입구가 보이지만 참고의 소박한 사무실보다 가구가 많은 큰 업무실로 확장되었다. 창밖은 밤이며 천장에는 형광등 두 개가 보인다. 비스듬한 높은 시점에서 전신 주변의 바닥을 넓게 보여주는 구도는 요청에 잘 맞는다.",
        "entities": "전경 남성의 짧은 회색 머리, 올리브색 군복, 검은 완장과 부츠는 참고와 대체로 맞지만 얼굴은 대부분 숨겨져 정확한 동일인 여부는 확인하기 어렵다. 뒤쪽 책상에 다른 성인 한 명이 앉아 있어 박철진만 등장해야 한다는 조건과 다르다. 박철진이 수화기를 들고 있는데 오른쪽 전화기 본체에도 별도의 수화기가 놓여 있다. 본체의 버튼과 화면 배치도 소품 참고와 다르다. 완장의 '헌병'과 'MP'가 읽힌다.",
        "hard_violations": [
         "뒤쪽 책상에 지시되지 않은 인물이 한 명 추가되어, 텅 빈 사무실에 박철진만 등장해야 한다는 조건을 위반한다.",
         "통화 중인 수화기 외에 같은 전화기 받침에도 수화기가 남아 있어 소품이 중복되었다.",
         "완장에 읽을 수 있는 '헌병'과 'MP'가 있어 문자 금지 조건을 위반한다."
        ],
        "physics": "양쪽 부츠가 바닥에 닿아 몸을 지지하며, 다리를 세우고 허리와 등을 깊이 굽힌 자세는 실제로 가능하다. 수화기는 손으로 잡고 본체는 책상에 놓여 있으며 전화선은 두 끝 사이에서 처진다. 배경 인물도 책상 뒤에 앉아 있어 공중에 뜬 신체는 없다. 다만 전화기 받침의 추가 수화기는 통화 소품의 상태를 모순되게 만든다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "혼자 허리를 숙여 통화하는 핵심 행동은 더 충실하지만, 읽히는 완장 문자와 바닥 여백이 부족한 타이트한 구도로 요구를 충족하지 못한다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "넓은 바닥과 깊이 숙인 전신은 요구에 가깝지만, 배경의 추가 인물과 중복 수화기, 읽히는 완장 문자 때문에 사용할 수 없다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박철진의 얼굴과 시선은 아래쪽 바닥을 향한다. 한 손으로 검은 수화기를 귀와 입 옆에 붙이고 있어 통화 방향은 자연스럽다. 손가락과 손등이 드러나며 수화기가 손을 가리지 않는다. 조준하거나 이동하는 대상은 없다.",
        "built_space": "왼쪽 가장자리, 뒤쪽, 오른쪽 전경에 책상 세 개가 보이며 녹색 의자는 세 개가 보인다. 뒤에는 큰 창 하나와 왼쪽에 일부 잘린 창, 열린 출입구가 있고 오른쪽 벽은 목재 패널이다. 창밖은 어둡다. 참고의 목재 벽·금속 창틀·녹색 의자·회의용 전화기는 이어지지만 책상과 서류가 늘어난 배치이고 벽에 박힌 칼은 확인되지 않는다. 인물이 화면 대부분의 높이를 차지하고 발이 잘려, 비스듬히 내려다본 넓은 바닥 여백이 충분하지 않다.",
        "entities": "인물은 한 명이며 짧은 회색 머리, 중년 남성의 체격, 올리브색 군복과 검은 완장이 참고와 대체로 맞는다. 숙인 얼굴 때문에 정확한 얼굴 일치와 민족적 외양은 제한적으로만 확인된다. 손에 든 검은 유선 수화기와 오른쪽 책상 위 본체는 소품의 종류에 맞지만 본체의 정확한 형태는 참고와 다르다. 뒤쪽에는 이전 장면의 회의용 전화기도 보인다. 완장의 '헌병'과 'MP'는 읽을 수 있다.",
        "hard_violations": [
         "완장에 '헌병'과 'MP'가 뚜렷하게 읽혀, 읽을 수 있는 문자를 전면 금지한 조건을 위반한다."
        ],
        "physics": "다리는 아래로 곧게 내려가고 상체는 허리에서 깊이 접혀 있어 실제로 가능한 절 자세다. 발의 접지점은 화면 밖이지만 공중에 떠 있다는 증거는 없다. 수화기는 손가락으로 확실히 잡고 있으며 본체는 책상에 놓여 있다. 전화선은 손에 든 수화기와 본체 사이에 늘어져 지지 관계가 자연스럽다. 손의 떨림 자체는 정지 화면에서 명확히 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "박철진은 왼쪽 아래 바닥을 향해 머리를 낮추고 수화기를 귀 옆에 대고 있다. 수화기를 잡은 손이 보이며 통화 자세의 방향은 맞는다. 뒤쪽에 추가된 인물은 자기 책상 위를 내려다본다.",
        "built_space": "책상은 왼쪽 전경부터 뒤쪽과 오른쪽까지 일곱 개가 보인다. 오른쪽 큰 창 아래에는 대기 의자 세 개가 나란히 있고 각 책상 주변에도 의자들이 있다. 목재 패널 벽, 여러 창과 오른쪽 출입구가 보이지만 참고의 소박한 사무실보다 가구가 많은 큰 업무실로 확장되었다. 창밖은 밤이며 천장에는 형광등 두 개가 보인다. 비스듬한 높은 시점에서 전신 주변의 바닥을 넓게 보여주는 구도는 요청에 잘 맞는다.",
        "entities": "전경 남성의 짧은 회색 머리, 올리브색 군복, 검은 완장과 부츠는 참고와 대체로 맞지만 얼굴은 대부분 숨겨져 정확한 동일인 여부는 확인하기 어렵다. 뒤쪽 책상에 다른 성인 한 명이 앉아 있어 박철진만 등장해야 한다는 조건과 다르다. 박철진이 수화기를 들고 있는데 오른쪽 전화기 본체에도 별도의 수화기가 놓여 있다. 본체의 버튼과 화면 배치도 소품 참고와 다르다. 완장의 '헌병'과 'MP'가 읽힌다.",
        "hard_violations": [
         "뒤쪽 책상에 지시되지 않은 인물이 한 명 추가되어, 텅 빈 사무실에 박철진만 등장해야 한다는 조건을 위반한다.",
         "통화 중인 수화기 외에 같은 전화기 받침에도 수화기가 남아 있어 소품이 중복되었다.",
         "완장에 읽을 수 있는 '헌병'과 'MP'가 있어 문자 금지 조건을 위반한다."
        ],
        "physics": "양쪽 부츠가 바닥에 닿아 몸을 지지하며, 다리를 세우고 허리와 등을 깊이 굽힌 자세는 실제로 가능하다. 수화기는 손으로 잡고 본체는 책상에 놓여 있으며 전화선은 두 끝 사이에서 처진다. 배경 인물도 책상 뒤에 앉아 있어 공중에 뜬 신체는 없다. 다만 전화기 받침의 추가 수화기는 통화 소품의 상태를 모순되게 만든다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.762,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.512,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 텍스트에 명시되지 않은 추가 인물이 배경에 등장함 (invented people)",
     "[gpt-high] 뒤쪽 책상에 지시되지 않은 인물이 한 명 추가되어, 텅 빈 사무실에 박철진만 등장해야 한다는 조건을 위반한다.",
     "[gpt-high] 통화 중인 수화기 외에 같은 전화기 받침에도 수화기가 남아 있어 소품이 중복되었다.",
     "[gpt-high] 완장에 읽을 수 있는 '헌병'과 'MP'가 있어 문자 금지 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 완장에 '헌병'과 'MP'가 뚜렷하게 읽혀, 읽을 수 있는 문자를 전면 금지한 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 512
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "텅 빈 사무실에서 수화기를 들고 허리를 깊게 숙인 굴욕적인 자세를 프롬프트의 지시대로 충실히 구현했습니다.  ★위반: [gpt-high] 완장에 '헌병'과 'MP'가 뚜렷하게 읽혀, 읽을 수 있는 문자를 전면 금지한 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 512,
    "verdict_ko": "자세는 잘 표현되었으나, 프롬프트에 명시되지 않은 추가 인물이 배경에 등장하여 핵심 규칙을 위반했습니다.  ★위반: [gemini-pro] 텍스트에 명시되지 않은 추가 인물이 배경에 등장함 (invented people) / [gpt-high] 뒤쪽 책상에 지시되지 않은 인물이 한 명 추가되어, 텅 빈 사무실에 박철진만 등장해야 한다는 조건을 위반한다. / [gpt-high] 통화 중인 수화기 외에 같은 전화기 받침에도 수화기가 남아 있어 소품이 중복되었다. / [gpt-high] 완장에 읽을 수 있는 '헌병'과 'MP'가 있어 문자 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S18sh8_sel.png",
    "asset_id": "440192f4-3d9b-4e56-af46-e3b69019d3b2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 유선 전화기: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:930067>",
    "asset_id": "bfdc7020-0567-4234-b1bc-a71b7757c95f",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdc9-8b6a-7d3e-91a1-50d975826701",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S18sh8"
  }
 },
 "S43sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:39:20.728616+00:00",
  "fingerprint": "640c3d63a5d40d6f02a5a56db9ce425285722f314537c5b18dc9465fe8e0effc",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S43sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S43sh5_sel.png",
  "source_sha256": "dd86778752ec12f2f8ed2cd6a6a3a890e72684435b51b424a1818322ca715e14",
  "file": "S43sh5_cine.png",
  "staged_sha256": "97bb47059ce33cb055a0e8ef0baab24dc92bd9fa58cb74491ab8bd126c469e32",
  "latency_ms": 13484
 },
 "S43sh10::signage": {
  "fp": "3cf7c807e0e85978",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S43sh10": {
  "input_fingerprint": "246c8ad3b9db48d3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 두 팔을 양옆으로 활짝 벌린 채 얼굴에 핏대를 세우고 광기 어린 미소를 지은 구도.\n\nLOCATION (lock): In the working area of the militia commander's office after the call, under ordinary interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office interior (Visible behind his expanded gesture) — Viewed obliquely from the subordinate's side; used as Leaves uncluttered separation around the outstretched arms.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's neutral ambient illumination, using controlled contrast to keep the strained expression readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the office furnishings, wall finishes, and established lighting from the reference. EXCLUDE the telephone receiver as a held prop now that the call has ended.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 박철진: He has ended the telephone call and remains in his office, now wearing an ambitious smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 두 팔을 양옆으로 활짝 벌린 채 얼굴에 핏대를 세우고 광기 어린 미소를 지은 구도.\n\nLOCATION (lock): In the working area of the militia commander's office after the call, under ordinary interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office interior (Visible behind his expanded gesture) — Viewed obliquely from the subordinate's side; used as Leaves uncluttered separation around the outstretched arms.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's neutral ambient illumination, using controlled contrast to keep the strained expression readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the office furnishings, wall finishes, and established lighting from the reference. EXCLUDE the telephone receiver as a held prop now that the call has ended.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 박철진: He has ended the telephone call and remains in his office, now wearing an ambitious smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진이 두 팔을 양옆으로 활짝 벌린 채 얼굴에 핏대를 세우고 광기 어린 미소를 지은 구도.\n\nLOCATION (lock): In the working area of the militia commander's office after the call, under ordinary interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office interior (Visible behind his expanded gesture) — Viewed obliquely from the subordinate's side; used as Leaves uncluttered separation around the outstretched arms.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's neutral ambient illumination, using controlled contrast to keep the strained expression readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the office furnishings, wall finishes, and established lighting from the reference. EXCLUDE the telephone receiver as a held prop now that the call has ended.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 박철진: He has ended the telephone call and remains in his office, now wearing an ambitious smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 우측 바깥을 향하고 있으며, 두 팔은 양옆의 빈 공간을 향해 수평으로 뻗어 있습니다.",
    "built_space": "야간의 사무실 내부로, 좌우에 서류가 놓인 책상과 녹색 의자가 있으며 바닥에는 레퍼런스와 동일하게 서류들이 흩어져 있습니다.",
    "entities": "레퍼런스와 일치하는 얼굴, 헤어스타일, MP 완장이 달린 군복을 입은 박철진이 두 팔을 벌리고 활짝 웃고 있습니다.",
    "hard_violations": [
     "[gpt-high] 검은 완장의 흰 글자를 읽을 수 있어, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지를 위반한다."
    ],
    "physics": "캐릭터가 두 발로 바닥을 딛고 서 있으며, 양팔은 어깨를 축으로 자연스럽게 들려 있어 물리적 어색함이 없습니다."
   },
   {
    "label": "B",
    "direction": "시선은 화면 우측을 쏘아보듯 향하고 있으며, 두 팔은 양옆으로 넓게 펼쳐져 있습니다.",
    "built_space": "레퍼런스와 일치하는 사무실 배경으로, 책상, 벽면의 포스터, 바닥의 종이 배치 등이 동일하게 밤 시간대로 렌더링되었습니다.",
    "entities": "박철진 캐릭터가 일치하게 묘사되었고 얼굴에 핏대가 선 광기 어린 표정이 프롬프트의 지시와 매우 잘 맞습니다.",
    "hard_violations": [
     "[gemini-pro] 우측 책상 위 전화기의 수화기 선이 바닥으로 늘어지다가 캐릭터의 하복부(사타구니)에 직접 연결되어 있는 물리적으로 불가능한 형태",
     "[gpt-high] 검은 완장의 흰 글자를 읽을 수 있어, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지를 위반한다."
    ],
    "physics": "인물의 자세 자체는 자연스러우나, 사물(전화선)이 허공을 가로질러 인물의 몸에 기형적으로 융합되어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "배경과 인물의 디테일을 레퍼런스에 맞게 잘 구현하였고 구도도 정확하지만, 프롬프트가 요구한 '핏대를 세우고 광기 어린 미소'라기보다는 다소 밝은 웃음에 가깝습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트가 요구한 광기 어린 표정과 핏대는 훌륭하게 표현되었으나, 전화선이 몸에 연결되는 치명적인 물리적 오류로 인해 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 우측 바깥을 향하고 있으며, 두 팔은 양옆의 빈 공간을 향해 수평으로 뻗어 있습니다.",
        "built_space": "야간의 사무실 내부로, 좌우에 서류가 놓인 책상과 녹색 의자가 있으며 바닥에는 레퍼런스와 동일하게 서류들이 흩어져 있습니다.",
        "entities": "레퍼런스와 일치하는 얼굴, 헤어스타일, MP 완장이 달린 군복을 입은 박철진이 두 팔을 벌리고 활짝 웃고 있습니다.",
        "hard_violations": [],
        "physics": "캐릭터가 두 발로 바닥을 딛고 서 있으며, 양팔은 어깨를 축으로 자연스럽게 들려 있어 물리적 어색함이 없습니다."
       },
       {
        "label": "B",
        "direction": "시선은 화면 우측을 쏘아보듯 향하고 있으며, 두 팔은 양옆으로 넓게 펼쳐져 있습니다.",
        "built_space": "레퍼런스와 일치하는 사무실 배경으로, 책상, 벽면의 포스터, 바닥의 종이 배치 등이 동일하게 밤 시간대로 렌더링되었습니다.",
        "entities": "박철진 캐릭터가 일치하게 묘사되었고 얼굴에 핏대가 선 광기 어린 표정이 프롬프트의 지시와 매우 잘 맞습니다.",
        "hard_violations": [
         "우측 책상 위 전화기의 수화기 선이 바닥으로 늘어지다가 캐릭터의 하복부(사타구니)에 직접 연결되어 있는 물리적으로 불가능한 형태"
        ],
        "physics": "인물의 자세 자체는 자연스러우나, 사물(전화선)이 허공을 가로질러 인물의 몸에 기형적으로 융합되어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "배경과 인물의 디테일을 레퍼런스에 맞게 잘 구현하였고 구도도 정확하지만, 프롬프트가 요구한 '핏대를 세우고 광기 어린 미소'라기보다는 다소 밝은 웃음에 가깝습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트가 요구한 광기 어린 표정과 핏대는 훌륭하게 표현되었으나, 전화선이 몸에 연결되는 치명적인 물리적 오류로 인해 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 우측 바깥을 향하고 있으며, 두 팔은 양옆의 빈 공간을 향해 수평으로 뻗어 있습니다.",
        "built_space": "야간의 사무실 내부로, 좌우에 서류가 놓인 책상과 녹색 의자가 있으며 바닥에는 레퍼런스와 동일하게 서류들이 흩어져 있습니다.",
        "entities": "레퍼런스와 일치하는 얼굴, 헤어스타일, MP 완장이 달린 군복을 입은 박철진이 두 팔을 벌리고 활짝 웃고 있습니다.",
        "hard_violations": [],
        "physics": "캐릭터가 두 발로 바닥을 딛고 서 있으며, 양팔은 어깨를 축으로 자연스럽게 들려 있어 물리적 어색함이 없습니다."
       },
       {
        "label": "B",
        "direction": "시선은 화면 우측을 쏘아보듯 향하고 있으며, 두 팔은 양옆으로 넓게 펼쳐져 있습니다.",
        "built_space": "레퍼런스와 일치하는 사무실 배경으로, 책상, 벽면의 포스터, 바닥의 종이 배치 등이 동일하게 밤 시간대로 렌더링되었습니다.",
        "entities": "박철진 캐릭터가 일치하게 묘사되었고 얼굴에 핏대가 선 광기 어린 표정이 프롬프트의 지시와 매우 잘 맞습니다.",
        "hard_violations": [
         "우측 책상 위 전화기의 수화기 선이 바닥으로 늘어지다가 캐릭터의 하복부(사타구니)에 직접 연결되어 있는 물리적으로 불가능한 형태"
        ],
        "physics": "인물의 자세 자체는 자연스러우나, 사물(전화선)이 허공을 가로질러 인물의 몸에 기형적으로 융합되어 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 팔을 활짝 벌린 동작과 일그러진 광기 어린 미소, 기존 사무실의 배치가 더 충실하지만, 완장의 읽을 수 있는 글자는 명시적 금지 위반이다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "팔을 벌린 동작과 목의 긴장은 맞지만 표정이 광기보다는 환희에 가깝고 배경 연속성이 덜 정확하며, 완장의 읽을 수 있는 글자도 금지 위반이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 카메라 오른쪽의 화면 밖 상대 방향을 향한다. 특정 상대는 보이지 않으며 지시문도 시선의 목표를 고정하지 않는다. 두 팔은 몸 양옆으로 거의 수평으로 펼쳐지고 손바닥은 카메라 쪽으로 비스듬히 열린다. 무기나 겨누는 물체는 없다.",
        "built_space": "왼쪽 책상 하나, 뒤쪽 책상 하나, 오른쪽 전경 책상 하나로 총 세 책상이 보인다. 녹색 의자 등받이는 왼쪽과 오른쪽에 각각 하나씩 보이고, 뒤쪽 의자 자리는 인물과 가구에 가려진다. 중앙의 어두운 창, 왼쪽 일부 창, 열린 회색 문 하나, 오른쪽 목재 벽과 지도·사진, 오염된 투톤 벽과 타일 바닥이 이전 장면과 대체로 이어진다. 인물은 책상 사이 통로에 서 있고 팔이 가구와 충돌하지 않는다. 뒤쪽 회의용 전화기와 전경 유선전화기가 각각 하나씩 있다. 불가능한 거울 반사는 없다.",
        "entities": "박철진으로 제시된 중년 남성 한 명만 등장한다. 짧게 깎은 회색 섞인 머리, 회색 수염, 얼굴과 체격은 참조와 대체로 맞고, 이전 장면처럼 모자가 없다. 올리브색 군복과 검은 완장도 유지된다. 이를 드러낸 미소, 좁아진 눈, 코와 볼의 강한 긴장이 광기 어린 연기에 가깝다. 얼굴의 핏대 자체는 뚜렷하지 않다. 수화기는 손에 없고 전경 전화기 위에 놓여 있어 통화 종료 상태가 읽힌다. 완장의 흰 글자는 식별 가능하다.",
        "hard_violations": [
         "검은 완장의 흰 글자를 읽을 수 있어, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지를 위반한다."
        ],
        "physics": "몸통은 직립하고 허벅지가 프레임 아래로 이어지는 정상적인 서 있는 자세다. 발은 구도 밖이므로 접지는 직접 확인되지 않지만 공중에 뜬 징후는 없다. 양팔과 열린 손은 어깨·팔꿈치·손목으로 자연스럽게 연결된다. 전화기와 수화기, 서류는 책상에 놓여 있고 바닥 종이는 바닥에 닿아 있다. 전화선은 책상 가장자리에서 아래로 늘어져 있으며 지지 없는 부유 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "얼굴은 카메라 오른쪽으로 돌아가 있고 눈은 그쪽의 약간 높은 화면 밖 지점을 바라본다. 두 팔은 양옆으로 넓게 펴지고 손가락과 손바닥은 앞쪽으로 열린다. 팔의 방향은 요청과 맞으며, 지시문이 지정한 별도의 시선 목표나 겨눌 대상은 없다.",
        "built_space": "왼쪽 뒤, 오른쪽 뒤, 오른쪽 전경의 책상 세 개 외에 왼쪽 전경 책상 일부가 보여 총 네 책상이 보인다. 녹색 의자 등받이는 세 개가 보인다. 어두운 창과 상부 블라인드, 열린 회색 문 하나, 목재 벽의 지도·사진, 투톤 벽과 타일 바닥은 같은 종류의 공간이다. 이전 장면보다 왼쪽 공간이 더 노출되어 책상과 밝은색 전화기가 추가로 보이지만, 참조에서 가려진 영역이므로 고정 설비의 복제라고 단정할 수는 없다. 인물은 중앙 통로에 서 있고 팔 주변에 여유가 있다. 전경 유선전화기의 수화기는 본체 위에서 보이지 않는다. 불가능한 반사는 없다.",
        "entities": "중년 남성 한 명만 있고, 짧은 회색 머리와 수염, 올리브색 군복, 검은 완장은 박철진 참조와 대체로 맞는다. 눈을 크게 뜨고 입을 벌린 미소와 목의 힘줄은 보이지만, 얼굴을 일그러뜨린 광기보다는 기뻐하는 표정에 가깝다. 양손에는 수화기나 다른 물건이 없다. 기존 검은 전화기 두 종류와 서류가 보이며, 왼쪽 끝에는 밝은색 전화기가 일부 보인다. 완장의 흰 글자는 식별 가능하다.",
        "hard_violations": [
         "검은 완장의 흰 글자를 읽을 수 있어, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지를 위반한다."
        ],
        "physics": "직립한 몸통과 아래로 이어지는 골반은 통상적인 기립 자세이며, 발은 프레임 밖이다. 팔을 어깨 높이로 벌리는 동작과 양손의 연결에는 명백한 해부학적 불가능성이 없다. 전화기와 서류는 책상에, 흩어진 종이는 바닥에 지지된다. 전경 전화선은 책상 가장자리로 처져 있다. 수화기의 위치는 확인되지 않지만 공중에 떠 있는 수화기나 지지 없는 물체가 보이는 것은 아니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 팔을 활짝 벌린 동작과 일그러진 광기 어린 미소, 기존 사무실의 배치가 더 충실하지만, 완장의 읽을 수 있는 글자는 명시적 금지 위반이다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "팔을 벌린 동작과 목의 긴장은 맞지만 표정이 광기보다는 환희에 가깝고 배경 연속성이 덜 정확하며, 완장의 읽을 수 있는 글자도 금지 위반이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 카메라 오른쪽의 화면 밖 상대 방향을 향한다. 특정 상대는 보이지 않으며 지시문도 시선의 목표를 고정하지 않는다. 두 팔은 몸 양옆으로 거의 수평으로 펼쳐지고 손바닥은 카메라 쪽으로 비스듬히 열린다. 무기나 겨누는 물체는 없다.",
        "built_space": "왼쪽 책상 하나, 뒤쪽 책상 하나, 오른쪽 전경 책상 하나로 총 세 책상이 보인다. 녹색 의자 등받이는 왼쪽과 오른쪽에 각각 하나씩 보이고, 뒤쪽 의자 자리는 인물과 가구에 가려진다. 중앙의 어두운 창, 왼쪽 일부 창, 열린 회색 문 하나, 오른쪽 목재 벽과 지도·사진, 오염된 투톤 벽과 타일 바닥이 이전 장면과 대체로 이어진다. 인물은 책상 사이 통로에 서 있고 팔이 가구와 충돌하지 않는다. 뒤쪽 회의용 전화기와 전경 유선전화기가 각각 하나씩 있다. 불가능한 거울 반사는 없다.",
        "entities": "박철진으로 제시된 중년 남성 한 명만 등장한다. 짧게 깎은 회색 섞인 머리, 회색 수염, 얼굴과 체격은 참조와 대체로 맞고, 이전 장면처럼 모자가 없다. 올리브색 군복과 검은 완장도 유지된다. 이를 드러낸 미소, 좁아진 눈, 코와 볼의 강한 긴장이 광기 어린 연기에 가깝다. 얼굴의 핏대 자체는 뚜렷하지 않다. 수화기는 손에 없고 전경 전화기 위에 놓여 있어 통화 종료 상태가 읽힌다. 완장의 흰 글자는 식별 가능하다.",
        "hard_violations": [
         "검은 완장의 흰 글자를 읽을 수 있어, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지를 위반한다."
        ],
        "physics": "몸통은 직립하고 허벅지가 프레임 아래로 이어지는 정상적인 서 있는 자세다. 발은 구도 밖이므로 접지는 직접 확인되지 않지만 공중에 뜬 징후는 없다. 양팔과 열린 손은 어깨·팔꿈치·손목으로 자연스럽게 연결된다. 전화기와 수화기, 서류는 책상에 놓여 있고 바닥 종이는 바닥에 닿아 있다. 전화선은 책상 가장자리에서 아래로 늘어져 있으며 지지 없는 부유 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "얼굴은 카메라 오른쪽으로 돌아가 있고 눈은 그쪽의 약간 높은 화면 밖 지점을 바라본다. 두 팔은 양옆으로 넓게 펴지고 손가락과 손바닥은 앞쪽으로 열린다. 팔의 방향은 요청과 맞으며, 지시문이 지정한 별도의 시선 목표나 겨눌 대상은 없다.",
        "built_space": "왼쪽 뒤, 오른쪽 뒤, 오른쪽 전경의 책상 세 개 외에 왼쪽 전경 책상 일부가 보여 총 네 책상이 보인다. 녹색 의자 등받이는 세 개가 보인다. 어두운 창과 상부 블라인드, 열린 회색 문 하나, 목재 벽의 지도·사진, 투톤 벽과 타일 바닥은 같은 종류의 공간이다. 이전 장면보다 왼쪽 공간이 더 노출되어 책상과 밝은색 전화기가 추가로 보이지만, 참조에서 가려진 영역이므로 고정 설비의 복제라고 단정할 수는 없다. 인물은 중앙 통로에 서 있고 팔 주변에 여유가 있다. 전경 유선전화기의 수화기는 본체 위에서 보이지 않는다. 불가능한 반사는 없다.",
        "entities": "중년 남성 한 명만 있고, 짧은 회색 머리와 수염, 올리브색 군복, 검은 완장은 박철진 참조와 대체로 맞는다. 눈을 크게 뜨고 입을 벌린 미소와 목의 힘줄은 보이지만, 얼굴을 일그러뜨린 광기보다는 기뻐하는 표정에 가깝다. 양손에는 수화기나 다른 물건이 없다. 기존 검은 전화기 두 종류와 서류가 보이며, 왼쪽 끝에는 밝은색 전화기가 일부 보인다. 완장의 흰 글자는 식별 가능하다.",
        "hard_violations": [
         "검은 완장의 흰 글자를 읽을 수 있어, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지를 위반한다."
        ],
        "physics": "직립한 몸통과 아래로 이어지는 골반은 통상적인 기립 자세이며, 발은 프레임 밖이다. 팔을 어깨 높이로 벌리는 동작과 양손의 연결에는 명백한 해부학적 불가능성이 없다. 전화기와 서류는 책상에, 흩어진 종이는 바닥에 지지된다. 전경 전화선은 책상 가장자리로 처져 있다. 수화기의 위치는 확인되지 않지만 공중에 떠 있는 수화기나 지지 없는 물체가 보이는 것은 아니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.179
   },
   "violations": {
    "B": [
     "[gemini-pro] 우측 책상 위 전화기의 수화기 선이 바닥으로 늘어지다가 캐릭터의 하복부(사타구니)에 직접 연결되어 있는 물리적으로 불가능한 형태",
     "[gpt-high] 검은 완장의 흰 글자를 읽을 수 있어, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지를 위반한다."
    ],
    "A": [
     "[gpt-high] 검은 완장의 흰 글자를 읽을 수 있어, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1500,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "배경과 인물의 디테일을 레퍼런스에 맞게 잘 구현하였고 구도도 정확하지만, 프롬프트가 요구한 '핏대를 세우고 광기 어린 미소'라기보다는 다소 밝은 웃음에 가깝습니다.  ★위반: [gpt-high] 검은 완장의 흰 글자를 읽을 수 있어, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지를 위반한다."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "프롬프트가 요구한 광기 어린 표정과 핏대는 훌륭하게 표현되었으나, 전화선이 몸에 연결되는 치명적인 물리적 오류로 인해 감점되었습니다.  ★위반: [gemini-pro] 우측 책상 위 전화기의 수화기 선이 바닥으로 늘어지다가 캐릭터의 하복부(사타구니)에 직접 연결되어 있는 물리적으로 불가능한 형태 / [gpt-high] 검은 완장의 흰 글자를 읽을 수 있어, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S43sh5_sel.png",
    "asset_id": "f03ca58b-d537-4428-949b-47cf6afc31a9",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdca-68c6-792f-9cf7-87ebcae18468",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S43sh5"
  }
 },
 "S43sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:39:34.559991+00:00",
  "fingerprint": "a876e5407db6e6a724517524f4fe0a0779097f7546159ed0e3bcea25c571d618",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S43sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S43sh10_sel.png",
  "source_sha256": "8062cf1de9bbaca326d9ada9c574bf9a338ef09a68249e87d80b4f4c17a70ed4",
  "file": "S43sh10_cine.png",
  "staged_sha256": "e20af14207f72a7be6bc4bb9a1ba2955ab63b360c8903d9dffccc15fa69c6d43",
  "latency_ms": 12343
 },
 "S44sh3::signage": {
  "fp": "c5346b09f7fb6292",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S44sh3": {
  "input_fingerprint": "d98a1b4dcde3262c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 셔터가 반쯤 열린 어두운 창고 안, 낡지만 매끈한 소형 캠핑카의 전면 외관.\n\nLOCATION (lock): Inside a dark storage shed beside an abandoned factory, just behind the half-raised shutter where a small camper is parked. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lower edge of the half-open shutter in the upper-center of the frame, foreground; Revealed camper front beneath the shutter in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Warehouse shutter (Halfway open and being raised) — The entrance-facing surface and lower edge are visible above the camper; used as Creates an upper masking edge for the reveal; Small camper (Old, with a smooth exterior, parked inside the warehouse) — Front and one adjoining side are visible; used as Primary reveal subject with surrounding space retained for scale; Warehouse entrance (Partly exposed beneath the rising shutter) — Viewed from outside at an offset angle; used as Frames the vehicle and establishes the path of the continuing approach.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the warehouse darkness with restrained exposure that gradually separates the camper's front from its surroundings without specifying a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter is being raised, revealing a small camper in the dark interior. The packed supplies and cash have been brought along, and Charlie retains the blanket obtained at the shop.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 셔터가 반쯤 열린 어두운 창고 안, 낡지만 매끈한 소형 캠핑카의 전면 외관.\n\nLOCATION (lock): Inside a dark storage shed beside an abandoned factory, just behind the half-raised shutter where a small camper is parked. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lower edge of the half-open shutter in the upper-center of the frame, foreground; Revealed camper front beneath the shutter in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Warehouse shutter (Halfway open and being raised) — The entrance-facing surface and lower edge are visible above the camper; used as Creates an upper masking edge for the reveal; Small camper (Old, with a smooth exterior, parked inside the warehouse) — Front and one adjoining side are visible; used as Primary reveal subject with surrounding space retained for scale; Warehouse entrance (Partly exposed beneath the rising shutter) — Viewed from outside at an offset angle; used as Frames the vehicle and establishes the path of the continuing approach.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the warehouse darkness with restrained exposure that gradually separates the camper's front from its surroundings without specifying a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter is being raised, revealing a small camper in the dark interior. The packed supplies and cash have been brought along, and Charlie retains the blanket obtained at the shop.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 셔터가 반쯤 열린 어두운 창고 안, 낡지만 매끈한 소형 캠핑카의 전면 외관.\n\nLOCATION (lock): Inside a dark storage shed beside an abandoned factory, just behind the half-raised shutter where a small camper is parked. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lower edge of the half-open shutter in the upper-center of the frame, foreground; Revealed camper front beneath the shutter in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Warehouse shutter (Halfway open and being raised) — The entrance-facing surface and lower edge are visible above the camper; used as Creates an upper masking edge for the reveal; Small camper (Old, with a smooth exterior, parked inside the warehouse) — Front and one adjoining side are visible; used as Primary reveal subject with surrounding space retained for scale; Warehouse entrance (Partly exposed beneath the rising shutter) — Viewed from outside at an offset angle; used as Frames the vehicle and establishes the path of the continuing approach.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the warehouse darkness with restrained exposure that gradually separates the camper's front from its surroundings without specifying a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter is being raised, revealing a small camper in the dark interior. The packed supplies and cash have been brought along, and Charlie retains the blanket obtained at the shop.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 셔터 아래를 지나 창고 중앙의 소형 캠핑카 정면을 향함.",
    "built_space": "상단에 반쯤 열린 금속 셔터가 있고 어두운 창고 내부에 캠핑카가 배치됨.",
    "entities": "낡고 매끈한 흰색 캠핑카, 바닥에 놓인 담요, 인물 없음.",
    "hard_violations": [],
    "physics": "캠핑카는 바퀴와 전면 지지대로 바닥에 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 창고 내부의 캠핑카를 향하며 우측 전경에 구조물과 손이 보임.",
    "built_space": "상단 셔터와 내부 장비, 적재물들이 있는 창고 공간.",
    "entities": "캠핑카, 전경에 지시문에 없는 사람의 손과 두꺼운 소매.",
    "hard_violations": [
     "[gemini-pro] 지시문에 묘사되지 않은 인물의 신체(손과 팔)가 등장하여 인물 배제(No people) 규칙을 위반함",
     "[gpt-high] 오른쪽 전경에 셔터를 올리는 동작과 무관한 인물의 손·팔과 신체 일부를 추가했다. 이는 셔터 조작 손의 예외에 해당하지 않는다.",
     "[gpt-high] 장소 설명에 없는 공작기계성 설비, 냉방기, 선반과 다량의 적재 자재를 추가하여 명시된 범위 밖의 작업장 구성을 만들었다."
    ],
    "physics": "캠핑카는 바닥에 지지되어 있고 손은 금속 프레임 위에 얹혀 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 프레이밍, 조명, 인물 배제 규칙을 정확히 준수하여 샷 텍스트를 충실히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물 등장 불가 지시에도 불구하고 우측 전경에 사람의 신체가 포함되어 심각한 위반을 범함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 셔터 아래를 지나 창고 중앙의 소형 캠핑카 정면을 향함.",
        "built_space": "상단에 반쯤 열린 금속 셔터가 있고 어두운 창고 내부에 캠핑카가 배치됨.",
        "entities": "낡고 매끈한 흰색 캠핑카, 바닥에 놓인 담요, 인물 없음.",
        "hard_violations": [],
        "physics": "캠핑카는 바퀴와 전면 지지대로 바닥에 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 창고 내부의 캠핑카를 향하며 우측 전경에 구조물과 손이 보임.",
        "built_space": "상단 셔터와 내부 장비, 적재물들이 있는 창고 공간.",
        "entities": "캠핑카, 전경에 지시문에 없는 사람의 손과 두꺼운 소매.",
        "hard_violations": [
         "지시문에 묘사되지 않은 인물의 신체(손과 팔)가 등장하여 인물 배제(No people) 규칙을 위반함"
        ],
        "physics": "캠핑카는 바닥에 지지되어 있고 손은 금속 프레임 위에 얹혀 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 프레이밍, 조명, 인물 배제 규칙을 정확히 준수하여 샷 텍스트를 충실히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물 등장 불가 지시에도 불구하고 우측 전경에 사람의 신체가 포함되어 심각한 위반을 범함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 셔터 아래를 지나 창고 중앙의 소형 캠핑카 정면을 향함.",
        "built_space": "상단에 반쯤 열린 금속 셔터가 있고 어두운 창고 내부에 캠핑카가 배치됨.",
        "entities": "낡고 매끈한 흰색 캠핑카, 바닥에 놓인 담요, 인물 없음.",
        "hard_violations": [],
        "physics": "캠핑카는 바퀴와 전면 지지대로 바닥에 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 창고 내부의 캠핑카를 향하며 우측 전경에 구조물과 손이 보임.",
        "built_space": "상단 셔터와 내부 장비, 적재물들이 있는 창고 공간.",
        "entities": "캠핑카, 전경에 지시문에 없는 사람의 손과 두꺼운 소매.",
        "hard_violations": [
         "지시문에 묘사되지 않은 인물의 신체(손과 팔)가 등장하여 인물 배제(No people) 규칙을 위반함"
        ],
        "physics": "캠핑카는 바닥에 지지되어 있고 손은 금속 프레임 위에 얹혀 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "셔터 아래 캠핑 차량을 보여주는 와이드 구도는 맞지만, 셔터를 조작하지 않는 인물 일부와 불필요한 작업장 집기를 추가해 장면 제한을 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "인물 없이 반쯤 열린 셔터 아래 차량 전면과 한쪽 측면을 어둡게 드러내어 지정된 외부 사선 와이드 구도에 가장 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "차량 전면과 견인 연결부가 출입구 바깥의 화면 왼쪽 앞을 향하며 오른쪽 측면이 보인다. 카메라는 출입구 밖에서 비스듬히 안을 본다. 오른쪽 아래 손은 셔터가 아니라 짐으로 보이는 물체 위에 놓여 있다. 확인 가능한 시선이나 무기는 없다.",
        "built_space": "출입구 하나와 그 양옆 기둥, 상단을 가리는 금속 셔터 하나가 보인다. 셔터 하단은 화면 상부를 가로지르고 차량 한 대는 중앙 중경에 놓여 있어 주요 배치는 맞는다. 그러나 뒤쪽 창 두 개, 왼쪽 공작기계처럼 보이는 설비, 오른쪽 세로형 냉방기와 선반, 여러 자재 더미까지 추가되어 단순한 어두운 보관 창고보다 구체적인 작업장으로 설정되었다.",
        "entities": "낡고 비교적 매끈한 소형 캠핑 차량 한 대가 있으나, 운전석과 전조등이 있는 자주식 캠핑카보다는 견인식 카라반이다. 오른쪽 전경에는 사람의 손과 팔, 담요 같은 천 및 짐 일부가 보이며 얼굴은 없다. 손만으로 나이·성별·민족적 정체성을 판별할 수 없다. 상자와 포대는 보이지만 현금은 식별되지 않는다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "오른쪽 전경에 셔터를 올리는 동작과 무관한 인물의 손·팔과 신체 일부를 추가했다. 이는 셔터 조작 손의 예외에 해당하지 않는다.",
         "장소 설명에 없는 공작기계성 설비, 냉방기, 선반과 다량의 적재 자재를 추가하여 명시된 범위 밖의 작업장 구성을 만들었다."
        ],
        "physics": "차량은 측면 바퀴와 전방 지지대로 바닥에 지지된다. 셔터는 출입구의 레일과 상부 구조에 연결되어 있어 중간 높이에 있는 상태가 물리적으로 가능하지만, 실제 상승 동작은 정지 화면만으로 확인하기 어렵다. 상자와 자재는 바닥이나 적재물 위에 놓여 있고, 전경 손은 짐 위에 닿아 있다. 명백히 무지지 상태로 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "차량 전면과 견인봉은 출입구 바깥의 화면 왼쪽 앞을 향하고, 오른쪽으로 이어지는 측면이 보인다. 카메라는 창고 밖 사선 위치에서 안을 바라본다. 셔터의 바깥 면과 하단이 차량 위를 가려 지정된 공개 방향에 맞는다. 사람의 시선이나 겨냥하는 물체는 없다.",
        "built_space": "출입구 하나, 양옆 수직 문틀, 반쯤 내려와 있는 셔터 하나가 보인다. 셔터 하단은 전경 상부 중앙을 지나며 차량 한 대가 그 아래 중앙 중경에 있다. 차량 앞쪽 바닥과 주변 여백을 남긴 와이드 구도이며, 창고 내부는 어둠에 묻혀 불필요한 고정 설비가 두드러지지 않는다. 왼쪽에 상자류, 차량 오른쪽 아래에 짐과 천이 보인다.",
        "entities": "낡고 매끈한 밝은색 소형 캠핑 차량 한 대가 보이며, 이 역시 견인식 카라반 형태다. 사람이나 얼굴은 없다. 차량 옆에는 포장된 짐으로 읽히는 가방과 상자류가 있고, 담요처럼 보이는 천은 바닥 가까이 덮여 있다. 그 천이 찰리가 보유한 담요인지는 확인할 수 없으며, 현금도 보이지 않는다. 읽을 수 있는 글자나 로고는 없다.",
        "hard_violations": [],
        "physics": "차량은 측면 바퀴와 견인봉 아래 보조 바퀴, 하부 지지대에 의해 바닥에 지지된다. 셔터는 문틀의 레일에 연결되어 있어 별도의 손이 보이지 않아도 해당 높이에 있을 수 있다. 상승 중이라는 운동 자체는 명확히 드러나지 않는다. 짐은 바닥에 놓여 있고 천은 낮은 짐과 바닥 위로 처져 있어, 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "셔터 아래 캠핑 차량을 보여주는 와이드 구도는 맞지만, 셔터를 조작하지 않는 인물 일부와 불필요한 작업장 집기를 추가해 장면 제한을 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "인물 없이 반쯤 열린 셔터 아래 차량 전면과 한쪽 측면을 어둡게 드러내어 지정된 외부 사선 와이드 구도에 가장 충실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "차량 전면과 견인 연결부가 출입구 바깥의 화면 왼쪽 앞을 향하며 오른쪽 측면이 보인다. 카메라는 출입구 밖에서 비스듬히 안을 본다. 오른쪽 아래 손은 셔터가 아니라 짐으로 보이는 물체 위에 놓여 있다. 확인 가능한 시선이나 무기는 없다.",
        "built_space": "출입구 하나와 그 양옆 기둥, 상단을 가리는 금속 셔터 하나가 보인다. 셔터 하단은 화면 상부를 가로지르고 차량 한 대는 중앙 중경에 놓여 있어 주요 배치는 맞는다. 그러나 뒤쪽 창 두 개, 왼쪽 공작기계처럼 보이는 설비, 오른쪽 세로형 냉방기와 선반, 여러 자재 더미까지 추가되어 단순한 어두운 보관 창고보다 구체적인 작업장으로 설정되었다.",
        "entities": "낡고 비교적 매끈한 소형 캠핑 차량 한 대가 있으나, 운전석과 전조등이 있는 자주식 캠핑카보다는 견인식 카라반이다. 오른쪽 전경에는 사람의 손과 팔, 담요 같은 천 및 짐 일부가 보이며 얼굴은 없다. 손만으로 나이·성별·민족적 정체성을 판별할 수 없다. 상자와 포대는 보이지만 현금은 식별되지 않는다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "오른쪽 전경에 셔터를 올리는 동작과 무관한 인물의 손·팔과 신체 일부를 추가했다. 이는 셔터 조작 손의 예외에 해당하지 않는다.",
         "장소 설명에 없는 공작기계성 설비, 냉방기, 선반과 다량의 적재 자재를 추가하여 명시된 범위 밖의 작업장 구성을 만들었다."
        ],
        "physics": "차량은 측면 바퀴와 전방 지지대로 바닥에 지지된다. 셔터는 출입구의 레일과 상부 구조에 연결되어 있어 중간 높이에 있는 상태가 물리적으로 가능하지만, 실제 상승 동작은 정지 화면만으로 확인하기 어렵다. 상자와 자재는 바닥이나 적재물 위에 놓여 있고, 전경 손은 짐 위에 닿아 있다. 명백히 무지지 상태로 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "차량 전면과 견인봉은 출입구 바깥의 화면 왼쪽 앞을 향하고, 오른쪽으로 이어지는 측면이 보인다. 카메라는 창고 밖 사선 위치에서 안을 바라본다. 셔터의 바깥 면과 하단이 차량 위를 가려 지정된 공개 방향에 맞는다. 사람의 시선이나 겨냥하는 물체는 없다.",
        "built_space": "출입구 하나, 양옆 수직 문틀, 반쯤 내려와 있는 셔터 하나가 보인다. 셔터 하단은 전경 상부 중앙을 지나며 차량 한 대가 그 아래 중앙 중경에 있다. 차량 앞쪽 바닥과 주변 여백을 남긴 와이드 구도이며, 창고 내부는 어둠에 묻혀 불필요한 고정 설비가 두드러지지 않는다. 왼쪽에 상자류, 차량 오른쪽 아래에 짐과 천이 보인다.",
        "entities": "낡고 매끈한 밝은색 소형 캠핑 차량 한 대가 보이며, 이 역시 견인식 카라반 형태다. 사람이나 얼굴은 없다. 차량 옆에는 포장된 짐으로 읽히는 가방과 상자류가 있고, 담요처럼 보이는 천은 바닥 가까이 덮여 있다. 그 천이 찰리가 보유한 담요인지는 확인할 수 없으며, 현금도 보이지 않는다. 읽을 수 있는 글자나 로고는 없다.",
        "hard_violations": [],
        "physics": "차량은 측면 바퀴와 견인봉 아래 보조 바퀴, 하부 지지대에 의해 바닥에 지지된다. 셔터는 문틀의 레일에 연결되어 있어 별도의 손이 보이지 않아도 해당 높이에 있을 수 있다. 상승 중이라는 운동 자체는 명확히 드러나지 않는다. 짐은 바닥에 놓여 있고 천은 낮은 짐과 바닥 위로 처져 있어, 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.804
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.554
   },
   "violations": {
    "B": [
     "[gemini-pro] 지시문에 묘사되지 않은 인물의 신체(손과 팔)가 등장하여 인물 배제(No people) 규칙을 위반함",
     "[gpt-high] 오른쪽 전경에 셔터를 올리는 동작과 무관한 인물의 손·팔과 신체 일부를 추가했다. 이는 셔터 조작 손의 예외에 해당하지 않는다.",
     "[gpt-high] 장소 설명에 없는 공작기계성 설비, 냉방기, 선반과 다량의 적재 자재를 추가하여 명시된 범위 밖의 작업장 구성을 만들었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 554
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 프레이밍, 조명, 인물 배제 규칙을 정확히 준수하여 샷 텍스트를 충실히 구현함."
   },
   {
    "label": "B",
    "score": 554,
    "verdict_ko": "인물 등장 불가 지시에도 불구하고 우측 전경에 사람의 신체가 포함되어 심각한 위반을 범함.  ★위반: [gemini-pro] 지시문에 묘사되지 않은 인물의 신체(손과 팔)가 등장하여 인물 배제(No people) 규칙을 위반함 / [gpt-high] 오른쪽 전경에 셔터를 올리는 동작과 무관한 인물의 손·팔과 신체 일부를 추가했다. 이는 셔터 조작 손의 예외에 해당하지 않는다. / [gpt-high] 장소 설명에 없는 공작기계성 설비, 냉방기, 선반과 다량의 적재 자재를 추가하여 명시된 범위 밖의 작업장 구성을 만들었다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdcb-33a3-7e4a-a708-40b83467e09b",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S44sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:39:47.235788+00:00",
  "fingerprint": "e08e597f71816bbf75313e661cdb2dbc5f1db6ebdbb6bfb07c09a33b6d152de4",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S44sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S44sh3_sel.png",
  "source_sha256": "360c10f2d1e64e128e055bd6bc24c9721e8cae482608c4d8f000a95919cead76",
  "file": "S44sh3_cine.png",
  "staged_sha256": "d583315e5f4da7752a74f9031a14efa50ba621606981f969fcac521cbbf5cfb0",
  "latency_ms": 12588
 },
 "S44sh6::signage": {
  "fp": "0b69dbfe2d9db342",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S44sh6": {
  "input_fingerprint": "ad9632f46a4ae1c3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 캠핑카 문 안쪽에 푹신한 침대와 정돈된 내부가 희미한 빛 아래 아늑하게 드러난 구도.\n\nLOCATION (lock): Inside the camper's rear living compartment as seen through its open door within the dark warehouse, with a bed visible in faint ambient light. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camper rear doorway (Open) — Seen obliquely from the threshold, with the opening unobstructed; used as Provides narrow foreground framing around the interior; Bed (Soft and visible inside the camper) — Its upper surface is seen at a gentle downward angle; used as Primary interior detail, retaining surrounding space for scale; Camper interior (Orderly and spacious enough for four people) — Recedes diagonally from the rear opening; used as Makes the usable interior layout legible beyond the bed.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Faint illumination reveals the bed and orderly interior with gentle tonal separation, preserving the described coziness without assigning a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter is raised, and the camper's rear door is now open, revealing an interior with a bed and room for four. The packed supplies and cash remain available for the journey, and Charlie retains the shop blanket.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 캠핑카 문 안쪽에 푹신한 침대와 정돈된 내부가 희미한 빛 아래 아늑하게 드러난 구도.\n\nLOCATION (lock): Inside the camper's rear living compartment as seen through its open door within the dark warehouse, with a bed visible in faint ambient light. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camper rear doorway (Open) — Seen obliquely from the threshold, with the opening unobstructed; used as Provides narrow foreground framing around the interior; Bed (Soft and visible inside the camper) — Its upper surface is seen at a gentle downward angle; used as Primary interior detail, retaining surrounding space for scale; Camper interior (Orderly and spacious enough for four people) — Recedes diagonally from the rear opening; used as Makes the usable interior layout legible beyond the bed.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Faint illumination reveals the bed and orderly interior with gentle tonal separation, preserving the described coziness without assigning a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter is raised, and the camper's rear door is now open, revealing an interior with a bed and room for four. The packed supplies and cash remain available for the journey, and Charlie retains the shop blanket.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 캠핑카 문 안쪽에 푹신한 침대와 정돈된 내부가 희미한 빛 아래 아늑하게 드러난 구도.\n\nLOCATION (lock): Inside the camper's rear living compartment as seen through its open door within the dark warehouse, with a bed visible in faint ambient light. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camper rear doorway (Open) — Seen obliquely from the threshold, with the opening unobstructed; used as Provides narrow foreground framing around the interior; Bed (Soft and visible inside the camper) — Its upper surface is seen at a gentle downward angle; used as Primary interior detail, retaining surrounding space for scale; Camper interior (Orderly and spacious enough for four people) — Recedes diagonally from the rear opening; used as Makes the usable interior layout legible beyond the bed.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Faint illumination reveals the bed and orderly interior with gentle tonal separation, preserving the described coziness without assigning a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter is raised, and the camper's rear door is now open, revealing an interior with a bed and room for four. The packed supplies and cash remain available for the journey, and Charlie retains the shop blanket.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 캠핑카의 후면 문을 사선이 아닌 정면에서 똑바로 바라보며 시선이 중앙 통로를 향함.",
    "built_space": "어두운 창고 내부. 양쪽으로 열린 캠핑카 후면 문 안으로 좌측에 소파, 우측에 테이블이 있고 맨 안쪽 끝에 침대가 위치함. 후면 범퍼 위에 물건들이 놓여 있음.",
    "entities": "사람은 등장하지 않음. 캠핑카, 침대, 지폐 뭉치와 캔 등 보급품이 묘사됨.",
    "hard_violations": [],
    "physics": "현금과 보급품은 범퍼 표면 위에 안정적으로 놓여 있으며, 공중에 떠 있는 물체는 없음."
   },
   {
    "label": "B",
    "direction": "카메라가 캠핑카 후면을 비스듬한 사선 각도(oblique)로 바라보며, 시선이 전면의 침대를 지나 내부 주방 쪽으로 향함.",
    "built_space": "어두운 창고 내부. 좌측으로 크게 열린 단일 후면 문을 통해 진입 공간 전체를 차지하는 넓은 침대가 보이며, 그 너머로 주방과 운전석 등 내부 구조가 깊숙이 전개됨.",
    "entities": "사람은 등장하지 않음. 푹신한 침구류와 캠핑카 내부는 묘사되었으나 프롬프트에 명시된 현금과 보급품은 보이지 않음.",
    "hard_violations": [
     "[gpt-high] 화면 하단 번호판에 판독 가능한 숫자가 남아 있어 읽을 수 있는 문자를 전면 금지한 조건을 위반한다."
    ],
    "physics": "침대와 내부 가구들은 캠핑카 차체 구조에 맞게 배치되어 지지되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 비스듬한 카메라 각도(oblique)를 정확히 구현했으며, 침대를 전면에 두고 그 너머로 내부 배치가 이어지도록 한 프레이밍 최우선 지침을 완벽히 충족했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "현금과 보급품 소품은 표현되었으나, 카메라가 정면을 향하고 침대가 맨 안쪽에 배치되어 숏 텍스트의 앵글 및 구도 우선순위를 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 캠핑카의 후면 문을 사선이 아닌 정면에서 똑바로 바라보며 시선이 중앙 통로를 향함.",
        "built_space": "어두운 창고 내부. 양쪽으로 열린 캠핑카 후면 문 안으로 좌측에 소파, 우측에 테이블이 있고 맨 안쪽 끝에 침대가 위치함. 후면 범퍼 위에 물건들이 놓여 있음.",
        "entities": "사람은 등장하지 않음. 캠핑카, 침대, 지폐 뭉치와 캔 등 보급품이 묘사됨.",
        "hard_violations": [],
        "physics": "현금과 보급품은 범퍼 표면 위에 안정적으로 놓여 있으며, 공중에 떠 있는 물체는 없음."
       },
       {
        "label": "B",
        "direction": "카메라가 캠핑카 후면을 비스듬한 사선 각도(oblique)로 바라보며, 시선이 전면의 침대를 지나 내부 주방 쪽으로 향함.",
        "built_space": "어두운 창고 내부. 좌측으로 크게 열린 단일 후면 문을 통해 진입 공간 전체를 차지하는 넓은 침대가 보이며, 그 너머로 주방과 운전석 등 내부 구조가 깊숙이 전개됨.",
        "entities": "사람은 등장하지 않음. 푹신한 침구류와 캠핑카 내부는 묘사되었으나 프롬프트에 명시된 현금과 보급품은 보이지 않음.",
        "hard_violations": [],
        "physics": "침대와 내부 가구들은 캠핑카 차체 구조에 맞게 배치되어 지지되고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 비스듬한 카메라 각도(oblique)를 정확히 구현했으며, 침대를 전면에 두고 그 너머로 내부 배치가 이어지도록 한 프레이밍 최우선 지침을 완벽히 충족했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "현금과 보급품 소품은 표현되었으나, 카메라가 정면을 향하고 침대가 맨 안쪽에 배치되어 숏 텍스트의 앵글 및 구도 우선순위를 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 캠핑카의 후면 문을 사선이 아닌 정면에서 똑바로 바라보며 시선이 중앙 통로를 향함.",
        "built_space": "어두운 창고 내부. 양쪽으로 열린 캠핑카 후면 문 안으로 좌측에 소파, 우측에 테이블이 있고 맨 안쪽 끝에 침대가 위치함. 후면 범퍼 위에 물건들이 놓여 있음.",
        "entities": "사람은 등장하지 않음. 캠핑카, 침대, 지폐 뭉치와 캔 등 보급품이 묘사됨.",
        "hard_violations": [],
        "physics": "현금과 보급품은 범퍼 표면 위에 안정적으로 놓여 있으며, 공중에 떠 있는 물체는 없음."
       },
       {
        "label": "B",
        "direction": "카메라가 캠핑카 후면을 비스듬한 사선 각도(oblique)로 바라보며, 시선이 전면의 침대를 지나 내부 주방 쪽으로 향함.",
        "built_space": "어두운 창고 내부. 좌측으로 크게 열린 단일 후면 문을 통해 진입 공간 전체를 차지하는 넓은 침대가 보이며, 그 너머로 주방과 운전석 등 내부 구조가 깊숙이 전개됨.",
        "entities": "사람은 등장하지 않음. 푹신한 침구류와 캠핑카 내부는 묘사되었으나 프롬프트에 명시된 현금과 보급품은 보이지 않음.",
        "hard_violations": [],
        "physics": "침대와 내부 가구들은 캠핑카 차체 구조에 맞게 배치되어 지지되고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "침대 중심의 가까운 사선 구도와 희미한 실내광은 더 충실하지만, 화면 하단 번호판의 판독 가능한 숫자가 문자 금지 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "열린 후문과 침대, 비어 있는 생활공간은 충실하지만, 문턱의 사선 시점 대신 창고 밖으로 물러난 정면 구도로 촬영했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "사람이나 시선, 겨누는 물체는 없다. 카메라는 열린 후문을 약간 비스듬히 바라보며 침대 윗면을 완만하게 내려다본다. 침대 뒤 생활공간도 안쪽으로 이어져 요청한 시점에 비교적 가깝다.",
        "built_space": "왼쪽으로 열린 문짝 하나와 오른쪽 가장자리의 문짝 일부가 보인다. 입구 바로 안쪽에 침대 하나, 왼쪽과 위쪽에 수납장들, 안쪽 중앙에 싱크대 하나, 오른쪽에 좌석 공간, 더 깊은 곳에 운전석 등받이가 보인다. 창고 벽과 바닥도 일부 포함된다. 침대가 개구부 너비 대부분을 차지해 내부 통로와 네 명의 사용 공간은 충분히 읽히지 않는다. 불가능한 반사나 명백히 중복된 설비는 없다.",
        "entities": "푹신한 침대와 베개, 구김이 있는 침구, 열린 캠핑카 후문, 어두운 창고가 보이며 사람은 없다. 식별할 인물이나 참조 사진은 주어지지 않았다. 보급품과 현금, 창고 셔터는 이 구도에서 확인되지 않으며 이를 보여주려고 확대할 필요는 없다. 다만 화면 맨 아래 번호판에 '70'으로 읽히는 숫자 일부가 노출되어 있다.",
        "hard_violations": [
         "화면 하단 번호판에 판독 가능한 숫자가 남아 있어 읽을 수 있는 문자를 전면 금지한 조건을 위반한다."
        ],
        "physics": "침대는 하부 받침에 놓여 있고 베개와 이불은 매트리스에 지지된다. 이불은 침대 가장자리에서 자연스럽게 늘어진다. 열린 문은 경첩으로 차체에 연결되어 있으며 수납장과 좌석도 고정되어 있다. 지지 없이 떠 있는 물체나 신체는 없다."
       },
       {
        "label": "B",
        "direction": "사람이나 시선, 겨누는 물체는 없다. 카메라는 캠핑카 후문과 통로를 거의 정면으로 바라본다. 침대는 통로 끝에 작게 보이며, 문턱에서 비스듬히 침대 윗면을 내려다보는 요청과 달리 축이 중앙으로 곧게 뻗는다.",
        "built_space": "위로 올라간 창고 셔터 하나가 화면 상단에 있고, 캠핑카의 후문 두 짝은 좌우로 열려 있다. 중앙 통로 끝에 침대 하나, 왼쪽에 벤치 하나, 오른쪽에 싱크대 하나와 작은 탁자 및 좌석, 양쪽 상부에 수납장들이 보인다. 통로가 비어 있어 생활공간의 배치는 명료하다. 그러나 창고 문틀과 차체 외부가 크게 들어와 후문이 좁은 전경 테두리 역할을 하지 못한다. 불가능한 반사나 설비 중복은 보이지 않는다.",
        "entities": "사람 없이 열린 캠핑카, 침대와 베개 및 침구, 정돈된 수납공간, 어두운 창고와 올라간 셔터를 보여준다. 후방 발판에는 지폐 묶음과 용기 등 보급품으로 읽히는 물건이 놓여 있다. 지폐의 국가와 액면은 확정할 만큼 선명하지 않으며 프롬프트도 이를 지정하지 않았다. 명확히 읽히는 문자는 없다. 천장 조명들의 발광과 따뜻한 색조는 광원과 색을 특정하지 않은 희미한 주변광 요청보다 구체적이다.",
        "hard_violations": [],
        "physics": "침대와 좌석은 하부 구조에 지지되고 침구는 그 위에 놓이거나 가장자리로 늘어진다. 현금과 용기들은 후방 발판 위에 놓여 있다. 양쪽 문은 경첩으로 연결되고 차체 아래에는 지지 다리도 보인다. 지지 없이 떠 있는 물체나 불가능한 동작은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "침대 중심의 가까운 사선 구도와 희미한 실내광은 더 충실하지만, 화면 하단 번호판의 판독 가능한 숫자가 문자 금지 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "열린 후문과 침대, 비어 있는 생활공간은 충실하지만, 문턱의 사선 시점 대신 창고 밖으로 물러난 정면 구도로 촬영했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "사람이나 시선, 겨누는 물체는 없다. 카메라는 열린 후문을 약간 비스듬히 바라보며 침대 윗면을 완만하게 내려다본다. 침대 뒤 생활공간도 안쪽으로 이어져 요청한 시점에 비교적 가깝다.",
        "built_space": "왼쪽으로 열린 문짝 하나와 오른쪽 가장자리의 문짝 일부가 보인다. 입구 바로 안쪽에 침대 하나, 왼쪽과 위쪽에 수납장들, 안쪽 중앙에 싱크대 하나, 오른쪽에 좌석 공간, 더 깊은 곳에 운전석 등받이가 보인다. 창고 벽과 바닥도 일부 포함된다. 침대가 개구부 너비 대부분을 차지해 내부 통로와 네 명의 사용 공간은 충분히 읽히지 않는다. 불가능한 반사나 명백히 중복된 설비는 없다.",
        "entities": "푹신한 침대와 베개, 구김이 있는 침구, 열린 캠핑카 후문, 어두운 창고가 보이며 사람은 없다. 식별할 인물이나 참조 사진은 주어지지 않았다. 보급품과 현금, 창고 셔터는 이 구도에서 확인되지 않으며 이를 보여주려고 확대할 필요는 없다. 다만 화면 맨 아래 번호판에 '70'으로 읽히는 숫자 일부가 노출되어 있다.",
        "hard_violations": [
         "화면 하단 번호판에 판독 가능한 숫자가 남아 있어 읽을 수 있는 문자를 전면 금지한 조건을 위반한다."
        ],
        "physics": "침대는 하부 받침에 놓여 있고 베개와 이불은 매트리스에 지지된다. 이불은 침대 가장자리에서 자연스럽게 늘어진다. 열린 문은 경첩으로 차체에 연결되어 있으며 수납장과 좌석도 고정되어 있다. 지지 없이 떠 있는 물체나 신체는 없다."
       },
       {
        "label": "A",
        "direction": "사람이나 시선, 겨누는 물체는 없다. 카메라는 캠핑카 후문과 통로를 거의 정면으로 바라본다. 침대는 통로 끝에 작게 보이며, 문턱에서 비스듬히 침대 윗면을 내려다보는 요청과 달리 축이 중앙으로 곧게 뻗는다.",
        "built_space": "위로 올라간 창고 셔터 하나가 화면 상단에 있고, 캠핑카의 후문 두 짝은 좌우로 열려 있다. 중앙 통로 끝에 침대 하나, 왼쪽에 벤치 하나, 오른쪽에 싱크대 하나와 작은 탁자 및 좌석, 양쪽 상부에 수납장들이 보인다. 통로가 비어 있어 생활공간의 배치는 명료하다. 그러나 창고 문틀과 차체 외부가 크게 들어와 후문이 좁은 전경 테두리 역할을 하지 못한다. 불가능한 반사나 설비 중복은 보이지 않는다.",
        "entities": "사람 없이 열린 캠핑카, 침대와 베개 및 침구, 정돈된 수납공간, 어두운 창고와 올라간 셔터를 보여준다. 후방 발판에는 지폐 묶음과 용기 등 보급품으로 읽히는 물건이 놓여 있다. 지폐의 국가와 액면은 확정할 만큼 선명하지 않으며 프롬프트도 이를 지정하지 않았다. 명확히 읽히는 문자는 없다. 천장 조명들의 발광과 따뜻한 색조는 광원과 색을 특정하지 않은 희미한 주변광 요청보다 구체적이다.",
        "hard_violations": [],
        "physics": "침대와 좌석은 하부 구조에 지지되고 침구는 그 위에 놓이거나 가장자리로 늘어진다. 현금과 용기들은 후방 발판 위에 놓여 있다. 양쪽 문은 경첩으로 연결되고 차체 아래에는 지지 다리도 보인다. 지지 없이 떠 있는 물체나 불가능한 동작은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gpt-high] 화면 하단 번호판에 판독 가능한 숫자가 남아 있어 읽을 수 있는 문자를 전면 금지한 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1250,
   "A": 1571
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "지시된 비스듬한 카메라 각도(oblique)를 정확히 구현했으며, 침대를 전면에 두고 그 너머로 내부 배치가 이어지도록 한 프레이밍 최우선 지침을 완벽히 충족했습니다.  ★위반: [gpt-high] 화면 하단 번호판에 판독 가능한 숫자가 남아 있어 읽을 수 있는 문자를 전면 금지한 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "현금과 보급품 소품은 표현되었으나, 카메라가 정면을 향하고 침대가 맨 안쪽에 배치되어 숏 텍스트의 앵글 및 구도 우선순위를 위반했습니다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdcc-02a5-741d-98be-802c261e466b",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S45sh1::signage": {
  "fp": "c2a6ffdde2bcccde",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S45sh1": {
  "input_fingerprint": "20f0cacfc8dea522",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 어두운 외곽 도로 위를 헤드라이트가 꺼진 낡은 캠핑카가 바퀴로 빗물을 튀기며 나아가는 mid-action 외관.\n\nLOCATION (lock): On a nearly deserted road at the city's outskirts at night, where the camper travels through rain without headlights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open road ahead of the rightward-moving camper in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Camper truck (Old and moving with headlights switched off) — Passenger side and a foreshortened forward section face the camera; used as Moving subject held at a stable size during the parallel track; Outskirts road (Rain-covered with almost no traffic) — Extends diagonally ahead toward screen right; used as Supplies forward space and motion reference; Rainwater at the wheels (Displaced outward by the turning wheels); used as Captures the vehicle's movement in a single frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the rainy nighttime darkness and unlit headlights, allowing only enough ambient exposure to distinguish the moving body and displaced water.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper travels through night rain with its headlights switched off and only a small rear-cabin lamp lit; the packed luggage and supplies are aboard. Charlie is seated in the rear with the shop blanket, and no repair to his old exterior has occurred.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 어두운 외곽 도로 위를 헤드라이트가 꺼진 낡은 캠핑카가 바퀴로 빗물을 튀기며 나아가는 mid-action 외관.\n\nLOCATION (lock): On a nearly deserted road at the city's outskirts at night, where the camper travels through rain without headlights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open road ahead of the rightward-moving camper in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Camper truck (Old and moving with headlights switched off) — Passenger side and a foreshortened forward section face the camera; used as Moving subject held at a stable size during the parallel track; Outskirts road (Rain-covered with almost no traffic) — Extends diagonally ahead toward screen right; used as Supplies forward space and motion reference; Rainwater at the wheels (Displaced outward by the turning wheels); used as Captures the vehicle's movement in a single frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the rainy nighttime darkness and unlit headlights, allowing only enough ambient exposure to distinguish the moving body and displaced water.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper travels through night rain with its headlights switched off and only a small rear-cabin lamp lit; the packed luggage and supplies are aboard. Charlie is seated in the rear with the shop blanket, and no repair to his old exterior has occurred.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 어두운 외곽 도로 위를 헤드라이트가 꺼진 낡은 캠핑카가 바퀴로 빗물을 튀기며 나아가는 mid-action 외관.\n\nLOCATION (lock): On a nearly deserted road at the city's outskirts at night, where the camper travels through rain without headlights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open road ahead of the rightward-moving camper in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Camper truck (Old and moving with headlights switched off) — Passenger side and a foreshortened forward section face the camera; used as Moving subject held at a stable size during the parallel track; Outskirts road (Rain-covered with almost no traffic) — Extends diagonally ahead toward screen right; used as Supplies forward space and motion reference; Rainwater at the wheels (Displaced outward by the turning wheels); used as Captures the vehicle's movement in a single frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the rainy nighttime darkness and unlit headlights, allowing only enough ambient exposure to distinguish the moving body and displaced water.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper travels through night rain with its headlights switched off and only a small rear-cabin lamp lit; the packed luggage and supplies are aboard. Charlie is seated in the rear with the shop blanket, and no repair to his old exterior has occurred.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "차량의 앞머리는 화면 왼쪽 아래와 카메라 쪽을 향한다. 오른쪽으로 멀어지는 도로는 차량의 앞이 아니라 뒤에 있으므로, 오른쪽으로 주행하며 전방 공간을 확보하라는 지시와 반대다. 앞유리 안 인물은 얼굴이 어두워 시선 방향을 확인할 수 없다.",
    "built_space": "왼쪽 가드레일 한 줄, 오른쪽 연석과 풀밭, 위쪽 전선과 전신주들, 오른쪽 원경으로 이어지는 가로등 열이 있어 참조 장소의 주요 구조를 유지한다. 캠핑카에는 상부 전면창 하나, 측면 출입문 하나와 그 창, 측면 상단의 작은 창 하나 및 뒤쪽 창 두 개가 보인다. 차량은 화면 중앙에서 왼쪽에 걸쳐 크게 놓여 있으며, 지시된 중간 오른쪽 배경의 작은 주행 피사체가 아니다. 젖은 노면의 가로등 반사는 가능한 위치에 있다.",
    "entities": "낡고 얼룩진 캠핑카 한 대, 꺼진 전조등, 비, 젖은 도로와 바퀴 주변 물보라가 보인다. 측면 창 하나에서 작은 따뜻한 실내등이 보인다. 판독 가능한 글자는 보이지 않는다. 다만 앞유리에 사람의 머리와 상체로 읽히는 어두운 형상이 추가되어 있다. 얼굴·나이·성별·민족성은 식별할 수 없으며, 이 외관 숏에서 보이도록 지정된 인물도 아니다. 찰리와 담요, 짐은 실내에 가려져 확인할 수 없다.",
    "hard_violations": [
     "사람을 추가하지 말라는 지시에도 앞유리 안에 인물의 머리와 상체 형상이 보인다."
    ],
    "physics": "차체는 노면에 닿은 바퀴들이 지지하며, 공중에 뜬 차량은 아니다. 바퀴와 노면의 접촉 부근에서 물이 양옆으로 튀어 올라 주행 중 물웅덩이를 가르는 동작으로 설명된다. 물보라가 크지만 물리적으로 불가능하다고 볼 근거는 없다. 실내 인물의 하체와 좌석은 가려져 있어 지지 상태를 직접 확인할 수 없으나, 공중에 떠 있는 모습은 아니다."
   },
   {
    "label": "A",
    "direction": "캠핑카는 화면 왼쪽 아래를 향한 전면과 측면을 드러낸다. 도로의 오른쪽 원경은 차량 뒤쪽으로 이어지므로, 요구된 오른쪽 진행 방향 및 그 앞의 열린 도로 관계를 충족하지 못한다. 식별 가능한 인물이나 시선은 없다.",
    "built_space": "왼쪽 가드레일 한 줄과 오른쪽 연석, 풀밭, 전신주 및 전선, 원경으로 이어지는 가로등 열이 참조 장소와 대응한다. 캠핑카의 상부 전면창 하나, 측면 출입문 하나와 문창, 측면 상단창 하나 및 뒤쪽 창 두 개가 보인다. 차량은 중간 오른쪽 배경이 아니라 중앙에서 왼쪽에 걸쳐 큰 크기로 배치되어 있다. 도로 위 광원 반사는 공간적으로 자연스럽다.",
    "entities": "오염과 낡은 외장을 유지한 캠핑카 한 대가 있으며 전조등은 꺼져 있다. 비와 젖은 도로, 바퀴 옆의 작은 물보라가 보이고, 뒤쪽 측면 창들에는 따뜻한 실내 빛이 비친다. 여러 창의 빛만으로 실내등이 여러 개라고 단정할 수는 없다. 전면 번호판에서는 문자와 숫자가 판독된다. 사람은 식별되지 않으며, 실내의 찰리·담요·짐은 이 외관에서 확인할 수 없다.",
    "hard_violations": [
     "전면 번호판의 문자와 숫자가 읽혀, 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반한다."
    ],
    "physics": "보이는 앞뒤 바퀴가 노면에 접촉하여 차량을 지지한다. 타이어 접촉 부근의 물 튀김은 회전하는 바퀴가 빗물을 밀어내는 주행 동작과 양립한다. 지지 없이 떠 있는 물체나 불가능한 차체 자세는 보이지 않는다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "진행 방향과 차량 배치가 틀리고 불필요한 인물까지 보이지만, 판독 가능한 글자 없이 소등 주행과 바퀴의 물보라를 더 명확히 구현해 상대적으로 우세하다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "진행 방향과 차량 배치가 요구와 반대이며, 전면 번호판의 글자가 읽혀 명시적인 무문자 조건까지 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "차량의 앞머리는 화면 왼쪽 아래와 카메라 쪽을 향한다. 오른쪽으로 멀어지는 도로는 차량의 앞이 아니라 뒤에 있으므로, 오른쪽으로 주행하며 전방 공간을 확보하라는 지시와 반대다. 앞유리 안 인물은 얼굴이 어두워 시선 방향을 확인할 수 없다.",
        "built_space": "왼쪽 가드레일 한 줄, 오른쪽 연석과 풀밭, 위쪽 전선과 전신주들, 오른쪽 원경으로 이어지는 가로등 열이 있어 참조 장소의 주요 구조를 유지한다. 캠핑카에는 상부 전면창 하나, 측면 출입문 하나와 그 창, 측면 상단의 작은 창 하나 및 뒤쪽 창 두 개가 보인다. 차량은 화면 중앙에서 왼쪽에 걸쳐 크게 놓여 있으며, 지시된 중간 오른쪽 배경의 작은 주행 피사체가 아니다. 젖은 노면의 가로등 반사는 가능한 위치에 있다.",
        "entities": "낡고 얼룩진 캠핑카 한 대, 꺼진 전조등, 비, 젖은 도로와 바퀴 주변 물보라가 보인다. 측면 창 하나에서 작은 따뜻한 실내등이 보인다. 판독 가능한 글자는 보이지 않는다. 다만 앞유리에 사람의 머리와 상체로 읽히는 어두운 형상이 추가되어 있다. 얼굴·나이·성별·민족성은 식별할 수 없으며, 이 외관 숏에서 보이도록 지정된 인물도 아니다. 찰리와 담요, 짐은 실내에 가려져 확인할 수 없다.",
        "hard_violations": [
         "사람을 추가하지 말라는 지시에도 앞유리 안에 인물의 머리와 상체 형상이 보인다."
        ],
        "physics": "차체는 노면에 닿은 바퀴들이 지지하며, 공중에 뜬 차량은 아니다. 바퀴와 노면의 접촉 부근에서 물이 양옆으로 튀어 올라 주행 중 물웅덩이를 가르는 동작으로 설명된다. 물보라가 크지만 물리적으로 불가능하다고 볼 근거는 없다. 실내 인물의 하체와 좌석은 가려져 있어 지지 상태를 직접 확인할 수 없으나, 공중에 떠 있는 모습은 아니다."
       },
       {
        "label": "B",
        "direction": "캠핑카는 화면 왼쪽 아래를 향한 전면과 측면을 드러낸다. 도로의 오른쪽 원경은 차량 뒤쪽으로 이어지므로, 요구된 오른쪽 진행 방향 및 그 앞의 열린 도로 관계를 충족하지 못한다. 식별 가능한 인물이나 시선은 없다.",
        "built_space": "왼쪽 가드레일 한 줄과 오른쪽 연석, 풀밭, 전신주 및 전선, 원경으로 이어지는 가로등 열이 참조 장소와 대응한다. 캠핑카의 상부 전면창 하나, 측면 출입문 하나와 문창, 측면 상단창 하나 및 뒤쪽 창 두 개가 보인다. 차량은 중간 오른쪽 배경이 아니라 중앙에서 왼쪽에 걸쳐 큰 크기로 배치되어 있다. 도로 위 광원 반사는 공간적으로 자연스럽다.",
        "entities": "오염과 낡은 외장을 유지한 캠핑카 한 대가 있으며 전조등은 꺼져 있다. 비와 젖은 도로, 바퀴 옆의 작은 물보라가 보이고, 뒤쪽 측면 창들에는 따뜻한 실내 빛이 비친다. 여러 창의 빛만으로 실내등이 여러 개라고 단정할 수는 없다. 전면 번호판에서는 문자와 숫자가 판독된다. 사람은 식별되지 않으며, 실내의 찰리·담요·짐은 이 외관에서 확인할 수 없다.",
        "hard_violations": [
         "전면 번호판의 문자와 숫자가 읽혀, 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "보이는 앞뒤 바퀴가 노면에 접촉하여 차량을 지지한다. 타이어 접촉 부근의 물 튀김은 회전하는 바퀴가 빗물을 밀어내는 주행 동작과 양립한다. 지지 없이 떠 있는 물체나 불가능한 차체 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "진행 방향과 차량 배치가 틀리고 불필요한 인물까지 보이지만, 판독 가능한 글자 없이 소등 주행과 바퀴의 물보라를 더 명확히 구현해 상대적으로 우세하다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "진행 방향과 차량 배치가 요구와 반대이며, 전면 번호판의 글자가 읽혀 명시적인 무문자 조건까지 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "차량의 앞머리는 화면 왼쪽 아래와 카메라 쪽을 향한다. 오른쪽으로 멀어지는 도로는 차량의 앞이 아니라 뒤에 있으므로, 오른쪽으로 주행하며 전방 공간을 확보하라는 지시와 반대다. 앞유리 안 인물은 얼굴이 어두워 시선 방향을 확인할 수 없다.",
        "built_space": "왼쪽 가드레일 한 줄, 오른쪽 연석과 풀밭, 위쪽 전선과 전신주들, 오른쪽 원경으로 이어지는 가로등 열이 있어 참조 장소의 주요 구조를 유지한다. 캠핑카에는 상부 전면창 하나, 측면 출입문 하나와 그 창, 측면 상단의 작은 창 하나 및 뒤쪽 창 두 개가 보인다. 차량은 화면 중앙에서 왼쪽에 걸쳐 크게 놓여 있으며, 지시된 중간 오른쪽 배경의 작은 주행 피사체가 아니다. 젖은 노면의 가로등 반사는 가능한 위치에 있다.",
        "entities": "낡고 얼룩진 캠핑카 한 대, 꺼진 전조등, 비, 젖은 도로와 바퀴 주변 물보라가 보인다. 측면 창 하나에서 작은 따뜻한 실내등이 보인다. 판독 가능한 글자는 보이지 않는다. 다만 앞유리에 사람의 머리와 상체로 읽히는 어두운 형상이 추가되어 있다. 얼굴·나이·성별·민족성은 식별할 수 없으며, 이 외관 숏에서 보이도록 지정된 인물도 아니다. 찰리와 담요, 짐은 실내에 가려져 확인할 수 없다.",
        "hard_violations": [
         "사람을 추가하지 말라는 지시에도 앞유리 안에 인물의 머리와 상체 형상이 보인다."
        ],
        "physics": "차체는 노면에 닿은 바퀴들이 지지하며, 공중에 뜬 차량은 아니다. 바퀴와 노면의 접촉 부근에서 물이 양옆으로 튀어 올라 주행 중 물웅덩이를 가르는 동작으로 설명된다. 물보라가 크지만 물리적으로 불가능하다고 볼 근거는 없다. 실내 인물의 하체와 좌석은 가려져 있어 지지 상태를 직접 확인할 수 없으나, 공중에 떠 있는 모습은 아니다."
       },
       {
        "label": "A",
        "direction": "캠핑카는 화면 왼쪽 아래를 향한 전면과 측면을 드러낸다. 도로의 오른쪽 원경은 차량 뒤쪽으로 이어지므로, 요구된 오른쪽 진행 방향 및 그 앞의 열린 도로 관계를 충족하지 못한다. 식별 가능한 인물이나 시선은 없다.",
        "built_space": "왼쪽 가드레일 한 줄과 오른쪽 연석, 풀밭, 전신주 및 전선, 원경으로 이어지는 가로등 열이 참조 장소와 대응한다. 캠핑카의 상부 전면창 하나, 측면 출입문 하나와 문창, 측면 상단창 하나 및 뒤쪽 창 두 개가 보인다. 차량은 중간 오른쪽 배경이 아니라 중앙에서 왼쪽에 걸쳐 큰 크기로 배치되어 있다. 도로 위 광원 반사는 공간적으로 자연스럽다.",
        "entities": "오염과 낡은 외장을 유지한 캠핑카 한 대가 있으며 전조등은 꺼져 있다. 비와 젖은 도로, 바퀴 옆의 작은 물보라가 보이고, 뒤쪽 측면 창들에는 따뜻한 실내 빛이 비친다. 여러 창의 빛만으로 실내등이 여러 개라고 단정할 수는 없다. 전면 번호판에서는 문자와 숫자가 판독된다. 사람은 식별되지 않으며, 실내의 찰리·담요·짐은 이 외관에서 확인할 수 없다.",
        "hard_violations": [
         "전면 번호판의 문자와 숫자가 읽혀, 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "보이는 앞뒤 바퀴가 노면에 접촉하여 차량을 지지한다. 타이어 접촉 부근의 물 튀김은 회전하는 바퀴가 빗물을 밀어내는 주행 동작과 양립한다. 지지 없이 떠 있는 물체나 불가능한 차체 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "totals": {
   "B": 3,
   "A": 2
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "진행 방향과 차량 배치가 틀리고 불필요한 인물까지 보이지만, 판독 가능한 글자 없이 소등 주행과 바퀴의 물보라를 더 명확히 구현해 상대적으로 우세하다."
   },
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "진행 방향과 차량 배치가 요구와 반대이며, 전면 번호판의 글자가 읽혀 명시적인 무문자 조건까지 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L54B02.png",
    "asset_id": "a1d54e19-8e26-4308-9098-5caf37466919",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdcc-de2e-78e1-a5b5-65fa999f644c",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S45sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:40:13.898273+00:00",
  "fingerprint": "49ca9ecdf307a19397ef8f9413ea24593c20461fd1995fa337261e28db79cddb",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S45sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S45sh1_sel.png",
  "source_sha256": "95bf44dc78fcb320b534a718bfc1db0eeda31413c905de8b6ceeca61a61eeaac",
  "file": "S45sh1_cine.png",
  "staged_sha256": "b3adca3a517fbebeb9ed31c6163f52edaab2ffa68ddfc60d47e8958af4e13c39",
  "latency_ms": 15831
 },
 "S45sh8::signage": {
  "fp": "c5cfaff6c2c3ea55",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::27d40b4a0770a059": {
  "subjects": [],
  "subject_text": "이현우 일행의 캠핑카 내부\n운전석과 조수석 뒤로 생활공간이 이어지는 낡은 소형 캠핑카 실내. 침대와 뒷좌석, 작은 전등, 창문과 짐을 둘 공간이 있다.",
  "identity": "canonical",
  "scope_id": "L54",
  "scope_role": "location_interior",
  "scope_sha": "95d30874e96744b1"
 },
 "S45sh8::bgfirst_bg": {
  "input_fingerprint": "b2d0d3207ca64fba",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 이현우를 향해 두 눈을 부릅뜨고 턱을 치켜든 항의하는 표정.\n\nLOCATION (lock): Inside the camper's rear seating area, lit by a single small lamp while rain darkens the windows.\n\nTIME OF DAY (lock): night, rainy.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rear seating area (Occupied by 앰버) — Seen from the aisle leading back from the front seats; used as Provides limited context behind her close-framed expression; Driver's seat edge (Occupied by 이현우) — A narrow rear edge is visible near the foreground shoulder; used as Anchors the driver's forward position without obstructing 앰버.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The small rear-seat light gives 앰버's expression restrained visibility against the dark nighttime interior, without adding a color or another source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 이현우를 향해 두 눈을 부릅뜨고 턱을 치켜든 항의하는 표정.\n\nLOCATION (lock): Inside the camper's rear seating area, lit by a single small lamp while rain darkens the windows.\n\nTIME OF DAY (lock): night, rainy.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rear seating area (Occupied by 앰버) — Seen from the aisle leading back from the front seats; used as Provides limited context behind her close-framed expression; Driver's seat edge (Occupied by 이현우) — A narrow rear edge is visible near the foreground shoulder; used as Anchors the driver's forward position without obstructing 앰버.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The small rear-seat light gives 앰버's expression restrained visibility against the dark nighttime interior, without adding a color or another source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S45sh8__bgfirst_bg.png",
  "asset_id": "9889945c-0242-4826-8452-c4aa835a707b",
  "input_asset_ids": [
   "3e08fc8e-7449-4d9c-a1f2-92f0872af317",
   "a14ad9c7-848f-4d9b-8f63-0cd9cf1a9b1e"
  ]
 },
 "S45sh8": {
  "input_fingerprint": "cc6169017c560623",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 앰버가 이현우를 향해 두 눈을 부릅뜨고 턱을 치켜든 항의하는 표정.\n\nLOCATION (lock): Inside the camper's rear seating area, lit by a single small lamp while rain darkens the windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rear seating area (Occupied by 앰버) — Seen from the aisle leading back from the front seats; used as Provides limited context behind her close-framed expression; Driver's seat edge (Occupied by 이현우) — A narrow rear edge is visible near the foreground shoulder; used as Anchors the driver's forward position without obstructing 앰버.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The small rear-seat light gives 앰버's expression restrained visibility against the dark nighttime interior, without adding a color or another source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy rain surrounds the moving camper; its headlights remain off while the small rear lamp stays on. The packed supplies remain aboard, and Charlie stays seated in the rear with the blanket. 앰버: She remains seated in the rear cabin as she argues about the destination.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 앰버가 이현우를 향해 두 눈을 부릅뜨고 턱을 치켜든 항의하는 표정.\n\nLOCATION (lock): Inside the camper's rear seating area, lit by a single small lamp while rain darkens the windows. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rear seating area (Occupied by 앰버) — Seen from the aisle leading back from the front seats; used as Provides limited context behind her close-framed expression; Driver's seat edge (Occupied by 이현우) — A narrow rear edge is visible near the foreground shoulder; used as Anchors the driver's forward position without obstructing 앰버.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The small rear-seat light gives 앰버's expression restrained visibility against the dark nighttime interior, without adding a color or another source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy rain surrounds the moving camper; its headlights remain off while the small rear lamp stays on. The packed supplies remain aboard, and Charlie stays seated in the rear with the blanket. 앰버: She remains seated in the rear cabin as she argues about the destination.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 앰버가 이현우를 향해 두 눈을 부릅뜨고 턱을 치켜든 항의하는 표정.\n\nLOCATION (lock): Inside the camper's rear seating area, lit by a single small lamp while rain darkens the windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rear seating area (Occupied by 앰버) — Seen from the aisle leading back from the front seats; used as Provides limited context behind her close-framed expression; Driver's seat edge (Occupied by 이현우) — A narrow rear edge is visible near the foreground shoulder; used as Anchors the driver's forward position without obstructing 앰버.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The small rear-seat light gives 앰버's expression restrained visibility against the dark nighttime interior, without adding a color or another source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy rain surrounds the moving camper; its headlights remain off while the small rear lamp stays on. The packed supplies remain aboard, and Charlie stays seated in the rear with the blanket. 앰버: She remains seated in the rear cabin as she argues about the destination.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S45sh8__bgfirst_bg.png",
     "asset_id": "9889945c-0242-4826-8452-c4aa835a707b",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S45sh8.png",
     "asset_id": "3e08fc8e-7449-4d9c-a1f2-92f0872af317",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L54B03.png",
     "asset_id": "a14ad9c7-848f-4d9b-8f63-0cd9cf1a9b1e",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "앰버는 왼쪽 전경에 위치한 인물을 향해 매서운 시선을 보냄.",
    "built_space": "카메라가 캠핑카 전면을 향하고 있어 배경에 앞유리와 대시보드가 보임. 두 앞좌석은 비어 있음.",
    "entities": "앰버의 외모, 지정된 의상, 턱을 치켜든 항의하는 표정이 매우 잘 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 명시된 카메라 뷰(앞에서 뒤를 바라봄)를 정반대로 뒤집어 전면부가 배경이 됨.",
     "[gemini-pro] 운전석을 배경에 빈 채로 두고 전경 인물을 잘못된 위치에 배치함.",
     "[gpt-high] 앞쪽 운전석에 있어야 할 이현우를 후방의 가까운 전경에 배치했다. 실제 운전대와 비어 있는 운전석이 배경에 따로 보여 동일한 운전자 위치로 해석할 수 없다.",
     "[gpt-high] 앞좌석에서 뒤쪽 좌석으로 이어지는 통로의 시점 대신, 후방에서 앞유리와 대시보드를 바라보는 반대 방향으로 장면을 구성했다."
    ],
    "physics": "좌석에 앉아 체중이 자연스럽게 시트에 지지됨."
   },
   {
    "label": "B",
    "direction": "앰버는 오른쪽 전경의 인물(운전석)을 향해 고개를 들고 시선을 맞춤.",
    "built_space": "카메라가 뒷좌석을 바라보는 구조를 정확히 구현했으며, 우측 전경에 운전석 등받이가 알맞게 배치됨.",
    "entities": "앰버의 인상착의와 복장은 일치하나, 항의하는 표정이라기엔 다소 온화함.",
    "hard_violations": [],
    "physics": "벤치 가장자리에 두 손을 짚고 앉아 체중을 싣고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "카메라 방향과 공간 배치를 정확히 따랐으나, 지정된 클로즈업 대신 와이드 샷으로 피사체를 작게 담은 점이 감점 요인임."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "표정과 클로즈업 프레이밍은 우수하나, 카메라 방향을 반대로 렌더링하고 운전석을 빈 채로 배경에 두는 치명적인 연출 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 왼쪽 전경에 위치한 인물을 향해 매서운 시선을 보냄.",
        "built_space": "카메라가 캠핑카 전면을 향하고 있어 배경에 앞유리와 대시보드가 보임. 두 앞좌석은 비어 있음.",
        "entities": "앰버의 외모, 지정된 의상, 턱을 치켜든 항의하는 표정이 매우 잘 묘사됨.",
        "hard_violations": [
         "명시된 카메라 뷰(앞에서 뒤를 바라봄)를 정반대로 뒤집어 전면부가 배경이 됨.",
         "운전석을 배경에 빈 채로 두고 전경 인물을 잘못된 위치에 배치함."
        ],
        "physics": "좌석에 앉아 체중이 자연스럽게 시트에 지지됨."
       },
       {
        "label": "B",
        "direction": "앰버는 오른쪽 전경의 인물(운전석)을 향해 고개를 들고 시선을 맞춤.",
        "built_space": "카메라가 뒷좌석을 바라보는 구조를 정확히 구현했으며, 우측 전경에 운전석 등받이가 알맞게 배치됨.",
        "entities": "앰버의 인상착의와 복장은 일치하나, 항의하는 표정이라기엔 다소 온화함.",
        "hard_violations": [],
        "physics": "벤치 가장자리에 두 손을 짚고 앉아 체중을 싣고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "카메라 방향과 공간 배치를 정확히 따랐으나, 지정된 클로즈업 대신 와이드 샷으로 피사체를 작게 담은 점이 감점 요인임."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "표정과 클로즈업 프레이밍은 우수하나, 카메라 방향을 반대로 렌더링하고 운전석을 빈 채로 배경에 두는 치명적인 연출 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 왼쪽 전경에 위치한 인물을 향해 매서운 시선을 보냄.",
        "built_space": "카메라가 캠핑카 전면을 향하고 있어 배경에 앞유리와 대시보드가 보임. 두 앞좌석은 비어 있음.",
        "entities": "앰버의 외모, 지정된 의상, 턱을 치켜든 항의하는 표정이 매우 잘 묘사됨.",
        "hard_violations": [
         "명시된 카메라 뷰(앞에서 뒤를 바라봄)를 정반대로 뒤집어 전면부가 배경이 됨.",
         "운전석을 배경에 빈 채로 두고 전경 인물을 잘못된 위치에 배치함."
        ],
        "physics": "좌석에 앉아 체중이 자연스럽게 시트에 지지됨."
       },
       {
        "label": "B",
        "direction": "앰버는 오른쪽 전경의 인물(운전석)을 향해 고개를 들고 시선을 맞춤.",
        "built_space": "카메라가 뒷좌석을 바라보는 구조를 정확히 구현했으며, 우측 전경에 운전석 등받이가 알맞게 배치됨.",
        "entities": "앰버의 인상착의와 복장은 일치하나, 항의하는 표정이라기엔 다소 온화함.",
        "hard_violations": [],
        "physics": "벤치 가장자리에 두 손을 짚고 앉아 체중을 싣고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "클로즈업을 지키지 못했지만, 뒤에 앉은 앰버가 앞쪽 운전자를 향해 눈을 부릅뜨고 턱을 드는 공간·행동 관계는 B보다 정확하다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "항의하는 얼굴은 더 크게 담았으나, 운전석이 배경에 비어 있는데 이현우를 후방 전경에 배치하여 명시된 운전자 위치와 촬영 방향을 어겼다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 화면 오른쪽 위의 운전자 쪽으로 얼굴과 눈을 향한다. 눈을 크게 뜨고 턱을 올렸으며 입술을 굳게 다문 항의 표정이다. 오른쪽 전경 남성은 차량 앞쪽을 향하고 있어 앰버와 마주 보지는 않지만, 앰버의 시선 목표는 그 남성으로 읽힌다. 무기나 방향을 확인할 휴대 소품은 없다.",
        "built_space": "앞좌석 사이에서 후방을 보는 구도다. 전경 양쪽에 앞좌석 등받이 두 개가 있고, 오른쪽 좌석에 운전자로 읽히는 남성 일부가 보인다. 앰버는 왼쪽 벽을 따른 천 벤치에 앉아 있다. 앞뒤가 뒤집힌 관찰 방향을 고려하면 참조의 오른쪽 측면 벤치와 위치가 양립한다. 후면 가로 벤치 하나, 후면 창 하나, 왼쪽 측면 창 두 구획, 상부 목재 수납장과 켜진 작은 등 하나가 보인다. 후면 가로 벤치는 참조가 보여주지 않은 영역이라 동일성을 확인할 수 없다. 운전석 가장자리는 요구보다 넓게 보이고, 실내와 앰버의 다리까지 드러나 클로즈업이 아니다.",
        "entities": "금발의 어린 여자아이 한 명과 전경 운전자 일부가 보인다. 앰버의 나이와 체격, 큰 눈, 카키색 겉옷과 회색 안쪽 옷은 대체로 맞는다. 한국계·백인 혼혈이라는 정체성을 외모만으로 확정할 수는 없으며 참조와 얼굴의 완전한 일치도 확인하기 어렵다. 참조의 흰 마스크는 없다. 짐가방, 빗물 묻은 어두운 창, 작은 실내등이 보이며 읽을 수 있는 글자는 없다. 찰리와 담요는 화면에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "앰버의 엉덩이는 벤치 쿠션에 놓이고 몸통은 앞으로 기울어 있다. 한 손은 좌석 가장자리 부근에 내려와 있으며 무릎은 자연스럽게 굽혀져 있다. 발의 바닥 접촉은 하단에서 잘리지만 앉은 몸의 지지는 분명하다. 운전자는 앞좌석에 지지되어 있고 가방은 바닥에 놓여 있다. 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "앰버는 화면 왼쪽 전경 남성의 얼굴을 올려다본다. 찌푸린 눈썹과 벌어진 입은 항의를 명확하게 나타내지만, 두 눈을 부릅뜬 정도와 턱을 치켜든 정도는 A보다 약하다. 남성의 머리도 앰버 쪽을 향한다. 시선은 상대에게 도달하지만 그 상대가 요구된 앞쪽 운전석 위치에 있지 않다.",
        "built_space": "참조처럼 뒤에서 차량 앞쪽을 보는 공간 배치다. 오른쪽에는 앰버가 앉은 측면 벤치, 왼쪽에는 목재 수납대와 바닥 짐, 배경에는 운전대와 비어 있는 운전석 하나가 뚜렷하다. 반대편 앞좌석 영역은 앰버에게 가려져 있다. 천장등 하나와 비에 젖은 측면 창들이 보인다. 이현우로 읽히는 남성은 왼쪽 가까운 전경에 있어 배경 운전석과 떨어져 있다. 따라서 전경의 좁은 운전석 뒤 모서리로 앞쪽 운전자 위치를 고정한다는 구성과 맞지 않는다. 앰버도 얼굴만 밀착한 구도보다는 상반신과 주변 공간이 상당히 포함된 구도다.",
        "entities": "금발의 어린 여자아이와 왼쪽 전경 성인 남성 일부가 보인다. 앰버의 어린 나이, 큰 눈, 카키색 재킷과 회색 안쪽 옷은 대체로 부합한다. 참조의 흰 마스크는 없고 얼굴의 정확한 동일성은 제한적으로만 확인된다. 벤치의 갈색 직물, 목재 수납대, 짐가방과 작은 등은 참조 장소에 가깝다. 창에는 빗방울이 있고 바깥은 어둡다. 읽을 수 있는 글자나 별도의 인물은 없다. 찰리와 담요는 보이지 않는다.",
        "hard_violations": [
         "앞쪽 운전석에 있어야 할 이현우를 후방의 가까운 전경에 배치했다. 실제 운전대와 비어 있는 운전석이 배경에 따로 보여 동일한 운전자 위치로 해석할 수 없다.",
         "앞좌석에서 뒤쪽 좌석으로 이어지는 통로의 시점 대신, 후방에서 앞유리와 대시보드를 바라보는 반대 방향으로 장면을 구성했다."
        ],
        "physics": "앰버의 하체는 잘려 있지만 몸통은 벤치 위 착석 위치와 연결되고 등 뒤에는 벤치 등받이가 있다. 공중에 떠 있다고 볼 근거는 없다. 전경 남성은 머리와 어깨만 보여 하체 지지 여부를 판정할 수 없으며, 그 자체로 부유는 아니다. 가방은 바닥에 놓이고 수납대와 조명은 구조물에 고정되어 있다. 물리적 지지 문제보다 인물의 공간 배치가 문제다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "클로즈업을 지키지 못했지만, 뒤에 앉은 앰버가 앞쪽 운전자를 향해 눈을 부릅뜨고 턱을 드는 공간·행동 관계는 B보다 정확하다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "항의하는 얼굴은 더 크게 담았으나, 운전석이 배경에 비어 있는데 이현우를 후방 전경에 배치하여 명시된 운전자 위치와 촬영 방향을 어겼다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 화면 오른쪽 위의 운전자 쪽으로 얼굴과 눈을 향한다. 눈을 크게 뜨고 턱을 올렸으며 입술을 굳게 다문 항의 표정이다. 오른쪽 전경 남성은 차량 앞쪽을 향하고 있어 앰버와 마주 보지는 않지만, 앰버의 시선 목표는 그 남성으로 읽힌다. 무기나 방향을 확인할 휴대 소품은 없다.",
        "built_space": "앞좌석 사이에서 후방을 보는 구도다. 전경 양쪽에 앞좌석 등받이 두 개가 있고, 오른쪽 좌석에 운전자로 읽히는 남성 일부가 보인다. 앰버는 왼쪽 벽을 따른 천 벤치에 앉아 있다. 앞뒤가 뒤집힌 관찰 방향을 고려하면 참조의 오른쪽 측면 벤치와 위치가 양립한다. 후면 가로 벤치 하나, 후면 창 하나, 왼쪽 측면 창 두 구획, 상부 목재 수납장과 켜진 작은 등 하나가 보인다. 후면 가로 벤치는 참조가 보여주지 않은 영역이라 동일성을 확인할 수 없다. 운전석 가장자리는 요구보다 넓게 보이고, 실내와 앰버의 다리까지 드러나 클로즈업이 아니다.",
        "entities": "금발의 어린 여자아이 한 명과 전경 운전자 일부가 보인다. 앰버의 나이와 체격, 큰 눈, 카키색 겉옷과 회색 안쪽 옷은 대체로 맞는다. 한국계·백인 혼혈이라는 정체성을 외모만으로 확정할 수는 없으며 참조와 얼굴의 완전한 일치도 확인하기 어렵다. 참조의 흰 마스크는 없다. 짐가방, 빗물 묻은 어두운 창, 작은 실내등이 보이며 읽을 수 있는 글자는 없다. 찰리와 담요는 화면에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "앰버의 엉덩이는 벤치 쿠션에 놓이고 몸통은 앞으로 기울어 있다. 한 손은 좌석 가장자리 부근에 내려와 있으며 무릎은 자연스럽게 굽혀져 있다. 발의 바닥 접촉은 하단에서 잘리지만 앉은 몸의 지지는 분명하다. 운전자는 앞좌석에 지지되어 있고 가방은 바닥에 놓여 있다. 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "앰버는 화면 왼쪽 전경 남성의 얼굴을 올려다본다. 찌푸린 눈썹과 벌어진 입은 항의를 명확하게 나타내지만, 두 눈을 부릅뜬 정도와 턱을 치켜든 정도는 A보다 약하다. 남성의 머리도 앰버 쪽을 향한다. 시선은 상대에게 도달하지만 그 상대가 요구된 앞쪽 운전석 위치에 있지 않다.",
        "built_space": "참조처럼 뒤에서 차량 앞쪽을 보는 공간 배치다. 오른쪽에는 앰버가 앉은 측면 벤치, 왼쪽에는 목재 수납대와 바닥 짐, 배경에는 운전대와 비어 있는 운전석 하나가 뚜렷하다. 반대편 앞좌석 영역은 앰버에게 가려져 있다. 천장등 하나와 비에 젖은 측면 창들이 보인다. 이현우로 읽히는 남성은 왼쪽 가까운 전경에 있어 배경 운전석과 떨어져 있다. 따라서 전경의 좁은 운전석 뒤 모서리로 앞쪽 운전자 위치를 고정한다는 구성과 맞지 않는다. 앰버도 얼굴만 밀착한 구도보다는 상반신과 주변 공간이 상당히 포함된 구도다.",
        "entities": "금발의 어린 여자아이와 왼쪽 전경 성인 남성 일부가 보인다. 앰버의 어린 나이, 큰 눈, 카키색 재킷과 회색 안쪽 옷은 대체로 부합한다. 참조의 흰 마스크는 없고 얼굴의 정확한 동일성은 제한적으로만 확인된다. 벤치의 갈색 직물, 목재 수납대, 짐가방과 작은 등은 참조 장소에 가깝다. 창에는 빗방울이 있고 바깥은 어둡다. 읽을 수 있는 글자나 별도의 인물은 없다. 찰리와 담요는 보이지 않는다.",
        "hard_violations": [
         "앞쪽 운전석에 있어야 할 이현우를 후방의 가까운 전경에 배치했다. 실제 운전대와 비어 있는 운전석이 배경에 따로 보여 동일한 운전자 위치로 해석할 수 없다.",
         "앞좌석에서 뒤쪽 좌석으로 이어지는 통로의 시점 대신, 후방에서 앞유리와 대시보드를 바라보는 반대 방향으로 장면을 구성했다."
        ],
        "physics": "앰버의 하체는 잘려 있지만 몸통은 벤치 위 착석 위치와 연결되고 등 뒤에는 벤치 등받이가 있다. 공중에 떠 있다고 볼 근거는 없다. 전경 남성은 머리와 어깨만 보여 하체 지지 여부를 판정할 수 없으며, 그 자체로 부유는 아니다. 가방은 바닥에 놓이고 수납대와 조명은 구조물에 고정되어 있다. 물리적 지지 문제보다 인물의 공간 배치가 문제다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.75,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 명시된 카메라 뷰(앞에서 뒤를 바라봄)를 정반대로 뒤집어 전면부가 배경이 됨.",
     "[gemini-pro] 운전석을 배경에 빈 채로 두고 전경 인물을 잘못된 위치에 배치함.",
     "[gpt-high] 앞쪽 운전석에 있어야 할 이현우를 후방의 가까운 전경에 배치했다. 실제 운전대와 비어 있는 운전석이 배경에 따로 보여 동일한 운전자 위치로 해석할 수 없다.",
     "[gpt-high] 앞좌석에서 뒤쪽 좌석으로 이어지는 통로의 시점 대신, 후방에서 앞유리와 대시보드를 바라보는 반대 방향으로 장면을 구성했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "카메라 방향과 공간 배치를 정확히 따랐으나, 지정된 클로즈업 대신 와이드 샷으로 피사체를 작게 담은 점이 감점 요인임."
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "표정과 클로즈업 프레이밍은 우수하나, 카메라 방향을 반대로 렌더링하고 운전석을 빈 채로 배경에 두는 치명적인 연출 오류가 발생함.  ★위반: [gemini-pro] 명시된 카메라 뷰(앞에서 뒤를 바라봄)를 정반대로 뒤집어 전면부가 배경이 됨. / [gemini-pro] 운전석을 배경에 빈 채로 두고 전경 인물을 잘못된 위치에 배치함. / [gpt-high] 앞쪽 운전석에 있어야 할 이현우를 후방의 가까운 전경에 배치했다. 실제 운전대와 비어 있는 운전석이 배경에 따로 보여 동일한 운전자 위치로 해석할 수 없다. / [gpt-high] 앞좌석에서 뒤쪽 좌석으로 이어지는 통로의 시점 대신, 후방에서 앞유리와 대시보드를 바라보는 반대 방향으로 장면을 구성했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L54B03.png",
    "asset_id": "a14ad9c7-848f-4d9b-8f63-0cd9cf1a9b1e",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdcd-e12f-79fc-853c-dfc1935a69dd",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S45sh8__bgfirst_bg.png",
   "bg_asset_id": "9889945c-0242-4826-8452-c4aa835a707b",
   "bg_record_key": "S45sh8::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S45sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:40:30.165888+00:00",
  "fingerprint": "07365ef81a0ee45dc1a55347fbdd00c03dd181815546349c765697052d58ef0e",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S45sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S45sh8_sel.png",
  "source_sha256": "8d9e1a77a32e5887debc889d74e1e6fe89b10011d1dbab87ce491ea9fd1e6488",
  "file": "S45sh8_cine.png",
  "staged_sha256": "dcb9844bacfb0d249592d0e3808d99ed8e209f63c9f65b8b71d0b36d97d550cf",
  "latency_ms": 11444
 },
 "S45sh14::signage": {
  "fp": "37a9c51eaa42742b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S45sh14": {
  "input_fingerprint": "a971e6519e2f098a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 라울과 앰버가 쏟아지는 빗물 아래서 양팔을 허공으로 뻗은 채 얼굴을 찡그린 당황한 구도.\n\nLOCATION (lock): Inside the camper's leaking rear living area, beneath the roof and beside the window frames under the single small rear lamp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Ceiling leak above 라울 in the upper-left of the frame, midground; Leaking window frame beside 앰버 in the upper-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Camper ceiling (Leaking rainwater) — Its underside is visible above 라울; used as Establishes the target of his upward reach; Adjacent window frame (Leaking rainwater into the cabin) — Interior-facing edge visible beside 앰버; used as Separates her attempted repair from 라울's overhead effort; Clothing used to block leaks (Pressed against the ceiling and window frame) — Held between each person's raised hands and the leaking surface; used as Makes the purpose of the raised arms explicit.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the small rear-seat light within the dark cabin, preserving enough contrast to read the raised arms, strained faces, and entering rainwater.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rainwater is leaking through the camper roof and window frames, with clothing pressed against the leaks but failing to stop them. The headlights remain off, the small rear lamp remains lit, and Charlie is still in the rear cabin with the packed luggage. 라울: He has taken clothing from the rear luggage and is holding it against the leaking ceiling. 앰버: She is pressing clothing against a leaking window frame.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 라울과 앰버가 쏟아지는 빗물 아래서 양팔을 허공으로 뻗은 채 얼굴을 찡그린 당황한 구도.\n\nLOCATION (lock): Inside the camper's leaking rear living area, beneath the roof and beside the window frames under the single small rear lamp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Ceiling leak above 라울 in the upper-left of the frame, midground; Leaking window frame beside 앰버 in the upper-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Camper ceiling (Leaking rainwater) — Its underside is visible above 라울; used as Establishes the target of his upward reach; Adjacent window frame (Leaking rainwater into the cabin) — Interior-facing edge visible beside 앰버; used as Separates her attempted repair from 라울's overhead effort; Clothing used to block leaks (Pressed against the ceiling and window frame) — Held between each person's raised hands and the leaking surface; used as Makes the purpose of the raised arms explicit.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the small rear-seat light within the dark cabin, preserving enough contrast to read the raised arms, strained faces, and entering rainwater.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rainwater is leaking through the camper roof and window frames, with clothing pressed against the leaks but failing to stop them. The headlights remain off, the small rear lamp remains lit, and Charlie is still in the rear cabin with the packed luggage. 라울: He has taken clothing from the rear luggage and is holding it against the leaking ceiling. 앰버: She is pressing clothing against a leaking window frame.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 라울과 앰버가 쏟아지는 빗물 아래서 양팔을 허공으로 뻗은 채 얼굴을 찡그린 당황한 구도.\n\nLOCATION (lock): Inside the camper's leaking rear living area, beneath the roof and beside the window frames under the single small rear lamp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Ceiling leak above 라울 in the upper-left of the frame, midground; Leaking window frame beside 앰버 in the upper-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Camper ceiling (Leaking rainwater) — Its underside is visible above 라울; used as Establishes the target of his upward reach; Adjacent window frame (Leaking rainwater into the cabin) — Interior-facing edge visible beside 앰버; used as Separates her attempted repair from 라울's overhead effort; Clothing used to block leaks (Pressed against the ceiling and window frame) — Held between each person's raised hands and the leaking surface; used as Makes the purpose of the raised arms explicit.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the small rear-seat light within the dark cabin, preserving enough contrast to read the raised arms, strained faces, and entering rainwater.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rainwater is leaking through the camper roof and window frames, with clothing pressed against the leaks but failing to stop them. The headlights remain off, the small rear lamp remains lit, and Charlie is still in the rear cabin with the packed luggage. 라울: He has taken clothing from the rear luggage and is holding it against the leaking ceiling. 앰버: She is pressing clothing against a leaking window frame.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "왼쪽 라울은 좌상단의 젖은 천장 가장자리로 두 팔을 뻗고, 옷 뭉치와 누수 지점을 올려다봅니다. 오른쪽 앰버는 오른쪽 창의 세로 프레임에 옷을 대고 양손으로 누르며 그 부근을 바라봅니다. 두 사람의 손이 서로 다른 수리 대상을 향한다는 관계가 명확합니다.",
    "built_space": "목재 상부 수납장, 왼쪽 작업대, 오른쪽 회색 벤치, 오른쪽의 인접한 창 두 구획과 왼쪽 측창이 보입니다. 중앙 뒤로 앞유리와 운전석 설비가 보여 차량 뒤에서 앞을 보는 구도입니다. 작은 점등등은 하나뿐입니다. 벤치가 참조의 반대 화면 쪽에 보이는 것은 역방향 카메라로 설명되지만, 참조에서 후면 창 위에 있던 등이 여기서는 운전실 쪽 상부에 보여 고정 조명의 위치 연속성이 맞지 않습니다.",
    "entities": "아이 두 명만 보입니다. 라울은 참조와 유사한 갈색 피부, 어린 얼굴, 뒤로 모은 곱슬머리와 해진 올리브색 티셔츠를 지녔고, 지정된 혼혈 남자아이 외형에 대체로 부합합니다. 앰버는 참조와 유사한 어린 얼굴과 긴 금발, 회녹색 재킷과 회색 옷을 입었으며 이전 장면처럼 마스크가 없습니다. 두 사람의 찡그린 표정, 누수 차단용 옷 두 뭉치, 젖은 창과 천장, 여러 여행가방이 보입니다. 민족적 배경 자체는 외형만으로 확정할 수 없습니다. 다른 사람이나 읽을 수 있는 글자는 보이지 않습니다.",
    "hard_violations": [],
    "physics": "라울의 손들이 옷 뭉치를 받치며 천장 가장자리로 밀고 있고, 앰버의 손들은 천을 창틀에 압착합니다. 옷은 손과 접촉면에 지지되며 뜬 물체는 없습니다. 두 몸통은 하단 프레임 밖으로 이어지는 직립 자세이고, 발이 잘렸다는 이유로 공중 부양으로 볼 근거는 없습니다. 천장과 창틀에서 물이 아래로 떨어지는 방향도 자연스럽습니다."
   },
   {
    "label": "A",
    "direction": "라울은 상단 중앙의 천장 개구부에서 늘어진 옷을 향해 팔을 뻗고 그곳을 올려다봅니다. 명확하게 보이는 손은 늘어진 천의 중간에 닿아 있어, 양손으로 천장 누수면을 누르는 동작은 읽기 어렵습니다. 앰버는 오른쪽 창틀 부근의 천을 양손으로 누르고 그쪽으로 얼굴을 돌렸습니다. 라울의 목표가 지정된 좌상단보다 중앙에 치우칩니다.",
    "built_space": "목재 수납장과 왼쪽 작업대, 오른쪽 회색 벤치, 좌우 측창, 중앙 뒤의 운전실이 보입니다. 작은 점등등 하나와 천장 개구부 하나가 있습니다. 차량 뒤에서 앞을 보는 공간 관계는 성립하지만, 후면 창 위에 있어야 할 참조의 등이 운전실 위쪽에 나타나는 연속성 문제는 A와 같습니다. 허벅지와 통로 바닥까지 포함하여 지정된 미디엄 숏보다 넓게 잡혔습니다.",
    "entities": "추가 인물 없이 아이 두 명이 보입니다. 라울의 피부색, 어린 얼굴, 뒤로 묶은 곱슬머리, 해진 올리브색 티셔츠와 데님 반바지는 참조에 대체로 부합합니다. 앰버의 금발, 어린 얼굴, 회녹색 재킷과 회색 원피스도 이전 장면과 대체로 이어집니다. 지정된 혼혈 정체성과 충돌하는 뚜렷한 외형 차이는 없지만 외형만으로 혈통을 확인할 수는 없습니다. 천장에 끼운 옷들과 창가의 옷 뭉치, 여행가방, 빗물이 보이고 두 사람 모두 얼굴을 찡그리고 있습니다. 읽을 수 있는 글자는 없습니다.",
    "hard_violations": [],
    "physics": "천장 쪽 옷은 개구부 테두리에 끼워져 매달리고 라울의 손이 늘어진 부분에 닿아 있어 무지지 부유는 아닙니다. 다만 손과 천장 사이 거리가 있어 직접 압착하는 힘의 전달은 A보다 불명확합니다. 앰버의 천은 양손과 창틀에 지지되고 아래로 늘어집니다. 두 아이는 몸을 기울여 서 있는 자세이며 다리는 하단으로 이어집니다. 바닥 접점은 잘렸지만 비행하거나 떠 있는 자세는 아니고, 물방울은 아래로 떨어집니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "미디엄 숏 안에서 라울의 좌상단 천장 막기와 앰버의 우상단 창틀 막기, 찡그린 표정을 더 정확히 구현하지만 후방등의 위치 연속성은 어긋납니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "누수를 막는 상황과 인물 외형은 맞지만 프레임이 더 넓고 천장 누수가 상단 중앙으로 옮겨졌으며, 라울이 양손으로 옷을 천장에 압착하는 동작이 불명확합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 라울은 좌상단의 젖은 천장 가장자리로 두 팔을 뻗고, 옷 뭉치와 누수 지점을 올려다봅니다. 오른쪽 앰버는 오른쪽 창의 세로 프레임에 옷을 대고 양손으로 누르며 그 부근을 바라봅니다. 두 사람의 손이 서로 다른 수리 대상을 향한다는 관계가 명확합니다.",
        "built_space": "목재 상부 수납장, 왼쪽 작업대, 오른쪽 회색 벤치, 오른쪽의 인접한 창 두 구획과 왼쪽 측창이 보입니다. 중앙 뒤로 앞유리와 운전석 설비가 보여 차량 뒤에서 앞을 보는 구도입니다. 작은 점등등은 하나뿐입니다. 벤치가 참조의 반대 화면 쪽에 보이는 것은 역방향 카메라로 설명되지만, 참조에서 후면 창 위에 있던 등이 여기서는 운전실 쪽 상부에 보여 고정 조명의 위치 연속성이 맞지 않습니다.",
        "entities": "아이 두 명만 보입니다. 라울은 참조와 유사한 갈색 피부, 어린 얼굴, 뒤로 모은 곱슬머리와 해진 올리브색 티셔츠를 지녔고, 지정된 혼혈 남자아이 외형에 대체로 부합합니다. 앰버는 참조와 유사한 어린 얼굴과 긴 금발, 회녹색 재킷과 회색 옷을 입었으며 이전 장면처럼 마스크가 없습니다. 두 사람의 찡그린 표정, 누수 차단용 옷 두 뭉치, 젖은 창과 천장, 여러 여행가방이 보입니다. 민족적 배경 자체는 외형만으로 확정할 수 없습니다. 다른 사람이나 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "라울의 손들이 옷 뭉치를 받치며 천장 가장자리로 밀고 있고, 앰버의 손들은 천을 창틀에 압착합니다. 옷은 손과 접촉면에 지지되며 뜬 물체는 없습니다. 두 몸통은 하단 프레임 밖으로 이어지는 직립 자세이고, 발이 잘렸다는 이유로 공중 부양으로 볼 근거는 없습니다. 천장과 창틀에서 물이 아래로 떨어지는 방향도 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "라울은 상단 중앙의 천장 개구부에서 늘어진 옷을 향해 팔을 뻗고 그곳을 올려다봅니다. 명확하게 보이는 손은 늘어진 천의 중간에 닿아 있어, 양손으로 천장 누수면을 누르는 동작은 읽기 어렵습니다. 앰버는 오른쪽 창틀 부근의 천을 양손으로 누르고 그쪽으로 얼굴을 돌렸습니다. 라울의 목표가 지정된 좌상단보다 중앙에 치우칩니다.",
        "built_space": "목재 수납장과 왼쪽 작업대, 오른쪽 회색 벤치, 좌우 측창, 중앙 뒤의 운전실이 보입니다. 작은 점등등 하나와 천장 개구부 하나가 있습니다. 차량 뒤에서 앞을 보는 공간 관계는 성립하지만, 후면 창 위에 있어야 할 참조의 등이 운전실 위쪽에 나타나는 연속성 문제는 A와 같습니다. 허벅지와 통로 바닥까지 포함하여 지정된 미디엄 숏보다 넓게 잡혔습니다.",
        "entities": "추가 인물 없이 아이 두 명이 보입니다. 라울의 피부색, 어린 얼굴, 뒤로 묶은 곱슬머리, 해진 올리브색 티셔츠와 데님 반바지는 참조에 대체로 부합합니다. 앰버의 금발, 어린 얼굴, 회녹색 재킷과 회색 원피스도 이전 장면과 대체로 이어집니다. 지정된 혼혈 정체성과 충돌하는 뚜렷한 외형 차이는 없지만 외형만으로 혈통을 확인할 수는 없습니다. 천장에 끼운 옷들과 창가의 옷 뭉치, 여행가방, 빗물이 보이고 두 사람 모두 얼굴을 찡그리고 있습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "천장 쪽 옷은 개구부 테두리에 끼워져 매달리고 라울의 손이 늘어진 부분에 닿아 있어 무지지 부유는 아닙니다. 다만 손과 천장 사이 거리가 있어 직접 압착하는 힘의 전달은 A보다 불명확합니다. 앰버의 천은 양손과 창틀에 지지되고 아래로 늘어집니다. 두 아이는 몸을 기울여 서 있는 자세이며 다리는 하단으로 이어집니다. 바닥 접점은 잘렸지만 비행하거나 떠 있는 자세는 아니고, 물방울은 아래로 떨어집니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "미디엄 숏 안에서 라울의 좌상단 천장 막기와 앰버의 우상단 창틀 막기, 찡그린 표정을 더 정확히 구현하지만 후방등의 위치 연속성은 어긋납니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "누수를 막는 상황과 인물 외형은 맞지만 프레임이 더 넓고 천장 누수가 상단 중앙으로 옮겨졌으며, 라울이 양손으로 옷을 천장에 압착하는 동작이 불명확합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 라울은 좌상단의 젖은 천장 가장자리로 두 팔을 뻗고, 옷 뭉치와 누수 지점을 올려다봅니다. 오른쪽 앰버는 오른쪽 창의 세로 프레임에 옷을 대고 양손으로 누르며 그 부근을 바라봅니다. 두 사람의 손이 서로 다른 수리 대상을 향한다는 관계가 명확합니다.",
        "built_space": "목재 상부 수납장, 왼쪽 작업대, 오른쪽 회색 벤치, 오른쪽의 인접한 창 두 구획과 왼쪽 측창이 보입니다. 중앙 뒤로 앞유리와 운전석 설비가 보여 차량 뒤에서 앞을 보는 구도입니다. 작은 점등등은 하나뿐입니다. 벤치가 참조의 반대 화면 쪽에 보이는 것은 역방향 카메라로 설명되지만, 참조에서 후면 창 위에 있던 등이 여기서는 운전실 쪽 상부에 보여 고정 조명의 위치 연속성이 맞지 않습니다.",
        "entities": "아이 두 명만 보입니다. 라울은 참조와 유사한 갈색 피부, 어린 얼굴, 뒤로 모은 곱슬머리와 해진 올리브색 티셔츠를 지녔고, 지정된 혼혈 남자아이 외형에 대체로 부합합니다. 앰버는 참조와 유사한 어린 얼굴과 긴 금발, 회녹색 재킷과 회색 옷을 입었으며 이전 장면처럼 마스크가 없습니다. 두 사람의 찡그린 표정, 누수 차단용 옷 두 뭉치, 젖은 창과 천장, 여러 여행가방이 보입니다. 민족적 배경 자체는 외형만으로 확정할 수 없습니다. 다른 사람이나 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "라울의 손들이 옷 뭉치를 받치며 천장 가장자리로 밀고 있고, 앰버의 손들은 천을 창틀에 압착합니다. 옷은 손과 접촉면에 지지되며 뜬 물체는 없습니다. 두 몸통은 하단 프레임 밖으로 이어지는 직립 자세이고, 발이 잘렸다는 이유로 공중 부양으로 볼 근거는 없습니다. 천장과 창틀에서 물이 아래로 떨어지는 방향도 자연스럽습니다."
       },
       {
        "label": "A",
        "direction": "라울은 상단 중앙의 천장 개구부에서 늘어진 옷을 향해 팔을 뻗고 그곳을 올려다봅니다. 명확하게 보이는 손은 늘어진 천의 중간에 닿아 있어, 양손으로 천장 누수면을 누르는 동작은 읽기 어렵습니다. 앰버는 오른쪽 창틀 부근의 천을 양손으로 누르고 그쪽으로 얼굴을 돌렸습니다. 라울의 목표가 지정된 좌상단보다 중앙에 치우칩니다.",
        "built_space": "목재 수납장과 왼쪽 작업대, 오른쪽 회색 벤치, 좌우 측창, 중앙 뒤의 운전실이 보입니다. 작은 점등등 하나와 천장 개구부 하나가 있습니다. 차량 뒤에서 앞을 보는 공간 관계는 성립하지만, 후면 창 위에 있어야 할 참조의 등이 운전실 위쪽에 나타나는 연속성 문제는 A와 같습니다. 허벅지와 통로 바닥까지 포함하여 지정된 미디엄 숏보다 넓게 잡혔습니다.",
        "entities": "추가 인물 없이 아이 두 명이 보입니다. 라울의 피부색, 어린 얼굴, 뒤로 묶은 곱슬머리, 해진 올리브색 티셔츠와 데님 반바지는 참조에 대체로 부합합니다. 앰버의 금발, 어린 얼굴, 회녹색 재킷과 회색 원피스도 이전 장면과 대체로 이어집니다. 지정된 혼혈 정체성과 충돌하는 뚜렷한 외형 차이는 없지만 외형만으로 혈통을 확인할 수는 없습니다. 천장에 끼운 옷들과 창가의 옷 뭉치, 여행가방, 빗물이 보이고 두 사람 모두 얼굴을 찡그리고 있습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "천장 쪽 옷은 개구부 테두리에 끼워져 매달리고 라울의 손이 늘어진 부분에 닿아 있어 무지지 부유는 아닙니다. 다만 손과 천장 사이 거리가 있어 직접 압착하는 힘의 전달은 A보다 불명확합니다. 앰버의 천은 양손과 창틀에 지지되고 아래로 늘어집니다. 두 아이는 몸을 기울여 서 있는 자세이며 다리는 하단으로 이어집니다. 바닥 접점은 잘렸지만 비행하거나 떠 있는 자세는 아니고, 물방울은 아래로 떨어집니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "totals": {
   "B": 8,
   "A": 6
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "미디엄 숏 안에서 라울의 좌상단 천장 막기와 앰버의 우상단 창틀 막기, 찡그린 표정을 더 정확히 구현하지만 후방등의 위치 연속성은 어긋납니다."
   },
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "누수를 막는 상황과 인물 외형은 맞지만 프레임이 더 넓고 천장 누수가 상단 중앙으로 옮겨졌으며, 라울이 양손으로 옷을 천장에 압착하는 동작이 불명확합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S45sh8_sel.png",
    "asset_id": "7654150f-eb85-45a8-a00a-115e66afac13",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:943979>",
    "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdce-9f4a-7936-a5de-345bf63b8490",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S45sh8"
  }
 },
 "S45sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:40:41.968262+00:00",
  "fingerprint": "5f1da8d950429ecd62f3e1a4ed9e8a258e62ba5461dc9b973297998ae8a22ec4",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S45sh14_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S45sh14_sel.png",
  "source_sha256": "bb632da7437ccb0e072f8fa7c71a2607e153cd87d7959fa61c0f8f9e6c5d10cc",
  "file": "S45sh14_cine.png",
  "staged_sha256": "eedac9328eb12a3cde2f4c3037d531412144677aa6b034a2dae00a40c314f31e",
  "latency_ms": 12980
 },
 "S46sh1::signage": {
  "fp": "309fce6643a0ee99",
  "inscriptions": [
   {
    "text_native": "미추홀도서관, 인천",
    "source": "scene_text_quoted",
    "reason_ko": "벽면에 걸린 간판의 문구로 장면에 직접 명시되어 있습니다.",
    "source_quote": "미추홀도서관, 인천"
   }
  ],
  "cues": [],
  "dropped": []
 },
 "S46sh1": {
  "input_fingerprint": "1bbb554b150a8de5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 어둠 속, 마른 넝쿨과 굵은 금이 간 벽면의 '미추홀도서관, 인천' 간판 앞으로 낡은 캠핑카가 멈춰 선 외관.\n\nLOCATION (lock): Outside an abandoned library, at its cracked, vine-covered facade beside the camper stopped in nighttime rain. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Library sign with readable destination text in the upper-center of the frame, background; Stopped camper below the sign in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Old camper (Coming to a stop in front of the library) — Seen from a three-quarter angle offset from its approach; used as Foreground arrival anchor beneath the building identification; Library sign (Old, bearing the text '미추홀도서관, 인천') — The text-bearing face is visible above and behind the camper; used as Identifies the destination without being blocked by the vehicle; Library exterior wall (Covered in attached dry vines and large cracks) — The entrance-side façade is visible obliquely; used as Carries the visible evidence of the building's condition; Library entrance (Visible beyond the stopped camper) — Set into the façade along the camera's continuing approach; used as Establishes the destination of the next camera movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the rainy night subdued while preserving readable tonal separation across the camper, cracked façade, and old sign without inventing an exterior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The leaking camper is stopped outside the abandoned library in night rain, beneath the old sign reading “미추홀도서관, 인천”; dry vines cling to heavily cracked exterior walls. Packed provisions and luggage remain aboard, and Charlie retains the blanket taken from the shop.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"미추홀도서관, 인천\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 어둠 속, 마른 넝쿨과 굵은 금이 간 벽면의 '미추홀도서관, 인천' 간판 앞으로 낡은 캠핑카가 멈춰 선 외관.\n\nLOCATION (lock): Outside an abandoned library, at its cracked, vine-covered facade beside the camper stopped in nighttime rain. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Library sign with readable destination text in the upper-center of the frame, background; Stopped camper below the sign in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Old camper (Coming to a stop in front of the library) — Seen from a three-quarter angle offset from its approach; used as Foreground arrival anchor beneath the building identification; Library sign (Old, bearing the text '미추홀도서관, 인천') — The text-bearing face is visible above and behind the camper; used as Identifies the destination without being blocked by the vehicle; Library exterior wall (Covered in attached dry vines and large cracks) — The entrance-side façade is visible obliquely; used as Carries the visible evidence of the building's condition; Library entrance (Visible beyond the stopped camper) — Set into the façade along the camera's continuing approach; used as Establishes the destination of the next camera movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the rainy night subdued while preserving readable tonal separation across the camper, cracked façade, and old sign without inventing an exterior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The leaking camper is stopped outside the abandoned library in night rain, beneath the old sign reading “미추홀도서관, 인천”; dry vines cling to heavily cracked exterior walls. Packed provisions and luggage remain aboard, and Charlie retains the blanket taken from the shop.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"미추홀도서관, 인천\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 어둠 속, 마른 넝쿨과 굵은 금이 간 벽면의 '미추홀도서관, 인천' 간판 앞으로 낡은 캠핑카가 멈춰 선 외관.\n\nLOCATION (lock): Outside an abandoned library, at its cracked, vine-covered facade beside the camper stopped in nighttime rain. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Library sign with readable destination text in the upper-center of the frame, background; Stopped camper below the sign in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Old camper (Coming to a stop in front of the library) — Seen from a three-quarter angle offset from its approach; used as Foreground arrival anchor beneath the building identification; Library sign (Old, bearing the text '미추홀도서관, 인천') — The text-bearing face is visible above and behind the camper; used as Identifies the destination without being blocked by the vehicle; Library exterior wall (Covered in attached dry vines and large cracks) — The entrance-side façade is visible obliquely; used as Carries the visible evidence of the building's condition; Library entrance (Visible beyond the stopped camper) — Set into the façade along the camera's continuing approach; used as Establishes the destination of the next camera movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the rainy night subdued while preserving readable tonal separation across the camper, cracked façade, and old sign without inventing an exterior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The leaking camper is stopped outside the abandoned library in night rain, beneath the old sign reading “미추홀도서관, 인천”; dry vines cling to heavily cracked exterior walls. Packed provisions and luggage remain aboard, and Charlie retains the blanket taken from the shop.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"미추홀도서관, 인천\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 도서관 건물을 정면 빗각으로 바라보며, 캠핑카는 건물을 배경으로 배치됨.",
    "built_space": "레퍼런스 이미지의 건물 외벽, 금 간 자국, 넝쿨, 그리고 좌측의 가로등까지 정확한 위치에 구현됨.",
    "entities": "지정된 간판 텍스트('미추홀도서관, 인천')가 오타 없이 정확하게 적혀 있으며, 낡은 캠핑카의 묘사도 적절함. 불필요한 추가 문자가 없음.",
    "hard_violations": [],
    "physics": "캠핑카의 바퀴가 젖은 아스팔트 바닥에 안정적으로 접지되어 있으며, 빗물 고인 바닥의 반사가 자연스러움."
   },
   {
    "label": "B",
    "direction": "카메라가 도서관 건물을 정면 빗각으로 바라보며, 캠핑카가 건물 진입로 쪽에 정차해 있음.",
    "built_space": "건물 외벽과 간판의 구조는 레퍼런스와 일치하나, 좌측에 있던 가로등 기둥이 삭제됨.",
    "entities": "간판 텍스트('미추홀도서관, 인천')는 정확하나, 차량 전면 번호판에 읽을 수 있는 문자/숫자가 추가되어 프롬프트 제약을 위반함.",
    "hard_violations": [],
    "physics": "캠핑카가 노면에 정상적으로 닿아 있으며, 비 내리는 환경에 따른 바닥 질감과 반사가 물리적으로 타당함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 장소(좌측 가로등 등)를 온전히 유지하고, 추가적인 텍스트 없이 지정된 간판 문자만 정확하게 구현하여 지시사항을 훌륭히 따름."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "요구된 텍스트는 정확히 렌더링했으나, 레퍼런스의 가로등을 임의로 삭제하여 장소 일관성을 훼손했고 금지된 읽을 수 있는 텍스트(차량 번호판)를 추가함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 도서관 건물을 정면 빗각으로 바라보며, 캠핑카는 건물을 배경으로 배치됨.",
        "built_space": "레퍼런스 이미지의 건물 외벽, 금 간 자국, 넝쿨, 그리고 좌측의 가로등까지 정확한 위치에 구현됨.",
        "entities": "지정된 간판 텍스트('미추홀도서관, 인천')가 오타 없이 정확하게 적혀 있으며, 낡은 캠핑카의 묘사도 적절함. 불필요한 추가 문자가 없음.",
        "hard_violations": [],
        "physics": "캠핑카의 바퀴가 젖은 아스팔트 바닥에 안정적으로 접지되어 있으며, 빗물 고인 바닥의 반사가 자연스러움."
       },
       {
        "label": "B",
        "direction": "카메라가 도서관 건물을 정면 빗각으로 바라보며, 캠핑카가 건물 진입로 쪽에 정차해 있음.",
        "built_space": "건물 외벽과 간판의 구조는 레퍼런스와 일치하나, 좌측에 있던 가로등 기둥이 삭제됨.",
        "entities": "간판 텍스트('미추홀도서관, 인천')는 정확하나, 차량 전면 번호판에 읽을 수 있는 문자/숫자가 추가되어 프롬프트 제약을 위반함.",
        "hard_violations": [],
        "physics": "캠핑카가 노면에 정상적으로 닿아 있으며, 비 내리는 환경에 따른 바닥 질감과 반사가 물리적으로 타당함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 장소(좌측 가로등 등)를 온전히 유지하고, 추가적인 텍스트 없이 지정된 간판 문자만 정확하게 구현하여 지시사항을 훌륭히 따름."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "요구된 텍스트는 정확히 렌더링했으나, 레퍼런스의 가로등을 임의로 삭제하여 장소 일관성을 훼손했고 금지된 읽을 수 있는 텍스트(차량 번호판)를 추가함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 도서관 건물을 정면 빗각으로 바라보며, 캠핑카는 건물을 배경으로 배치됨.",
        "built_space": "레퍼런스 이미지의 건물 외벽, 금 간 자국, 넝쿨, 그리고 좌측의 가로등까지 정확한 위치에 구현됨.",
        "entities": "지정된 간판 텍스트('미추홀도서관, 인천')가 오타 없이 정확하게 적혀 있으며, 낡은 캠핑카의 묘사도 적절함. 불필요한 추가 문자가 없음.",
        "hard_violations": [],
        "physics": "캠핑카의 바퀴가 젖은 아스팔트 바닥에 안정적으로 접지되어 있으며, 빗물 고인 바닥의 반사가 자연스러움."
       },
       {
        "label": "B",
        "direction": "카메라가 도서관 건물을 정면 빗각으로 바라보며, 캠핑카가 건물 진입로 쪽에 정차해 있음.",
        "built_space": "건물 외벽과 간판의 구조는 레퍼런스와 일치하나, 좌측에 있던 가로등 기둥이 삭제됨.",
        "entities": "간판 텍스트('미추홀도서관, 인천')는 정확하나, 차량 전면 번호판에 읽을 수 있는 문자/숫자가 추가되어 프롬프트 제약을 위반함.",
        "hard_violations": [],
        "physics": "캠핑카가 노면에 정상적으로 닿아 있으며, 비 내리는 환경에 따른 바닥 질감과 반사가 물리적으로 타당함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "간판과 정차한 캠핑카의 관계는 맞지만, 차량이 전경에 가깝게 크게 잡혀 중경 배치에서 벗어나고 참조의 왼쪽 가로등도 보이지 않는다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "캠핑카를 좌하단 중경에 둔 와이드 구도와 가려지지 않은 간판·출입구, 참조의 가로등과 젖은 외벽을 더 충실하게 재현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카 앞부분은 화면 오른쪽을 향하며 앞면과 왼쪽 측면이 함께 보이는 사선 구도다. 차량은 출입구 앞에 멈춰 있고, 간판의 글자 면은 카메라에 보인다. 사람의 시선이나 겨냥하는 물체는 없다.",
        "built_space": "낡은 2층 외벽에 상층 창들이 반복되고, 출입구 위에는 긴 간판 겸 차양 한 개가 있다. 그 아래 두 구획의 유리 출입문이 있으며 왼쪽 구획 일부는 차량에 가린다. 오른쪽 난간 한 줄과 양쪽 낮은 화단이 보인다. 참조의 왼쪽 가로등은 보이지 않는다. 간판은 상단 중앙에서 오른쪽에 걸쳐 있고 차량은 좌하단이지만, 차체가 커서 지정된 중경보다 전경에 가깝다.",
        "entities": "낡고 오염된 캠핑카 한 대, 마른 넝쿨이 붙은 폐도서관 외벽, 유리 출입구와 금속 간판이 보인다. 간판에는 요청한 ‘미추홀도서관, 인천’이 읽힌다. 외벽의 박락과 넝쿨은 뚜렷하지만 굵은 균열은 그보다 덜 명확하다. 사람은 보이지 않는다. 차량 내부에는 적재물처럼 보이는 물건이 있으나 식량·짐·담요의 종류는 확인할 수 없다. 차량 번호판과 작은 표식은 있으나 정확한 추가 문구는 확정하기 어렵다.",
        "hard_violations": [],
        "physics": "차량의 보이는 앞뒤 타이어가 젖은 노면에 닿아 차체를 지탱한다. 정차 자세가 자연스럽고 떠 있는 물체는 없다. 간판은 출입구 구조물에 고정되고 넝쿨은 외벽에 붙어 있다. 빗줄기와 노면 반사는 물리적으로 가능하지만, 왼쪽의 따뜻한 조명을 만드는 광원은 화면에서 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "캠핑카는 화면 오른쪽을 향해 정차해 있으며 앞면과 왼쪽 측면이 함께 보인다. 차량 뒤쪽으로 출입구가 이어지고, 그 위 간판의 글자 면이 카메라에 노출된다. 간판을 차량이 가리지 않는다. 사람의 시선이나 겨냥하는 물체는 없다.",
        "built_space": "상층의 반복 창 배열, 긴 간판 겸 차양 한 개, 그 아래 나란한 두 유리 출입문 구획이 보인다. 오른쪽에는 난간 한 줄, 출입구 양쪽에는 낮은 화단이 있고 왼쪽에는 참조와 같은 가로등 한 개가 서 있다. 차량은 좌하단 중경에 놓여 출입구와 젖은 진입 공간이 충분히 드러난다. 간판은 상단 중앙에서 약간 오른쪽에 위치한다.",
        "entities": "때와 녹이 밴 오래된 캠핑카 한 대와 폐도서관의 콘크리트 외벽, 마른 넝쿨, 창문, 출입구가 보인다. 간판에는 ‘미추홀도서관, 인천’이 읽힌다. 넝쿨과 표면 손상은 분명하지만 굵은 균열의 강조는 다소 약하다. 추가 인물은 없다. 실내가 어두워 적재 식량·짐과 찰리의 담요는 확인할 수 없으며 이 외관 구도에서 드러날 필요는 없다. 번호판과 출입문 부착물의 세부 글자는 확정하기 어렵다.",
        "hard_violations": [],
        "physics": "보이는 앞뒤 바퀴가 노면에 접촉하여 차량을 지탱하고, 차체는 자연스러운 정차 상태다. 간판과 난간은 건물 및 지면에 고정되어 있다. 왼쪽 가로등의 따뜻한 빛이 비와 젖은 노면에 반사되는 관계가 자연스럽고 참조의 기존 광원과도 일치한다. 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "간판과 정차한 캠핑카의 관계는 맞지만, 차량이 전경에 가깝게 크게 잡혀 중경 배치에서 벗어나고 참조의 왼쪽 가로등도 보이지 않는다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "캠핑카를 좌하단 중경에 둔 와이드 구도와 가려지지 않은 간판·출입구, 참조의 가로등과 젖은 외벽을 더 충실하게 재현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카 앞부분은 화면 오른쪽을 향하며 앞면과 왼쪽 측면이 함께 보이는 사선 구도다. 차량은 출입구 앞에 멈춰 있고, 간판의 글자 면은 카메라에 보인다. 사람의 시선이나 겨냥하는 물체는 없다.",
        "built_space": "낡은 2층 외벽에 상층 창들이 반복되고, 출입구 위에는 긴 간판 겸 차양 한 개가 있다. 그 아래 두 구획의 유리 출입문이 있으며 왼쪽 구획 일부는 차량에 가린다. 오른쪽 난간 한 줄과 양쪽 낮은 화단이 보인다. 참조의 왼쪽 가로등은 보이지 않는다. 간판은 상단 중앙에서 오른쪽에 걸쳐 있고 차량은 좌하단이지만, 차체가 커서 지정된 중경보다 전경에 가깝다.",
        "entities": "낡고 오염된 캠핑카 한 대, 마른 넝쿨이 붙은 폐도서관 외벽, 유리 출입구와 금속 간판이 보인다. 간판에는 요청한 ‘미추홀도서관, 인천’이 읽힌다. 외벽의 박락과 넝쿨은 뚜렷하지만 굵은 균열은 그보다 덜 명확하다. 사람은 보이지 않는다. 차량 내부에는 적재물처럼 보이는 물건이 있으나 식량·짐·담요의 종류는 확인할 수 없다. 차량 번호판과 작은 표식은 있으나 정확한 추가 문구는 확정하기 어렵다.",
        "hard_violations": [],
        "physics": "차량의 보이는 앞뒤 타이어가 젖은 노면에 닿아 차체를 지탱한다. 정차 자세가 자연스럽고 떠 있는 물체는 없다. 간판은 출입구 구조물에 고정되고 넝쿨은 외벽에 붙어 있다. 빗줄기와 노면 반사는 물리적으로 가능하지만, 왼쪽의 따뜻한 조명을 만드는 광원은 화면에서 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "캠핑카는 화면 오른쪽을 향해 정차해 있으며 앞면과 왼쪽 측면이 함께 보인다. 차량 뒤쪽으로 출입구가 이어지고, 그 위 간판의 글자 면이 카메라에 노출된다. 간판을 차량이 가리지 않는다. 사람의 시선이나 겨냥하는 물체는 없다.",
        "built_space": "상층의 반복 창 배열, 긴 간판 겸 차양 한 개, 그 아래 나란한 두 유리 출입문 구획이 보인다. 오른쪽에는 난간 한 줄, 출입구 양쪽에는 낮은 화단이 있고 왼쪽에는 참조와 같은 가로등 한 개가 서 있다. 차량은 좌하단 중경에 놓여 출입구와 젖은 진입 공간이 충분히 드러난다. 간판은 상단 중앙에서 약간 오른쪽에 위치한다.",
        "entities": "때와 녹이 밴 오래된 캠핑카 한 대와 폐도서관의 콘크리트 외벽, 마른 넝쿨, 창문, 출입구가 보인다. 간판에는 ‘미추홀도서관, 인천’이 읽힌다. 넝쿨과 표면 손상은 분명하지만 굵은 균열의 강조는 다소 약하다. 추가 인물은 없다. 실내가 어두워 적재 식량·짐과 찰리의 담요는 확인할 수 없으며 이 외관 구도에서 드러날 필요는 없다. 번호판과 출입문 부착물의 세부 글자는 확정하기 어렵다.",
        "hard_violations": [],
        "physics": "보이는 앞뒤 바퀴가 노면에 접촉하여 차량을 지탱하고, 차체는 자연스러운 정차 상태다. 간판과 난간은 건물 및 지면에 고정되어 있다. 왼쪽 가로등의 따뜻한 빛이 비와 젖은 노면에 반사되는 관계가 자연스럽고 참조의 기존 광원과도 일치한다. 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.492
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.492
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1492
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스의 장소(좌측 가로등 등)를 온전히 유지하고, 추가적인 텍스트 없이 지정된 간판 문자만 정확하게 구현하여 지시사항을 훌륭히 따름."
   },
   {
    "label": "B",
    "score": 1492,
    "verdict_ko": "요구된 텍스트는 정확히 렌더링했으나, 레퍼런스의 가로등을 임의로 삭제하여 장소 일관성을 훼손했고 금지된 읽을 수 있는 텍스트(차량 번호판)를 추가함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L57B01.png",
    "asset_id": "9d242e06-fe86-44ea-8d1e-34ce22a741dd",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdcf-74bb-7ca1-a0d5-6be9b3d1b728",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S46sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:40:55.309314+00:00",
  "fingerprint": "2280aa45bb32acc4944097cd02505028eed105cff8b6dbe41af87a44d213068b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S46sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S46sh1_sel.png",
  "source_sha256": "562dfed185a1ce355e24d68a7f0332a63f30dd6f4847a4a3d8a646da5c7036e8",
  "file": "S46sh1_cine.png",
  "staged_sha256": "77baba0408461280dcc5efbab3928afe458166afed3abd4f355e62eb0bb46acd",
  "latency_ms": 11473
 },
 "S46sh21::signage": {
  "fp": "e71a11430ff8681b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::8faf47aeed83f60e": {
  "subjects": [],
  "subject_text": "미추홀도서관 열람실과 서가\n먼지와 거미줄이 낀 오래 방치된 열람실. 낡은 책들이 서가에 빽빽하며 일부 책장은 어지럽고 구형 컴퓨터가 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L57",
  "scope_role": "location_interior",
  "scope_sha": "c9699e751dbc2c56"
 },
 "S46sh21::bgfirst_bg": {
  "input_fingerprint": "cbe544cad7843274",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 눈물 고인 눈으로 이현우를 향해 두 주먹을 불끈 쥐고 소리치는 격앙된 얼굴 클로즈업.\n\nLOCATION (lock): At the group's makeshift eating spot inside the dusty abandoned library reading room, in faint nighttime ambient light.\n\nTIME OF DAY (lock): night, rainy.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reading-room shelving (Disordered and dust-covered) — Only indistinct portions remain behind 앰버; used as Retains the abandoned-library context without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the reading room, keeping the tears and facial tension readable without inventing a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 눈물 고인 눈으로 이현우를 향해 두 주먹을 불끈 쥐고 소리치는 격앙된 얼굴 클로즈업.\n\nLOCATION (lock): At the group's makeshift eating spot inside the dusty abandoned library reading room, in faint nighttime ambient light.\n\nTIME OF DAY (lock): night, rainy.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reading-room shelving (Disordered and dust-covered) — Only indistinct portions remain behind 앰버; used as Retains the abandoned-library context without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the reading room, keeping the tears and facial tension readable without inventing a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S46sh21__bgfirst_bg.png",
  "asset_id": "d98ebfeb-91b2-4251-8842-2465c7a28f27",
  "input_asset_ids": [
   "3c9bba97-24b3-45b0-8fbe-9aab9b556d77",
   "91a5b479-7f25-41c3-bb98-f9a4555173f3"
  ]
 },
 "S46sh21": {
  "input_fingerprint": "2ebabb58ac83beaf",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 앰버가 눈물 고인 눈으로 이현우를 향해 두 주먹을 불끈 쥐고 소리치는 격앙된 얼굴 클로즈업.\n\nLOCATION (lock): At the group's makeshift eating spot inside the dusty abandoned library reading room, in faint nighttime ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reading-room shelving (Disordered and dust-covered) — Only indistinct portions remain behind 앰버; used as Retains the abandoned-library context without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the reading room, keeping the tears and facial tension readable without inventing a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The library interior remains thick with dust and cobwebs, with disordered shelves; the UBIC book thrown aside is now on the floor, and canned food and luggage occupy the makeshift eating spot. Rain continues outside the parked, leaking camper, while Charlie remains nearby with his old body and blanket. 앰버: She has refused her canned food and is now tearful and angry.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 앰버가 눈물 고인 눈으로 이현우를 향해 두 주먹을 불끈 쥐고 소리치는 격앙된 얼굴 클로즈업.\n\nLOCATION (lock): At the group's makeshift eating spot inside the dusty abandoned library reading room, in faint nighttime ambient light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reading-room shelving (Disordered and dust-covered) — Only indistinct portions remain behind 앰버; used as Retains the abandoned-library context without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the reading room, keeping the tears and facial tension readable without inventing a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The library interior remains thick with dust and cobwebs, with disordered shelves; the UBIC book thrown aside is now on the floor, and canned food and luggage occupy the makeshift eating spot. Rain continues outside the parked, leaking camper, while Charlie remains nearby with his old body and blanket. 앰버: She has refused her canned food and is now tearful and angry.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 앰버가 눈물 고인 눈으로 이현우를 향해 두 주먹을 불끈 쥐고 소리치는 격앙된 얼굴 클로즈업.\n\nLOCATION (lock): At the group's makeshift eating spot inside the dusty abandoned library reading room, in faint nighttime ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Reading-room shelving (Disordered and dust-covered) — Only indistinct portions remain behind 앰버; used as Retains the abandoned-library context without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the reading room, keeping the tears and facial tension readable without inventing a source or hue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The library interior remains thick with dust and cobwebs, with disordered shelves; the UBIC book thrown aside is now on the floor, and canned food and luggage occupy the makeshift eating spot. Rain continues outside the parked, leaking camper, while Charlie remains nearby with his old body and blanket. 앰버: She has refused her canned food and is now tearful and angry.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S46sh21__bgfirst_bg.png",
     "asset_id": "d98ebfeb-91b2-4251-8842-2465c7a28f27",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S46sh21.png",
     "asset_id": "3c9bba97-24b3-45b0-8fbe-9aab9b556d77",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L57B02.png",
     "asset_id": "91a5b479-7f25-41c3-bb98-f9a4555173f3",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "앰버의 시선과 외침이 전경에 등 돌린 남성을 명확히 향하고 있음.",
    "built_space": "기준 이미지의 도서관 서가, 비 오는 창문, 테이블이 올바른 위치에 배치됨.",
    "entities": "앰버의 인상착의(금발, 회색 옷)가 기준 이미지와 일치하며 눈물 흘리는 모습이 명확함.",
    "hard_violations": [
     "[gpt-high] 화면에 허용된 인물은 앰버뿐인데 왼쪽 전경에 별도 남성의 머리와 상체를 크게 추가했다."
    ],
    "physics": "들어 올린 두 주먹과 팔이 몸통에 의해 자연스럽게 지탱됨."
   },
   {
    "label": "B",
    "direction": "앰버가 우측 전경에 있는 인물을 향해 시선을 두고 있음.",
    "built_space": "도서관 서가가 흐릿하게 배경에 나타나 위치를 암시함.",
    "entities": "앰버의 인상(금발, 녹색 재킷)이 기준과 일치하며 눈물을 흘림.",
    "hard_violations": [
     "[gpt-high] 화면에 허용된 인물은 앰버뿐인데 오른쪽 전경에 별도 남성의 머리와 어깨를 추가했다."
    ],
    "physics": "주먹 쥔 두 손이 가슴 상단 옷깃 부근에 기대어 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트가 요구한 '소리치는 격앙된 얼굴'과 '두 주먹을 불끈 쥔' 행동을 기준 장소에 맞춰 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "눈물은 보이나 지시된 '소리치는' 모습이 아니며, 주먹을 쥔 자세와 표정이 프롬프트의 격앙된 감정과 거리가 멉니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선과 외침이 전경에 등 돌린 남성을 명확히 향하고 있음.",
        "built_space": "기준 이미지의 도서관 서가, 비 오는 창문, 테이블이 올바른 위치에 배치됨.",
        "entities": "앰버의 인상착의(금발, 회색 옷)가 기준 이미지와 일치하며 눈물 흘리는 모습이 명확함.",
        "hard_violations": [],
        "physics": "들어 올린 두 주먹과 팔이 몸통에 의해 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "앰버가 우측 전경에 있는 인물을 향해 시선을 두고 있음.",
        "built_space": "도서관 서가가 흐릿하게 배경에 나타나 위치를 암시함.",
        "entities": "앰버의 인상(금발, 녹색 재킷)이 기준과 일치하며 눈물을 흘림.",
        "hard_violations": [],
        "physics": "주먹 쥔 두 손이 가슴 상단 옷깃 부근에 기대어 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트가 요구한 '소리치는 격앙된 얼굴'과 '두 주먹을 불끈 쥔' 행동을 기준 장소에 맞춰 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "눈물은 보이나 지시된 '소리치는' 모습이 아니며, 주먹을 쥔 자세와 표정이 프롬프트의 격앙된 감정과 거리가 멉니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선과 외침이 전경에 등 돌린 남성을 명확히 향하고 있음.",
        "built_space": "기준 이미지의 도서관 서가, 비 오는 창문, 테이블이 올바른 위치에 배치됨.",
        "entities": "앰버의 인상착의(금발, 회색 옷)가 기준 이미지와 일치하며 눈물 흘리는 모습이 명확함.",
        "hard_violations": [],
        "physics": "들어 올린 두 주먹과 팔이 몸통에 의해 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "앰버가 우측 전경에 있는 인물을 향해 시선을 두고 있음.",
        "built_space": "도서관 서가가 흐릿하게 배경에 나타나 위치를 암시함.",
        "entities": "앰버의 인상(금발, 녹색 재킷)이 기준과 일치하며 눈물을 흘림.",
        "hard_violations": [],
        "physics": "주먹 쥔 두 손이 가슴 상단 옷깃 부근에 기대어 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "얼굴 클로즈업과 눈물은 맞지만, 시선이 오른쪽 남성을 향하지 않고 격앙된 외침보다 슬픈 호소로 보이며 허용되지 않은 인물이 추가됐다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "상대를 향한 시선·외침·불끈 쥔 두 주먹은 더 정확하지만, 남성을 추가하고 상반신과 배경까지 넓혀 필수 얼굴 클로즈업을 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 얼굴과 눈은 화면 왼쪽의 프레임 밖을 향한다. 오른쪽 전경에 있는 남성은 앰버 쪽으로 고개를 기울이지만 앰버의 시선은 그에게 닿지 않는다. 이 남성을 이현우로 해석하면 요구된 대면 방향과 어긋난다. 두 주먹은 상대 쪽으로 뻗지 않고 자기 목 아래에 모여 있다.",
        "built_space": "왼쪽과 얼굴 뒤로 흐릿한 목재 서가 한 구간이 보이고 오른쪽 아래에는 의자 등받이 일부가 있다. 얼굴 중심의 좁은 구도와 흐릿한 서가 처리는 요구에 부합한다. 식탁·창문·서가 전체 수는 프레임 밖이므로 검증할 수 없다. 참고 장소와 재료는 유사하지만 두꺼운 먼지와 거미줄은 뚜렷하지 않다.",
        "entities": "금발의 큰 눈을 가진 약 10세 여자아이 한 명이 있으며 눈물이 양 볼에 흐른다. 카키색 겉옷과 회색 안쪽 옷은 참고 의상에 가깝다. 얼굴의 정확한 동일성과 한국계·백인 혼혈 설정은 외형만으로 확정하기 어렵다. 입은 조금 벌어져 있어 고함보다는 울먹이는 호소에 가깝다. 오른쪽에는 검은 머리 남성의 얼굴·귀·어깨 일부가 추가되어 있다. 읽을 수 있는 글자는 없다. 바닥의 책과 식사 소품은 올바른 좁은 구도 밖이므로 누락으로 판단하지 않는다.",
        "hard_violations": [
         "화면에 허용된 인물은 앰버뿐인데 오른쪽 전경에 별도 남성의 머리와 어깨를 추가했다."
        ],
        "physics": "두 손은 소매와 팔에 자연스럽게 연결되어 턱 아래에서 주먹을 쥐고 있다. 목과 어깨가 머리를 정상적으로 지지하며 눈물은 볼을 따라 아래로 흐른다. 하체와 좌면은 잘려 있어 서 있는지 앉아 있는지는 확정할 수 없지만 부유나 불가능한 지지는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "앰버는 화면 왼쪽 전경 남성의 얼굴 쪽을 올려다보며 입을 크게 벌리고 소리친다. 보이는 대화 상대를 이현우로 해석하면 시선과 외침의 방향이 맞는다. 두 주먹은 가슴 앞에 각각 들려 있고 상대를 때리거나 물건을 겨누는 동작은 아니다.",
        "built_space": "중앙부터 오른쪽까지 독립 서가의 끝판 세 개와 오른쪽 벽면 서가 일부, 뒤쪽 큰 창문, 오른쪽 식탁 한 개가 보인다. 왼쪽 중경과 식탁 뒤에는 검은 등받이 목재 의자가 있고 오른쪽 뒤에는 금속 운반대 한 대가 있다. 참고 장소의 서가 배열과 식탁 재료는 상당히 가깝다. 다만 앰버의 상반신과 넓은 열람실을 보여 주는 대면 구도로, 얼굴 클로즈업과 배경을 불분명한 일부로만 남기라는 지시를 어긴다.",
        "entities": "눈물이 흐르는 금발의 큰 눈을 가진 여자아이 한 명이 두 주먹을 꽉 쥐고 고함치는 모습이다. 나이대는 대체로 맞지만 정확한 참고 인물 동일성과 혼혈 설정은 확정하기 어렵다. 회색 반소매 옷만 보이며 참고의 카키색 겉옷은 없다. 왼쪽 전경에는 검은 머리 남성의 머리와 넓은 어깨가 추가됐다. 식탁에는 흰 컵, 책, 투명한 뚜껑과 검은 짐이 보인다. 글자는 읽히지 않는다. 창문과 어두운 실내는 밤 분위기에 맞지만 푸른 조명이 참고보다 강하다.",
        "hard_violations": [
         "화면에 허용된 인물은 앰버뿐인데 왼쪽 전경에 별도 남성의 머리와 상체를 크게 추가했다."
        ],
        "physics": "들어 올린 두 주먹은 굽힌 팔꿈치와 팔에 연결되어 있으며 긴장한 어깨와 고함치는 자세가 물리적으로 가능하다. 인물의 하체와 좌면은 프레임 밖이라 체중 지지는 직접 확인되지 않지만 떠 있는 징후는 없다. 컵·책·뚜껑은 식탁에 놓여 있고 의자와 서가는 바닥 위에 정상적으로 배치되어 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "얼굴 클로즈업과 눈물은 맞지만, 시선이 오른쪽 남성을 향하지 않고 격앙된 외침보다 슬픈 호소로 보이며 허용되지 않은 인물이 추가됐다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "상대를 향한 시선·외침·불끈 쥔 두 주먹은 더 정확하지만, 남성을 추가하고 상반신과 배경까지 넓혀 필수 얼굴 클로즈업을 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 얼굴과 눈은 화면 왼쪽의 프레임 밖을 향한다. 오른쪽 전경에 있는 남성은 앰버 쪽으로 고개를 기울이지만 앰버의 시선은 그에게 닿지 않는다. 이 남성을 이현우로 해석하면 요구된 대면 방향과 어긋난다. 두 주먹은 상대 쪽으로 뻗지 않고 자기 목 아래에 모여 있다.",
        "built_space": "왼쪽과 얼굴 뒤로 흐릿한 목재 서가 한 구간이 보이고 오른쪽 아래에는 의자 등받이 일부가 있다. 얼굴 중심의 좁은 구도와 흐릿한 서가 처리는 요구에 부합한다. 식탁·창문·서가 전체 수는 프레임 밖이므로 검증할 수 없다. 참고 장소와 재료는 유사하지만 두꺼운 먼지와 거미줄은 뚜렷하지 않다.",
        "entities": "금발의 큰 눈을 가진 약 10세 여자아이 한 명이 있으며 눈물이 양 볼에 흐른다. 카키색 겉옷과 회색 안쪽 옷은 참고 의상에 가깝다. 얼굴의 정확한 동일성과 한국계·백인 혼혈 설정은 외형만으로 확정하기 어렵다. 입은 조금 벌어져 있어 고함보다는 울먹이는 호소에 가깝다. 오른쪽에는 검은 머리 남성의 얼굴·귀·어깨 일부가 추가되어 있다. 읽을 수 있는 글자는 없다. 바닥의 책과 식사 소품은 올바른 좁은 구도 밖이므로 누락으로 판단하지 않는다.",
        "hard_violations": [
         "화면에 허용된 인물은 앰버뿐인데 오른쪽 전경에 별도 남성의 머리와 어깨를 추가했다."
        ],
        "physics": "두 손은 소매와 팔에 자연스럽게 연결되어 턱 아래에서 주먹을 쥐고 있다. 목과 어깨가 머리를 정상적으로 지지하며 눈물은 볼을 따라 아래로 흐른다. 하체와 좌면은 잘려 있어 서 있는지 앉아 있는지는 확정할 수 없지만 부유나 불가능한 지지는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "앰버는 화면 왼쪽 전경 남성의 얼굴 쪽을 올려다보며 입을 크게 벌리고 소리친다. 보이는 대화 상대를 이현우로 해석하면 시선과 외침의 방향이 맞는다. 두 주먹은 가슴 앞에 각각 들려 있고 상대를 때리거나 물건을 겨누는 동작은 아니다.",
        "built_space": "중앙부터 오른쪽까지 독립 서가의 끝판 세 개와 오른쪽 벽면 서가 일부, 뒤쪽 큰 창문, 오른쪽 식탁 한 개가 보인다. 왼쪽 중경과 식탁 뒤에는 검은 등받이 목재 의자가 있고 오른쪽 뒤에는 금속 운반대 한 대가 있다. 참고 장소의 서가 배열과 식탁 재료는 상당히 가깝다. 다만 앰버의 상반신과 넓은 열람실을 보여 주는 대면 구도로, 얼굴 클로즈업과 배경을 불분명한 일부로만 남기라는 지시를 어긴다.",
        "entities": "눈물이 흐르는 금발의 큰 눈을 가진 여자아이 한 명이 두 주먹을 꽉 쥐고 고함치는 모습이다. 나이대는 대체로 맞지만 정확한 참고 인물 동일성과 혼혈 설정은 확정하기 어렵다. 회색 반소매 옷만 보이며 참고의 카키색 겉옷은 없다. 왼쪽 전경에는 검은 머리 남성의 머리와 넓은 어깨가 추가됐다. 식탁에는 흰 컵, 책, 투명한 뚜껑과 검은 짐이 보인다. 글자는 읽히지 않는다. 창문과 어두운 실내는 밤 분위기에 맞지만 푸른 조명이 참고보다 강하다.",
        "hard_violations": [
         "화면에 허용된 인물은 앰버뿐인데 왼쪽 전경에 별도 남성의 머리와 상체를 크게 추가했다."
        ],
        "physics": "들어 올린 두 주먹은 굽힌 팔꿈치와 팔에 연결되어 있으며 긴장한 어깨와 고함치는 자세가 물리적으로 가능하다. 인물의 하체와 좌면은 프레임 밖이라 체중 지지는 직접 확인되지 않지만 떠 있는 징후는 없다. 컵·책·뚜껑은 식탁에 놓여 있고 의자와 서가는 바닥 위에 정상적으로 배치되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.167
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.917
   },
   "violations": {
    "B": [
     "[gpt-high] 화면에 허용된 인물은 앰버뿐인데 오른쪽 전경에 별도 남성의 머리와 어깨를 추가했다."
    ],
    "A": [
     "[gpt-high] 화면에 허용된 인물은 앰버뿐인데 왼쪽 전경에 별도 남성의 머리와 상체를 크게 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 917
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "프롬프트가 요구한 '소리치는 격앙된 얼굴'과 '두 주먹을 불끈 쥔' 행동을 기준 장소에 맞춰 완벽하게 구현했습니다.  ★위반: [gpt-high] 화면에 허용된 인물은 앰버뿐인데 왼쪽 전경에 별도 남성의 머리와 상체를 크게 추가했다."
   },
   {
    "label": "B",
    "score": 917,
    "verdict_ko": "눈물은 보이나 지시된 '소리치는' 모습이 아니며, 주먹을 쥔 자세와 표정이 프롬프트의 격앙된 감정과 거리가 멉니다.  ★위반: [gpt-high] 화면에 허용된 인물은 앰버뿐인데 오른쪽 전경에 별도 남성의 머리와 어깨를 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L57B02.png",
    "asset_id": "91a5b479-7f25-41c3-bb98-f9a4555173f3",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd0-31ec-78e6-b9b1-b703b501d166",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S46sh21__bgfirst_bg.png",
   "bg_asset_id": "d98ebfeb-91b2-4251-8842-2465c7a28f27",
   "bg_record_key": "S46sh21::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S46sh21::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:41:07.211914+00:00",
  "fingerprint": "935917608a3e53c2d5f36c0d8f755b3cb62c1e7b2df340f3a49468647d5f67d1",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S46sh21_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S46sh21_sel.png",
  "source_sha256": "72b79093301094a7bd3f39eaa7ddcbf38a002f2d105a15fb4f8959f4663dea19",
  "file": "S46sh21_cine.png",
  "staged_sha256": "6d4c7ff7ebfd1367bacd9271f58618756abcac0eeb4e7a41931d5e1743c720b4",
  "latency_ms": 12226
 },
 "S46sh26::signage": {
  "fp": "d98e4ef6f7042161",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "책 페이지"
   }
  ],
  "dropped": []
 },
 "S46sh26": {
  "input_fingerprint": "2a9ec68e710887c9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 먼지 쌓인 서가 사이에서 찰리가 두 눈의 푸른빛을 책 페이지 위에 눈부시게 비추는 열중한 모습.\n\nLOCATION (lock): Between dusty book stacks inside the abandoned library, where the robot's blue eye light falls brightly on an open page. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Open book (Open and being read intently) — The interior page faces are visible obliquely; no particular wording or illustration is specified; used as Shares focus with the inclined face and makes his reading action legible; Library shelves and books (Dust-covered) — Shelf fronts and side edges recede along the aisle beside 찰리; used as Provide depth brackets and a quiet enclosure around the reader.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: 찰리's described blue eye-light shines brightly onto the pages, providing a selective vivid accent while the surrounding library illumination remains restrained and unspecified.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dust-covered shelves and books surround Charlie as he reads; no mural has yet been established at this point. In one corner, boxes and books serve as makeshift bedding, and the camper remains parked outside the abandoned library.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 먼지 쌓인 서가 사이에서 찰리가 두 눈의 푸른빛을 책 페이지 위에 눈부시게 비추는 열중한 모습.\n\nLOCATION (lock): Between dusty book stacks inside the abandoned library, where the robot's blue eye light falls brightly on an open page. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Open book (Open and being read intently) — The interior page faces are visible obliquely; no particular wording or illustration is specified; used as Shares focus with the inclined face and makes his reading action legible; Library shelves and books (Dust-covered) — Shelf fronts and side edges recede along the aisle beside 찰리; used as Provide depth brackets and a quiet enclosure around the reader.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: 찰리's described blue eye-light shines brightly onto the pages, providing a selective vivid accent while the surrounding library illumination remains restrained and unspecified.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dust-covered shelves and books surround Charlie as he reads; no mural has yet been established at this point. In one corner, boxes and books serve as makeshift bedding, and the camper remains parked outside the abandoned library.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, rainy.\n\nSHOT TEXT (authoritative, Korean): 먼지 쌓인 서가 사이에서 찰리가 두 눈의 푸른빛을 책 페이지 위에 눈부시게 비추는 열중한 모습.\n\nLOCATION (lock): Between dusty book stacks inside the abandoned library, where the robot's blue eye light falls brightly on an open page. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Open book (Open and being read intently) — The interior page faces are visible obliquely; no particular wording or illustration is specified; used as Shares focus with the inclined face and makes his reading action legible; Library shelves and books (Dust-covered) — Shelf fronts and side edges recede along the aisle beside 찰리; used as Provide depth brackets and a quiet enclosure around the reader.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: 찰리's described blue eye-light shines brightly onto the pages, providing a selective vivid accent while the surrounding library illumination remains restrained and unspecified.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dust-covered shelves and books surround Charlie as he reads; no mural has yet been established at this point. In one corner, boxes and books serve as makeshift bedding, and the camper remains parked outside the abandoned library.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선과 눈에서 나오는 파란 광선이 펼쳐진 책 페이지를 직접적으로 향하고 있음.",
    "built_space": "도서관 내부. 양측에 책이 꽂힌 서가가 늘어서 있고 배경의 창문으로 비 내리는 모습이 확인됨.",
    "entities": "고릴라형 기계 머리와 회색 망토를 두른 찰리의 외형이 일치하며, 텍스트가 뭉개진 펼쳐진 책이 있음.",
    "hard_violations": [],
    "physics": "찰리의 로봇 손이 아래에서 책을 단단히 쥐고 지탱함."
   },
   {
    "label": "B",
    "direction": "시선은 책을 향하나, 눈에서 발광하는 파란빛이 책 페이지를 전혀 비추지 않음.",
    "built_space": "도서관 내부. 화면 전면에 서가의 책들이 배치되어 그 틈새로 인물을 바라보는 구도임.",
    "entities": "찰리의 외형은 레퍼런스와 일치하나, 펼쳐진 책 페이지에 읽을 수 있는 세로쓰기 문자가 명확히 보임.",
    "hard_violations": [
     "[gemini-pro] 읽을 수 있는 문자 포함 (No readable writing 지침 위반)"
    ],
    "physics": "찰리의 두 로봇 손이 책 하단을 잡고 자세를 유지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 눈에서 뿜어져 나오는 파란빛이 책 페이지를 눈부시게 비추는 핵심 연출을 정확히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "책 페이지에 파란빛이 닿지 않으며, 금지된 식별 가능한 문자가 책에 나타나 지침을 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선과 눈에서 나오는 파란 광선이 펼쳐진 책 페이지를 직접적으로 향하고 있음.",
        "built_space": "도서관 내부. 양측에 책이 꽂힌 서가가 늘어서 있고 배경의 창문으로 비 내리는 모습이 확인됨.",
        "entities": "고릴라형 기계 머리와 회색 망토를 두른 찰리의 외형이 일치하며, 텍스트가 뭉개진 펼쳐진 책이 있음.",
        "hard_violations": [],
        "physics": "찰리의 로봇 손이 아래에서 책을 단단히 쥐고 지탱함."
       },
       {
        "label": "B",
        "direction": "시선은 책을 향하나, 눈에서 발광하는 파란빛이 책 페이지를 전혀 비추지 않음.",
        "built_space": "도서관 내부. 화면 전면에 서가의 책들이 배치되어 그 틈새로 인물을 바라보는 구도임.",
        "entities": "찰리의 외형은 레퍼런스와 일치하나, 펼쳐진 책 페이지에 읽을 수 있는 세로쓰기 문자가 명확히 보임.",
        "hard_violations": [
         "읽을 수 있는 문자 포함 (No readable writing 지침 위반)"
        ],
        "physics": "찰리의 두 로봇 손이 책 하단을 잡고 자세를 유지함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 눈에서 뿜어져 나오는 파란빛이 책 페이지를 눈부시게 비추는 핵심 연출을 정확히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "책 페이지에 파란빛이 닿지 않으며, 금지된 식별 가능한 문자가 책에 나타나 지침을 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선과 눈에서 나오는 파란 광선이 펼쳐진 책 페이지를 직접적으로 향하고 있음.",
        "built_space": "도서관 내부. 양측에 책이 꽂힌 서가가 늘어서 있고 배경의 창문으로 비 내리는 모습이 확인됨.",
        "entities": "고릴라형 기계 머리와 회색 망토를 두른 찰리의 외형이 일치하며, 텍스트가 뭉개진 펼쳐진 책이 있음.",
        "hard_violations": [],
        "physics": "찰리의 로봇 손이 아래에서 책을 단단히 쥐고 지탱함."
       },
       {
        "label": "B",
        "direction": "시선은 책을 향하나, 눈에서 발광하는 파란빛이 책 페이지를 전혀 비추지 않음.",
        "built_space": "도서관 내부. 화면 전면에 서가의 책들이 배치되어 그 틈새로 인물을 바라보는 구도임.",
        "entities": "찰리의 외형은 레퍼런스와 일치하나, 펼쳐진 책 페이지에 읽을 수 있는 세로쓰기 문자가 명확히 보임.",
        "hard_violations": [
         "읽을 수 있는 문자 포함 (No readable writing 지침 위반)"
        ],
        "physics": "찰리의 두 로봇 손이 책 하단을 잡고 자세를 유지함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "찰리의 외형과 서가 환경은 맞지만, 정면에 가까운 눈 방향과 표지 위주의 책 배치 때문에 페이지에 푸른 눈빛을 비추며 읽는 핵심 순간이 드러나지 않는다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "중간 숏 안에서 숙인 얼굴, 비스듬히 보이는 펼친 페이지, 그 위에 도달하는 강한 푸른 눈빛과 책을 받치는 기계 손을 가장 충실하게 구현했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 조금 숙였지만 두 눈의 발광면은 책보다 카메라 쪽을 향하는 인상이 강하다. 눈 자체는 파랗게 빛나지만 페이지에 도달하는 빛이나 밝은 청색 조명 영역은 확인되지 않는다. 책의 안쪽은 찰리 쪽을 향해 사용 방향은 맞으나, 카메라에는 주로 겉표지와 종이 가장자리만 보인다.",
        "built_space": "뒤쪽 좌우에 목재 서가 열이 하나씩 보이고 중앙 통로 끝에는 큰 창이 있다. 전경에는 위아래 선반과 양옆 책으로 둘러싸인 서가 한 칸이 추가로 보이며, 찰리는 그 너머 통로에 있다. 목재와 책, 차가운 야간 창빛은 장소 참조와 부합하지만, 전경 서가가 화면을 크게 가려 통로보다 책장 틈을 통한 구도가 강조된다. 불가능한 반사나 명백한 시설 중복은 없다.",
        "entities": "등장 주체는 기계형 찰리 한 명이다. 각진 회색 머리 외장, 흰 고릴라형 얼굴판, 작은 푸른 눈 두 개와 선형 입이 명시된 특징에 맞는다. 펼친 낡은 책 한 권, 관절이 있는 검은 기계 손 두 개, 먼지 낀 서가와 책들이 보인다. 손 참조에서는 손 자체가 잘 드러나지 않아 세부 일치까지 확인하기는 어렵다. 식별 가능한 문구나 다른 인물, 벽화는 보이지 않는다. 침구와 외부 캠퍼는 이 구도 밖이다.",
        "hard_violations": [],
        "physics": "책은 한 손이 아래쪽을 받치고 다른 손이 옆 가장자리를 잡아 지지한다. 손과 팔은 몸으로 이어지고 책의 무게를 받는 관계도 자연스럽다. 전경 책들은 아래 선반에 놓여 있다. 하체는 화면 밖이며 공중에 떠 있는 신체나 지지 없는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "찰리의 머리와 두 눈은 아래쪽 펼친 책을 향한다. 화면 왼쪽 눈에서 특히 뚜렷한 청색 광선이 내려가 왼쪽 페이지와 책 중앙 접힘 부근의 밝은 청색 영역에 도달한다. 다른 눈도 파랗게 발광하지만 별도의 광선 경로는 덜 분명하다. 책 안쪽은 찰리의 눈을 향하면서 카메라에도 비스듬히 보여 독서 방향이 성립한다.",
        "built_space": "왼쪽에는 가까운 목재 서가 열 하나가 통로를 따라 후퇴하고, 오른쪽에도 책이 채워진 서가 열 하나가 보인다. 찰리 뒤에는 서가 끝판 하나가 드러나며 통로 끝에는 큰 격자 창이 있다. 찰리는 두 서가 사이에 서서 책을 들고 있다. 낡은 목재, 조밀한 책, 어두운 창빛과 창의 빗물 흔적이 참조의 도서관 및 비 오는 밤 설정에 부합한다. 시설의 불가능한 중복이나 반사는 보이지 않는다.",
        "entities": "찰리 한 명만 등장하며 각진 금속 머리, 흰 고릴라형 얼굴판, 작은 푸른 눈 두 개와 가는 입이 요구에 맞는다. 검은 관절형 기계 손들이 낡은 펼친 책을 받치고, 먼지와 마모가 있는 서가와 책들이 주변을 채운다. 페이지에는 인쇄 줄 같은 흔적이 있지만 확실히 읽을 수 있는 문구는 식별되지 않는다. 다른 인물이나 벽화는 없고, 침구와 외부 캠퍼는 중간 숏의 범위 밖이다.",
        "hard_violations": [],
        "physics": "두 기계 손과 팔이 책 밑면을 받쳐 무게를 지지한다. 위로 들린 종이 한 장은 책의 제본부에 연결되어 있어 독립적으로 떠 있는 물체가 아니다. 몸은 직립한 채 고개를 숙이고 있으며 하체는 화면 밖이다. 푸른빛은 눈에서 책 쪽으로 이어지고 먼지가 있는 공기 중에서 드러나는 형태로 표현되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "찰리의 외형과 서가 환경은 맞지만, 정면에 가까운 눈 방향과 표지 위주의 책 배치 때문에 페이지에 푸른 눈빛을 비추며 읽는 핵심 순간이 드러나지 않는다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "중간 숏 안에서 숙인 얼굴, 비스듬히 보이는 펼친 페이지, 그 위에 도달하는 강한 푸른 눈빛과 책을 받치는 기계 손을 가장 충실하게 구현했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 조금 숙였지만 두 눈의 발광면은 책보다 카메라 쪽을 향하는 인상이 강하다. 눈 자체는 파랗게 빛나지만 페이지에 도달하는 빛이나 밝은 청색 조명 영역은 확인되지 않는다. 책의 안쪽은 찰리 쪽을 향해 사용 방향은 맞으나, 카메라에는 주로 겉표지와 종이 가장자리만 보인다.",
        "built_space": "뒤쪽 좌우에 목재 서가 열이 하나씩 보이고 중앙 통로 끝에는 큰 창이 있다. 전경에는 위아래 선반과 양옆 책으로 둘러싸인 서가 한 칸이 추가로 보이며, 찰리는 그 너머 통로에 있다. 목재와 책, 차가운 야간 창빛은 장소 참조와 부합하지만, 전경 서가가 화면을 크게 가려 통로보다 책장 틈을 통한 구도가 강조된다. 불가능한 반사나 명백한 시설 중복은 없다.",
        "entities": "등장 주체는 기계형 찰리 한 명이다. 각진 회색 머리 외장, 흰 고릴라형 얼굴판, 작은 푸른 눈 두 개와 선형 입이 명시된 특징에 맞는다. 펼친 낡은 책 한 권, 관절이 있는 검은 기계 손 두 개, 먼지 낀 서가와 책들이 보인다. 손 참조에서는 손 자체가 잘 드러나지 않아 세부 일치까지 확인하기는 어렵다. 식별 가능한 문구나 다른 인물, 벽화는 보이지 않는다. 침구와 외부 캠퍼는 이 구도 밖이다.",
        "hard_violations": [],
        "physics": "책은 한 손이 아래쪽을 받치고 다른 손이 옆 가장자리를 잡아 지지한다. 손과 팔은 몸으로 이어지고 책의 무게를 받는 관계도 자연스럽다. 전경 책들은 아래 선반에 놓여 있다. 하체는 화면 밖이며 공중에 떠 있는 신체나 지지 없는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "찰리의 머리와 두 눈은 아래쪽 펼친 책을 향한다. 화면 왼쪽 눈에서 특히 뚜렷한 청색 광선이 내려가 왼쪽 페이지와 책 중앙 접힘 부근의 밝은 청색 영역에 도달한다. 다른 눈도 파랗게 발광하지만 별도의 광선 경로는 덜 분명하다. 책 안쪽은 찰리의 눈을 향하면서 카메라에도 비스듬히 보여 독서 방향이 성립한다.",
        "built_space": "왼쪽에는 가까운 목재 서가 열 하나가 통로를 따라 후퇴하고, 오른쪽에도 책이 채워진 서가 열 하나가 보인다. 찰리 뒤에는 서가 끝판 하나가 드러나며 통로 끝에는 큰 격자 창이 있다. 찰리는 두 서가 사이에 서서 책을 들고 있다. 낡은 목재, 조밀한 책, 어두운 창빛과 창의 빗물 흔적이 참조의 도서관 및 비 오는 밤 설정에 부합한다. 시설의 불가능한 중복이나 반사는 보이지 않는다.",
        "entities": "찰리 한 명만 등장하며 각진 금속 머리, 흰 고릴라형 얼굴판, 작은 푸른 눈 두 개와 가는 입이 요구에 맞는다. 검은 관절형 기계 손들이 낡은 펼친 책을 받치고, 먼지와 마모가 있는 서가와 책들이 주변을 채운다. 페이지에는 인쇄 줄 같은 흔적이 있지만 확실히 읽을 수 있는 문구는 식별되지 않는다. 다른 인물이나 벽화는 없고, 침구와 외부 캠퍼는 중간 숏의 범위 밖이다.",
        "hard_violations": [],
        "physics": "두 기계 손과 팔이 책 밑면을 받쳐 무게를 지지한다. 위로 들린 종이 한 장은 책의 제본부에 연결되어 있어 독립적으로 떠 있는 물체가 아니다. 몸은 직립한 채 고개를 숙이고 있으며 하체는 화면 밖이다. 푸른빛은 눈에서 책 쪽으로 이어지고 먼지가 있는 공기 중에서 드러나는 형태로 표현되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.984
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.734
   },
   "violations": {
    "B": [
     "[gemini-pro] 읽을 수 있는 문자 포함 (No readable writing 지침 위반)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 734
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "두 눈에서 뿜어져 나오는 파란빛이 책 페이지를 눈부시게 비추는 핵심 연출을 정확히 구현함."
   },
   {
    "label": "B",
    "score": 734,
    "verdict_ko": "책 페이지에 파란빛이 닿지 않으며, 금지된 식별 가능한 문자가 책에 나타나 지침을 위반함.  ★위반: [gemini-pro] 읽을 수 있는 문자 포함 (No readable writing 지침 위반)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S46sh21_sel.png",
    "asset_id": "b5911fae-05c8-4dbc-9de9-b14aff9057cd",
    "role": "prev_still"
   },
   {
    "label": "HAND OWNER REFERENCE — 찰리: the person whose hand is on the object in this shot. Use this photograph ONLY to get that hand right — its size, build, skin tone and texture, apparent age, grooming, nails, sleeve and anything worn on the wrist or fingers. Do NOT bring their face, body or clothing into the frame; the framing stays on the object and reaches no further than the wrist and forearm.",
    "path": "<bytes:948699>",
    "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd0-fc92-7f5d-8af7-8cf7f52a39aa",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S46sh21"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S46sh26::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:41:19.799011+00:00",
  "fingerprint": "284101f2c2b0aa182acee3c8920eecc9e37d9eb66e75654617f8b60051258df2",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S46sh26_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S46sh26_sel.png",
  "source_sha256": "ef8fd0cc10243e05d3b645629effd553f2cff11d2ae40a865c6c1ec2c861d7da",
  "file": "S46sh26_cine.png",
  "staged_sha256": "ccb58337b4fb896a745c628288781884b44b870ab035affb9140a466384e642a",
  "latency_ms": 11343
 },
 "S47sh5::signage": {
  "fp": "7341830a8e600e41",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S47sh5": {
  "input_fingerprint": "2657ccac9cebc1d2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 찰리가 벽면 한쪽에 웅크려 앉아 파란색 크레파스로 선을 긋는 도중의 뒷모습.\n\nLOCATION (lock): Along a mural-covered wall deeper inside the abandoned library, in early-morning ambient daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Illustrated library wall (Covered with crayon drawings, with a blue line being added) — The drawing-bearing face is visible obliquely beside 찰리, showing a portion of the sea illustration; used as Provides the destination of the drawing gesture and a narrow plane of color beside his back; Blue crayon (Held against the wall during a drawing stroke) — Seen obliquely along the drawing hand, with its tip contacting the wall; used as Small focal detail linking the hand to the newly drawn line.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the early-morning library preserves the blue crayon's color and quiet tonal separation without introducing a visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dusty library walls carry a blue-crayon sea and a colorful Haenam mural depicting Charlie, Hyunwoo, Raul and Amber arriving by camping truck, with Pedro, Miyeon and the priest among marine life and palm trees. Charlie continues drawing with crayons; the old camper remains parked outside the cracked, vine-covered library.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 찰리가 벽면 한쪽에 웅크려 앉아 파란색 크레파스로 선을 긋는 도중의 뒷모습.\n\nLOCATION (lock): Along a mural-covered wall deeper inside the abandoned library, in early-morning ambient daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Illustrated library wall (Covered with crayon drawings, with a blue line being added) — The drawing-bearing face is visible obliquely beside 찰리, showing a portion of the sea illustration; used as Provides the destination of the drawing gesture and a narrow plane of color beside his back; Blue crayon (Held against the wall during a drawing stroke) — Seen obliquely along the drawing hand, with its tip contacting the wall; used as Small focal detail linking the hand to the newly drawn line.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the early-morning library preserves the blue crayon's color and quiet tonal separation without introducing a visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dusty library walls carry a blue-crayon sea and a colorful Haenam mural depicting Charlie, Hyunwoo, Raul and Amber arriving by camping truck, with Pedro, Miyeon and the priest among marine life and palm trees. Charlie continues drawing with crayons; the old camper remains parked outside the cracked, vine-covered library.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 찰리가 벽면 한쪽에 웅크려 앉아 파란색 크레파스로 선을 긋는 도중의 뒷모습.\n\nLOCATION (lock): Along a mural-covered wall deeper inside the abandoned library, in early-morning ambient daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Illustrated library wall (Covered with crayon drawings, with a blue line being added) — The drawing-bearing face is visible obliquely beside 찰리, showing a portion of the sea illustration; used as Provides the destination of the drawing gesture and a narrow plane of color beside his back; Blue crayon (Held against the wall during a drawing stroke) — Seen obliquely along the drawing hand, with its tip contacting the wall; used as Small focal detail linking the hand to the newly drawn line.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the early-morning library preserves the blue crayon's color and quiet tonal separation without introducing a visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dusty library walls carry a blue-crayon sea and a colorful Haenam mural depicting Charlie, Hyunwoo, Raul and Amber arriving by camping truck, with Pedro, Miyeon and the priest among marine life and palm trees. Charlie continues drawing with crayons; the old camper remains parked outside the cracked, vine-covered library.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 시선과 파란색 크레파스를 쥔 손이 벽면의 벽화(고래 위쪽 파도)를 향하고 있습니다.",
    "built_space": "제공된 로케이션 레퍼런스와 일치합니다. 벽화가 있는 긴 벽면과 우측으로 꺾이는 복도, 그리고 그 끝의 창문이 올바른 위치에 배치되어 있습니다.",
    "entities": "찰리의 기계 머리 뒷모습, 어두운 망토, 로봇 팔이 레퍼런스와 일치하며, 손에 쥔 파란색 크레파스와 지정된 벽화 요소들이 정확히 묘사되었습니다.",
    "hard_violations": [],
    "physics": "웅크린 자세로 망토가 바닥을 향해 자연스럽게 늘어져 있으며, 들어 올린 팔과 벽에 닿은 크레파스가 안정적으로 지탱되고 있습니다."
   },
   {
    "label": "B",
    "direction": "크레파스를 쥔 로봇 손이 벽면의 파도 그림을 향해 선을 긋고 있습니다.",
    "built_space": "레퍼런스와 상이합니다. 벽화가 있는 벽이 창문으로 곧게 이어지는 복도의 우측에 배치되어 원래의 '우측으로 꺾이는' 공간 구조를 완전히 무시했습니다.",
    "entities": "망토를 두른 찰리와 기계 팔, 파란색 크레파스는 확인되나, 벽화의 디테일과 공간 배치가 레퍼런스와 다릅니다.",
    "hard_violations": [
     "[gemini-pro] 지정된 로케이션 레퍼런스의 건축적 공간 구조(복도의 방향과 창문의 위치)를 전혀 다르게 렌더링한 심각한 공간 위반"
    ],
    "physics": "바닥에 웅크린 자세와 팔의 움직임, 손에 쥐어진 크레파스의 접촉은 물리적으로 기능합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 로케이션의 공간 구조와 벽화를 정확하게 반영하며 찰리의 뒷모습과 작화 동작을 훌륭하게 연출했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 외형과 동작은 무난하나, 고정된 로케이션 레퍼런스의 핵심적인 공간 구조를 완전히 다르게 왜곡했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선과 파란색 크레파스를 쥔 손이 벽면의 벽화(고래 위쪽 파도)를 향하고 있습니다.",
        "built_space": "제공된 로케이션 레퍼런스와 일치합니다. 벽화가 있는 긴 벽면과 우측으로 꺾이는 복도, 그리고 그 끝의 창문이 올바른 위치에 배치되어 있습니다.",
        "entities": "찰리의 기계 머리 뒷모습, 어두운 망토, 로봇 팔이 레퍼런스와 일치하며, 손에 쥔 파란색 크레파스와 지정된 벽화 요소들이 정확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "웅크린 자세로 망토가 바닥을 향해 자연스럽게 늘어져 있으며, 들어 올린 팔과 벽에 닿은 크레파스가 안정적으로 지탱되고 있습니다."
       },
       {
        "label": "B",
        "direction": "크레파스를 쥔 로봇 손이 벽면의 파도 그림을 향해 선을 긋고 있습니다.",
        "built_space": "레퍼런스와 상이합니다. 벽화가 있는 벽이 창문으로 곧게 이어지는 복도의 우측에 배치되어 원래의 '우측으로 꺾이는' 공간 구조를 완전히 무시했습니다.",
        "entities": "망토를 두른 찰리와 기계 팔, 파란색 크레파스는 확인되나, 벽화의 디테일과 공간 배치가 레퍼런스와 다릅니다.",
        "hard_violations": [
         "지정된 로케이션 레퍼런스의 건축적 공간 구조(복도의 방향과 창문의 위치)를 전혀 다르게 렌더링한 심각한 공간 위반"
        ],
        "physics": "바닥에 웅크린 자세와 팔의 움직임, 손에 쥐어진 크레파스의 접촉은 물리적으로 기능합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 로케이션의 공간 구조와 벽화를 정확하게 반영하며 찰리의 뒷모습과 작화 동작을 훌륭하게 연출했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 외형과 동작은 무난하나, 고정된 로케이션 레퍼런스의 핵심적인 공간 구조를 완전히 다르게 왜곡했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선과 파란색 크레파스를 쥔 손이 벽면의 벽화(고래 위쪽 파도)를 향하고 있습니다.",
        "built_space": "제공된 로케이션 레퍼런스와 일치합니다. 벽화가 있는 긴 벽면과 우측으로 꺾이는 복도, 그리고 그 끝의 창문이 올바른 위치에 배치되어 있습니다.",
        "entities": "찰리의 기계 머리 뒷모습, 어두운 망토, 로봇 팔이 레퍼런스와 일치하며, 손에 쥔 파란색 크레파스와 지정된 벽화 요소들이 정확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "웅크린 자세로 망토가 바닥을 향해 자연스럽게 늘어져 있으며, 들어 올린 팔과 벽에 닿은 크레파스가 안정적으로 지탱되고 있습니다."
       },
       {
        "label": "B",
        "direction": "크레파스를 쥔 로봇 손이 벽면의 파도 그림을 향해 선을 긋고 있습니다.",
        "built_space": "레퍼런스와 상이합니다. 벽화가 있는 벽이 창문으로 곧게 이어지는 복도의 우측에 배치되어 원래의 '우측으로 꺾이는' 공간 구조를 완전히 무시했습니다.",
        "entities": "망토를 두른 찰리와 기계 팔, 파란색 크레파스는 확인되나, 벽화의 디테일과 공간 배치가 레퍼런스와 다릅니다.",
        "hard_violations": [
         "지정된 로케이션 레퍼런스의 건축적 공간 구조(복도의 방향과 창문의 위치)를 전혀 다르게 렌더링한 심각한 공간 위반"
        ],
        "physics": "바닥에 웅크린 자세와 팔의 움직임, 손에 쥐어진 크레파스의 접촉은 물리적으로 기능합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "등과 그리는 손을 담은 미디엄 숏 및 크레파스의 벽 접촉은 충족하지만, 벽화가 넓게 펼쳐지고 배경 공간이 참조 장소와 덜 일치한다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "찰리의 뒷모습 옆에 비스듬한 벽화 면과 크레파스 접점을 배치하고, 참조의 벽 끝·오른쪽 개구부·창문 관계까지 더 충실히 유지한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 머리와 오른손이 오른쪽 벽화를 향한다. 눈은 보이지 않아 정확한 시선은 확인할 수 없다. 손에 쥔 파란 크레파스는 오른쪽 위로 향하며, 끝이 돌고래 위쪽의 파란 물결 부근에 닿아 있다.",
        "built_space": "오른쪽에 긴 벽화 벽과 하단 걸레받이, 왼쪽 뒤에 복도와 이분할 창문 하나, 선반 하나, 수납 설비가 보인다. 참조와 유사한 낡은 벽·바닥 재질이지만, 참조의 벽 끝과 그 오른쪽 개구부 관계는 확인되지 않는다. 벽화 면이 화면 오른쪽을 넓게 차지해 요청한 좁은 색면보다 펼쳐져 보인다. 반사는 없다.",
        "entities": "실물 인물은 찰리 한 명뿐이다. 금속제 기계 머리, 일부 드러난 흰 얼굴판, 관절형 금속 손과 전완, 낡은 회색 천이 보인다. 머리 외장은 설명보다 둥근 편이며, 뒷모습이라 점 눈과 선 입은 확인할 수 없다. 파란 크레파스와 바다·고래·돌고래·배·캠핑 차량 벽화가 보인다. 벽화 속 특정 인물들의 정체는 판별되지 않으며, 실외 캠핑 차량은 프레임 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "크레파스는 오른손 손가락에 잡혀 있고 끝은 벽 표면에 닿아 있어 그리는 동작이 가능하다. 팔은 몸에 자연스럽게 연결되고 천은 아래로 처진다. 하체와 바닥 접점은 프레임 밖이므로 웅크려 앉은 다리 자세까지 확인되지는 않지만, 공중에 떠 있다는 단서는 없다."
       },
       {
        "label": "B",
        "direction": "찰리의 머리는 오른쪽 앞의 벽과 그리는 손 쪽으로 돌아가 있다. 눈 자체는 보이지 않는다. 오른손의 파란 크레파스 끝은 야자수 아래, 고래 위쪽의 파란 물결선에 닿으며 벽을 향한다. 새 선과 기존 물결선의 구분은 뚜렷하지 않다.",
        "built_space": "벽화 벽의 끝 모서리 하나, 그 오른쪽 큰 개구부 하나, 안쪽 출입구 하나와 이분할 창문 하나가 보인다. 오른쪽 뒤에는 참조와 유사한 선반·상자·청소용품이 있고 바닥 매트 하나도 보인다. 회색 하부 벽면과 검은 걸레받이, 마모된 바닥 및 고래·야자수의 위치가 참조 공간과 잘 맞는다. 찰리는 벽화 면 앞에 있고 손이 모서리 안쪽의 그림에 도달한다. 불가능한 반사는 없다.",
        "entities": "찰리 한 명의 등과 기계 머리, 흰 얼굴판 일부, 금속제 오른손과 전완이 보인다. 머리는 설명보다 둥글지만 기계적 정체성은 유지하며 눈과 입은 가려져 있다. 손은 몸과 같은 금속 구조로 이어진다. 파란 크레파스, 바다와 고래, 야자수, 캠핑 차량 및 작은 인물 형태의 벽화가 보인다. 벽화 인물 각각의 이름까지는 확인할 수 없다. 실외 차량은 이 구도에서 보이지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "오른손이 크레파스를 잡고 손목과 팔꿈치가 벽으로 뻗어 있어 실제로 선을 그을 수 있는 자세다. 크레파스 끝은 벽에 접촉한다. 몸은 낮아진 자세로 보이고 천은 등을 따라 아래로 늘어진다. 하체 지지점은 화면 밖이므로 정확한 쪼그림은 확인할 수 없지만, 무지지 부유나 불가능한 관절 배치는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "등과 그리는 손을 담은 미디엄 숏 및 크레파스의 벽 접촉은 충족하지만, 벽화가 넓게 펼쳐지고 배경 공간이 참조 장소와 덜 일치한다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "찰리의 뒷모습 옆에 비스듬한 벽화 면과 크레파스 접점을 배치하고, 참조의 벽 끝·오른쪽 개구부·창문 관계까지 더 충실히 유지한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 머리와 오른손이 오른쪽 벽화를 향한다. 눈은 보이지 않아 정확한 시선은 확인할 수 없다. 손에 쥔 파란 크레파스는 오른쪽 위로 향하며, 끝이 돌고래 위쪽의 파란 물결 부근에 닿아 있다.",
        "built_space": "오른쪽에 긴 벽화 벽과 하단 걸레받이, 왼쪽 뒤에 복도와 이분할 창문 하나, 선반 하나, 수납 설비가 보인다. 참조와 유사한 낡은 벽·바닥 재질이지만, 참조의 벽 끝과 그 오른쪽 개구부 관계는 확인되지 않는다. 벽화 면이 화면 오른쪽을 넓게 차지해 요청한 좁은 색면보다 펼쳐져 보인다. 반사는 없다.",
        "entities": "실물 인물은 찰리 한 명뿐이다. 금속제 기계 머리, 일부 드러난 흰 얼굴판, 관절형 금속 손과 전완, 낡은 회색 천이 보인다. 머리 외장은 설명보다 둥근 편이며, 뒷모습이라 점 눈과 선 입은 확인할 수 없다. 파란 크레파스와 바다·고래·돌고래·배·캠핑 차량 벽화가 보인다. 벽화 속 특정 인물들의 정체는 판별되지 않으며, 실외 캠핑 차량은 프레임 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "크레파스는 오른손 손가락에 잡혀 있고 끝은 벽 표면에 닿아 있어 그리는 동작이 가능하다. 팔은 몸에 자연스럽게 연결되고 천은 아래로 처진다. 하체와 바닥 접점은 프레임 밖이므로 웅크려 앉은 다리 자세까지 확인되지는 않지만, 공중에 떠 있다는 단서는 없다."
       },
       {
        "label": "A",
        "direction": "찰리의 머리는 오른쪽 앞의 벽과 그리는 손 쪽으로 돌아가 있다. 눈 자체는 보이지 않는다. 오른손의 파란 크레파스 끝은 야자수 아래, 고래 위쪽의 파란 물결선에 닿으며 벽을 향한다. 새 선과 기존 물결선의 구분은 뚜렷하지 않다.",
        "built_space": "벽화 벽의 끝 모서리 하나, 그 오른쪽 큰 개구부 하나, 안쪽 출입구 하나와 이분할 창문 하나가 보인다. 오른쪽 뒤에는 참조와 유사한 선반·상자·청소용품이 있고 바닥 매트 하나도 보인다. 회색 하부 벽면과 검은 걸레받이, 마모된 바닥 및 고래·야자수의 위치가 참조 공간과 잘 맞는다. 찰리는 벽화 면 앞에 있고 손이 모서리 안쪽의 그림에 도달한다. 불가능한 반사는 없다.",
        "entities": "찰리 한 명의 등과 기계 머리, 흰 얼굴판 일부, 금속제 오른손과 전완이 보인다. 머리는 설명보다 둥글지만 기계적 정체성은 유지하며 눈과 입은 가려져 있다. 손은 몸과 같은 금속 구조로 이어진다. 파란 크레파스, 바다와 고래, 야자수, 캠핑 차량 및 작은 인물 형태의 벽화가 보인다. 벽화 인물 각각의 이름까지는 확인할 수 없다. 실외 차량은 이 구도에서 보이지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "오른손이 크레파스를 잡고 손목과 팔꿈치가 벽으로 뻗어 있어 실제로 선을 그을 수 있는 자세다. 크레파스 끝은 벽에 접촉한다. 몸은 낮아진 자세로 보이고 천은 등을 따라 아래로 늘어진다. 하체 지지점은 화면 밖이므로 정확한 쪼그림은 확인할 수 없지만, 무지지 부유나 불가능한 관절 배치는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.206
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.956
   },
   "violations": {
    "B": [
     "[gemini-pro] 지정된 로케이션 레퍼런스의 건축적 공간 구조(복도의 방향과 창문의 위치)를 전혀 다르게 렌더링한 심각한 공간 위반"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 956
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 로케이션의 공간 구조와 벽화를 정확하게 반영하며 찰리의 뒷모습과 작화 동작을 훌륭하게 연출했습니다."
   },
   {
    "label": "B",
    "score": 956,
    "verdict_ko": "찰리의 외형과 동작은 무난하나, 고정된 로케이션 레퍼런스의 핵심적인 공간 구조를 완전히 다르게 왜곡했습니다.  ★위반: [gemini-pro] 지정된 로케이션 레퍼런스의 건축적 공간 구조(복도의 방향과 창문의 위치)를 전혀 다르게 렌더링한 심각한 공간 위반"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L58B01.png",
    "asset_id": "e381ac75-3b70-4df7-83a9-f70bc3b4868a",
    "role": "location_plate"
   },
   {
    "label": "HAND OWNER REFERENCE — 찰리: the person whose hand is on the object in this shot. Use this photograph ONLY to get that hand right — its size, build, skin tone and texture, apparent age, grooming, nails, sleeve and anything worn on the wrist or fingers. Do NOT bring their face, body or clothing into the frame; the framing stays on the object and reaches no further than the wrist and forearm.",
    "path": "<bytes:948699>",
    "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd1-b7a1-74ce-a0b5-4efb49575084",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S47sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:41:31.489193+00:00",
  "fingerprint": "3dac17f3efeb29d4710dc240bbfbad9a8a406405a3870535f2afb26b8c19a396",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S47sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S47sh5_sel.png",
  "source_sha256": "3169d986406f9e2fdec59cffeee98e9a8480975e777e3f6e06133993b5dfe705",
  "file": "S47sh5_cine.png",
  "staged_sha256": "0b67752bb20f8278d0e95e7d79203147da51da1806ee7338fec70a1b5c65fa4c",
  "latency_ms": 29476
 },
 "S47sh15::signage": {
  "fp": "3ac0171e1c2dde7f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::7a115c4c1ea76edc": {
  "subjects": [],
  "subject_text": "미추홀도서관 복도와 외진 구석\n오래 방치된 도서관 안쪽으로 이어지는 복도. 먼지가 쌓인 바닥과 낡은 벽면이 이어지며 끝에는 외진 구석 공간이 있다.",
  "identity": "canonical",
  "scope_id": "L58",
  "scope_role": "location_interior",
  "scope_sha": "f7e2dc3c75a4468a"
 },
 "S47sh15::bgfirst_bg": {
  "input_fingerprint": "d576f01eac3ad818",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 일회용 핸드폰을 귓가에 바짝 댄 채 미간을 깊게 찌푸린 초조한 얼굴.\n\nLOCATION (lock): In a secluded outdoor corner beside the abandoned library, away from the others in the morning.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Secluded library corner (Separated from the gathering elsewhere in the library) — The corner is seen obliquely behind his phone-side shoulder; used as A softly resolved spatial boundary reinforces the private call without adding furnishings; Disposable phone (Pressed tightly against his ear during the call) — Its outer edge is visible beside his cheek; the screen is not presented to the camera; used as Connects his facial tension to the ongoing conversation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the library keeps the furrowed brow readable while leaving the secluded background subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 일회용 핸드폰을 귓가에 바짝 댄 채 미간을 깊게 찌푸린 초조한 얼굴.\n\nLOCATION (lock): In a secluded outdoor corner beside the abandoned library, away from the others in the morning.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Secluded library corner (Separated from the gathering elsewhere in the library) — The corner is seen obliquely behind his phone-side shoulder; used as A softly resolved spatial boundary reinforces the private call without adding furnishings; Disposable phone (Pressed tightly against his ear during the call) — Its outer edge is visible beside his cheek; the screen is not presented to the camera; used as Connects his facial tension to the ongoing conversation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the library keeps the furrowed brow readable while leaving the secluded background subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S47sh15__bgfirst_bg.png",
  "asset_id": "88792161-b554-44d6-ba47-854436f1f5ca",
  "input_asset_ids": [
   "ef232a1d-7546-491d-aa82-56013414b27b",
   "e381ac75-3b70-4df7-83a9-f70bc3b4868a"
  ]
 },
 "S47sh15": {
  "input_fingerprint": "f162bed058352310",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 일회용 핸드폰을 귓가에 바짝 댄 채 미간을 깊게 찌푸린 초조한 얼굴.\n\nLOCATION (lock): In a secluded outdoor corner beside the abandoned library, away from the others in the morning. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Secluded library corner (Separated from the gathering elsewhere in the library) — The corner is seen obliquely behind his phone-side shoulder; used as A softly resolved spatial boundary reinforces the private call without adding furnishings; Disposable phone (Pressed tightly against his ear during the call) — Its outer edge is visible beside his cheek; the screen is not presented to the camera; used as Connects his facial tension to the ongoing conversation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the library keeps the furrowed brow readable while leaving the secluded background subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crayon sea, colorful Haenam mural and handwritten message remain on the dusty library walls, and the old camper remains outside. Charlie has left the drawing area and is approaching the secluded corner. 이현우: He stands in a secluded library corner using the disposable phone, with the intact contact card now removed from his shoe. His facial bruising and untreated leg injury persist.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 일회용 핸드폰을 귓가에 바짝 댄 채 미간을 깊게 찌푸린 초조한 얼굴.\n\nLOCATION (lock): In a secluded outdoor corner beside the abandoned library, away from the others in the morning. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Secluded library corner (Separated from the gathering elsewhere in the library) — The corner is seen obliquely behind his phone-side shoulder; used as A softly resolved spatial boundary reinforces the private call without adding furnishings; Disposable phone (Pressed tightly against his ear during the call) — Its outer edge is visible beside his cheek; the screen is not presented to the camera; used as Connects his facial tension to the ongoing conversation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the library keeps the furrowed brow readable while leaving the secluded background subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crayon sea, colorful Haenam mural and handwritten message remain on the dusty library walls, and the old camper remains outside. Charlie has left the drawing area and is approaching the secluded corner. 이현우: He stands in a secluded library corner using the disposable phone, with the intact contact card now removed from his shoe. His facial bruising and untreated leg injury persist.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 일회용 핸드폰을 귓가에 바짝 댄 채 미간을 깊게 찌푸린 초조한 얼굴.\n\nLOCATION (lock): In a secluded outdoor corner beside the abandoned library, away from the others in the morning. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Secluded library corner (Separated from the gathering elsewhere in the library) — The corner is seen obliquely behind his phone-side shoulder; used as A softly resolved spatial boundary reinforces the private call without adding furnishings; Disposable phone (Pressed tightly against his ear during the call) — Its outer edge is visible beside his cheek; the screen is not presented to the camera; used as Connects his facial tension to the ongoing conversation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the library keeps the furrowed brow readable while leaving the secluded background subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crayon sea, colorful Haenam mural and handwritten message remain on the dusty library walls, and the old camper remains outside. Charlie has left the drawing area and is approaching the secluded corner. 이현우: He stands in a secluded library corner using the disposable phone, with the intact contact card now removed from his shoe. His facial bruising and untreated leg injury persist.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S47sh15__bgfirst_bg.png",
     "asset_id": "88792161-b554-44d6-ba47-854436f1f5ca",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S47sh15.png",
     "asset_id": "ef232a1d-7546-491d-aa82-56013414b27b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L58B01.png",
     "asset_id": "e381ac75-3b70-4df7-83a9-f70bc3b4868a",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 약간 왼쪽 아래를 향하고 있으며, 전화기는 오른쪽 귀에 밀착되어 있음.",
    "built_space": "실내 복도 공간. 오른쪽 벽면에는 레퍼런스와 일치하는 벽화(고래, 파도, 야자수 등)가 그려져 있고, 복도 끝에는 창문이 보임.",
    "entities": "이현우(레퍼런스와 외모 및 복장이 일치하나 얼굴의 멍이 없음), 일회용 핸드폰(스마트폰 형태, 화면이 귀 쪽을 향해 올바르게 들려 있음).",
    "hard_violations": [],
    "physics": "서 있는 자세이며, 오른손이 전화기를 쥐고 귀에 안정적으로 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "시선은 아래를 향하고 있으며, 전화기를 왼쪽 귀에 대고 있음.",
    "built_space": "건물 외벽 옆의 야외 공간. 배경에 캠핑카와 야외 구조물들이 보이며, 레퍼런스의 실내 장소와 완전히 불일치함.",
    "entities": "이현우(레퍼런스와 일치하며 얼굴에 찰과상과 멍이 있음), 핸드폰(화면이 바깥쪽 카메라를 향하고 있음).",
    "hard_violations": [
     "[gemini-pro] 전화기의 화면(기능면)이 사용자의 귀가 아닌 바깥쪽(카메라)을 향하도록 거꾸로 들고 있음 (물리적/기능적 오류)."
    ],
    "physics": "서 있는 자세로 왼손이 전화기를 지탱하고 있으나, 기기를 앞뒤가 뒤집힌 비정상적인 형태로 쥐고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "얼굴의 멍과 상처가 누락되었으나, 지정된 장소 레퍼런스(실내 벽화 앞)를 정확히 반영하고 소품(전화기)을 올바른 방향으로 파지하여 프롬프트를 가장 잘 충족함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 얼굴 상처는 잘 묘사되었으나, 지정된 실내 복도가 아닌 야외를 배경으로 삼았으며, 전화기 화면이 바깥을 향하게 거꾸로 들고 있는 치명적인 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 약간 왼쪽 아래를 향하고 있으며, 전화기는 오른쪽 귀에 밀착되어 있음.",
        "built_space": "실내 복도 공간. 오른쪽 벽면에는 레퍼런스와 일치하는 벽화(고래, 파도, 야자수 등)가 그려져 있고, 복도 끝에는 창문이 보임.",
        "entities": "이현우(레퍼런스와 외모 및 복장이 일치하나 얼굴의 멍이 없음), 일회용 핸드폰(스마트폰 형태, 화면이 귀 쪽을 향해 올바르게 들려 있음).",
        "hard_violations": [],
        "physics": "서 있는 자세이며, 오른손이 전화기를 쥐고 귀에 안정적으로 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "시선은 아래를 향하고 있으며, 전화기를 왼쪽 귀에 대고 있음.",
        "built_space": "건물 외벽 옆의 야외 공간. 배경에 캠핑카와 야외 구조물들이 보이며, 레퍼런스의 실내 장소와 완전히 불일치함.",
        "entities": "이현우(레퍼런스와 일치하며 얼굴에 찰과상과 멍이 있음), 핸드폰(화면이 바깥쪽 카메라를 향하고 있음).",
        "hard_violations": [
         "전화기의 화면(기능면)이 사용자의 귀가 아닌 바깥쪽(카메라)을 향하도록 거꾸로 들고 있음 (물리적/기능적 오류)."
        ],
        "physics": "서 있는 자세로 왼손이 전화기를 지탱하고 있으나, 기기를 앞뒤가 뒤집힌 비정상적인 형태로 쥐고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "얼굴의 멍과 상처가 누락되었으나, 지정된 장소 레퍼런스(실내 벽화 앞)를 정확히 반영하고 소품(전화기)을 올바른 방향으로 파지하여 프롬프트를 가장 잘 충족함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 얼굴 상처는 잘 묘사되었으나, 지정된 실내 복도가 아닌 야외를 배경으로 삼았으며, 전화기 화면이 바깥을 향하게 거꾸로 들고 있는 치명적인 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 약간 왼쪽 아래를 향하고 있으며, 전화기는 오른쪽 귀에 밀착되어 있음.",
        "built_space": "실내 복도 공간. 오른쪽 벽면에는 레퍼런스와 일치하는 벽화(고래, 파도, 야자수 등)가 그려져 있고, 복도 끝에는 창문이 보임.",
        "entities": "이현우(레퍼런스와 외모 및 복장이 일치하나 얼굴의 멍이 없음), 일회용 핸드폰(스마트폰 형태, 화면이 귀 쪽을 향해 올바르게 들려 있음).",
        "hard_violations": [],
        "physics": "서 있는 자세이며, 오른손이 전화기를 쥐고 귀에 안정적으로 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "시선은 아래를 향하고 있으며, 전화기를 왼쪽 귀에 대고 있음.",
        "built_space": "건물 외벽 옆의 야외 공간. 배경에 캠핑카와 야외 구조물들이 보이며, 레퍼런스의 실내 장소와 완전히 불일치함.",
        "entities": "이현우(레퍼런스와 일치하며 얼굴에 찰과상과 멍이 있음), 핸드폰(화면이 바깥쪽 카메라를 향하고 있음).",
        "hard_violations": [
         "전화기의 화면(기능면)이 사용자의 귀가 아닌 바깥쪽(카메라)을 향하도록 거꾸로 들고 있음 (물리적/기능적 오류)."
        ],
        "physics": "서 있는 자세로 왼손이 전화기를 지탱하고 있으나, 기기를 앞뒤가 뒤집힌 비정상적인 형태로 쥐고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "밀착 클로즈업과 초조한 표정, 얼굴 상처는 충실하지만 휴대폰 화면이 카메라 쪽으로 드러나며 배경도 지정된 도서관 참조 장소와 다르다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "휴대폰의 올바른 방향과 깊이 찌푸린 미간, 참조 도서관의 벽화·재질을 더 충실히 구현했지만 실외 구석이라는 조건과 뚜렷한 얼굴 멍은 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 왼쪽 아래의 프레임 밖을 향하며 렌즈를 보지 않는다. 휴대폰은 화면 오른쪽 귀에 밀착되어 있지만, 검은 화면으로 보이는 전면과 상단 수화부가 카메라 쪽으로 드러난다. 화면을 보이지 말라는 조건 및 통화 시 기능면의 방향과 맞지 않는다.",
        "built_space": "전화기 쪽 어깨 뒤로 회색 외벽, 창틀 일부, 수직 배관 한 줄이 보인다. 반대쪽 먼 배경에는 캠핑카처럼 보이는 차량 한 대와 낮은 건물들이 있고, 아래에는 실외기처럼 보이는 상자형 설비 하나가 있다. 실외 배치는 맞지만 참조 사진의 크레용 벽화, 회백색 이중 도장 벽, 복도 개구부와 연결되는 정확한 장소임을 확인할 수 없다. 인물은 얼굴과 어깨를 크게 담은 클로즈업이다.",
        "entities": "한 명의 젊은 한국계 남성으로 묘사되며 헝클어진 검은 머리와 낡은 짙은 회색 후드가 참조와 대체로 일치한다. 코와 볼의 찰과상, 입술 상처가 보이며 미간을 깊이 찌푸리고 있다. 손의 피부와 소매도 인물에 자연스럽게 연결된다. 은색 구형 휴대폰은 일회용 통화용 단말로 해석할 수 있으나 실제 일회용 여부는 확인할 수 없다. 다리 부상과 연락처 카드는 올바르게 프레임 밖에 있어 평가하지 않는다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "휴대폰은 손가락과 손바닥으로 잡고 귀와 볼 옆에 눌러 지지한다. 손목과 팔은 후드 소매로 이어져 별개의 손처럼 보이지 않는다. 머리와 상체의 자세는 통화 중 가능한 자세이며 공중에 떠 있는 물체는 없다. 하체와 발의 지지는 클로즈업 밖이므로 확인할 수 없다."
       },
       {
        "label": "B",
        "direction": "시선은 화면 오른쪽 아래의 프레임 밖으로 내려가 있다. 화면 왼쪽 귀에 댄 검은 휴대폰은 후면 카메라가 관객 쪽으로 보이고 기능면은 귀 쪽을 향한다. 화면을 관객에게 제시하지 않고 통화한다는 방향 조건에 맞는다.",
        "built_space": "회백색 복도 벽과 파란 크레용 바다, 야자수와 해 그림이 보이며 참조 장소의 재질과 벽화 요소를 유지한다. 전화기 쪽 어깨 뒤에는 복도 벽의 경계와 먼 창 하나가 보이고, 천장에는 흐릿한 조명 일부가 있다. 추가 가구나 중복 설비는 보이지 않는다. 다만 명백한 실내 복도이므로 글로 명시된 실외 구석은 아니다. 얼굴과 상부 가슴을 담아 A보다 조금 느슨한 클로즈업이다.",
        "entities": "한 명의 젊은 한국계 남성으로 묘사되며 검은 헝클어진 머리, 얼굴 형태, 해진 짙은 회색 후드가 참조에 대체로 부합한다. 깊게 잡힌 미간과 아래로 향한 눈이 통화 중 초조함을 나타낸다. 얼굴에는 피부 얼룩이 있으나 지속되어야 할 멍과 상처는 A만큼 명확하지 않다. 검은 소형 휴대폰의 후면이 보이며 일회용 여부는 외형만으로 확정할 수 없다. 손과 소매는 해당 인물에 일치한다. 연락처 카드와 다리, 캠핑카는 이 구도 밖이므로 누락으로 평가하지 않는다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "휴대폰은 귀 옆에서 손가락과 손바닥에 잡혀 있고, 손목과 팔이 소매를 통해 상체에 자연스럽게 연결된다. 귀에 기기를 누른 상태와 약간 숙인 머리는 실제 통화 동작으로 가능하다. 떠 있는 신체나 물체는 없으며 발과 지면의 접촉은 프레임 밖이다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "밀착 클로즈업과 초조한 표정, 얼굴 상처는 충실하지만 휴대폰 화면이 카메라 쪽으로 드러나며 배경도 지정된 도서관 참조 장소와 다르다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "휴대폰의 올바른 방향과 깊이 찌푸린 미간, 참조 도서관의 벽화·재질을 더 충실히 구현했지만 실외 구석이라는 조건과 뚜렷한 얼굴 멍은 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "시선은 화면 왼쪽 아래의 프레임 밖을 향하며 렌즈를 보지 않는다. 휴대폰은 화면 오른쪽 귀에 밀착되어 있지만, 검은 화면으로 보이는 전면과 상단 수화부가 카메라 쪽으로 드러난다. 화면을 보이지 말라는 조건 및 통화 시 기능면의 방향과 맞지 않는다.",
        "built_space": "전화기 쪽 어깨 뒤로 회색 외벽, 창틀 일부, 수직 배관 한 줄이 보인다. 반대쪽 먼 배경에는 캠핑카처럼 보이는 차량 한 대와 낮은 건물들이 있고, 아래에는 실외기처럼 보이는 상자형 설비 하나가 있다. 실외 배치는 맞지만 참조 사진의 크레용 벽화, 회백색 이중 도장 벽, 복도 개구부와 연결되는 정확한 장소임을 확인할 수 없다. 인물은 얼굴과 어깨를 크게 담은 클로즈업이다.",
        "entities": "한 명의 젊은 한국계 남성으로 묘사되며 헝클어진 검은 머리와 낡은 짙은 회색 후드가 참조와 대체로 일치한다. 코와 볼의 찰과상, 입술 상처가 보이며 미간을 깊이 찌푸리고 있다. 손의 피부와 소매도 인물에 자연스럽게 연결된다. 은색 구형 휴대폰은 일회용 통화용 단말로 해석할 수 있으나 실제 일회용 여부는 확인할 수 없다. 다리 부상과 연락처 카드는 올바르게 프레임 밖에 있어 평가하지 않는다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "휴대폰은 손가락과 손바닥으로 잡고 귀와 볼 옆에 눌러 지지한다. 손목과 팔은 후드 소매로 이어져 별개의 손처럼 보이지 않는다. 머리와 상체의 자세는 통화 중 가능한 자세이며 공중에 떠 있는 물체는 없다. 하체와 발의 지지는 클로즈업 밖이므로 확인할 수 없다."
       },
       {
        "label": "A",
        "direction": "시선은 화면 오른쪽 아래의 프레임 밖으로 내려가 있다. 화면 왼쪽 귀에 댄 검은 휴대폰은 후면 카메라가 관객 쪽으로 보이고 기능면은 귀 쪽을 향한다. 화면을 관객에게 제시하지 않고 통화한다는 방향 조건에 맞는다.",
        "built_space": "회백색 복도 벽과 파란 크레용 바다, 야자수와 해 그림이 보이며 참조 장소의 재질과 벽화 요소를 유지한다. 전화기 쪽 어깨 뒤에는 복도 벽의 경계와 먼 창 하나가 보이고, 천장에는 흐릿한 조명 일부가 있다. 추가 가구나 중복 설비는 보이지 않는다. 다만 명백한 실내 복도이므로 글로 명시된 실외 구석은 아니다. 얼굴과 상부 가슴을 담아 A보다 조금 느슨한 클로즈업이다.",
        "entities": "한 명의 젊은 한국계 남성으로 묘사되며 검은 헝클어진 머리, 얼굴 형태, 해진 짙은 회색 후드가 참조에 대체로 부합한다. 깊게 잡힌 미간과 아래로 향한 눈이 통화 중 초조함을 나타낸다. 얼굴에는 피부 얼룩이 있으나 지속되어야 할 멍과 상처는 A만큼 명확하지 않다. 검은 소형 휴대폰의 후면이 보이며 일회용 여부는 외형만으로 확정할 수 없다. 손과 소매는 해당 인물에 일치한다. 연락처 카드와 다리, 캠핑카는 이 구도 밖이므로 누락으로 평가하지 않는다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "휴대폰은 귀 옆에서 손가락과 손바닥에 잡혀 있고, 손목과 팔이 소매를 통해 상체에 자연스럽게 연결된다. 귀에 기기를 누른 상태와 약간 숙인 머리는 실제 통화 동작으로 가능하다. 떠 있는 신체나 물체는 없으며 발과 지면의 접촉은 프레임 밖이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.143
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.893
   },
   "violations": {
    "B": [
     "[gemini-pro] 전화기의 화면(기능면)이 사용자의 귀가 아닌 바깥쪽(카메라)을 향하도록 거꾸로 들고 있음 (물리적/기능적 오류)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 893
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "얼굴의 멍과 상처가 누락되었으나, 지정된 장소 레퍼런스(실내 벽화 앞)를 정확히 반영하고 소품(전화기)을 올바른 방향으로 파지하여 프롬프트를 가장 잘 충족함."
   },
   {
    "label": "B",
    "score": 893,
    "verdict_ko": "캐릭터의 얼굴 상처는 잘 묘사되었으나, 지정된 실내 복도가 아닌 야외를 배경으로 삼았으며, 전화기 화면이 바깥을 향하게 거꾸로 들고 있는 치명적인 오류가 발생함.  ★위반: [gemini-pro] 전화기의 화면(기능면)이 사용자의 귀가 아닌 바깥쪽(카메라)을 향하도록 거꾸로 들고 있음 (물리적/기능적 오류)."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L58B01.png",
    "asset_id": "e381ac75-3b70-4df7-83a9-f70bc3b4868a",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd3-94c0-78a6-83d3-0671c5096aac",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S47sh15__bgfirst_bg.png",
   "bg_asset_id": "88792161-b554-44d6-ba47-854436f1f5ca",
   "bg_record_key": "S47sh15::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S47sh15::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:42:01.381792+00:00",
  "fingerprint": "778cca45ccda71fe5d147f4ffdd28fac906adc15dde5596a74a445b94cc07378",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S47sh15_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S47sh15_sel.png",
  "source_sha256": "7af8ef27ad5a390e6cb666cbe8123d9bcc76abddcd131bab3484c52387513e28",
  "file": "S47sh15_cine.png",
  "staged_sha256": "6baa173c3d39c4dd98d086c25b38562e77e70f87c9fedf4e3154a22d686e2a9a",
  "latency_ms": 27594
 },
 "S47sh17::signage": {
  "fp": "380998441eb33369",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S47sh17": {
  "input_fingerprint": "580904e0dc516ea2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 이현우의 등 뒤에 거대한 찰리가 우두커니 선 채 앞을 응시하는 구도.\n\nLOCATION (lock): In the same secluded outdoor corner beside the abandoned library, where the caller turns to find the robot nearby. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Secluded library corner (Now visible around both figures) — The same corner recedes behind the laterally separated figures; used as Maintains continuity with the call and establishes that 찰리 is physically behind 이현우, not in a separate space; Library floor (Visible between the two figures); used as Provides an uninterrupted depth reference between the foreground figure and 찰리.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the call's restrained library illumination so the revelation comes from newly visible space rather than a dramatic lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The completed crayon mural and its handwritten message remain on the library walls; the old camper is still parked outside. Charlie now stands beside the secluded calling spot. 이현우: He has finished the call and is turning away, still carrying the disposable phone and the intact contact card. His facial bruises and untreated leg injury remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 이현우의 등 뒤에 거대한 찰리가 우두커니 선 채 앞을 응시하는 구도.\n\nLOCATION (lock): In the same secluded outdoor corner beside the abandoned library, where the caller turns to find the robot nearby. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Secluded library corner (Now visible around both figures) — The same corner recedes behind the laterally separated figures; used as Maintains continuity with the call and establishes that 찰리 is physically behind 이현우, not in a separate space; Library floor (Visible between the two figures); used as Provides an uninterrupted depth reference between the foreground figure and 찰리.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the call's restrained library illumination so the revelation comes from newly visible space rather than a dramatic lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The completed crayon mural and its handwritten message remain on the library walls; the old camper is still parked outside. Charlie now stands beside the secluded calling spot. 이현우: He has finished the call and is turning away, still carrying the disposable phone and the intact contact card. His facial bruises and untreated leg injury remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 이현우의 등 뒤에 거대한 찰리가 우두커니 선 채 앞을 응시하는 구도.\n\nLOCATION (lock): In the same secluded outdoor corner beside the abandoned library, where the caller turns to find the robot nearby. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Secluded library corner (Now visible around both figures) — The same corner recedes behind the laterally separated figures; used as Maintains continuity with the call and establishes that 찰리 is physically behind 이현우, not in a separate space; Library floor (Visible between the two figures); used as Provides an uninterrupted depth reference between the foreground figure and 찰리.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the call's restrained library illumination so the revelation comes from newly visible space rather than a dramatic lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The completed crayon mural and its handwritten message remain on the library walls; the old camper is still parked outside. Charlie now stands beside the secluded calling spot. 이현우: He has finished the call and is turning away, still carrying the disposable phone and the intact contact card. His facial bruises and untreated leg injury remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리는 정면 카메라를 응시하고, 이현우는 카메라를 등진 채 약간 오른쪽 바깥을 향해 서 있다.",
    "built_space": "복도 공간과 왼쪽의 크레용 벽화, 밖의 캠핑카 등은 잘 구현되었으나, 오른쪽 벽에 지침을 정면으로 위반하는 명확히 읽을 수 있는 한글 문장들이 나타난다.",
    "entities": "찰리는 망토를 포함해 참조와 일치하는 모습이다. 이현우는 전화기만 들고 있고 명함이 보이지 않으며, 얼굴이 가려져 상처를 확인할 수 없다.",
    "hard_violations": [
     "[gemini-pro] 화면 내 읽을 수 있는 글씨 렌더링 금지 지침을 정면으로 위반하여 오른쪽 벽에 명확한 한글 문장 생성",
     "[gpt-high] 오른쪽 벽에 읽을 수 있는 한글 문장들이 노출되어 글자 판독 금지 조건을 위반한다.",
     "[gpt-high] 이현우의 몸이 찰리를 향해 있어 찰리가 그의 등 뒤에 서야 하는 핵심 위치 관계가 반대로 배치되었다."
    ],
    "physics": "두 인물 모두 바닥에 온전히 발을 딛고 서 있으며 불안정한 요소는 없다."
   },
   {
    "label": "B",
    "direction": "찰리는 정면 카메라 방향을 우두커니 응시하고 있으며, 이현우는 등을 돌린 상태에서 고개를 돌려 카메라 쪽을 바라보고 있다.",
    "built_space": "왼쪽 벽에는 바다 테마의 크레용 벽화가 연속성을 유지하며, 오른쪽 벽의 글귀는 지침대로 읽을 수 없게 뭉개져 있다. 배경 야외에는 캠핑카가 잘 배치되어 있다.",
    "entities": "이현우는 참조 이미지와 일치하는 복장이며, 오른손에 전화기와 온전한 명함을 모두 들고 있고 얼굴의 상처도 확인된다. 찰리는 고릴라 형태의 기계 머리와 외장을 갖췄으나, 참조 이미지의 회색 망토를 입고 있지 않다.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥을 딛고 자연스럽고 무게감 있게 서 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "찰리의 망토가 누락되었으나, 명함과 전화기를 든 상태 및 얼굴 상처를 잘 표현했으며 읽을 수 없는 텍스트 지침을 훌륭히 준수했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 외형은 참조와 일치하나, 화면 내 읽을 수 있는 텍스트 금지 지침을 심각하게 위반하였고 명함과 얼굴 상처가 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 정면 카메라를 응시하고, 이현우는 카메라를 등진 채 약간 오른쪽 바깥을 향해 서 있다.",
        "built_space": "복도 공간과 왼쪽의 크레용 벽화, 밖의 캠핑카 등은 잘 구현되었으나, 오른쪽 벽에 지침을 정면으로 위반하는 명확히 읽을 수 있는 한글 문장들이 나타난다.",
        "entities": "찰리는 망토를 포함해 참조와 일치하는 모습이다. 이현우는 전화기만 들고 있고 명함이 보이지 않으며, 얼굴이 가려져 상처를 확인할 수 없다.",
        "hard_violations": [
         "화면 내 읽을 수 있는 글씨 렌더링 금지 지침을 정면으로 위반하여 오른쪽 벽에 명확한 한글 문장 생성"
        ],
        "physics": "두 인물 모두 바닥에 온전히 발을 딛고 서 있으며 불안정한 요소는 없다."
       },
       {
        "label": "B",
        "direction": "찰리는 정면 카메라 방향을 우두커니 응시하고 있으며, 이현우는 등을 돌린 상태에서 고개를 돌려 카메라 쪽을 바라보고 있다.",
        "built_space": "왼쪽 벽에는 바다 테마의 크레용 벽화가 연속성을 유지하며, 오른쪽 벽의 글귀는 지침대로 읽을 수 없게 뭉개져 있다. 배경 야외에는 캠핑카가 잘 배치되어 있다.",
        "entities": "이현우는 참조 이미지와 일치하는 복장이며, 오른손에 전화기와 온전한 명함을 모두 들고 있고 얼굴의 상처도 확인된다. 찰리는 고릴라 형태의 기계 머리와 외장을 갖췄으나, 참조 이미지의 회색 망토를 입고 있지 않다.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥을 딛고 자연스럽고 무게감 있게 서 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "찰리의 망토가 누락되었으나, 명함과 전화기를 든 상태 및 얼굴 상처를 잘 표현했으며 읽을 수 없는 텍스트 지침을 훌륭히 준수했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 외형은 참조와 일치하나, 화면 내 읽을 수 있는 텍스트 금지 지침을 심각하게 위반하였고 명함과 얼굴 상처가 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 정면 카메라를 응시하고, 이현우는 카메라를 등진 채 약간 오른쪽 바깥을 향해 서 있다.",
        "built_space": "복도 공간과 왼쪽의 크레용 벽화, 밖의 캠핑카 등은 잘 구현되었으나, 오른쪽 벽에 지침을 정면으로 위반하는 명확히 읽을 수 있는 한글 문장들이 나타난다.",
        "entities": "찰리는 망토를 포함해 참조와 일치하는 모습이다. 이현우는 전화기만 들고 있고 명함이 보이지 않으며, 얼굴이 가려져 상처를 확인할 수 없다.",
        "hard_violations": [
         "화면 내 읽을 수 있는 글씨 렌더링 금지 지침을 정면으로 위반하여 오른쪽 벽에 명확한 한글 문장 생성"
        ],
        "physics": "두 인물 모두 바닥에 온전히 발을 딛고 서 있으며 불안정한 요소는 없다."
       },
       {
        "label": "B",
        "direction": "찰리는 정면 카메라 방향을 우두커니 응시하고 있으며, 이현우는 등을 돌린 상태에서 고개를 돌려 카메라 쪽을 바라보고 있다.",
        "built_space": "왼쪽 벽에는 바다 테마의 크레용 벽화가 연속성을 유지하며, 오른쪽 벽의 글귀는 지침대로 읽을 수 없게 뭉개져 있다. 배경 야외에는 캠핑카가 잘 배치되어 있다.",
        "entities": "이현우는 참조 이미지와 일치하는 복장이며, 오른손에 전화기와 온전한 명함을 모두 들고 있고 얼굴의 상처도 확인된다. 찰리는 고릴라 형태의 기계 머리와 외장을 갖췄으나, 참조 이미지의 회색 망토를 입고 있지 않다.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥을 딛고 자연스럽고 무게감 있게 서 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리가 이현우의 뒤쪽에 서서 앞을 응시하고 두 인물 사이 바닥이 드러나는 구도가 더 충실하지만, 찰리의 회색 망토가 빠졌다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "오른쪽 벽의 읽을 수 있는 한글이 금지 조건을 위반하며, 두 인물이 중앙에서 겹치고 찰리가 이현우의 등 뒤가 아닌 몸 앞쪽에 놓인다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 몸을 화면 오른쪽으로 틀면서 얼굴과 시선을 카메라 쪽으로 돌린다. 찰리는 그보다 뒤쪽 오른편에서 카메라가 있는 전방을 응시한다. 찰리가 이현우를 내려다보는 장면은 아니며, 앞을 응시하라는 지시에는 부합한다. 내려 든 전화기의 화면을 관객에게 의도적으로 내보이지 않는다.",
        "built_space": "양쪽에 낡은 벽과 청회색 하부 도장이 있고, 왼쪽 벽에 바다·야자수 벽화가 있다. 왼쪽 측면 출입구 하나, 오른쪽 측면 출입구 하나, 뒤쪽 외부 개구부 하나가 보이며 밖에는 캠핑카 한 대가 있다. 이현우는 왼쪽 전경, 찰리는 오른쪽 후경에 있어 일부 겹치지만 옆으로 분리되고 사이의 바닥이 연속된다. 참고 사진의 마모된 재료와 절제된 아침빛을 대체로 유지한다.",
        "entities": "인물은 이현우와 찰리뿐이다. 이현우는 젊은 한국계 남성으로 보이고 검은 헝클어진 머리, 해진 어두운 후드와 청바지, 얼굴의 멍이 확인된다. 한 손에 검은 전화기와 흰 카드가 보이며 카드가 찢어진 흔적은 없다. 찰리는 큰 금속 몸체, 각진 고릴라형 머리, 흰 얼굴판, 점형 눈과 선형 입을 갖췄지만 참고 이미지의 회색 망토가 없다. 벽화와 캠핑카는 보이고 벽의 글자 흔적은 명확히 읽히지 않는다. 다리 부상은 이 구도에서 뚜렷하게 확인되지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 양발은 바닥에 닿아 무게를 지탱하며 팔은 몸 옆으로 내려와 있다. 이현우의 발은 화면 밖이지만 몸통과 다리는 자연스럽게 아래로 이어져 서 있는 자세로 읽히며 부유의 증거는 없다. 전화기와 카드는 손으로 잡고 있다. 캠핑카는 바퀴로 지면에 지지된다."
       },
       {
        "label": "B",
        "direction": "찰리는 카메라 쪽 정면을 응시한다. 이현우는 카메라에 등을 보인 채 몸은 찰리 쪽을 향하고 고개만 화면 오른쪽으로 돌린다. 따라서 찰리는 화면상 후경이지만 이현우의 신체 방향을 기준으로는 등 뒤가 아니라 앞쪽에 있다.",
        "built_space": "낡은 벽 두 면 사이로 외부 개구부 하나가 열리고, 왼쪽에는 바다와 야자수 벽화, 오른쪽에는 여러 줄의 한글 문장이 보인다. 외부에는 캠핑카 한 대와 그 출입문 앞 발판 하나가 있다. 두 인물이 중앙의 같은 축에 겹쳐 서 있어 요구된 좌우 분리가 없고, 이현우가 둘 사이 바닥의 상당 부분을 가린다. 바닥 자체는 실내외로 이어지며 낡은 도장과 부드러운 자연광은 참고 장소와 대체로 어울린다.",
        "entities": "인물은 젊은 한국계 남성으로 보이는 이현우와 기계형 찰리 둘이다. 이현우의 검은 머리, 해진 후드, 청바지와 낡은 신발은 참고와 대체로 맞고 옆얼굴에 상처 흔적이 있다. 오른손 주변에 검은 전화기와 흰 카드로 보이는 물건이 있으나 카드 전체 상태는 가려져 있다. 찰리의 흰 얼굴판, 각진 머리, 금속 몸체와 회색 망토는 참고에 가깝다. 벽화와 오래된 캠핑카도 있지만 오른쪽 벽의 한글은 읽을 수 있다. 다리의 미처리 부상은 뚜렷하지 않다.",
        "hard_violations": [
         "오른쪽 벽에 읽을 수 있는 한글 문장들이 노출되어 글자 판독 금지 조건을 위반한다.",
         "이현우의 몸이 찰리를 향해 있어 찰리가 그의 등 뒤에 서야 하는 핵심 위치 관계가 반대로 배치되었다."
        ],
        "physics": "이현우의 두 신발과 찰리의 두 금속 발이 모두 바닥에 닿아 있다. 두 몸은 각자의 다리로 지지되고, 망토는 찰리의 어깨에서 아래로 늘어진다. 오른손의 물건에는 손의 접촉이 보이며 캠핑카와 발판도 지면에 놓여 있다. 떠 있거나 지지 없이 매달린 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리가 이현우의 뒤쪽에 서서 앞을 응시하고 두 인물 사이 바닥이 드러나는 구도가 더 충실하지만, 찰리의 회색 망토가 빠졌다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "오른쪽 벽의 읽을 수 있는 한글이 금지 조건을 위반하며, 두 인물이 중앙에서 겹치고 찰리가 이현우의 등 뒤가 아닌 몸 앞쪽에 놓인다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 몸을 화면 오른쪽으로 틀면서 얼굴과 시선을 카메라 쪽으로 돌린다. 찰리는 그보다 뒤쪽 오른편에서 카메라가 있는 전방을 응시한다. 찰리가 이현우를 내려다보는 장면은 아니며, 앞을 응시하라는 지시에는 부합한다. 내려 든 전화기의 화면을 관객에게 의도적으로 내보이지 않는다.",
        "built_space": "양쪽에 낡은 벽과 청회색 하부 도장이 있고, 왼쪽 벽에 바다·야자수 벽화가 있다. 왼쪽 측면 출입구 하나, 오른쪽 측면 출입구 하나, 뒤쪽 외부 개구부 하나가 보이며 밖에는 캠핑카 한 대가 있다. 이현우는 왼쪽 전경, 찰리는 오른쪽 후경에 있어 일부 겹치지만 옆으로 분리되고 사이의 바닥이 연속된다. 참고 사진의 마모된 재료와 절제된 아침빛을 대체로 유지한다.",
        "entities": "인물은 이현우와 찰리뿐이다. 이현우는 젊은 한국계 남성으로 보이고 검은 헝클어진 머리, 해진 어두운 후드와 청바지, 얼굴의 멍이 확인된다. 한 손에 검은 전화기와 흰 카드가 보이며 카드가 찢어진 흔적은 없다. 찰리는 큰 금속 몸체, 각진 고릴라형 머리, 흰 얼굴판, 점형 눈과 선형 입을 갖췄지만 참고 이미지의 회색 망토가 없다. 벽화와 캠핑카는 보이고 벽의 글자 흔적은 명확히 읽히지 않는다. 다리 부상은 이 구도에서 뚜렷하게 확인되지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 양발은 바닥에 닿아 무게를 지탱하며 팔은 몸 옆으로 내려와 있다. 이현우의 발은 화면 밖이지만 몸통과 다리는 자연스럽게 아래로 이어져 서 있는 자세로 읽히며 부유의 증거는 없다. 전화기와 카드는 손으로 잡고 있다. 캠핑카는 바퀴로 지면에 지지된다."
       },
       {
        "label": "A",
        "direction": "찰리는 카메라 쪽 정면을 응시한다. 이현우는 카메라에 등을 보인 채 몸은 찰리 쪽을 향하고 고개만 화면 오른쪽으로 돌린다. 따라서 찰리는 화면상 후경이지만 이현우의 신체 방향을 기준으로는 등 뒤가 아니라 앞쪽에 있다.",
        "built_space": "낡은 벽 두 면 사이로 외부 개구부 하나가 열리고, 왼쪽에는 바다와 야자수 벽화, 오른쪽에는 여러 줄의 한글 문장이 보인다. 외부에는 캠핑카 한 대와 그 출입문 앞 발판 하나가 있다. 두 인물이 중앙의 같은 축에 겹쳐 서 있어 요구된 좌우 분리가 없고, 이현우가 둘 사이 바닥의 상당 부분을 가린다. 바닥 자체는 실내외로 이어지며 낡은 도장과 부드러운 자연광은 참고 장소와 대체로 어울린다.",
        "entities": "인물은 젊은 한국계 남성으로 보이는 이현우와 기계형 찰리 둘이다. 이현우의 검은 머리, 해진 후드, 청바지와 낡은 신발은 참고와 대체로 맞고 옆얼굴에 상처 흔적이 있다. 오른손 주변에 검은 전화기와 흰 카드로 보이는 물건이 있으나 카드 전체 상태는 가려져 있다. 찰리의 흰 얼굴판, 각진 머리, 금속 몸체와 회색 망토는 참고에 가깝다. 벽화와 오래된 캠핑카도 있지만 오른쪽 벽의 한글은 읽을 수 있다. 다리의 미처리 부상은 뚜렷하지 않다.",
        "hard_violations": [
         "오른쪽 벽에 읽을 수 있는 한글 문장들이 노출되어 글자 판독 금지 조건을 위반한다.",
         "이현우의 몸이 찰리를 향해 있어 찰리가 그의 등 뒤에 서야 하는 핵심 위치 관계가 반대로 배치되었다."
        ],
        "physics": "이현우의 두 신발과 찰리의 두 금속 발이 모두 바닥에 닿아 있다. 두 몸은 각자의 다리로 지지되고, 망토는 찰리의 어깨에서 아래로 늘어진다. 오른손의 물건에는 손의 접촉이 보이며 캠핑카와 발판도 지면에 놓여 있다. 떠 있거나 지지 없이 매달린 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.661,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.411,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 화면 내 읽을 수 있는 글씨 렌더링 금지 지침을 정면으로 위반하여 오른쪽 벽에 명확한 한글 문장 생성",
     "[gpt-high] 오른쪽 벽에 읽을 수 있는 한글 문장들이 노출되어 글자 판독 금지 조건을 위반한다.",
     "[gpt-high] 이현우의 몸이 찰리를 향해 있어 찰리가 그의 등 뒤에 서야 하는 핵심 위치 관계가 반대로 배치되었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 411
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "찰리의 망토가 누락되었으나, 명함과 전화기를 든 상태 및 얼굴 상처를 잘 표현했으며 읽을 수 없는 텍스트 지침을 훌륭히 준수했습니다."
   },
   {
    "label": "A",
    "score": 411,
    "verdict_ko": "찰리의 외형은 참조와 일치하나, 화면 내 읽을 수 있는 텍스트 금지 지침을 심각하게 위반하였고 명함과 얼굴 상처가 누락되었습니다.  ★위반: [gemini-pro] 화면 내 읽을 수 있는 글씨 렌더링 금지 지침을 정면으로 위반하여 오른쪽 벽에 명확한 한글 문장 생성 / [gpt-high] 오른쪽 벽에 읽을 수 있는 한글 문장들이 노출되어 글자 판독 금지 조건을 위반한다. / [gpt-high] 이현우의 몸이 찰리를 향해 있어 찰리가 그의 등 뒤에 서야 하는 핵심 위치 관계가 반대로 배치되었다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S47sh15_sel.png",
    "asset_id": "bd66636e-a246-40aa-a63c-ed511eec2f0c",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:948699>",
    "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd5-54fb-7289-a862-554f2575fdcc",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S47sh15"
  }
 },
 "S47sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:42:29.324047+00:00",
  "fingerprint": "a2cf75a19f17f1d38854d07da9921bfce18ca3b1846bac3330201881981d8c9c",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S47sh17_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S47sh17_sel.png",
  "source_sha256": "5394c374af4720f8922991d70d252575074fefa415fa06f98f750a3b4818cb26",
  "file": "S47sh17_cine.png",
  "staged_sha256": "63d97a0dc71a155baac97b0f75288f4aa8eafa7bbb73ae23ba9d18185c422f00",
  "latency_ms": 11058
 },
 "S48sh5::signage": {
  "fp": "b06b372f40fc6e2e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S8sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:18:31.415982+00:00",
  "fingerprint": "450d44183d1074853d6b96e70a93a8ef7b7f278cb95fa63605729b20708a0088",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S8sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S8sh1_sel.png",
  "source_sha256": "5db104a821fe1b874cc9ac7a3b446ce3d04a448b9b9ac341b29022bc0a1256bb",
  "file": "S8sh1_cine.png",
  "staged_sha256": "8a6d799830328d629f4d726c7ce65598ccc0cba23d30cd0451b08df667c1695b",
  "latency_ms": 11235
 },
 "S44sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:40:00.173433+00:00",
  "fingerprint": "365de4bb9295d111b5cd335c3f30a1c4042161cfbe79ec9665594e7f5416d318",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S44sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S44sh6_sel.png",
  "source_sha256": "d71d1552a1779560abb8d62013d790ddb988b3fe1391ca32e92f3d5374baf313",
  "file": "S44sh6_cine.png",
  "staged_sha256": "2e3678037be1d73e6d177ce17f19c7460657f44f4dc290131ea933a805a6b9e0",
  "latency_ms": 13372
 },
 "S48sh5": {
  "input_fingerprint": "f934060ad2815c23",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 캠핑카가 검문소를 앞두고 산길 쪽으로 차체가 급격히 꺾인 채 타이어에서 흙먼지가 뿜어져 나오는 mid-action 역동적인 순간.\n\nLOCATION (lock): At the junction where the main road turns onto a mountain forestry track before a checkpoint. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Mountain-road entrance in the upper-right of the frame, background; Checkpoint on the abandoned route in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Camper (Caught in a sharp turn toward the mountain road) — Rear quarter and one flank face the camera, with the nose directed toward upper-right; used as Primary action subject, kept small enough for the road junction to explain the evasive maneuver; Mountain-road entrance (The camper is turning onto it) — The entrance opens diagonally away toward upper-right; used as Visible destination of the turn; Distant checkpoint (Ahead on the original road) — Seen distantly along the route the camper is abandoning; used as Small background evidence of the reason for the maneuver; Earth dust (Bursting from the tires during the turn); used as Traces the wheels' abrupt directional change without concealing the vehicle.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light and controlled contrast keep the turning vehicle and earth dust legible without adding colored or artificial illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old camper carries the collected supplies and has unresolved roof and window-frame leaks from the previous rainstorm. A checkpoint lies ahead amid barren fields, dead trees and abandoned houses; Charlie occupies the rear seating area.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 캠핑카가 검문소를 앞두고 산길 쪽으로 차체가 급격히 꺾인 채 타이어에서 흙먼지가 뿜어져 나오는 mid-action 역동적인 순간.\n\nLOCATION (lock): At the junction where the main road turns onto a mountain forestry track before a checkpoint. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Mountain-road entrance in the upper-right of the frame, background; Checkpoint on the abandoned route in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Camper (Caught in a sharp turn toward the mountain road) — Rear quarter and one flank face the camera, with the nose directed toward upper-right; used as Primary action subject, kept small enough for the road junction to explain the evasive maneuver; Mountain-road entrance (The camper is turning onto it) — The entrance opens diagonally away toward upper-right; used as Visible destination of the turn; Distant checkpoint (Ahead on the original road) — Seen distantly along the route the camper is abandoning; used as Small background evidence of the reason for the maneuver; Earth dust (Bursting from the tires during the turn); used as Traces the wheels' abrupt directional change without concealing the vehicle.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light and controlled contrast keep the turning vehicle and earth dust legible without adding colored or artificial illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old camper carries the collected supplies and has unresolved roof and window-frame leaks from the previous rainstorm. A checkpoint lies ahead amid barren fields, dead trees and abandoned houses; Charlie occupies the rear seating area.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 캠핑카가 검문소를 앞두고 산길 쪽으로 차체가 급격히 꺾인 채 타이어에서 흙먼지가 뿜어져 나오는 mid-action 역동적인 순간.\n\nLOCATION (lock): At the junction where the main road turns onto a mountain forestry track before a checkpoint. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Mountain-road entrance in the upper-right of the frame, background; Checkpoint on the abandoned route in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Camper (Caught in a sharp turn toward the mountain road) — Rear quarter and one flank face the camera, with the nose directed toward upper-right; used as Primary action subject, kept small enough for the road junction to explain the evasive maneuver; Mountain-road entrance (The camper is turning onto it) — The entrance opens diagonally away toward upper-right; used as Visible destination of the turn; Distant checkpoint (Ahead on the original road) — Seen distantly along the route the camper is abandoning; used as Small background evidence of the reason for the maneuver; Earth dust (Bursting from the tires during the turn); used as Traces the wheels' abrupt directional change without concealing the vehicle.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light and controlled contrast keep the turning vehicle and earth dust legible without adding colored or artificial illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old camper carries the collected supplies and has unresolved roof and window-frame leaks from the previous rainstorm. A checkpoint lies ahead amid barren fields, dead trees and abandoned houses; Charlie occupies the rear seating area.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "캠핑카 전면부가 우측 상단의 산길 입구를 향해 꺾여 있음.",
    "built_space": "주도로는 좌측 상단으로 뻗어 있고 멀리 차단기가 있는 검문소가 있음. 우측 상단으로 산길이 이어지며 주변에 빈 집과 마른 나무가 배치됨.",
    "entities": "캠핑카 내부에 짐이 있고 표면에 녹슨 자국이 보임. 측면 창문에 흰 얼굴판과 점 형태의 눈을 가진 기계형 찰리가 나타남. 차량 번호판에 문자가 읽힘.",
    "hard_violations": [
     "[gpt-high] 샷 텍스트가 인물을 보여주도록 지정하지 않았는데 측면 창문에 찰리를 명확히 노출했다."
    ],
    "physics": "타이어에서 흙먼지가 튀어오르며 지면에 밀착된 채 급격한 회전 동작을 지지함."
   },
   {
    "label": "B",
    "direction": "캠핑카 전면부가 우측 상단의 흙길을 향해 꺾여 있음.",
    "built_space": "좌측 상단의 주도로에 바리케이드가 있는 검문소가 뚜렷하며, 우측 상단으로 산길이 이어짐. 주변에 버려진 집과 마른 나무가 있음.",
    "entities": "캠핑카 지붕에 짐이 묶여 있음. 뒷유리창의 찰리는 기계가 아닌 생물학적 고릴라의 모습을 하고 있음.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 없는 인물(검문소 경비원들) 추가 (invented people)",
     "[gpt-high] 샷 텍스트가 가시적 인물을 요구하지 않았는데 검문소에 다수의 인원을 추가했다.",
     "[gpt-high] 장면 수준의 인물 목록만으로 후면 창문에 찰리로 보이는 인물을 노출해, 이 샷에 인물을 추가하지 말라는 지시를 위반했다."
    ],
    "physics": "뒷바퀴에서 흙먼지가 날리며 회전하는 차량의 하중과 움직임을 지지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 구도를 정확히 따랐으며 찰리의 기계적인 얼굴 특징(흰 얼굴판, 점 눈)을 잘 구현했으나, 번호판에 일부 읽을 수 있는 문자가 노출되어 감점됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "검문소에 샷 텍스트에 없는 인물(경비원)을 추가하는 치명적 위반을 범했고, 찰리가 명시된 기계 형태가 아닌 생물학적 고릴라로 그려짐."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카 전면부가 우측 상단의 산길 입구를 향해 꺾여 있음.",
        "built_space": "주도로는 좌측 상단으로 뻗어 있고 멀리 차단기가 있는 검문소가 있음. 우측 상단으로 산길이 이어지며 주변에 빈 집과 마른 나무가 배치됨.",
        "entities": "캠핑카 내부에 짐이 있고 표면에 녹슨 자국이 보임. 측면 창문에 흰 얼굴판과 점 형태의 눈을 가진 기계형 찰리가 나타남. 차량 번호판에 문자가 읽힘.",
        "hard_violations": [],
        "physics": "타이어에서 흙먼지가 튀어오르며 지면에 밀착된 채 급격한 회전 동작을 지지함."
       },
       {
        "label": "B",
        "direction": "캠핑카 전면부가 우측 상단의 흙길을 향해 꺾여 있음.",
        "built_space": "좌측 상단의 주도로에 바리케이드가 있는 검문소가 뚜렷하며, 우측 상단으로 산길이 이어짐. 주변에 버려진 집과 마른 나무가 있음.",
        "entities": "캠핑카 지붕에 짐이 묶여 있음. 뒷유리창의 찰리는 기계가 아닌 생물학적 고릴라의 모습을 하고 있음.",
        "hard_violations": [
         "샷 텍스트에 없는 인물(검문소 경비원들) 추가 (invented people)"
        ],
        "physics": "뒷바퀴에서 흙먼지가 날리며 회전하는 차량의 하중과 움직임을 지지함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 구도를 정확히 따랐으며 찰리의 기계적인 얼굴 특징(흰 얼굴판, 점 눈)을 잘 구현했으나, 번호판에 일부 읽을 수 있는 문자가 노출되어 감점됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "검문소에 샷 텍스트에 없는 인물(경비원)을 추가하는 치명적 위반을 범했고, 찰리가 명시된 기계 형태가 아닌 생물학적 고릴라로 그려짐."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카 전면부가 우측 상단의 산길 입구를 향해 꺾여 있음.",
        "built_space": "주도로는 좌측 상단으로 뻗어 있고 멀리 차단기가 있는 검문소가 있음. 우측 상단으로 산길이 이어지며 주변에 빈 집과 마른 나무가 배치됨.",
        "entities": "캠핑카 내부에 짐이 있고 표면에 녹슨 자국이 보임. 측면 창문에 흰 얼굴판과 점 형태의 눈을 가진 기계형 찰리가 나타남. 차량 번호판에 문자가 읽힘.",
        "hard_violations": [],
        "physics": "타이어에서 흙먼지가 튀어오르며 지면에 밀착된 채 급격한 회전 동작을 지지함."
       },
       {
        "label": "B",
        "direction": "캠핑카 전면부가 우측 상단의 흙길을 향해 꺾여 있음.",
        "built_space": "좌측 상단의 주도로에 바리케이드가 있는 검문소가 뚜렷하며, 우측 상단으로 산길이 이어짐. 주변에 버려진 집과 마른 나무가 있음.",
        "entities": "캠핑카 지붕에 짐이 묶여 있음. 뒷유리창의 찰리는 기계가 아닌 생물학적 고릴라의 모습을 하고 있음.",
        "hard_violations": [
         "샷 텍스트에 없는 인물(검문소 경비원들) 추가 (invented people)"
        ],
        "physics": "뒷바퀴에서 흙먼지가 날리며 회전하는 차량의 하중과 움직임을 지지함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "검문소와 산길의 배치는 맞지만 차량 앞머리가 산길 반대쪽을 향하며, 지시하지 않은 검문 인원과 창문 속 인물을 추가했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "차량의 앞머리와 흙먼지가 우상단 산길로 급선회하는 동작을 제대로 설명하지만, 인물이 보이지 않아야 하는 샷에 찰리를 노출했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "운전석과 앞바퀴가 있는 차량 앞부분은 화면 왼쪽의 포장도로를 향하고, 후면은 오른쪽 아래를 향한다. 목적지인 산길 입구는 오른쪽 위에 있어 차량의 전진 방향과 맞지 않는다. 후면 창문의 인물은 카메라 쪽으로 고개를 돌리고 있다. 검문소 인원들의 정확한 시선은 거리상 확인하기 어렵다.",
        "built_space": "포장도로는 왼쪽 위 검문소로 이어지고 비포장 산길 하나는 오른쪽 위로 올라간다. 검문소에는 양쪽 초소 두 동과 노란 차단시설들이 보인다. 주변에는 황량한 밭, 잎 없는 나무, 낡은 주택과 전신주가 있다. 높은 외부 시점은 허용된 구도지만 차량이 화면 오른쪽 절반을 크게 차지해 교차로보다 강조된다.",
        "entities": "낡고 얼룩진 캠핑카, 지붕에 묶인 보급품, 타이어 주변 흙먼지, 먼 검문소와 산길이 보인다. 후면 창문에는 고릴라 같은 얼굴과 기계식 몸체를 가진 인물이 있으나, 지정된 흰 얼굴판·점 형태 눈·선 형태 입보다는 사실적인 고릴라 얼굴에 가깝다. 검문소에는 최소 일곱 명가량의 인원이 추가되어 있다. 앰버는 보이지 않으며 이 구도에서는 문제가 아니다. 낮 조명은 충족한다.",
        "hard_violations": [
         "샷 텍스트가 가시적 인물을 요구하지 않았는데 검문소에 다수의 인원을 추가했다.",
         "장면 수준의 인물 목록만으로 후면 창문에 찰리로 보이는 인물을 노출해, 이 샷에 인물을 추가하지 말라는 지시를 위반했다."
        ],
        "physics": "차량은 보이는 앞뒤 타이어로 지면에 지지되고, 보급품은 지붕 적재대와 끈으로 고정되어 있다. 창문 속 인물의 하체는 차체에 가려져 있으므로 공중에 떠 있다고 볼 근거는 없다. 타이어 접지부에서 먼지가 일어나는 것은 가능하지만, 차체 방향은 요청된 산길 진입이 아니라 포장도로 쪽으로 빠져나오는 동작에 가깝다."
       },
       {
        "label": "B",
        "direction": "차량의 뒤가 왼쪽 아래, 운전실과 앞머리가 오른쪽 위를 향한다. 앞머리 바로 앞에 비포장 산길이 이어져 목적지가 명확하다. 포기한 포장도로는 왼쪽 위의 먼 검문소로 이어진다. 측면 창문의 흰 얼굴은 카메라 쪽을 보고 있다.",
        "built_space": "하나의 포장도로에서 하나의 비포장 산길이 오른쪽 위로 갈라지는 구조다. 먼 검문소에는 작은 초소 한 동과 도로를 가로막는 차단시설이 보인다. 왼쪽 배경에는 주택 두 동, 황량한 밭과 잎 없는 나무가 있다. 차량의 후면과 한쪽 측면을 보여주는 외부 와이드 구도이며, 차량이 다소 크기는 해도 분기점과 두 경로가 모두 읽힌다.",
        "entities": "낡고 녹슨 캠핑카, 뒤쪽 창문 너머의 보급품, 흙먼지, 산길 입구와 먼 검문소가 보인다. 측면 창문에는 흰 얼굴판과 어두운 외장을 가진 찰리로 보이는 존재가 있으며, A보다 지정된 기계 머리에 가깝다. 앰버는 보이지 않는다. 누수의 지속 여부와 정확한 뒷좌석 착석 상태는 외부 구도만으로 확인할 수 없으며, 이를 누락으로 판단할 수는 없다. 낮의 자연광은 맞는다.",
        "hard_violations": [
         "샷 텍스트가 인물을 보여주도록 지정하지 않았는데 측면 창문에 찰리를 명확히 노출했다."
        ],
        "physics": "차량은 앞뒤 타이어로 노면에 지지되고, 앞바퀴는 산길 진입 방향으로 돌아가 있다. 뒤쪽 접지부에서 왼쪽 아래로 퍼지는 흙먼지가 차량의 급격한 방향 전환과 연결된다. 보급품은 실내에 놓여 있으며 떠 있는 물체는 보이지 않는다. 창문 속 인물의 하체는 실내에 가려져 있어 지지 없는 부양으로 판단할 근거가 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "검문소와 산길의 배치는 맞지만 차량 앞머리가 산길 반대쪽을 향하며, 지시하지 않은 검문 인원과 창문 속 인물을 추가했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "차량의 앞머리와 흙먼지가 우상단 산길로 급선회하는 동작을 제대로 설명하지만, 인물이 보이지 않아야 하는 샷에 찰리를 노출했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "운전석과 앞바퀴가 있는 차량 앞부분은 화면 왼쪽의 포장도로를 향하고, 후면은 오른쪽 아래를 향한다. 목적지인 산길 입구는 오른쪽 위에 있어 차량의 전진 방향과 맞지 않는다. 후면 창문의 인물은 카메라 쪽으로 고개를 돌리고 있다. 검문소 인원들의 정확한 시선은 거리상 확인하기 어렵다.",
        "built_space": "포장도로는 왼쪽 위 검문소로 이어지고 비포장 산길 하나는 오른쪽 위로 올라간다. 검문소에는 양쪽 초소 두 동과 노란 차단시설들이 보인다. 주변에는 황량한 밭, 잎 없는 나무, 낡은 주택과 전신주가 있다. 높은 외부 시점은 허용된 구도지만 차량이 화면 오른쪽 절반을 크게 차지해 교차로보다 강조된다.",
        "entities": "낡고 얼룩진 캠핑카, 지붕에 묶인 보급품, 타이어 주변 흙먼지, 먼 검문소와 산길이 보인다. 후면 창문에는 고릴라 같은 얼굴과 기계식 몸체를 가진 인물이 있으나, 지정된 흰 얼굴판·점 형태 눈·선 형태 입보다는 사실적인 고릴라 얼굴에 가깝다. 검문소에는 최소 일곱 명가량의 인원이 추가되어 있다. 앰버는 보이지 않으며 이 구도에서는 문제가 아니다. 낮 조명은 충족한다.",
        "hard_violations": [
         "샷 텍스트가 가시적 인물을 요구하지 않았는데 검문소에 다수의 인원을 추가했다.",
         "장면 수준의 인물 목록만으로 후면 창문에 찰리로 보이는 인물을 노출해, 이 샷에 인물을 추가하지 말라는 지시를 위반했다."
        ],
        "physics": "차량은 보이는 앞뒤 타이어로 지면에 지지되고, 보급품은 지붕 적재대와 끈으로 고정되어 있다. 창문 속 인물의 하체는 차체에 가려져 있으므로 공중에 떠 있다고 볼 근거는 없다. 타이어 접지부에서 먼지가 일어나는 것은 가능하지만, 차체 방향은 요청된 산길 진입이 아니라 포장도로 쪽으로 빠져나오는 동작에 가깝다."
       },
       {
        "label": "A",
        "direction": "차량의 뒤가 왼쪽 아래, 운전실과 앞머리가 오른쪽 위를 향한다. 앞머리 바로 앞에 비포장 산길이 이어져 목적지가 명확하다. 포기한 포장도로는 왼쪽 위의 먼 검문소로 이어진다. 측면 창문의 흰 얼굴은 카메라 쪽을 보고 있다.",
        "built_space": "하나의 포장도로에서 하나의 비포장 산길이 오른쪽 위로 갈라지는 구조다. 먼 검문소에는 작은 초소 한 동과 도로를 가로막는 차단시설이 보인다. 왼쪽 배경에는 주택 두 동, 황량한 밭과 잎 없는 나무가 있다. 차량의 후면과 한쪽 측면을 보여주는 외부 와이드 구도이며, 차량이 다소 크기는 해도 분기점과 두 경로가 모두 읽힌다.",
        "entities": "낡고 녹슨 캠핑카, 뒤쪽 창문 너머의 보급품, 흙먼지, 산길 입구와 먼 검문소가 보인다. 측면 창문에는 흰 얼굴판과 어두운 외장을 가진 찰리로 보이는 존재가 있으며, A보다 지정된 기계 머리에 가깝다. 앰버는 보이지 않는다. 누수의 지속 여부와 정확한 뒷좌석 착석 상태는 외부 구도만으로 확인할 수 없으며, 이를 누락으로 판단할 수는 없다. 낮의 자연광은 맞는다.",
        "hard_violations": [
         "샷 텍스트가 인물을 보여주도록 지정하지 않았는데 측면 창문에 찰리를 명확히 노출했다."
        ],
        "physics": "차량은 앞뒤 타이어로 노면에 지지되고, 앞바퀴는 산길 진입 방향으로 돌아가 있다. 뒤쪽 접지부에서 왼쪽 아래로 퍼지는 흙먼지가 차량의 급격한 방향 전환과 연결된다. 보급품은 실내에 놓여 있으며 떠 있는 물체는 보이지 않는다. 창문 속 인물의 하체는 실내에 가려져 있어 지지 없는 부양으로 판단할 근거가 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.829
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.579
   },
   "violations": {
    "B": [
     "[gemini-pro] 샷 텍스트에 없는 인물(검문소 경비원들) 추가 (invented people)",
     "[gpt-high] 샷 텍스트가 가시적 인물을 요구하지 않았는데 검문소에 다수의 인원을 추가했다.",
     "[gpt-high] 장면 수준의 인물 목록만으로 후면 창문에 찰리로 보이는 인물을 노출해, 이 샷에 인물을 추가하지 말라는 지시를 위반했다."
    ],
    "A": [
     "[gpt-high] 샷 텍스트가 인물을 보여주도록 지정하지 않았는데 측면 창문에 찰리를 명확히 노출했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 579
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지시된 구도를 정확히 따랐으며 찰리의 기계적인 얼굴 특징(흰 얼굴판, 점 눈)을 잘 구현했으나, 번호판에 일부 읽을 수 있는 문자가 노출되어 감점됨.  ★위반: [gpt-high] 샷 텍스트가 인물을 보여주도록 지정하지 않았는데 측면 창문에 찰리를 명확히 노출했다."
   },
   {
    "label": "B",
    "score": 579,
    "verdict_ko": "검문소에 샷 텍스트에 없는 인물(경비원)을 추가하는 치명적 위반을 범했고, 찰리가 명시된 기계 형태가 아닌 생물학적 고릴라로 그려짐.  ★위반: [gemini-pro] 샷 텍스트에 없는 인물(검문소 경비원들) 추가 (invented people) / [gpt-high] 샷 텍스트가 가시적 인물을 요구하지 않았는데 검문소에 다수의 인원을 추가했다. / [gpt-high] 장면 수준의 인물 목록만으로 후면 창문에 찰리로 보이는 인물을 노출해, 이 샷에 인물을 추가하지 말라는 지시를 위반했다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd6-0b68-732b-ab0f-41c202c9474a",
  "ref_mode": "lane(map_marker): 스케치",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S48sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:42:40.721805+00:00",
  "fingerprint": "9baa95419f3a1fc01b9b29b860633d3f632b2776248e1716a8c5570f078d35a1",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S48sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S48sh5_sel.png",
  "source_sha256": "3fbe5146e3bd7e9d02f18e02a1bb46dd8cd9579064f45ccd4d7f9f185046b914",
  "file": "S48sh5_cine.png",
  "staged_sha256": "f2cb651e7e816fe79c20895f3224eb59983ebfdd92cca9daf079482e8f34db39",
  "latency_ms": 12429
 },
 "S48sh15::signage": {
  "fp": "d72ae93a05ad0502",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S48sh15": {
  "input_fingerprint": "86822fc3fe8842cf",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 운전하는 캠핑카가 흙먼지를 내며 산길 아래 도로변을 향해 차체를 덜컹 기울인 채 나아가는 mid-action 후면 외관.\n\nLOCATION (lock): On the descending dirt mountain track leading back toward the roadside below. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camper (Descending with its body tilted mid-rock) — Rear and one flank are visible as it travels away toward upper-left; used as The receding exterior carries the departure; no artificial view through the body reveals the driver; Descending mountain path (Leading toward the roadside below) — Runs from the lower foreground toward upper-left in screen space; used as Provides a continuous depth line for the departure; Earth dust (Trailing the moving camper); used as Marks the route just traveled without obscuring the rear silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain consistent daytime ambient illumination, using restrained contrast to separate the rocking body from the descending path.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper's loose wheel bolts have been tightened, but its tire is still punctured and its fuel is nearly exhausted; the earlier roof and window-frame leaks remain unrepaired. The camper is descending toward the roadside while Charlie travels down the mountain path on foot.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 운전하는 캠핑카가 흙먼지를 내며 산길 아래 도로변을 향해 차체를 덜컹 기울인 채 나아가는 mid-action 후면 외관.\n\nLOCATION (lock): On the descending dirt mountain track leading back toward the roadside below. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camper (Descending with its body tilted mid-rock) — Rear and one flank are visible as it travels away toward upper-left; used as The receding exterior carries the departure; no artificial view through the body reveals the driver; Descending mountain path (Leading toward the roadside below) — Runs from the lower foreground toward upper-left in screen space; used as Provides a continuous depth line for the departure; Earth dust (Trailing the moving camper); used as Marks the route just traveled without obscuring the rear silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain consistent daytime ambient illumination, using restrained contrast to separate the rocking body from the descending path.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper's loose wheel bolts have been tightened, but its tire is still punctured and its fuel is nearly exhausted; the earlier roof and window-frame leaks remain unrepaired. The camper is descending toward the roadside while Charlie travels down the mountain path on foot.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 운전하는 캠핑카가 흙먼지를 내며 산길 아래 도로변을 향해 차체를 덜컹 기울인 채 나아가는 mid-action 후면 외관.\n\nLOCATION (lock): On the descending dirt mountain track leading back toward the roadside below. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camper (Descending with its body tilted mid-rock) — Rear and one flank are visible as it travels away toward upper-left; used as The receding exterior carries the departure; no artificial view through the body reveals the driver; Descending mountain path (Leading toward the roadside below) — Runs from the lower foreground toward upper-left in screen space; used as Provides a continuous depth line for the departure; Earth dust (Trailing the moving camper); used as Marks the route just traveled without obscuring the rear silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain consistent daytime ambient illumination, using restrained contrast to separate the rocking body from the descending path.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper's loose wheel bolts have been tightened, but its tire is still punctured and its fuel is nearly exhausted; the earlier roof and window-frame leaks remain unrepaired. The camper is descending toward the roadside while Charlie travels down the mountain path on foot.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "캠핑카가 목표인 우측 하단 도로변이 아니라 좌측 상단의 산 위쪽으로 이동함.",
    "built_space": "흙길이 화면 하단에서 좌측 상단으로 뻗어 있으며, 내리막이 아닌 명백한 오르막 지형을 이룸. 우측 하단에 포장도로가 위치함.",
    "entities": "캠핑카(후면 및 좌측면), 흙먼지. 펑크 난 타이어는 묘사되지 않음.",
    "hard_violations": [
     "[gemini-pro] 수직 방향 위반: 지시된 '내리막길(Descending)'과 반대로 오르막길을 렌더링함.",
     "[gpt-high] 하산해야 하는 진행 경로를 왼쪽 위 산비탈로 올라가는 오르막으로 구현했습니다.",
     "[gpt-high] 장소 설명에 없는 도시 건물군, 도로 교통과 도로변 시설을 추가했습니다."
    ],
    "physics": "캠핑카의 바퀴가 노면에 안착해 지탱되며 흙먼지를 일으킴. 울퉁불퉁한 지면으로 인해 차체가 우측으로 크게 기울어짐."
   },
   {
    "label": "B",
    "direction": "캠핑카가 아래쪽 도로변이 아닌 좌측 상단 오르막으로 향함. 찰리는 화면 앞을 향해 걸어 내려옴.",
    "built_space": "흙길이 좌측 상단으로 이어지는 오르막 지형. 우측 멀리 계곡 아래에 도로와 구조물들이 있음.",
    "entities": "캠핑카(좌측 뒷바퀴 펑크 묘사됨), 흙먼지. 샷 텍스트에 없는 인물인 찰리(고릴라 형태 기계 머리)가 존재함.",
    "hard_violations": [
     "[gemini-pro] 수직 방향 위반: 지시된 '내리막길'과 반대로 오르막길을 렌더링함.",
     "[gemini-pro] 발명된 인물: 샷 텍스트에 없는 인물(찰리)을 임의로 화면에 추가함.",
     "[gpt-high] 내려가야 하는 차량의 진행 경로를 왼쪽 위 능선으로 올라가는 오르막으로 구현했습니다.",
     "[gpt-high] 이 후면 외관 숏에서 보이도록 지정되지 않은 찰리 형태의 인물을 추가했습니다.",
     "[gpt-high] 장소 설명에 없는 도로 주변 건물들을 추가했습니다."
    ],
    "physics": "바람 빠진 뒷바퀴가 지면에 눌려 차체를 지탱함. 찰리의 두 발이 흙길을 딛고 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "샷 텍스트의 내리막길 지시를 어기고 오르막길로 표현한 심각한 오류가 있으나, 금지된 인물 추가를 피하여 B보다 낫습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "샷 텍스트에 명시되지 않은 인물(찰리)을 임의로 등장시켜 인물 렌더링 규칙을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카가 목표인 우측 하단 도로변이 아니라 좌측 상단의 산 위쪽으로 이동함.",
        "built_space": "흙길이 화면 하단에서 좌측 상단으로 뻗어 있으며, 내리막이 아닌 명백한 오르막 지형을 이룸. 우측 하단에 포장도로가 위치함.",
        "entities": "캠핑카(후면 및 좌측면), 흙먼지. 펑크 난 타이어는 묘사되지 않음.",
        "hard_violations": [
         "수직 방향 위반: 지시된 '내리막길(Descending)'과 반대로 오르막길을 렌더링함."
        ],
        "physics": "캠핑카의 바퀴가 노면에 안착해 지탱되며 흙먼지를 일으킴. 울퉁불퉁한 지면으로 인해 차체가 우측으로 크게 기울어짐."
       },
       {
        "label": "B",
        "direction": "캠핑카가 아래쪽 도로변이 아닌 좌측 상단 오르막으로 향함. 찰리는 화면 앞을 향해 걸어 내려옴.",
        "built_space": "흙길이 좌측 상단으로 이어지는 오르막 지형. 우측 멀리 계곡 아래에 도로와 구조물들이 있음.",
        "entities": "캠핑카(좌측 뒷바퀴 펑크 묘사됨), 흙먼지. 샷 텍스트에 없는 인물인 찰리(고릴라 형태 기계 머리)가 존재함.",
        "hard_violations": [
         "수직 방향 위반: 지시된 '내리막길'과 반대로 오르막길을 렌더링함.",
         "발명된 인물: 샷 텍스트에 없는 인물(찰리)을 임의로 화면에 추가함."
        ],
        "physics": "바람 빠진 뒷바퀴가 지면에 눌려 차체를 지탱함. 찰리의 두 발이 흙길을 딛고 서 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "샷 텍스트의 내리막길 지시를 어기고 오르막길로 표현한 심각한 오류가 있으나, 금지된 인물 추가를 피하여 B보다 낫습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "샷 텍스트에 명시되지 않은 인물(찰리)을 임의로 등장시켜 인물 렌더링 규칙을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카가 목표인 우측 하단 도로변이 아니라 좌측 상단의 산 위쪽으로 이동함.",
        "built_space": "흙길이 화면 하단에서 좌측 상단으로 뻗어 있으며, 내리막이 아닌 명백한 오르막 지형을 이룸. 우측 하단에 포장도로가 위치함.",
        "entities": "캠핑카(후면 및 좌측면), 흙먼지. 펑크 난 타이어는 묘사되지 않음.",
        "hard_violations": [
         "수직 방향 위반: 지시된 '내리막길(Descending)'과 반대로 오르막길을 렌더링함."
        ],
        "physics": "캠핑카의 바퀴가 노면에 안착해 지탱되며 흙먼지를 일으킴. 울퉁불퉁한 지면으로 인해 차체가 우측으로 크게 기울어짐."
       },
       {
        "label": "B",
        "direction": "캠핑카가 아래쪽 도로변이 아닌 좌측 상단 오르막으로 향함. 찰리는 화면 앞을 향해 걸어 내려옴.",
        "built_space": "흙길이 좌측 상단으로 이어지는 오르막 지형. 우측 멀리 계곡 아래에 도로와 구조물들이 있음.",
        "entities": "캠핑카(좌측 뒷바퀴 펑크 묘사됨), 흙먼지. 샷 텍스트에 없는 인물인 찰리(고릴라 형태 기계 머리)가 존재함.",
        "hard_violations": [
         "수직 방향 위반: 지시된 '내리막길'과 반대로 오르막길을 렌더링함.",
         "발명된 인물: 샷 텍스트에 없는 인물(찰리)을 임의로 화면에 추가함."
        ],
        "physics": "바람 빠진 뒷바퀴가 지면에 눌려 차체를 지탱함. 찰리의 두 발이 흙길을 딛고 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "후면·측면과 흙먼지는 맞지만, 차량이 내리막 대신 오르막을 향하고 금지된 추가 인물까지 보여 장면을 위반합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "와이드 구도와 덜컹 기울어진 차체는 더 충실하지만, 진행 경사가 반대이며 지정되지 않은 도시와 교통을 추가해 재촬영이 필요합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카의 앞부분은 화면 왼쪽 위로 향하고 후면은 카메라를 향합니다. 먼지는 뒤쪽인 오른쪽 아래로 이어집니다. 그러나 차량 앞의 길은 왼쪽 위 능선으로 올라가므로, 오른쪽 아래에 보이는 도로변으로 내려가는 진행이 아닙니다. 능선 위 기계형 인물은 대체로 카메라 쪽을 향합니다.",
        "built_space": "흙길이 아래 전경에서 차량을 지나 왼쪽 위 능선으로 연결됩니다. 포장도로는 오른쪽 골짜기에 있으며 흙길의 현재 진행 방향과 연결되는 하행 구간은 보이지 않습니다. 도로 주변에는 지정되지 않은 작은 건물 여러 채가 있습니다. 차량에는 큰 후면 창 하나, 측면 창 두 개, 좌우 후미등 묶음이 보이며 불가능한 투시나 반사는 없습니다.",
        "entities": "낡은 캠핑카 한 대와 흙먼지가 있으며 후면과 한쪽 측면이 보입니다. 화면에서 차량이 상당히 크게 잡혔습니다. 왼쪽 뒤 타이어는 아래가 눌려 펑크 상태에 부합합니다. 창틀과 차체의 얼룩은 보이나 누수 여부나 연료량, 볼트 조임은 확인할 수 없습니다. 운전자와 손은 외관 구도에 가려져 신원을 확인할 수 없습니다. 왼쪽 위에는 흰 얼굴판과 기계형 머리를 가진 찰리처럼 보이는 인물이 추가되어 있습니다. 낮이며 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "내려가야 하는 차량의 진행 경로를 왼쪽 위 능선으로 올라가는 오르막으로 구현했습니다.",
         "이 후면 외관 숏에서 보이도록 지정되지 않은 찰리 형태의 인물을 추가했습니다.",
         "장소 설명에 없는 도로 주변 건물들을 추가했습니다."
        ],
        "physics": "차량의 보이는 바퀴들은 울퉁불퉁한 지면에 닿아 차체를 지탱하고, 눌린 뒤 타이어도 지면과 접촉합니다. 먼지는 바퀴 뒤에서 발생해 이동 흔적으로 성립합니다. 차체의 좌우 기울기는 비교적 약하지만 부유하지는 않습니다. 추가 인물도 발이 길에 닿아 있으며 공중에 떠 있지는 않습니다."
       },
       {
        "label": "B",
        "direction": "캠핑카는 후면과 왼쪽 측면을 보이며 화면 왼쪽 위로 멀어집니다. 흙먼지는 오른쪽 아래 후방으로 흘러 이동 방향과 맞습니다. 다만 앞길은 왼쪽 위 산비탈을 올라가고, 목적지여야 할 낮은 도로는 오른쪽 아래에 있어 하산 방향과 맞지 않습니다.",
        "built_space": "아래 전경의 흙길이 캠핑카를 지나 왼쪽 위로 이어지는 와이드 구도입니다. 오른쪽 아래에는 포장도로가 있고 그 주변에 여러 전신주 또는 가로등, 차량, 건물과 아파트군이 배치되어 있습니다. 이 도시 배경은 장소 설명에 없습니다. 캠핑카의 후면 창 하나와 측면 창 두 개, 좌우 후미등이 보이며 중복된 차체나 불가능한 실내 투시는 없습니다.",
        "entities": "낡은 캠핑카 한 대가 중경에 있고 사람은 보이지 않습니다. 운전자 이현우와 손은 외관 프레이밍상 확인할 수 없으며 차체를 투시해 보여주지 않습니다. 차량 규모와 주변 공간은 와이드 숏에 더 잘 맞고 차체 기울기도 뚜렷합니다. 타이어의 확실한 펑크는 판별하기 어렵고, 볼트·연료량·누수 상태 역시 확인되지 않습니다. 낮의 자연광과 후면 윤곽을 가리지 않는 흙먼지는 맞습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "하산해야 하는 진행 경로를 왼쪽 위 산비탈로 올라가는 오르막으로 구현했습니다.",
         "장소 설명에 없는 도시 건물군, 도로 교통과 도로변 시설을 추가했습니다."
        ],
        "physics": "차량은 바퀴와 울퉁불퉁한 노면의 접촉으로 지지됩니다. 차체의 좌우 기울기는 바퀴 높이 차이와 서스펜션의 흔들림으로 가능한 자세이며, 차량 전체가 근거 없이 공중에 떠 있지는 않습니다. 바퀴 주변에서 후방으로 퍼지는 먼지도 주행으로 발생할 수 있습니다. 배경 차량들은 포장도로 위에 놓여 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "후면·측면과 흙먼지는 맞지만, 차량이 내리막 대신 오르막을 향하고 금지된 추가 인물까지 보여 장면을 위반합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "와이드 구도와 덜컹 기울어진 차체는 더 충실하지만, 진행 경사가 반대이며 지정되지 않은 도시와 교통을 추가해 재촬영이 필요합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카의 앞부분은 화면 왼쪽 위로 향하고 후면은 카메라를 향합니다. 먼지는 뒤쪽인 오른쪽 아래로 이어집니다. 그러나 차량 앞의 길은 왼쪽 위 능선으로 올라가므로, 오른쪽 아래에 보이는 도로변으로 내려가는 진행이 아닙니다. 능선 위 기계형 인물은 대체로 카메라 쪽을 향합니다.",
        "built_space": "흙길이 아래 전경에서 차량을 지나 왼쪽 위 능선으로 연결됩니다. 포장도로는 오른쪽 골짜기에 있으며 흙길의 현재 진행 방향과 연결되는 하행 구간은 보이지 않습니다. 도로 주변에는 지정되지 않은 작은 건물 여러 채가 있습니다. 차량에는 큰 후면 창 하나, 측면 창 두 개, 좌우 후미등 묶음이 보이며 불가능한 투시나 반사는 없습니다.",
        "entities": "낡은 캠핑카 한 대와 흙먼지가 있으며 후면과 한쪽 측면이 보입니다. 화면에서 차량이 상당히 크게 잡혔습니다. 왼쪽 뒤 타이어는 아래가 눌려 펑크 상태에 부합합니다. 창틀과 차체의 얼룩은 보이나 누수 여부나 연료량, 볼트 조임은 확인할 수 없습니다. 운전자와 손은 외관 구도에 가려져 신원을 확인할 수 없습니다. 왼쪽 위에는 흰 얼굴판과 기계형 머리를 가진 찰리처럼 보이는 인물이 추가되어 있습니다. 낮이며 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "내려가야 하는 차량의 진행 경로를 왼쪽 위 능선으로 올라가는 오르막으로 구현했습니다.",
         "이 후면 외관 숏에서 보이도록 지정되지 않은 찰리 형태의 인물을 추가했습니다.",
         "장소 설명에 없는 도로 주변 건물들을 추가했습니다."
        ],
        "physics": "차량의 보이는 바퀴들은 울퉁불퉁한 지면에 닿아 차체를 지탱하고, 눌린 뒤 타이어도 지면과 접촉합니다. 먼지는 바퀴 뒤에서 발생해 이동 흔적으로 성립합니다. 차체의 좌우 기울기는 비교적 약하지만 부유하지는 않습니다. 추가 인물도 발이 길에 닿아 있으며 공중에 떠 있지는 않습니다."
       },
       {
        "label": "A",
        "direction": "캠핑카는 후면과 왼쪽 측면을 보이며 화면 왼쪽 위로 멀어집니다. 흙먼지는 오른쪽 아래 후방으로 흘러 이동 방향과 맞습니다. 다만 앞길은 왼쪽 위 산비탈을 올라가고, 목적지여야 할 낮은 도로는 오른쪽 아래에 있어 하산 방향과 맞지 않습니다.",
        "built_space": "아래 전경의 흙길이 캠핑카를 지나 왼쪽 위로 이어지는 와이드 구도입니다. 오른쪽 아래에는 포장도로가 있고 그 주변에 여러 전신주 또는 가로등, 차량, 건물과 아파트군이 배치되어 있습니다. 이 도시 배경은 장소 설명에 없습니다. 캠핑카의 후면 창 하나와 측면 창 두 개, 좌우 후미등이 보이며 중복된 차체나 불가능한 실내 투시는 없습니다.",
        "entities": "낡은 캠핑카 한 대가 중경에 있고 사람은 보이지 않습니다. 운전자 이현우와 손은 외관 프레이밍상 확인할 수 없으며 차체를 투시해 보여주지 않습니다. 차량 규모와 주변 공간은 와이드 숏에 더 잘 맞고 차체 기울기도 뚜렷합니다. 타이어의 확실한 펑크는 판별하기 어렵고, 볼트·연료량·누수 상태 역시 확인되지 않습니다. 낮의 자연광과 후면 윤곽을 가리지 않는 흙먼지는 맞습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "하산해야 하는 진행 경로를 왼쪽 위 산비탈로 올라가는 오르막으로 구현했습니다.",
         "장소 설명에 없는 도시 건물군, 도로 교통과 도로변 시설을 추가했습니다."
        ],
        "physics": "차량은 바퀴와 울퉁불퉁한 노면의 접촉으로 지지됩니다. 차체의 좌우 기울기는 바퀴 높이 차이와 서스펜션의 흔들림으로 가능한 자세이며, 차량 전체가 근거 없이 공중에 떠 있지는 않습니다. 바퀴 주변에서 후방으로 퍼지는 먼지도 주행으로 발생할 수 있습니다. 배경 차량들은 포장도로 위에 놓여 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.417
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.167
   },
   "violations": {
    "A": [
     "[gemini-pro] 수직 방향 위반: 지시된 '내리막길(Descending)'과 반대로 오르막길을 렌더링함.",
     "[gpt-high] 하산해야 하는 진행 경로를 왼쪽 위 산비탈로 올라가는 오르막으로 구현했습니다.",
     "[gpt-high] 장소 설명에 없는 도시 건물군, 도로 교통과 도로변 시설을 추가했습니다."
    ],
    "B": [
     "[gemini-pro] 수직 방향 위반: 지시된 '내리막길'과 반대로 오르막길을 렌더링함.",
     "[gemini-pro] 발명된 인물: 샷 텍스트에 없는 인물(찰리)을 임의로 화면에 추가함.",
     "[gpt-high] 내려가야 하는 차량의 진행 경로를 왼쪽 위 능선으로 올라가는 오르막으로 구현했습니다.",
     "[gpt-high] 이 후면 외관 숏에서 보이도록 지정되지 않은 찰리 형태의 인물을 추가했습니다.",
     "[gpt-high] 장소 설명에 없는 도로 주변 건물들을 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1167
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "샷 텍스트의 내리막길 지시를 어기고 오르막길로 표현한 심각한 오류가 있으나, 금지된 인물 추가를 피하여 B보다 낫습니다.  ★위반: [gemini-pro] 수직 방향 위반: 지시된 '내리막길(Descending)'과 반대로 오르막길을 렌더링함. / [gpt-high] 하산해야 하는 진행 경로를 왼쪽 위 산비탈로 올라가는 오르막으로 구현했습니다. / [gpt-high] 장소 설명에 없는 도시 건물군, 도로 교통과 도로변 시설을 추가했습니다."
   },
   {
    "label": "B",
    "score": 1167,
    "verdict_ko": "샷 텍스트에 명시되지 않은 인물(찰리)을 임의로 등장시켜 인물 렌더링 규칙을 심각하게 위반했습니다.  ★위반: [gemini-pro] 수직 방향 위반: 지시된 '내리막길'과 반대로 오르막길을 렌더링함. / [gemini-pro] 발명된 인물: 샷 텍스트에 없는 인물(찰리)을 임의로 화면에 추가함. / [gpt-high] 내려가야 하는 차량의 진행 경로를 왼쪽 위 능선으로 올라가는 오르막으로 구현했습니다. / [gpt-high] 이 후면 외관 숏에서 보이도록 지정되지 않은 찰리 형태의 인물을 추가했습니다. / [gpt-high] 장소 설명에 없는 도로 주변 건물들을 추가했습니다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd6-d79a-73d2-a6b4-a8ceb1a1c04d",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S48sh15::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:42:53.485643+00:00",
  "fingerprint": "4a818a7a4cb5662951c847ea5ffef81baa11f48784b37993c061a1cdb4c5126e",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S48sh15_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S48sh15_sel.png",
  "source_sha256": "63dd3aaad91d09da7e88790fe146725e03e65391a6025b31946764cdc9544f52",
  "file": "S48sh15_cine.png",
  "staged_sha256": "768a9f05418f54c6519ace4566b2cdc9933cb4c32a170f911494c8d02970f438",
  "latency_ms": 11263
 },
 "S48sh19::signage": {
  "fp": "9f79dc2a40bc0194",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S48sh19": {
  "input_fingerprint": "4bf3503bb2a51598",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 굵고 둥근 금속 손바닥 위로 앰버의 작은 두 손이 단단히 포개어져 맞잡힌 손끝 클로즈업.\n\nLOCATION (lock): On the mountain footpath descending beside the camper's route toward the road. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Mountain path (Visible beneath the joined hands); used as A softly resolved strip maintains the walking scene's continuity without competing with the hand gesture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime ambient light gives gentle separation to 앰버's hands and 찰리's rounded metal palm without adding a sentimental color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is being taken down toward the roadside with tightened wheel bolts, a punctured tire and very little fuel. Charlie is on the descending mountain path with one hand extended for the promise. 앰버: She is on the mountain path with her hands extended for a hand-clasp promise, wearing the replacement shoes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 굵고 둥근 금속 손바닥 위로 앰버의 작은 두 손이 단단히 포개어져 맞잡힌 손끝 클로즈업.\n\nLOCATION (lock): On the mountain footpath descending beside the camper's route toward the road. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Mountain path (Visible beneath the joined hands); used as A softly resolved strip maintains the walking scene's continuity without competing with the hand gesture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime ambient light gives gentle separation to 앰버's hands and 찰리's rounded metal palm without adding a sentimental color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is being taken down toward the roadside with tightened wheel bolts, a punctured tire and very little fuel. Charlie is on the descending mountain path with one hand extended for the promise. 앰버: She is on the mountain path with her hands extended for a hand-clasp promise, wearing the replacement shoes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 굵고 둥근 금속 손바닥 위로 앰버의 작은 두 손이 단단히 포개어져 맞잡힌 손끝 클로즈업.\n\nLOCATION (lock): On the mountain footpath descending beside the camper's route toward the road. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Mountain path (Visible beneath the joined hands); used as A softly resolved strip maintains the walking scene's continuity without competing with the hand gesture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime ambient light gives gentle separation to 앰버's hands and 찰리's rounded metal palm without adding a sentimental color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is being taken down toward the roadside with tightened wheel bolts, a punctured tire and very little fuel. Charlie is on the descending mountain path with one hand extended for the promise. 앰버: She is on the mountain path with her hands extended for a hand-clasp promise, wearing the replacement shoes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인간의 손이 위를 향해 금속 손가락을 맞잡고 있음.",
    "built_space": "배경에 야외 산길(흙길)이 흐릿하게 보이며 특정 구조물은 없음.",
    "entities": "찰리의 금속 손, 앰버의 녹색 재킷 소매가 확인되나 인간의 손가락 형태가 심하게 붕괴됨.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (손가락 개수 초과 및 기형적인 융합)"
    ],
    "physics": "프레임 밖에서 이어진 팔이 손을 지지하며 공중에서 맞잡고 있음."
   },
   {
    "label": "B",
    "direction": "두 개의 인간 손이 금속 손바닥 위에 포개어져 놓여 있음.",
    "built_space": "흐릿하게 처리된 산길 배경이 샷 텍스트의 야외 환경과 일치함.",
    "entities": "찰리의 금속 팔, 앰버의 녹색 소매 팔이 확인되나 출처 불명의 맨팔(오른손)이 하나 더 존재함.",
    "hard_violations": [
     "[gemini-pro] 발명된 인물 및 물리적으로 불가능한 해부학 (앰버의 양손이어야 할 두 손이, 복장과 각도가 다르고 둘 다 오른손인 서로 다른 사람의 팔로 묘사됨)"
    ],
    "physics": "프레임 밖에서 이어진 팔들에 의해 지지된 두 손이 금속 손바닥 위에 얹혀 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "금속 손바닥 위로 두 손이 포개진 구도는 구현했으나, 두 팔이 서로 다른 인물의 오른손(복장과 각도 불일치)으로 렌더링되는 치명적인 해부학적 오류가 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "두 손이 포개진 샷의 핵심 지시를 따르지 못했으며, 하나의 손목에서 여러 손가락이 기형적으로 자라난 심각한 해부학적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인간의 손이 위를 향해 금속 손가락을 맞잡고 있음.",
        "built_space": "배경에 야외 산길(흙길)이 흐릿하게 보이며 특정 구조물은 없음.",
        "entities": "찰리의 금속 손, 앰버의 녹색 재킷 소매가 확인되나 인간의 손가락 형태가 심하게 붕괴됨.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (손가락 개수 초과 및 기형적인 융합)"
        ],
        "physics": "프레임 밖에서 이어진 팔이 손을 지지하며 공중에서 맞잡고 있음."
       },
       {
        "label": "B",
        "direction": "두 개의 인간 손이 금속 손바닥 위에 포개어져 놓여 있음.",
        "built_space": "흐릿하게 처리된 산길 배경이 샷 텍스트의 야외 환경과 일치함.",
        "entities": "찰리의 금속 팔, 앰버의 녹색 소매 팔이 확인되나 출처 불명의 맨팔(오른손)이 하나 더 존재함.",
        "hard_violations": [
         "발명된 인물 및 물리적으로 불가능한 해부학 (앰버의 양손이어야 할 두 손이, 복장과 각도가 다르고 둘 다 오른손인 서로 다른 사람의 팔로 묘사됨)"
        ],
        "physics": "프레임 밖에서 이어진 팔들에 의해 지지된 두 손이 금속 손바닥 위에 얹혀 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "금속 손바닥 위로 두 손이 포개진 구도는 구현했으나, 두 팔이 서로 다른 인물의 오른손(복장과 각도 불일치)으로 렌더링되는 치명적인 해부학적 오류가 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "두 손이 포개진 샷의 핵심 지시를 따르지 못했으며, 하나의 손목에서 여러 손가락이 기형적으로 자라난 심각한 해부학적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "인간의 손이 위를 향해 금속 손가락을 맞잡고 있음.",
        "built_space": "배경에 야외 산길(흙길)이 흐릿하게 보이며 특정 구조물은 없음.",
        "entities": "찰리의 금속 손, 앰버의 녹색 재킷 소매가 확인되나 인간의 손가락 형태가 심하게 붕괴됨.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (손가락 개수 초과 및 기형적인 융합)"
        ],
        "physics": "프레임 밖에서 이어진 팔이 손을 지지하며 공중에서 맞잡고 있음."
       },
       {
        "label": "B",
        "direction": "두 개의 인간 손이 금속 손바닥 위에 포개어져 놓여 있음.",
        "built_space": "흐릿하게 처리된 산길 배경이 샷 텍스트의 야외 환경과 일치함.",
        "entities": "찰리의 금속 팔, 앰버의 녹색 소매 팔이 확인되나 출처 불명의 맨팔(오른손)이 하나 더 존재함.",
        "hard_violations": [
         "발명된 인물 및 물리적으로 불가능한 해부학 (앰버의 양손이어야 할 두 손이, 복장과 각도가 다르고 둘 다 오른손인 서로 다른 사람의 팔로 묘사됨)"
        ],
        "physics": "프레임 밖에서 이어진 팔들에 의해 지지된 두 손이 금속 손바닥 위에 얹혀 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "찰리의 넓은 금속 손 위에 앰버의 두 손이 단단히 포개진 순간을 밀착 클로즈업으로 보여 주어 핵심 동작과 구도에 가장 충실하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "손의 접촉과 산길 배경은 맞지만, 세운 금속 손 안에서 서로 감싸 쥔 형태여서 손바닥 위로 두 손을 포개는 지정 동작이 A보다 덜 명확하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 금속 팔은 왼쪽에서 중앙으로 뻗고 손가락은 오른쪽을 향한다. 앰버의 두 손은 위쪽과 오른쪽에서 들어와 금속 손의 넓은 면 위에 아래 방향으로 포개진다. 두 손의 접촉 대상이 찰리의 손으로 명확하다. 얼굴과 시선은 화면 밖이다.",
        "built_space": "맞잡은 손 바로 아래로 흙과 작은 돌이 있는 산길이 흐릿하게 보인다. 이전 장면의 건조한 흙길 재질과 이어지며, 건축물이나 고정 설비는 보이지 않는다. 캠퍼와 도로는 손 클로즈업의 화면 밖이므로 누락으로 볼 수 없다.",
        "entities": "금속 손 하나와 사람의 손 두 개가 보인다. 찰리의 손과 팔에는 마모된 은회색 금속 외장과 관절이 있고, 앰버 쪽에는 참고 의상과 비슷한 연녹색 소매가 보인다. 사람 손은 금속 손보다 작지만 다소 두툼하게 표현되었다. 얼굴이 없어 앰버의 혼혈 정체성·금발·눈과 찰리의 머리 형태는 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "금속 손은 왼쪽의 손목과 팔에 연결되어 있고, 앰버의 두 손도 각각 손목과 팔로 이어진다. 아래쪽 사람 손은 금속 손에 닿고 위쪽 손은 그 손등을 누르므로 포개진 접촉과 지지가 자연스럽다. 떠 있는 물체나 불가능한 관절 배치는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 팔은 화면 아래에서 올라오고 금속 손가락은 위를 향해 굽어 있다. 앰버의 한 손은 왼쪽 아래에서 금속 손 안으로 들어가며 다른 손은 그 위를 가로질러 감싼다. 접촉 대상은 맞지만, 수평으로 내민 손바닥 위에 포개기보다 세운 손 안에서 움켜쥐는 동작으로 읽힌다. 시선은 보이지 않는다.",
        "built_space": "중앙 뒤로 흙과 자갈의 산길이 이어지고 양옆에 흐릿한 풀과 바위가 보인다. 장소의 재질은 이전 장면과 대체로 일치한다. 고정 설비나 구조물은 없으며, 배경 길의 노출 면적은 A보다 넓지만 손이 여전히 중심인 클로즈업이다.",
        "entities": "마모된 금속 손 하나와 작은 사람 손 두 개가 보이며, 왼쪽 아래의 연녹색 소매는 앰버의 참고 의상과 부합한다. 둥근 금속 손끝과 두꺼운 관절은 찰리의 기계 신체 설정에 맞는다. 손 주변에는 가는 실밥 같은 선도 보인다. 얼굴과 신발은 화면 밖이라 인종·머리·얼굴 특징과 교체 신발은 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "금속 손은 아래쪽 팔과 손목으로 지지되고 사람의 앞쪽 손은 보이는 팔에 연결된다. 다른 손의 손목은 겹친 손과 금속 손에 가려져 있지만, 보이는 손가락은 접촉면을 따라 자연스럽게 굽어 있다. 서로 감싸 잡는 자세는 물리적으로 가능하며, 가림만으로 손이 떠 있거나 분리되었다고 판단할 근거는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "찰리의 넓은 금속 손 위에 앰버의 두 손이 단단히 포개진 순간을 밀착 클로즈업으로 보여 주어 핵심 동작과 구도에 가장 충실하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "손의 접촉과 산길 배경은 맞지만, 세운 금속 손 안에서 서로 감싸 쥔 형태여서 손바닥 위로 두 손을 포개는 지정 동작이 A보다 덜 명확하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 금속 팔은 왼쪽에서 중앙으로 뻗고 손가락은 오른쪽을 향한다. 앰버의 두 손은 위쪽과 오른쪽에서 들어와 금속 손의 넓은 면 위에 아래 방향으로 포개진다. 두 손의 접촉 대상이 찰리의 손으로 명확하다. 얼굴과 시선은 화면 밖이다.",
        "built_space": "맞잡은 손 바로 아래로 흙과 작은 돌이 있는 산길이 흐릿하게 보인다. 이전 장면의 건조한 흙길 재질과 이어지며, 건축물이나 고정 설비는 보이지 않는다. 캠퍼와 도로는 손 클로즈업의 화면 밖이므로 누락으로 볼 수 없다.",
        "entities": "금속 손 하나와 사람의 손 두 개가 보인다. 찰리의 손과 팔에는 마모된 은회색 금속 외장과 관절이 있고, 앰버 쪽에는 참고 의상과 비슷한 연녹색 소매가 보인다. 사람 손은 금속 손보다 작지만 다소 두툼하게 표현되었다. 얼굴이 없어 앰버의 혼혈 정체성·금발·눈과 찰리의 머리 형태는 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "금속 손은 왼쪽의 손목과 팔에 연결되어 있고, 앰버의 두 손도 각각 손목과 팔로 이어진다. 아래쪽 사람 손은 금속 손에 닿고 위쪽 손은 그 손등을 누르므로 포개진 접촉과 지지가 자연스럽다. 떠 있는 물체나 불가능한 관절 배치는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 팔은 화면 아래에서 올라오고 금속 손가락은 위를 향해 굽어 있다. 앰버의 한 손은 왼쪽 아래에서 금속 손 안으로 들어가며 다른 손은 그 위를 가로질러 감싼다. 접촉 대상은 맞지만, 수평으로 내민 손바닥 위에 포개기보다 세운 손 안에서 움켜쥐는 동작으로 읽힌다. 시선은 보이지 않는다.",
        "built_space": "중앙 뒤로 흙과 자갈의 산길이 이어지고 양옆에 흐릿한 풀과 바위가 보인다. 장소의 재질은 이전 장면과 대체로 일치한다. 고정 설비나 구조물은 없으며, 배경 길의 노출 면적은 A보다 넓지만 손이 여전히 중심인 클로즈업이다.",
        "entities": "마모된 금속 손 하나와 작은 사람 손 두 개가 보이며, 왼쪽 아래의 연녹색 소매는 앰버의 참고 의상과 부합한다. 둥근 금속 손끝과 두꺼운 관절은 찰리의 기계 신체 설정에 맞는다. 손 주변에는 가는 실밥 같은 선도 보인다. 얼굴과 신발은 화면 밖이라 인종·머리·얼굴 특징과 교체 신발은 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "금속 손은 아래쪽 팔과 손목으로 지지되고 사람의 앞쪽 손은 보이는 팔에 연결된다. 다른 손의 손목은 겹친 손과 금속 손에 가려져 있지만, 보이는 손가락은 접촉면을 따라 자연스럽게 굽어 있다. 서로 감싸 잡는 자세는 물리적으로 가능하며, 가림만으로 손이 떠 있거나 분리되었다고 판단할 근거는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.528,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.278,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (손가락 개수 초과 및 기형적인 융합)"
    ],
    "B": [
     "[gemini-pro] 발명된 인물 및 물리적으로 불가능한 해부학 (앰버의 양손이어야 할 두 손이, 복장과 각도가 다르고 둘 다 오른손인 서로 다른 사람의 팔로 묘사됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 1278
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "금속 손바닥 위로 두 손이 포개진 구도는 구현했으나, 두 팔이 서로 다른 인물의 오른손(복장과 각도 불일치)으로 렌더링되는 치명적인 해부학적 오류가 있습니다.  ★위반: [gemini-pro] 발명된 인물 및 물리적으로 불가능한 해부학 (앰버의 양손이어야 할 두 손이, 복장과 각도가 다르고 둘 다 오른손인 서로 다른 사람의 팔로 묘사됨)"
   },
   {
    "label": "A",
    "score": 1278,
    "verdict_ko": "두 손이 포개진 샷의 핵심 지시를 따르지 못했으며, 하나의 손목에서 여러 손가락이 기형적으로 자라난 심각한 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (손가락 개수 초과 및 기형적인 융합)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S48sh15_sel.png",
    "asset_id": "22ef5ab9-e208-4324-81d1-92d10b7c811d",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:948699>",
    "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd7-9134-7289-8e7b-16582fb5c0a1",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S48sh15"
  }
 },
 "S48sh19::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:43:05.086764+00:00",
  "fingerprint": "19ec5c647957f62fd511c2593dc64841d3e4bf90ae3c4742d1b3916284fe1a66",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S48sh19_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S48sh19_sel.png",
  "source_sha256": "22d4ae4380cb912e88eb985f2e31be58ac5584ba4bb55f8442db4607982427e4",
  "file": "S48sh19_cine.png",
  "staged_sha256": "7e87269dd212f0445c7c1fe31f7ebb48a0266232555fc7b589dd8a6abb361eab",
  "latency_ms": 10463
 },
 "S49sh12::signage": {
  "fp": "2fe3679ac17c13cf",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::80e27c6d0d054474": {
  "subjects": [],
  "subject_text": "휴게소 마트 내부\n주유장 안쪽에 붙은 작은 마트. 상품 진열대와 약품 코너가 있고 테이블 주변의 텔레비전과 구석의 반사경이 보인다.",
  "identity": "canonical",
  "scope_id": "L61",
  "scope_role": "location_interior",
  "scope_sha": "45bfc2bc2eddbd19"
 },
 "S49sh12::bgfirst_bg": {
  "input_fingerprint": "55c1fcdc73691c25",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버, 라울, 찰리의 등 뒤로 조일광이 산탄총을 맹렬히 겨눈 채 한 발짝 내디딘 위압적인 구도.\n\nLOCATION (lock): Inside the roadside fuel stop's shop, in the medicine aisle under daytime retail ambient light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Store aisle and merchandise displays (Still intact before the struggle) — Display sides recede along the same oblique aisle line as the camera; used as Establishes the shared space between the separated foreground silhouettes and the approaching gunman; Medicine section (Stocked, where 찰리 has been searching) — A limited portion of the displayed products is visible beside 찰리; used as Keeps the innocent purpose of the visit present at the edge of the confrontation; Shotgun (Aimed toward the group during 조일광's forward step) — Seen obliquely between foreground silhouettes, with its barrel directed toward the group rather than the camera; used as Creates a readable threat line without exaggerated muzzle perspective.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination and controlled contrast preserve the shotgun's outline between the figures without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버, 라울, 찰리의 등 뒤로 조일광이 산탄총을 맹렬히 겨눈 채 한 발짝 내디딘 위압적인 구도.\n\nLOCATION (lock): Inside the roadside fuel stop's shop, in the medicine aisle under daytime retail ambient light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Store aisle and merchandise displays (Still intact before the struggle) — Display sides recede along the same oblique aisle line as the camera; used as Establishes the shared space between the separated foreground silhouettes and the approaching gunman; Medicine section (Stocked, where 찰리 has been searching) — A limited portion of the displayed products is visible beside 찰리; used as Keeps the innocent purpose of the visit present at the edge of the confrontation; Shotgun (Aimed toward the group during 조일광's forward step) — Seen obliquely between foreground silhouettes, with its barrel directed toward the group rather than the camera; used as Creates a readable threat line without exaggerated muzzle perspective.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination and controlled contrast preserve the shotgun's outline between the figures without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S49sh12__bgfirst_bg.png",
  "asset_id": "383ae210-3fe6-4311-9ed0-ec26fff64b47",
  "input_asset_ids": [
   "1ad4b69b-2c21-4e57-8eb7-f3b615c1c0dc",
   "756e5c78-4995-466d-a55c-5775930cd9e3"
  ]
 },
 "S49sh12": {
  "input_fingerprint": "364e8f520f229762",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버, 라울, 찰리의 등 뒤로 조일광이 산탄총을 맹렬히 겨눈 채 한 발짝 내디딘 위압적인 구도.\n\nLOCATION (lock): Inside the roadside fuel stop's shop, in the medicine aisle under daytime retail ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Store aisle and merchandise displays (Still intact before the struggle) — Display sides recede along the same oblique aisle line as the camera; used as Establishes the shared space between the separated foreground silhouettes and the approaching gunman; Medicine section (Stocked, where 찰리 has been searching) — A limited portion of the displayed products is visible beside 찰리; used as Keeps the innocent purpose of the visit present at the edge of the confrontation; Shotgun (Aimed toward the group during 조일광's forward step) — Seen obliquely between foreground silhouettes, with its barrel directed toward the group rather than the camera; used as Creates a readable threat line without exaggerated muzzle perspective.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination and controlled contrast preserve the shotgun's outline between the figures without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper stands at the fuel pump with its luggage reloaded; its punctured tire and earlier leaks have not been repaired. Inside the mart, Charlie holds the medicine bottle taken from the shelf, and the television has been showing the wanted group. 조일광: He wears a Marine Corps cap, military uniform and medals, and holds a loaded shotgun at the ready. 앰버: She stands inside the mart near the medicine section, still wearing the replacement shoes. 라울: He stands inside the mart near the medicine section.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조일광 right now, so 조일광's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조일광: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 조일광 (50대 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 과거의 낡은 해병대 위장 군복 상의에 수많은 훈장이 요란하게 달려 있으며, 낡은 팔각모를 쓴 복장. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버, 라울, 찰리의 등 뒤로 조일광이 산탄총을 맹렬히 겨눈 채 한 발짝 내디딘 위압적인 구도.\n\nLOCATION (lock): Inside the roadside fuel stop's shop, in the medicine aisle under daytime retail ambient light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Store aisle and merchandise displays (Still intact before the struggle) — Display sides recede along the same oblique aisle line as the camera; used as Establishes the shared space between the separated foreground silhouettes and the approaching gunman; Medicine section (Stocked, where 찰리 has been searching) — A limited portion of the displayed products is visible beside 찰리; used as Keeps the innocent purpose of the visit present at the edge of the confrontation; Shotgun (Aimed toward the group during 조일광's forward step) — Seen obliquely between foreground silhouettes, with its barrel directed toward the group rather than the camera; used as Creates a readable threat line without exaggerated muzzle perspective.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination and controlled contrast preserve the shotgun's outline between the figures without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper stands at the fuel pump with its luggage reloaded; its punctured tire and earlier leaks have not been repaired. Inside the mart, Charlie holds the medicine bottle taken from the shelf, and the television has been showing the wanted group. 조일광: He wears a Marine Corps cap, military uniform and medals, and holds a loaded shotgun at the ready. 앰버: She stands inside the mart near the medicine section, still wearing the replacement shoes. 라울: He stands inside the mart near the medicine section.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조일광 right now, so 조일광's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조일광: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 조일광 (50대 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 과거의 낡은 해병대 위장 군복 상의에 수많은 훈장이 요란하게 달려 있으며, 낡은 팔각모를 쓴 복장. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버, 라울, 찰리의 등 뒤로 조일광이 산탄총을 맹렬히 겨눈 채 한 발짝 내디딘 위압적인 구도.\n\nLOCATION (lock): Inside the roadside fuel stop's shop, in the medicine aisle under daytime retail ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Store aisle and merchandise displays (Still intact before the struggle) — Display sides recede along the same oblique aisle line as the camera; used as Establishes the shared space between the separated foreground silhouettes and the approaching gunman; Medicine section (Stocked, where 찰리 has been searching) — A limited portion of the displayed products is visible beside 찰리; used as Keeps the innocent purpose of the visit present at the edge of the confrontation; Shotgun (Aimed toward the group during 조일광's forward step) — Seen obliquely between foreground silhouettes, with its barrel directed toward the group rather than the camera; used as Creates a readable threat line without exaggerated muzzle perspective.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-appropriate ambient illumination and controlled contrast preserve the shotgun's outline between the figures without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper stands at the fuel pump with its luggage reloaded; its punctured tire and earlier leaks have not been repaired. Inside the mart, Charlie holds the medicine bottle taken from the shelf, and the television has been showing the wanted group. 조일광: He wears a Marine Corps cap, military uniform and medals, and holds a loaded shotgun at the ready. 앰버: She stands inside the mart near the medicine section, still wearing the replacement shoes. 라울: He stands inside the mart near the medicine section.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조일광 right now, so 조일광's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조일광: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 조일광 (50대 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 과거의 낡은 해병대 위장 군복 상의에 수많은 훈장이 요란하게 달려 있으며, 낡은 팔각모를 쓴 복장. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S49sh12__bgfirst_bg.png",
     "asset_id": "383ae210-3fe6-4311-9ed0-ec26fff64b47",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S49sh12.png",
     "asset_id": "1ad4b69b-2c21-4e57-8eb7-f3b615c1c0dc",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:948699>",
     "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L61B02.png",
     "asset_id": "756e5c78-4995-466d-a55c-5775930cd9e3",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:948699>",
     "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "조일광의 시선과 산탄총의 총구가 전경에 배치된 무리(라울과 찰리)를 명확히 향하고 있음.",
    "built_space": "주유소 마트 내부의 통로와 진열대가 원본 사진의 구조대로 깊이감 있게 묘사되었으며, 인물들이 통로 공간에 알맞게 서 있음.",
    "entities": "앰버(금발), 라울(묶은 머리), 찰리(기계 외장), 조일광(군복, 팔각모, 훈장)이 모두 등장함. 단, 찰리의 겉옷(망토)이 묘사되지 않음.",
    "hard_violations": [],
    "physics": "조일광이 한 발을 앞으로 내딛으며 체중을 이동하는 자세가 자연스럽고, 바닥에 닿은 발과 총을 쥔 양손의 파지가 안정적임."
   },
   {
    "label": "B",
    "direction": "조일광이 산탄총을 앞을 향해 맹렬히 겨누지 않고 대각선 아래로 편하게 들고 있음.",
    "built_space": "마트 진열대와 통로가 구현되어 있고 약품 코너 곁에 인물들이 서 있음.",
    "entities": "앰버와 라울, 조일광은 묘사되었으나 찰리가 화면에 전혀 등장하지 않음. 찰리가 들어야 할 약병을 라울이 들고 있음.",
    "hard_violations": [
     "[gpt-high] 약품 상자와 매장 표지에 읽을 수 있는 글자가 노출되어, 어떤 글자도 읽히면 안 된다는 명시적 금지 조건을 위반한다."
    ],
    "physics": "조일광의 걷는 자세와 바닥 지지, 라울이 약병을 가볍게 쥔 손의 물리적 상태는 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "앰버, 라울, 찰리의 뒷모습 너머로 산탄총을 겨누며 다가오는 조일광의 구도를 정확히 구현했으나, 찰리의 망토와 약병이 누락된 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에 명시된 주요 인물인 찰리가 완전히 누락되었고 조일광이 총을 겨누지 않고 있으며, 약병을 라울이 들고 있어 지시와 크게 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "조일광의 시선과 산탄총의 총구가 전경에 배치된 무리(라울과 찰리)를 명확히 향하고 있음.",
        "built_space": "주유소 마트 내부의 통로와 진열대가 원본 사진의 구조대로 깊이감 있게 묘사되었으며, 인물들이 통로 공간에 알맞게 서 있음.",
        "entities": "앰버(금발), 라울(묶은 머리), 찰리(기계 외장), 조일광(군복, 팔각모, 훈장)이 모두 등장함. 단, 찰리의 겉옷(망토)이 묘사되지 않음.",
        "hard_violations": [],
        "physics": "조일광이 한 발을 앞으로 내딛으며 체중을 이동하는 자세가 자연스럽고, 바닥에 닿은 발과 총을 쥔 양손의 파지가 안정적임."
       },
       {
        "label": "B",
        "direction": "조일광이 산탄총을 앞을 향해 맹렬히 겨누지 않고 대각선 아래로 편하게 들고 있음.",
        "built_space": "마트 진열대와 통로가 구현되어 있고 약품 코너 곁에 인물들이 서 있음.",
        "entities": "앰버와 라울, 조일광은 묘사되었으나 찰리가 화면에 전혀 등장하지 않음. 찰리가 들어야 할 약병을 라울이 들고 있음.",
        "hard_violations": [],
        "physics": "조일광의 걷는 자세와 바닥 지지, 라울이 약병을 가볍게 쥔 손의 물리적 상태는 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "앰버, 라울, 찰리의 뒷모습 너머로 산탄총을 겨누며 다가오는 조일광의 구도를 정확히 구현했으나, 찰리의 망토와 약병이 누락된 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에 명시된 주요 인물인 찰리가 완전히 누락되었고 조일광이 총을 겨누지 않고 있으며, 약병을 라울이 들고 있어 지시와 크게 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "조일광의 시선과 산탄총의 총구가 전경에 배치된 무리(라울과 찰리)를 명확히 향하고 있음.",
        "built_space": "주유소 마트 내부의 통로와 진열대가 원본 사진의 구조대로 깊이감 있게 묘사되었으며, 인물들이 통로 공간에 알맞게 서 있음.",
        "entities": "앰버(금발), 라울(묶은 머리), 찰리(기계 외장), 조일광(군복, 팔각모, 훈장)이 모두 등장함. 단, 찰리의 겉옷(망토)이 묘사되지 않음.",
        "hard_violations": [],
        "physics": "조일광이 한 발을 앞으로 내딛으며 체중을 이동하는 자세가 자연스럽고, 바닥에 닿은 발과 총을 쥔 양손의 파지가 안정적임."
       },
       {
        "label": "B",
        "direction": "조일광이 산탄총을 앞을 향해 맹렬히 겨누지 않고 대각선 아래로 편하게 들고 있음.",
        "built_space": "마트 진열대와 통로가 구현되어 있고 약품 코너 곁에 인물들이 서 있음.",
        "entities": "앰버와 라울, 조일광은 묘사되었으나 찰리가 화면에 전혀 등장하지 않음. 찰리가 들어야 할 약병을 라울이 들고 있음.",
        "hard_violations": [],
        "physics": "조일광의 걷는 자세와 바닥 지지, 라울이 약병을 가볍게 쥔 손의 물리적 상태는 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "찰리가 빠지고 약병을 라울이 들었으며, 산탄총도 맹렬한 견착 조준 대신 낮게 내려져 있고 읽을 수 있는 상품 글자가 노출된다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "세 인물의 등 너머로 조일광이 전진하며 일행을 조준하는 와이드 구도를 충실히 구현했지만, 찰리의 망토와 라울의 꽁지머리, 수배 방송은 맞지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버와 라울은 등을 보이며 조일광 쪽을 향하고, 조일광도 두 아이를 바라본다. 산탄총은 가슴 높이에서 오른쪽 아래로 내려가며, 연장선은 라울의 허리 아래쪽으로 향한다. 카메라를 직접 겨누지는 않지만, 요구한 맹렬한 조준보다 총을 낮춰 든 상태로 읽힌다.",
        "built_space": "왼쪽 유리 출입구와 창, 음료 냉장고 한 대, 뒤쪽 벽걸이 텔레비전 한 대, 책상과 의자 한 개, 오른쪽 약품 선반과 상단에 일부 보이는 볼록거울 한 개가 있다. 중앙 상품 진열대는 온전하며 참고 장소의 주요 재료와 배치는 상당 부분 유지된다. 전경 왼쪽에도 약품 선반이 크게 들어와 참고보다 통로가 좁게 보인다. 전경에는 두 아이만 있어 세 인물 사이로 총잡이를 보는 공간 관계가 성립하지 않는다.",
        "entities": "금발의 어린 여자아이와 녹색 겉옷은 앰버에 부합한다. 오른쪽의 어두운 피부와 곱슬머리, 낡은 녹색 티셔츠를 입은 어린 남자아이는 라울에 가깝지만 꽁지머리는 뚜렷하지 않다. 두 사람 모두 뒷모습이라 얼굴과 혼혈 정체성의 정확한 일치는 확인할 수 없다. 찰리는 없고 약병은 라울의 오른손에 있다. 조일광은 중년 남성으로 군모, 낡은 군복, 다수의 훈장과 산탄총을 갖췄다. 텔레비전에는 수배 일행 대신 뉴스 진행자가 보이며, 약품 상자의 '타이레놀', '판콜에이' 등은 읽힌다. 캠핑카와 아이들의 신발은 이 구도에서 확인할 수 없다.",
        "hard_violations": [
         "약품 상자와 매장 표지에 읽을 수 있는 글자가 노출되어, 어떤 글자도 읽히면 안 된다는 명시적 금지 조건을 위반한다."
        ],
        "physics": "조일광은 한쪽 군화로 바닥을 딛고 다른 발을 들어 앞으로 걷는 중이며, 떠 있는 몸은 아니다. 양손이 산탄총의 손잡이와 앞부분을 지지한다. 약병은 라울의 손에 잡혀 있어 물리적으로는 지지되지만 소지자가 틀렸다. 아이들의 발은 화면 밖이므로 접지는 직접 확인할 수 없으며, 상체에 부유나 불가능한 관절 징후는 없다."
       },
       {
        "label": "B",
        "direction": "앰버, 라울, 찰리는 전경에서 등을 보이며 조일광 쪽으로 몸과 머리를 돌리고 있다. 조일광은 산탄총에 시선을 맞춰 전경 일행을 겨눈다. 총열은 화면 오른쪽 전방의 라울·찰리 쪽으로 비스듬히 향하고, 총구가 렌즈를 정면으로 겨누는 과장된 원근은 없다.",
        "built_space": "왼쪽 유리 출입구와 창, 음료 냉장고 한 대, 뒤쪽 텔레비전 한 대, 책상과 의자 한 개, 천장형 냉난방기 한 대, 오른쪽 약품 선반과 볼록거울 한 개가 보여 참고 장소와 잘 맞는다. 온전한 중앙 진열대의 측면이 통로를 따라 후퇴한다. 앰버는 왼쪽 출입구 쪽, 라울은 중앙 전경, 찰리는 오른쪽 약품 선반 옆에 서고, 조일광은 열린 통로를 따라 다가온다. 볼록거울에 보이는 출입구와 진열대는 해당 위치에서 가능한 반사로 보이며 명백한 광학 모순은 없다.",
        "entities": "요구한 네 인물이 모두 있다. 앰버의 금발과 녹색 재킷, 라울의 어린 체격과 어두운 피부는 부합하지만 라울의 꽁지머리는 보이지 않는다. 뒷모습이어서 두 아이의 얼굴 정체성은 확인할 수 없다. 찰리는 각진 기계 머리와 금속 신체를 갖췄고 약병을 직접 들지만, 참고의 회색 망토가 없다. 흰 얼굴판과 눈·입은 뒤로 돌아선 구도 때문에 평가할 수 없다. 조일광은 중년 남성으로 팔각형 군모, 위장 군복, 여러 훈장을 착용하고 산탄총을 든다. 텔레비전은 수배 일행이 아닌 풍경을 보여준다. 명확히 읽히는 글자는 식별되지 않는다.",
        "hard_violations": [],
        "physics": "조일광은 뒤쪽 발로 바닥을 지지하면서 앞발을 내미는 보행 순간이며, 체중 이동과 착지 방향이 자연스럽다. 산탄총은 양손과 어깨에 의해 지지되어 조준 동작이 성립한다. 찰리의 약병은 손에 잡혀 있고 팔꿈치가 굽혀져 무게를 받친다. 전경 세 인물의 하체는 일부 잘렸지만 보이는 몸통과 팔에 지지 없는 부유나 불가능한 자세는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "찰리가 빠지고 약병을 라울이 들었으며, 산탄총도 맹렬한 견착 조준 대신 낮게 내려져 있고 읽을 수 있는 상품 글자가 노출된다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "세 인물의 등 너머로 조일광이 전진하며 일행을 조준하는 와이드 구도를 충실히 구현했지만, 찰리의 망토와 라울의 꽁지머리, 수배 방송은 맞지 않는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버와 라울은 등을 보이며 조일광 쪽을 향하고, 조일광도 두 아이를 바라본다. 산탄총은 가슴 높이에서 오른쪽 아래로 내려가며, 연장선은 라울의 허리 아래쪽으로 향한다. 카메라를 직접 겨누지는 않지만, 요구한 맹렬한 조준보다 총을 낮춰 든 상태로 읽힌다.",
        "built_space": "왼쪽 유리 출입구와 창, 음료 냉장고 한 대, 뒤쪽 벽걸이 텔레비전 한 대, 책상과 의자 한 개, 오른쪽 약품 선반과 상단에 일부 보이는 볼록거울 한 개가 있다. 중앙 상품 진열대는 온전하며 참고 장소의 주요 재료와 배치는 상당 부분 유지된다. 전경 왼쪽에도 약품 선반이 크게 들어와 참고보다 통로가 좁게 보인다. 전경에는 두 아이만 있어 세 인물 사이로 총잡이를 보는 공간 관계가 성립하지 않는다.",
        "entities": "금발의 어린 여자아이와 녹색 겉옷은 앰버에 부합한다. 오른쪽의 어두운 피부와 곱슬머리, 낡은 녹색 티셔츠를 입은 어린 남자아이는 라울에 가깝지만 꽁지머리는 뚜렷하지 않다. 두 사람 모두 뒷모습이라 얼굴과 혼혈 정체성의 정확한 일치는 확인할 수 없다. 찰리는 없고 약병은 라울의 오른손에 있다. 조일광은 중년 남성으로 군모, 낡은 군복, 다수의 훈장과 산탄총을 갖췄다. 텔레비전에는 수배 일행 대신 뉴스 진행자가 보이며, 약품 상자의 '타이레놀', '판콜에이' 등은 읽힌다. 캠핑카와 아이들의 신발은 이 구도에서 확인할 수 없다.",
        "hard_violations": [
         "약품 상자와 매장 표지에 읽을 수 있는 글자가 노출되어, 어떤 글자도 읽히면 안 된다는 명시적 금지 조건을 위반한다."
        ],
        "physics": "조일광은 한쪽 군화로 바닥을 딛고 다른 발을 들어 앞으로 걷는 중이며, 떠 있는 몸은 아니다. 양손이 산탄총의 손잡이와 앞부분을 지지한다. 약병은 라울의 손에 잡혀 있어 물리적으로는 지지되지만 소지자가 틀렸다. 아이들의 발은 화면 밖이므로 접지는 직접 확인할 수 없으며, 상체에 부유나 불가능한 관절 징후는 없다."
       },
       {
        "label": "A",
        "direction": "앰버, 라울, 찰리는 전경에서 등을 보이며 조일광 쪽으로 몸과 머리를 돌리고 있다. 조일광은 산탄총에 시선을 맞춰 전경 일행을 겨눈다. 총열은 화면 오른쪽 전방의 라울·찰리 쪽으로 비스듬히 향하고, 총구가 렌즈를 정면으로 겨누는 과장된 원근은 없다.",
        "built_space": "왼쪽 유리 출입구와 창, 음료 냉장고 한 대, 뒤쪽 텔레비전 한 대, 책상과 의자 한 개, 천장형 냉난방기 한 대, 오른쪽 약품 선반과 볼록거울 한 개가 보여 참고 장소와 잘 맞는다. 온전한 중앙 진열대의 측면이 통로를 따라 후퇴한다. 앰버는 왼쪽 출입구 쪽, 라울은 중앙 전경, 찰리는 오른쪽 약품 선반 옆에 서고, 조일광은 열린 통로를 따라 다가온다. 볼록거울에 보이는 출입구와 진열대는 해당 위치에서 가능한 반사로 보이며 명백한 광학 모순은 없다.",
        "entities": "요구한 네 인물이 모두 있다. 앰버의 금발과 녹색 재킷, 라울의 어린 체격과 어두운 피부는 부합하지만 라울의 꽁지머리는 보이지 않는다. 뒷모습이어서 두 아이의 얼굴 정체성은 확인할 수 없다. 찰리는 각진 기계 머리와 금속 신체를 갖췄고 약병을 직접 들지만, 참고의 회색 망토가 없다. 흰 얼굴판과 눈·입은 뒤로 돌아선 구도 때문에 평가할 수 없다. 조일광은 중년 남성으로 팔각형 군모, 위장 군복, 여러 훈장을 착용하고 산탄총을 든다. 텔레비전은 수배 일행이 아닌 풍경을 보여준다. 명확히 읽히는 글자는 식별되지 않는다.",
        "hard_violations": [],
        "physics": "조일광은 뒤쪽 발로 바닥을 지지하면서 앞발을 내미는 보행 순간이며, 체중 이동과 착지 방향이 자연스럽다. 산탄총은 양손과 어깨에 의해 지지되어 조준 동작이 성립한다. 찰리의 약병은 손에 잡혀 있고 팔꿈치가 굽혀져 무게를 받친다. 전경 세 인물의 하체는 일부 잘렸지만 보이는 몸통과 팔에 지지 없는 부유나 불가능한 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.679
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.429
   },
   "violations": {
    "B": [
     "[gpt-high] 약품 상자와 매장 표지에 읽을 수 있는 글자가 노출되어, 어떤 글자도 읽히면 안 된다는 명시적 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "앰버, 라울, 찰리의 뒷모습 너머로 산탄총을 겨누며 다가오는 조일광의 구도를 정확히 구현했으나, 찰리의 망토와 약병이 누락된 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 429,
    "verdict_ko": "프롬프트에 명시된 주요 인물인 찰리가 완전히 누락되었고 조일광이 총을 겨누지 않고 있으며, 약병을 라울이 들고 있어 지시와 크게 어긋납니다.  ★위반: [gpt-high] 약품 상자와 매장 표지에 읽을 수 있는 글자가 노출되어, 어떤 글자도 읽히면 안 된다는 명시적 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L61B02.png",
    "asset_id": "756e5c78-4995-466d-a55c-5775930cd9e3",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:943979>",
    "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:948699>",
    "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd8-3e08-7f42-9ec8-3c5ebf1c20e2",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S49sh12__bgfirst_bg.png",
   "bg_asset_id": "383ae210-3fe6-4311-9ed0-ec26fff64b47",
   "bg_record_key": "S49sh12::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S49sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:43:15.973041+00:00",
  "fingerprint": "01cfc5881ad3d9bacc778b4b94dd963231017a2435b0561f786118702c94dcfe",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S49sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S49sh12_sel.png",
  "source_sha256": "7558a718340da10c4ca336938f7eb748dfad7aae67466b3f762bd7ce5a01515b",
  "file": "S49sh12_cine.png",
  "staged_sha256": "7973b6715d59aa997a072d81c117e6677f265891c299682b9e79f308e4ca21b8",
  "latency_ms": 13586
 },
 "S49sh47::signage": {
  "fp": "6a47f228c521be6d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S49sh47": {
  "input_fingerprint": "f12a62f0a8a0488d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거꾸로 뒤집힌 채 찌그러진 캠핑카의 잔해에서 뽀얀 먼지와 짙은 연기가 뿜어져 나오는 처참한 전경.\n\nLOCATION (lock): At a roadside rubbish embankment where the camper has crashed and overturned, with smoke rising into the open air. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Overturned camper (Completely inverted and crushed after the collision) — Front-side quarter and upward-facing wheels are visible from above; used as Primary spatial evidence of the crash, shown in its entirety with surrounding ground; Trash mound (Beside the vehicle after the impact); used as Connects the wreck to the collision site and supplies a background scale reference; Dust and smoke (Pale dust and dark smoke escaping from the wreck); used as Carries the aftermath while leaving the inverted silhouette legible; Broken front-side window opening (Broken open) — Its opening is visible on the camera-facing flank of the inverted vehicle; used as Provides the spatial anchor for the subsequent crane descent and interior cut.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination and restrained contrast distinguish the pale dust, dark smoke and overturned body without introducing flames or additional light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is completely overturned and crushed after striking the garbage embankment, with smoke inside, shattered glazing and a dim interior bulb flickering on and off. Charlie lies unconscious inside with the unrepaired gunshot damage to the shoulder.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 조일광 (50대 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 과거의 낡은 해병대 위장 군복 상의에 수많은 훈장이 요란하게 달려 있으며, 낡은 팔각모를 쓴 복장.; 태진 (30세 여성, 단발머리, 젊은 성인의 얼굴); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 은영 (31세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — wearing: 헐렁하게 늘어난 회색 니트 스웨터와 편안한 청바지 차림.. That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거꾸로 뒤집힌 채 찌그러진 캠핑카의 잔해에서 뽀얀 먼지와 짙은 연기가 뿜어져 나오는 처참한 전경.\n\nLOCATION (lock): At a roadside rubbish embankment where the camper has crashed and overturned, with smoke rising into the open air. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Overturned camper (Completely inverted and crushed after the collision) — Front-side quarter and upward-facing wheels are visible from above; used as Primary spatial evidence of the crash, shown in its entirety with surrounding ground; Trash mound (Beside the vehicle after the impact); used as Connects the wreck to the collision site and supplies a background scale reference; Dust and smoke (Pale dust and dark smoke escaping from the wreck); used as Carries the aftermath while leaving the inverted silhouette legible; Broken front-side window opening (Broken open) — Its opening is visible on the camera-facing flank of the inverted vehicle; used as Provides the spatial anchor for the subsequent crane descent and interior cut.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination and restrained contrast distinguish the pale dust, dark smoke and overturned body without introducing flames or additional light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is completely overturned and crushed after striking the garbage embankment, with smoke inside, shattered glazing and a dim interior bulb flickering on and off. Charlie lies unconscious inside with the unrepaired gunshot damage to the shoulder.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 조일광 (50대 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 과거의 낡은 해병대 위장 군복 상의에 수많은 훈장이 요란하게 달려 있으며, 낡은 팔각모를 쓴 복장.; 태진 (30세 여성, 단발머리, 젊은 성인의 얼굴); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 은영 (31세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — wearing: 헐렁하게 늘어난 회색 니트 스웨터와 편안한 청바지 차림.. That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거꾸로 뒤집힌 채 찌그러진 캠핑카의 잔해에서 뽀얀 먼지와 짙은 연기가 뿜어져 나오는 처참한 전경.\n\nLOCATION (lock): At a roadside rubbish embankment where the camper has crashed and overturned, with smoke rising into the open air. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Overturned camper (Completely inverted and crushed after the collision) — Front-side quarter and upward-facing wheels are visible from above; used as Primary spatial evidence of the crash, shown in its entirety with surrounding ground; Trash mound (Beside the vehicle after the impact); used as Connects the wreck to the collision site and supplies a background scale reference; Dust and smoke (Pale dust and dark smoke escaping from the wreck); used as Carries the aftermath while leaving the inverted silhouette legible; Broken front-side window opening (Broken open) — Its opening is visible on the camera-facing flank of the inverted vehicle; used as Provides the spatial anchor for the subsequent crane descent and interior cut.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination and restrained contrast distinguish the pale dust, dark smoke and overturned body without introducing flames or additional light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is completely overturned and crushed after striking the garbage embankment, with smoke inside, shattered glazing and a dim interior bulb flickering on and off. Charlie lies unconscious inside with the unrepaired gunshot damage to the shoulder.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 조일광 (50대 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 과거의 낡은 해병대 위장 군복 상의에 수많은 훈장이 요란하게 달려 있으며, 낡은 팔각모를 쓴 복장.; 태진 (30세 여성, 단발머리, 젊은 성인의 얼굴); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 은영 (31세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — wearing: 헐렁하게 늘어난 회색 니트 스웨터와 편안한 청바지 차림.. That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 뒤집힌 캠핑카 측면을 정면으로 비추며, 차량 앞부분은 우측 쓰레기 더미를 향함.",
    "built_space": "좌측 도로와 우측 쓰레기 언덕이 레퍼런스의 지형 구조와 일치하며, 카메라를 향한 차량 측면에 깨진 창문 공간이 위치함.",
    "entities": "완전히 뒤집힌 캠핑카, 위를 향한 바퀴, 뽀얀 먼지와 검은 연기, 카메라를 향한 측면의 뚫린 창문 모두 프롬프트와 일치함.",
    "hard_violations": [],
    "physics": "차량이 쓰레기 더미에 물리적으로 자연스럽게 안착되어 지지받고 있으며, 먼지와 연기가 위로 퍼짐."
   },
   {
    "label": "B",
    "direction": "캠핑카 전면이 좌측 하단을 향하고 있으며, 카메라는 주로 전면 유리를 내려다봄.",
    "built_space": "도로가 배경에 가로로 배치되어 레퍼런스(좌측 종방향 도로)의 지형 구조와 어긋남.",
    "entities": "뒤집힌 캠핑카, 위를 향한 바퀴, 쓰레기, 연기가 존재하나 프롬프트가 요구한 '측면 창문' 대신 정면 앞유리가 깨져 있음.",
    "hard_violations": [],
    "physics": "차량이 잔해 위에 얹혀 있으나, 후방 도로와 주변 지형 간의 물리적 공간 연결이 레퍼런스와 다르게 구성됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 공간 구조를 유지하고 '카메라를 향한 측면의 깨진 창문'을 정확하게 묘사하여 프롬프트 지시를 충실히 따름."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "레퍼런스의 공간 배치(도로 위치)를 어겼으며, 지시된 측면 창문 대신 전면 유리창을 깨뜨려 핵심 세부 사항 구현에 실패함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 뒤집힌 캠핑카 측면을 정면으로 비추며, 차량 앞부분은 우측 쓰레기 더미를 향함.",
        "built_space": "좌측 도로와 우측 쓰레기 언덕이 레퍼런스의 지형 구조와 일치하며, 카메라를 향한 차량 측면에 깨진 창문 공간이 위치함.",
        "entities": "완전히 뒤집힌 캠핑카, 위를 향한 바퀴, 뽀얀 먼지와 검은 연기, 카메라를 향한 측면의 뚫린 창문 모두 프롬프트와 일치함.",
        "hard_violations": [],
        "physics": "차량이 쓰레기 더미에 물리적으로 자연스럽게 안착되어 지지받고 있으며, 먼지와 연기가 위로 퍼짐."
       },
       {
        "label": "B",
        "direction": "캠핑카 전면이 좌측 하단을 향하고 있으며, 카메라는 주로 전면 유리를 내려다봄.",
        "built_space": "도로가 배경에 가로로 배치되어 레퍼런스(좌측 종방향 도로)의 지형 구조와 어긋남.",
        "entities": "뒤집힌 캠핑카, 위를 향한 바퀴, 쓰레기, 연기가 존재하나 프롬프트가 요구한 '측면 창문' 대신 정면 앞유리가 깨져 있음.",
        "hard_violations": [],
        "physics": "차량이 잔해 위에 얹혀 있으나, 후방 도로와 주변 지형 간의 물리적 공간 연결이 레퍼런스와 다르게 구성됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 공간 구조를 유지하고 '카메라를 향한 측면의 깨진 창문'을 정확하게 묘사하여 프롬프트 지시를 충실히 따름."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "레퍼런스의 공간 배치(도로 위치)를 어겼으며, 지시된 측면 창문 대신 전면 유리창을 깨뜨려 핵심 세부 사항 구현에 실패함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 뒤집힌 캠핑카 측면을 정면으로 비추며, 차량 앞부분은 우측 쓰레기 더미를 향함.",
        "built_space": "좌측 도로와 우측 쓰레기 언덕이 레퍼런스의 지형 구조와 일치하며, 카메라를 향한 차량 측면에 깨진 창문 공간이 위치함.",
        "entities": "완전히 뒤집힌 캠핑카, 위를 향한 바퀴, 뽀얀 먼지와 검은 연기, 카메라를 향한 측면의 뚫린 창문 모두 프롬프트와 일치함.",
        "hard_violations": [],
        "physics": "차량이 쓰레기 더미에 물리적으로 자연스럽게 안착되어 지지받고 있으며, 먼지와 연기가 위로 퍼짐."
       },
       {
        "label": "B",
        "direction": "캠핑카 전면이 좌측 하단을 향하고 있으며, 카메라는 주로 전면 유리를 내려다봄.",
        "built_space": "도로가 배경에 가로로 배치되어 레퍼런스(좌측 종방향 도로)의 지형 구조와 어긋남.",
        "entities": "뒤집힌 캠핑카, 위를 향한 바퀴, 쓰레기, 연기가 존재하나 프롬프트가 요구한 '측면 창문' 대신 정면 앞유리가 깨져 있음.",
        "hard_violations": [],
        "physics": "차량이 잔해 위에 얹혀 있으나, 후방 도로와 주변 지형 간의 물리적 공간 연결이 레퍼런스와 다르게 구성됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "높은 전방 사선 와이드 구도와 깨진 창은 충실하지만, 차체가 비교적 온전하고 도로·쓰레기 비탈의 배치가 참조 장소와 크게 다릅니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "내려다보는 각도는 약하지만, 참조 장소의 도로변 배치와 완전히 전복되어 찌그러진 잔해, 전방 측면의 깨진 창 및 먼지·연기를 더 충실하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카 앞부분은 화면 왼쪽 아래를 향하고, 카메라는 앞과 오른쪽에 보이는 측면을 높은 사선에서 내려다봅니다. 네 바퀴가 위를 향하며, 깨진 전면 유리와 전방 측면 창의 개구부가 카메라 쪽으로 열려 있습니다. 검은 연기는 차량 앞쪽 뒤편에서 위로 올라가고 옅은 먼지는 차체 주변으로 퍼집니다. 사람이나 조준 대상은 없습니다.",
        "built_space": "차량 전체와 주변 지면이 들어오는 와이드입니다. 도로는 차량 뒤쪽과 화면 오른쪽에 있고 쓰레기 더미는 주로 왼쪽에 있어, 참조의 전경 도로와 오른쪽 쓰레기 비탈 관계와 다릅니다. 참조의 표지판 구조물은 보이지 않습니다. 차량에는 바퀴 네 개, 깨진 전면 창 하나, 전방 측면 창 하나와 그 뒤의 측면 창 세 개가 보입니다. 불가능한 반사나 중복된 고정 시설은 확인되지 않습니다.",
        "entities": "흰색 캠핑카 한 대, 쓰레기 더미, 파손 유리, 옅은 먼지와 짙은 연기가 있으며 사람과 불꽃은 없습니다. 낮의 산지 도로 환경은 맞지만 차량의 형태와 창 배치가 참조 차량과 다르고, 전면 모서리의 손상에 비해 긴 거주부는 상당히 온전하여 심하게 찌그러진 잔해라는 요구가 약합니다. 내부 인물과 전구는 식별되지 않으며 이 외부 구도에서 반드시 보여야 하는 요소는 아닙니다. 뚜렷하게 읽히는 문구는 확인되지 않습니다.",
        "hard_violations": [],
        "physics": "차량은 아래로 향한 지붕 가장자리와 차체를 흙·쓰레기 지면에 대고 놓여 있습니다. 위쪽 바퀴는 차축과 하부 구조에 연결되어 있고, 주변 파편은 지면이나 쓰레기 더미에 놓여 있습니다. 연기와 먼지의 상승·확산은 사고 직후 상황으로 설명됩니다. 지지 없이 떠 있는 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "캠핑카의 찌그러진 앞부분은 화면 오른쪽 쓰레기 비탈을 향하고, 측면과 일부 전방이 카메라를 향합니다. 바퀴는 위를 향하며 전방 측면의 깨진 창은 관객 쪽으로 열려 있습니다. 짙은 연기가 앞부분 뒤에서 위로 솟고 옅은 먼지는 차체 밑과 측면에서 퍼집니다. 사람의 시선이나 무기 방향은 해당하지 않습니다.",
        "built_space": "전경의 아스팔트와 흰 도로 가장자리 선, 오른쪽 쓰레기 비탈, 뒤의 수목 산지, 왼쪽 표지판 지주 하나가 참조 장소의 공간 관계를 유지합니다. 차량 전체와 주변 지면을 담았지만 높은 조감보다는 측면 위주의 완만한 사선 구도라 하부를 내려다보는 정도는 약합니다. 바퀴 두 개가 뚜렷하고 반대편 바퀴 하나가 부분적으로 보입니다. 측면 출입구 하나, 전방의 깨진 창 개구부 하나, 뒤쪽 작은 창과 환기구가 구별됩니다. 불가능한 반사나 시설 중복은 보이지 않습니다.",
        "entities": "참조와 유사한 흰색·회색 무늬의 캠핑카 한 대가 완전히 뒤집혀 있고, 앞부분은 쓰레기 비탈에 부딪혀 크게 찌그러져 있습니다. 파손된 전방 측면 창, 주변 잔해, 뽀얀 먼지와 짙은 연기가 모두 식별됩니다. 낮의 자연광이며 사람이나 불꽃은 없습니다. 내부 인물과 전구는 어두운 개구부에서 식별되지 않습니다. 작은 차체 배지는 있지만 확실히 판독할 수 있는 문구는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "뒤집힌 차체는 지붕 쪽으로 도로변 지면에 닿고, 무너진 앞부분은 쓰레기 비탈에 받쳐져 있습니다. 위로 향한 바퀴는 차량 하부에 연결되어 있으며 파손 창틀도 차체에 붙어 있습니다. 쓰레기와 금속 파편은 지면 또는 더미 위에 놓여 있고 연기는 열린 공중으로 상승합니다. 지지 없이 떠 있는 물체는 확인되지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "높은 전방 사선 와이드 구도와 깨진 창은 충실하지만, 차체가 비교적 온전하고 도로·쓰레기 비탈의 배치가 참조 장소와 크게 다릅니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "내려다보는 각도는 약하지만, 참조 장소의 도로변 배치와 완전히 전복되어 찌그러진 잔해, 전방 측면의 깨진 창 및 먼지·연기를 더 충실하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카 앞부분은 화면 왼쪽 아래를 향하고, 카메라는 앞과 오른쪽에 보이는 측면을 높은 사선에서 내려다봅니다. 네 바퀴가 위를 향하며, 깨진 전면 유리와 전방 측면 창의 개구부가 카메라 쪽으로 열려 있습니다. 검은 연기는 차량 앞쪽 뒤편에서 위로 올라가고 옅은 먼지는 차체 주변으로 퍼집니다. 사람이나 조준 대상은 없습니다.",
        "built_space": "차량 전체와 주변 지면이 들어오는 와이드입니다. 도로는 차량 뒤쪽과 화면 오른쪽에 있고 쓰레기 더미는 주로 왼쪽에 있어, 참조의 전경 도로와 오른쪽 쓰레기 비탈 관계와 다릅니다. 참조의 표지판 구조물은 보이지 않습니다. 차량에는 바퀴 네 개, 깨진 전면 창 하나, 전방 측면 창 하나와 그 뒤의 측면 창 세 개가 보입니다. 불가능한 반사나 중복된 고정 시설은 확인되지 않습니다.",
        "entities": "흰색 캠핑카 한 대, 쓰레기 더미, 파손 유리, 옅은 먼지와 짙은 연기가 있으며 사람과 불꽃은 없습니다. 낮의 산지 도로 환경은 맞지만 차량의 형태와 창 배치가 참조 차량과 다르고, 전면 모서리의 손상에 비해 긴 거주부는 상당히 온전하여 심하게 찌그러진 잔해라는 요구가 약합니다. 내부 인물과 전구는 식별되지 않으며 이 외부 구도에서 반드시 보여야 하는 요소는 아닙니다. 뚜렷하게 읽히는 문구는 확인되지 않습니다.",
        "hard_violations": [],
        "physics": "차량은 아래로 향한 지붕 가장자리와 차체를 흙·쓰레기 지면에 대고 놓여 있습니다. 위쪽 바퀴는 차축과 하부 구조에 연결되어 있고, 주변 파편은 지면이나 쓰레기 더미에 놓여 있습니다. 연기와 먼지의 상승·확산은 사고 직후 상황으로 설명됩니다. 지지 없이 떠 있는 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "캠핑카의 찌그러진 앞부분은 화면 오른쪽 쓰레기 비탈을 향하고, 측면과 일부 전방이 카메라를 향합니다. 바퀴는 위를 향하며 전방 측면의 깨진 창은 관객 쪽으로 열려 있습니다. 짙은 연기가 앞부분 뒤에서 위로 솟고 옅은 먼지는 차체 밑과 측면에서 퍼집니다. 사람의 시선이나 무기 방향은 해당하지 않습니다.",
        "built_space": "전경의 아스팔트와 흰 도로 가장자리 선, 오른쪽 쓰레기 비탈, 뒤의 수목 산지, 왼쪽 표지판 지주 하나가 참조 장소의 공간 관계를 유지합니다. 차량 전체와 주변 지면을 담았지만 높은 조감보다는 측면 위주의 완만한 사선 구도라 하부를 내려다보는 정도는 약합니다. 바퀴 두 개가 뚜렷하고 반대편 바퀴 하나가 부분적으로 보입니다. 측면 출입구 하나, 전방의 깨진 창 개구부 하나, 뒤쪽 작은 창과 환기구가 구별됩니다. 불가능한 반사나 시설 중복은 보이지 않습니다.",
        "entities": "참조와 유사한 흰색·회색 무늬의 캠핑카 한 대가 완전히 뒤집혀 있고, 앞부분은 쓰레기 비탈에 부딪혀 크게 찌그러져 있습니다. 파손된 전방 측면 창, 주변 잔해, 뽀얀 먼지와 짙은 연기가 모두 식별됩니다. 낮의 자연광이며 사람이나 불꽃은 없습니다. 내부 인물과 전구는 어두운 개구부에서 식별되지 않습니다. 작은 차체 배지는 있지만 확실히 판독할 수 있는 문구는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "뒤집힌 차체는 지붕 쪽으로 도로변 지면에 닿고, 무너진 앞부분은 쓰레기 비탈에 받쳐져 있습니다. 위로 향한 바퀴는 차량 하부에 연결되어 있으며 파손 창틀도 차체에 붙어 있습니다. 쓰레기와 금속 파편은 지면 또는 더미 위에 놓여 있고 연기는 열린 공중으로 상승합니다. 지지 없이 떠 있는 물체는 확인되지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.321
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.321
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1321
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 공간 구조를 유지하고 '카메라를 향한 측면의 깨진 창문'을 정확하게 묘사하여 프롬프트 지시를 충실히 따름."
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "레퍼런스의 공간 배치(도로 위치)를 어겼으며, 지시된 측면 창문 대신 전면 유리창을 깨뜨려 핵심 세부 사항 구현에 실패함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L61B03.png",
    "asset_id": "469b7f81-e00b-44fe-9a2b-e19d1ccff50b",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd9-1e3f-7855-8438-6c36339b4db2",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S49sh47::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:43:29.904006+00:00",
  "fingerprint": "f9c1798844b286d9e219f9c321fe7d7e7de1a02e94ad0c53347660b91f69ac38",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S49sh47_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S49sh47_sel.png",
  "source_sha256": "adee867eb12bb5dd7d0ad09d557a35f8425447d1e29e9237d8ab8397274213ba",
  "file": "S49sh47_cine.png",
  "staged_sha256": "69bd338161ffdcabc7d0293c45a3681c48a11a73890b82d63a1b7c58a6d41102",
  "latency_ms": 12859
 },
 "S49sh50::confined_fp_apt": {
  "applies": false,
  "reason_ko": "이 샷은 전복된 캠핑카 내부에서 창문 밖의 인물들을 비추는 장면으로 조종석이나 통제실과 같이 기기 조작 위치나 인물의 좌석 배치가 중요한 공간이 아니므로 평면도 레이아웃 보조가 필요하지 않습니다.",
  "input_fingerprint": "f3bc3f9c4db6e444"
 },
 "S49sh50::signage": {
  "fp": "b71ad1f01f02ea15",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S49sh50::bgfirst_bg": {
  "input_fingerprint": "6221837ca4d002d1",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 깜빡이는 캠핑카 전구 불빛 아래, 깨진 유리창 밖으로 태진과 은영이 몸을 굽혀 안을 빤히 들여다보는 시선.\n\nLOCATION (lock): Inside the overturned camper's crushed cabin, looking through broken window glass under a flickering bulb.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Broken window opening separating cabin interior from the women outside in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Broken camper window (Broken, with an open view outside) — Seen obliquely from the inverted cabin, with 태진 and 은영 visible beyond the opening; used as Thin foreground boundaries establish that the women are outside and the observer is inside; Cabin bulb (Still present inside the overturned cabin); used as A small physical continuity detail near the frame edge, not a foreground obstruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The cabin bulb's faint intermittent light modulates the interior edges against daytime ambient illumination outside, with no hallucination distortion or facial substitution.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 깜빡이는 캠핑카 전구 불빛 아래, 깨진 유리창 밖으로 태진과 은영이 몸을 굽혀 안을 빤히 들여다보는 시선.\n\nLOCATION (lock): Inside the overturned camper's crushed cabin, looking through broken window glass under a flickering bulb.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Broken window opening separating cabin interior from the women outside in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Broken camper window (Broken, with an open view outside) — Seen obliquely from the inverted cabin, with 태진 and 은영 visible beyond the opening; used as Thin foreground boundaries establish that the women are outside and the observer is inside; Cabin bulb (Still present inside the overturned cabin); used as A small physical continuity detail near the frame edge, not a foreground obstruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The cabin bulb's faint intermittent light modulates the interior edges against daytime ambient illumination outside, with no hallucination distortion or facial substitution.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S49sh50__bgfirst_bg.png",
  "asset_id": "c1d6b515-354b-4cec-aa06-35af6f547af5",
  "input_asset_ids": [
   "8a130eca-30b5-4604-96b0-c14106773cbb",
   "02e6a7d0-4405-47be-a590-9b6f84cbfb8c"
  ]
 },
 "S49sh50": {
  "input_fingerprint": "3ba0f0674308b029",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깜빡이는 캠핑카 전구 불빛 아래, 깨진 유리창 밖으로 태진과 은영이 몸을 굽혀 안을 빤히 들여다보는 시선.\n\nLOCATION (lock): Inside the overturned camper's crushed cabin, looking through broken window glass under a flickering bulb. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Broken window opening separating cabin interior from the women outside in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Broken camper window (Broken, with an open view outside) — Seen obliquely from the inverted cabin, with 태진 and 은영 visible beyond the opening; used as Thin foreground boundaries establish that the women are outside and the observer is inside; Cabin bulb (Still present inside the overturned cabin); used as A small physical continuity detail near the frame edge, not a foreground obstruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The cabin bulb's faint intermittent light modulates the interior edges against daytime ambient illumination outside, with no hallucination distortion or facial substitution.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is losing consciousness inside the overturned camper, his eyes partly open as he looks toward the women outside the broken window. His torso orientation, head support, arm and leg positions, and contact with the inverted interior are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crushed camper remains overturned, with broken windows, smoke inside and its interior bulb flickering weakly. Charlie remains unconscious with a damaged shoulder, while the wreck and its loose contents have not yet been recovered. 태진: She is outside the wreck, looking into its interior. 은영: She is outside the wreck, looking into its interior.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (30세 여성, 단발머리, 젊은 성인의 얼굴); 은영 (31세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — wearing: 헐렁하게 늘어난 회색 니트 스웨터와 편안한 청바지 차림. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot inside a tight, built interior. The FIRST attached image (SHOT BACKGROUND) is the finished empty interior of this shot, and in a space this cramped its geometry is the truth of the shot — keep it EXACTLY: its camera, perspective and every fixed fitting it has stay untouched, in the same place, at the same angle, in the same number.\n\nBefore you place anyone, count what the background shows: every fixed fitting this space has, how many of each, which way each faces, and what any reflective surface among them could return from this camera. Those counts and placements are what you must still be able to make after the people are in. Adding a second one of anything, moving a fitting or growing a new surface is a failure even when the person looks right.\n\nThe SECOND attached image (LAYOUT SKETCH) tells you where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Its background lines are not decoration — they are the same structure seen in line form, so use them to place each person correctly with respect to it: where the body rests and on what, which side of a fitting the hands are on, what the body passes in front of and what it passes behind. Where sketch and background disagree about the structure itself, the background wins.\n\nThe CHARACTER REFERENCE photographs show the real people.\n\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. A person may cover part of the structure — that is expected, and covering is not redrawing. What the body hides stays hidden; what remains visible stays exactly as the background had it. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깜빡이는 캠핑카 전구 불빛 아래, 깨진 유리창 밖으로 태진과 은영이 몸을 굽혀 안을 빤히 들여다보는 시선.\n\nLOCATION (lock): Inside the overturned camper's crushed cabin, looking through broken window glass under a flickering bulb. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Broken window opening separating cabin interior from the women outside in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Broken camper window (Broken, with an open view outside) — Seen obliquely from the inverted cabin, with 태진 and 은영 visible beyond the opening; used as Thin foreground boundaries establish that the women are outside and the observer is inside; Cabin bulb (Still present inside the overturned cabin); used as A small physical continuity detail near the frame edge, not a foreground obstruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The cabin bulb's faint intermittent light modulates the interior edges against daytime ambient illumination outside, with no hallucination distortion or facial substitution.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is losing consciousness inside the overturned camper, his eyes partly open as he looks toward the women outside the broken window. His torso orientation, head support, arm and leg positions, and contact with the inverted interior are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crushed camper remains overturned, with broken windows, smoke inside and its interior bulb flickering weakly. Charlie remains unconscious with a damaged shoulder, while the wreck and its loose contents have not yet been recovered. 태진: She is outside the wreck, looking into its interior. 은영: She is outside the wreck, looking into its interior.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (30세 여성, 단발머리, 젊은 성인의 얼굴); 은영 (31세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — wearing: 헐렁하게 늘어난 회색 니트 스웨터와 편안한 청바지 차림. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깜빡이는 캠핑카 전구 불빛 아래, 깨진 유리창 밖으로 태진과 은영이 몸을 굽혀 안을 빤히 들여다보는 시선.\n\nLOCATION (lock): Inside the overturned camper's crushed cabin, looking through broken window glass under a flickering bulb. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Broken window opening separating cabin interior from the women outside in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Broken camper window (Broken, with an open view outside) — Seen obliquely from the inverted cabin, with 태진 and 은영 visible beyond the opening; used as Thin foreground boundaries establish that the women are outside and the observer is inside; Cabin bulb (Still present inside the overturned cabin); used as A small physical continuity detail near the frame edge, not a foreground obstruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The cabin bulb's faint intermittent light modulates the interior edges against daytime ambient illumination outside, with no hallucination distortion or facial substitution.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is losing consciousness inside the overturned camper, his eyes partly open as he looks toward the women outside the broken window. His torso orientation, head support, arm and leg positions, and contact with the inverted interior are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crushed camper remains overturned, with broken windows, smoke inside and its interior bulb flickering weakly. Charlie remains unconscious with a damaged shoulder, while the wreck and its loose contents have not yet been recovered. 태진: She is outside the wreck, looking into its interior. 은영: She is outside the wreck, looking into its interior.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (30세 여성, 단발머리, 젊은 성인의 얼굴); 은영 (31세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — wearing: 헐렁하게 늘어난 회색 니트 스웨터와 편안한 청바지 차림. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S49sh50__bgfirst_bg.png",
     "asset_id": "c1d6b515-354b-4cec-aa06-35af6f547af5",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement AND the structure they sit inside — the sketched fixed fittings and openings are the same ones the background photograph shows, drawn as lines; read them to place each body correctly against that structure, and where the two disagree about the structure itself the background photograph wins)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S49sh50.png",
     "asset_id": "8a130eca-30b5-4604-96b0-c14106773cbb",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:942632>",
     "asset_id": "cb567190-df6f-4833-8024-26cec5633f22",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L61B01.png",
     "asset_id": "02e6a7d0-4405-47be-a590-9b6f84cbfb8c",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:942632>",
     "asset_id": "cb567190-df6f-4833-8024-26cec5633f22",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "B",
    "direction": "두 여성 모두 창밖에서 몸을 낮추고 창 안쪽의 낮은 카메라 위치를 응시한다. 시선은 서로나 외부 풍경이 아니라 실내 관찰자 쪽으로 향해 요구된 들여다보기 동작과 맞는다.",
    "built_space": "중앙에 깨진 주 창문 하나, 오른쪽 가장자리에 별도 창 일부, 주 창 양옆에 커튼 두 묶음이 보인다. 왼쪽에는 좌석 일부와 안전벨트, 위에는 목재 수납장, 아래에는 유리 조각과 떨어진 패널들이 있다. 두 여성은 창틀 뒤 외부에 있고 카메라는 실내의 낮은 위치에 있다. 참조의 금속 창틀·목재·회색 직물·파손된 실내 재질은 유사하지만 필수 전구는 보이지 않는다. 아래쪽 실내 잔해가 비교적 넓고 여성들의 무릎까지 드러나 요청한 중경보다 조금 넓게 읽힌다.",
    "entities": "젊은 성인 여성 두 명만 보인다. 오른쪽 여성은 태진에 해당하는 짧은 단발과 참조에 가까운 얼굴을 갖지만, 참조의 모자·청색 목수건·녹색 작업복 대신 회색 니트를 입었다. 왼쪽 여성은 머리를 뒤로 묶었으며 은영의 젊은 여성 인상, 헐렁한 회색 니트와 청바지 조건에 맞는다. 깨진 유리와 옅은 실내 연무는 보이지만 전구는 없다. 읽을 수 있는 글자는 없다.",
    "hard_violations": [],
    "physics": "두 여성은 무릎을 접고 쪼그린 자세로 몸을 앞으로 기울인다. 오른쪽 여성의 한 손은 올라온 무릎에 놓이고 다른 손은 아래로 내려와 있다. 왼쪽 여성도 접은 무릎 가까이에 손을 두고 반대 팔을 창 옆으로 뻗는다. 발과 지면 접점은 창턱에 가려지지만, 보이는 자세는 외부에서 쪼그려 앉아 유지할 수 있으며 공중에 떠 있다는 증거는 없다. 유리 파편은 창틀에 붙어 있거나 아래 실내 표면에 놓여 있다."
   },
   {
    "label": "A",
    "direction": "두 여성은 허리를 굽혀 창 안쪽 아래를 바라본다. 시선 방향은 카메라와 하단 전경의 누운 사람 쪽으로 모이며, 밖에서 내부를 살피는 동작 자체는 맞는다.",
    "built_space": "큰 깨진 창문 하나가 중앙을 차지하고, 왼쪽 위 실내 가장자리에 켜진 전구 하나가 있다. 왼쪽에는 커튼 한 묶음과 안전벨트, 목재 가구가 보이고 위에는 좌석 쿠션 일부가 걸쳐 보인다. 창틀과 유리 잔해가 실내외를 분리하며 두 여성은 외부에 서 있다. 인물은 허리 부근까지 보이는 중경이고 전구도 작은 주변 요소로 배치됐다. 다만 하단 전경을 추가 인체가 크게 차지한다.",
    "entities": "창밖에는 단발 여성과 긴 머리 여성 두 명이 있으며 모두 젊은 성인으로 보이고 회색 니트와 청바지를 입었다. 왼쪽 단발 여성은 태진의 머리와 얼굴 인상에 가깝지만 참조의 모자·목수건·작업복은 재현하지 않았다. 오른쪽 여성은 은영의 연령대와 의상 조건에 부합한다. 하단에는 흐릿하지만 옷 입은 어깨와 머리로 읽히는 제삼자의 신체가 추가되어 있다. 깨진 창과 물리적 전구는 있으며 읽을 수 있는 글자는 없다.",
    "hard_violations": [
     "하단 전경에 누운 제삼자의 어깨와 머리로 읽히는 신체를 추가했다. 이는 이 숏에서 태진과 은영 외에는 신체 일부도 보이지 않아야 한다는 인물 제한을 위반한다."
    ],
    "physics": "두 여성의 몸통은 아래로 이어지는 다리 위에서 앞으로 기울어 있고 팔은 창턱 가까이 내려와 있어 가능한 들여다보기 자세다. 발과 손의 정확한 접점 일부는 프레임에 가려진다. 전경의 추가 인체는 낮게 누워 있는 것으로 보이나 지지면과의 접촉은 흐림과 크롭 때문에 확인하기 어렵고, 공중에 뜬 모습으로 단정할 근거는 없다. 전구는 소켓에 고정되어 있으며 유리는 창틀에 남아 있거나 아래 표면에 쌓여 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "창밖의 두 여성만이 몸을 낮춰 내부를 응시하는 관계와 장소 재질은 충실하지만, 필수 전구가 보이지 않고 태진의 의상은 참조와 다르다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "중경 구도와 가장자리 전구는 잘 구현했지만, 하단 전경에 제삼자의 몸을 추가해 두 여성만 보여야 한다는 명시적 제한을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 여성 모두 창밖에서 몸을 낮추고 창 안쪽의 낮은 카메라 위치를 응시한다. 시선은 서로나 외부 풍경이 아니라 실내 관찰자 쪽으로 향해 요구된 들여다보기 동작과 맞는다.",
        "built_space": "중앙에 깨진 주 창문 하나, 오른쪽 가장자리에 별도 창 일부, 주 창 양옆에 커튼 두 묶음이 보인다. 왼쪽에는 좌석 일부와 안전벨트, 위에는 목재 수납장, 아래에는 유리 조각과 떨어진 패널들이 있다. 두 여성은 창틀 뒤 외부에 있고 카메라는 실내의 낮은 위치에 있다. 참조의 금속 창틀·목재·회색 직물·파손된 실내 재질은 유사하지만 필수 전구는 보이지 않는다. 아래쪽 실내 잔해가 비교적 넓고 여성들의 무릎까지 드러나 요청한 중경보다 조금 넓게 읽힌다.",
        "entities": "젊은 성인 여성 두 명만 보인다. 오른쪽 여성은 태진에 해당하는 짧은 단발과 참조에 가까운 얼굴을 갖지만, 참조의 모자·청색 목수건·녹색 작업복 대신 회색 니트를 입었다. 왼쪽 여성은 머리를 뒤로 묶었으며 은영의 젊은 여성 인상, 헐렁한 회색 니트와 청바지 조건에 맞는다. 깨진 유리와 옅은 실내 연무는 보이지만 전구는 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 여성은 무릎을 접고 쪼그린 자세로 몸을 앞으로 기울인다. 오른쪽 여성의 한 손은 올라온 무릎에 놓이고 다른 손은 아래로 내려와 있다. 왼쪽 여성도 접은 무릎 가까이에 손을 두고 반대 팔을 창 옆으로 뻗는다. 발과 지면 접점은 창턱에 가려지지만, 보이는 자세는 외부에서 쪼그려 앉아 유지할 수 있으며 공중에 떠 있다는 증거는 없다. 유리 파편은 창틀에 붙어 있거나 아래 실내 표면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "두 여성은 허리를 굽혀 창 안쪽 아래를 바라본다. 시선 방향은 카메라와 하단 전경의 누운 사람 쪽으로 모이며, 밖에서 내부를 살피는 동작 자체는 맞는다.",
        "built_space": "큰 깨진 창문 하나가 중앙을 차지하고, 왼쪽 위 실내 가장자리에 켜진 전구 하나가 있다. 왼쪽에는 커튼 한 묶음과 안전벨트, 목재 가구가 보이고 위에는 좌석 쿠션 일부가 걸쳐 보인다. 창틀과 유리 잔해가 실내외를 분리하며 두 여성은 외부에 서 있다. 인물은 허리 부근까지 보이는 중경이고 전구도 작은 주변 요소로 배치됐다. 다만 하단 전경을 추가 인체가 크게 차지한다.",
        "entities": "창밖에는 단발 여성과 긴 머리 여성 두 명이 있으며 모두 젊은 성인으로 보이고 회색 니트와 청바지를 입었다. 왼쪽 단발 여성은 태진의 머리와 얼굴 인상에 가깝지만 참조의 모자·목수건·작업복은 재현하지 않았다. 오른쪽 여성은 은영의 연령대와 의상 조건에 부합한다. 하단에는 흐릿하지만 옷 입은 어깨와 머리로 읽히는 제삼자의 신체가 추가되어 있다. 깨진 창과 물리적 전구는 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "하단 전경에 누운 제삼자의 어깨와 머리로 읽히는 신체를 추가했다. 이는 이 숏에서 태진과 은영 외에는 신체 일부도 보이지 않아야 한다는 인물 제한을 위반한다."
        ],
        "physics": "두 여성의 몸통은 아래로 이어지는 다리 위에서 앞으로 기울어 있고 팔은 창턱 가까이 내려와 있어 가능한 들여다보기 자세다. 발과 손의 정확한 접점 일부는 프레임에 가려진다. 전경의 추가 인체는 낮게 누워 있는 것으로 보이나 지지면과의 접촉은 흐림과 크롭 때문에 확인하기 어렵고, 공중에 뜬 모습으로 단정할 근거는 없다. 전구는 소켓에 고정되어 있으며 유리는 창틀에 남아 있거나 아래 표면에 쌓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "창밖의 두 여성만이 몸을 낮춰 내부를 응시하는 관계와 장소 재질은 충실하지만, 필수 전구가 보이지 않고 태진의 의상은 참조와 다르다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "중경 구도와 가장자리 전구는 잘 구현했지만, 하단 전경에 제삼자의 몸을 추가해 두 여성만 보여야 한다는 명시적 제한을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 여성 모두 창밖에서 몸을 낮추고 창 안쪽의 낮은 카메라 위치를 응시한다. 시선은 서로나 외부 풍경이 아니라 실내 관찰자 쪽으로 향해 요구된 들여다보기 동작과 맞는다.",
        "built_space": "중앙에 깨진 주 창문 하나, 오른쪽 가장자리에 별도 창 일부, 주 창 양옆에 커튼 두 묶음이 보인다. 왼쪽에는 좌석 일부와 안전벨트, 위에는 목재 수납장, 아래에는 유리 조각과 떨어진 패널들이 있다. 두 여성은 창틀 뒤 외부에 있고 카메라는 실내의 낮은 위치에 있다. 참조의 금속 창틀·목재·회색 직물·파손된 실내 재질은 유사하지만 필수 전구는 보이지 않는다. 아래쪽 실내 잔해가 비교적 넓고 여성들의 무릎까지 드러나 요청한 중경보다 조금 넓게 읽힌다.",
        "entities": "젊은 성인 여성 두 명만 보인다. 오른쪽 여성은 태진에 해당하는 짧은 단발과 참조에 가까운 얼굴을 갖지만, 참조의 모자·청색 목수건·녹색 작업복 대신 회색 니트를 입었다. 왼쪽 여성은 머리를 뒤로 묶었으며 은영의 젊은 여성 인상, 헐렁한 회색 니트와 청바지 조건에 맞는다. 깨진 유리와 옅은 실내 연무는 보이지만 전구는 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 여성은 무릎을 접고 쪼그린 자세로 몸을 앞으로 기울인다. 오른쪽 여성의 한 손은 올라온 무릎에 놓이고 다른 손은 아래로 내려와 있다. 왼쪽 여성도 접은 무릎 가까이에 손을 두고 반대 팔을 창 옆으로 뻗는다. 발과 지면 접점은 창턱에 가려지지만, 보이는 자세는 외부에서 쪼그려 앉아 유지할 수 있으며 공중에 떠 있다는 증거는 없다. 유리 파편은 창틀에 붙어 있거나 아래 실내 표면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "두 여성은 허리를 굽혀 창 안쪽 아래를 바라본다. 시선 방향은 카메라와 하단 전경의 누운 사람 쪽으로 모이며, 밖에서 내부를 살피는 동작 자체는 맞는다.",
        "built_space": "큰 깨진 창문 하나가 중앙을 차지하고, 왼쪽 위 실내 가장자리에 켜진 전구 하나가 있다. 왼쪽에는 커튼 한 묶음과 안전벨트, 목재 가구가 보이고 위에는 좌석 쿠션 일부가 걸쳐 보인다. 창틀과 유리 잔해가 실내외를 분리하며 두 여성은 외부에 서 있다. 인물은 허리 부근까지 보이는 중경이고 전구도 작은 주변 요소로 배치됐다. 다만 하단 전경을 추가 인체가 크게 차지한다.",
        "entities": "창밖에는 단발 여성과 긴 머리 여성 두 명이 있으며 모두 젊은 성인으로 보이고 회색 니트와 청바지를 입었다. 왼쪽 단발 여성은 태진의 머리와 얼굴 인상에 가깝지만 참조의 모자·목수건·작업복은 재현하지 않았다. 오른쪽 여성은 은영의 연령대와 의상 조건에 부합한다. 하단에는 흐릿하지만 옷 입은 어깨와 머리로 읽히는 제삼자의 신체가 추가되어 있다. 깨진 창과 물리적 전구는 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "하단 전경에 누운 제삼자의 어깨와 머리로 읽히는 신체를 추가했다. 이는 이 숏에서 태진과 은영 외에는 신체 일부도 보이지 않아야 한다는 인물 제한을 위반한다."
        ],
        "physics": "두 여성의 몸통은 아래로 이어지는 다리 위에서 앞으로 기울어 있고 팔은 창턱 가까이 내려와 있어 가능한 들여다보기 자세다. 발과 손의 정확한 접점 일부는 프레임에 가려진다. 전경의 추가 인체는 낮게 누워 있는 것으로 보이나 지지면과의 접촉은 흐림과 크롭 때문에 확인하기 어렵고, 공중에 뜬 모습으로 단정할 근거는 없다. 전구는 소켓에 고정되어 있으며 유리는 창틀에 남아 있거나 아래 표면에 쌓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "창밖의 두 여성만이 몸을 낮춰 내부를 응시하는 관계와 장소 재질은 충실하지만, 필수 전구가 보이지 않고 태진의 의상은 참조와 다르다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "중경 구도와 가장자리 전구는 잘 구현했지만, 하단 전경에 제삼자의 몸을 추가해 두 여성만 보여야 한다는 명시적 제한을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L61B01.png",
    "asset_id": "02e6a7d0-4405-47be-a590-9b6f84cbfb8c",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:942632>",
    "asset_id": "cb567190-df6f-4833-8024-26cec5633f22",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdd9-f192-72b7-a9f8-1cbeab4f1838",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S49sh50__bgfirst_bg.png",
   "bg_asset_id": "c1d6b515-354b-4cec-aa06-35af6f547af5",
   "bg_record_key": "S49sh50::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S49sh50::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:43:43.189450+00:00",
  "fingerprint": "1668cf346b251890a705dfa8a6cbdf3599f1e8c99443e6e5c3b038c849b22553",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S49sh50_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S49sh50_sel.png",
  "source_sha256": "8bc81bd181d995e6b79888f6ecc1ddf1c6e1e6d86a8ef3bcfd53c76e2c8de723",
  "file": "S49sh50_cine.png",
  "staged_sha256": "2e50a87374816925e9aeaa61c2db6dc01451e33ebf3ab931ed6daea12581bfe3",
  "latency_ms": 11159
 },
 "S50sh1::signage": {
  "fp": "1567a4c3667a9ac7",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "록밴드 포스터"
   }
  ],
  "dropped": []
 },
 "era_assess::9cfb31a686391167": {
  "subjects": [],
  "subject_text": "태진과 은영의 카센터 창고 내부\n벽에 여러 록밴드 포스터가 붙은 정비 창고. 거대한 스피커와 펜치 등 정비 공구, 수리용 용액 병이 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L67",
  "scope_role": "location_interior",
  "scope_sha": "a317875a2fb6727c"
 },
 "S50sh1::bgfirst_bg": {
  "input_fingerprint": "2183893ace76d721",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 록밴드 포스터가 가득 붙은 어두운 카센터 창고 안, 붕대가 감긴 다리로 침상에 걸터앉은 이현우가 두 눈을 번쩍 뜬 얼굴.\n\nLOCATION (lock): At a makeshift bed in the dim storage-room section of an auto repair garage, with rock-band posters covering the walls.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rock-band posters (Attached throughout the room) — Printed faces are visible in partial, oblique views behind 이현우; no specific band names are introduced; used as Establishes the unfamiliar room through surrounding graphic fragments rather than a separate establishing shot; Bed (Occupied at its edge by 이현우) — The near edge and a limited portion of its upper surface are visible beneath him; used as Anchors his newly raised seated posture and supports the lower-frame leg placement; Leg bandage (Wrapped around his treated leg) — Visible around the leg in the lower portion of the frame; used as Provides immediate evidence of treatment while the face remains the primary focus.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination in the dark storage room keeps his opened eyes and bandages readable without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 록밴드 포스터가 가득 붙은 어두운 카센터 창고 안, 붕대가 감긴 다리로 침상에 걸터앉은 이현우가 두 눈을 번쩍 뜬 얼굴.\n\nLOCATION (lock): At a makeshift bed in the dim storage-room section of an auto repair garage, with rock-band posters covering the walls.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rock-band posters (Attached throughout the room) — Printed faces are visible in partial, oblique views behind 이현우; no specific band names are introduced; used as Establishes the unfamiliar room through surrounding graphic fragments rather than a separate establishing shot; Bed (Occupied at its edge by 이현우) — The near edge and a limited portion of its upper surface are visible beneath him; used as Anchors his newly raised seated posture and supports the lower-frame leg placement; Leg bandage (Wrapped around his treated leg) — Visible around the leg in the lower portion of the frame; used as Provides immediate evidence of treatment while the face remains the primary focus.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination in the dark storage room keeps his opened eyes and bandages readable without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S50sh1__bgfirst_bg.png",
  "asset_id": "dbdd5fe6-62a1-4304-92a6-29ab6d4465e4",
  "input_asset_ids": [
   "83e4520d-3f4d-46c2-a71b-50ed5b1d7432",
   "38f979f5-28cd-471a-a942-4cbe22909078"
  ]
 },
 "S50sh1": {
  "input_fingerprint": "806d0008dc55bf9d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 록밴드 포스터가 가득 붙은 어두운 카센터 창고 안, 붕대가 감긴 다리로 침상에 걸터앉은 이현우가 두 눈을 번쩍 뜬 얼굴.\n\nLOCATION (lock): At a makeshift bed in the dim storage-room section of an auto repair garage, with rock-band posters covering the walls. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rock-band posters (Attached throughout the room) — Printed faces are visible in partial, oblique views behind 이현우; no specific band names are introduced; used as Establishes the unfamiliar room through surrounding graphic fragments rather than a separate establishing shot; Bed (Occupied at its edge by 이현우) — The near edge and a limited portion of its upper surface are visible beneath him; used as Anchors his newly raised seated posture and supports the lower-frame leg placement; Leg bandage (Wrapped around his treated leg) — Visible around the leg in the lower portion of the frame; used as Provides immediate evidence of treatment while the face remains the primary focus.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination in the dark storage room keeps his opened eyes and bandages readable without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rock-band posters cover the storage-room walls, a large pair of pliers lies within reach, and a large speaker stands in the repair area. The recovered camper is under repair with extensive engine and brake damage; Charlie's shoulder repair is established by the immediately following meal scene. 이현우: He has awakened in the storage room with his injured leg bandaged and visible signs of treatment elsewhere. He retains the contact card concealed in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 록밴드 포스터가 가득 붙은 어두운 카센터 창고 안, 붕대가 감긴 다리로 침상에 걸터앉은 이현우가 두 눈을 번쩍 뜬 얼굴.\n\nLOCATION (lock): At a makeshift bed in the dim storage-room section of an auto repair garage, with rock-band posters covering the walls. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rock-band posters (Attached throughout the room) — Printed faces are visible in partial, oblique views behind 이현우; no specific band names are introduced; used as Establishes the unfamiliar room through surrounding graphic fragments rather than a separate establishing shot; Bed (Occupied at its edge by 이현우) — The near edge and a limited portion of its upper surface are visible beneath him; used as Anchors his newly raised seated posture and supports the lower-frame leg placement; Leg bandage (Wrapped around his treated leg) — Visible around the leg in the lower portion of the frame; used as Provides immediate evidence of treatment while the face remains the primary focus.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination in the dark storage room keeps his opened eyes and bandages readable without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rock-band posters cover the storage-room walls, a large pair of pliers lies within reach, and a large speaker stands in the repair area. The recovered camper is under repair with extensive engine and brake damage; Charlie's shoulder repair is established by the immediately following meal scene. 이현우: He has awakened in the storage room with his injured leg bandaged and visible signs of treatment elsewhere. He retains the contact card concealed in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 록밴드 포스터가 가득 붙은 어두운 카센터 창고 안, 붕대가 감긴 다리로 침상에 걸터앉은 이현우가 두 눈을 번쩍 뜬 얼굴.\n\nLOCATION (lock): At a makeshift bed in the dim storage-room section of an auto repair garage, with rock-band posters covering the walls. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rock-band posters (Attached throughout the room) — Printed faces are visible in partial, oblique views behind 이현우; no specific band names are introduced; used as Establishes the unfamiliar room through surrounding graphic fragments rather than a separate establishing shot; Bed (Occupied at its edge by 이현우) — The near edge and a limited portion of its upper surface are visible beneath him; used as Anchors his newly raised seated posture and supports the lower-frame leg placement; Leg bandage (Wrapped around his treated leg) — Visible around the leg in the lower portion of the frame; used as Provides immediate evidence of treatment while the face remains the primary focus.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination in the dark storage room keeps his opened eyes and bandages readable without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rock-band posters cover the storage-room walls, a large pair of pliers lies within reach, and a large speaker stands in the repair area. The recovered camper is under repair with extensive engine and brake damage; Charlie's shoulder repair is established by the immediately following meal scene. 이현우: He has awakened in the storage room with his injured leg bandaged and visible signs of treatment elsewhere. He retains the contact card concealed in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S50sh1__bgfirst_bg.png",
     "asset_id": "dbdd5fe6-62a1-4304-92a6-29ab6d4465e4",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S50sh1.png",
     "asset_id": "83e4520d-3f4d-46c2-a71b-50ed5b1d7432",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L67B01.png",
     "asset_id": "38f979f5-28cd-471a-a942-4cbe22909078",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 천장을 향해 위쪽으로 쏠려 있습니다.",
    "built_space": "카센터 창고 내부의 구조(침상, 작업대, 뒤편의 캠핑카)가 참조 사진의 배치와 잘 맞습니다. 좌측 벽면에 록밴드 포스터들이 붙어 있습니다.",
    "entities": "인물의 얼굴과 머리 모양은 참조 이미지와 흡사하나, 의상이 짙은 회색 반팔 티셔츠와 밝은 반바지로 완전히 변경되었습니다. 왼쪽 종아리에 붕대가 감겨 있습니다.",
    "hard_violations": [
     "[gemini-pro] 캐릭터 참조 지침 위반: 참조 이미지에서 반드시 유지해야 할 의상(후드티와 청바지)을 무시하고 완전히 다른 형태의 의상(반팔 티셔츠와 반바지)을 입혔습니다."
    ],
    "physics": "침상 가장자리에 안정적으로 걸터앉아 있으며, 양손을 침상 바닥과 무릎에 올려두어 몸을 지탱하고 있습니다. 물리적으로 어긋난 부분은 없습니다."
   },
   {
    "label": "B",
    "direction": "인물의 시선은 화면 우측을 향하고 있으며, 두 눈을 크게 뜨고 주변을 살피는 듯한 뚜렷한 응시가 보입니다.",
    "built_space": "카센터 창고의 배경이 참조 사진과 매우 일치합니다. 침상, 작업대, 캠핑카가 올바른 위치에 있으며, 침상 앞의 작은 검은 의자가 붕대 감긴 다리를 받치는 용도로 적절히 활용되었습니다. 벽면과 좌측 구조물에 록밴드 포스터가 부착되어 있습니다.",
    "entities": "인물의 얼굴, 헤어스타일, 체형뿐만 아니라 참조 이미지에 명시된 낡은 어두운 후드티와 찢어진 청바지까지 완벽하게 일치합니다. 왼쪽 다리와 발에 붕대가 감겨 있으며, 작업대 위에는 플라이어가 놓여 있습니다.",
    "hard_violations": [
     "[gpt-high] 배경 포스터에 읽을 수 있는 영문 밴드 제목이 노출되어, 읽히는 글자와 특정 밴드명 도입을 금지한 조건을 위반한다."
    ],
    "physics": "침상 가장자리에 앉아 있으며, 오른쪽 발은 바닥을 향하고 왼쪽 다리는 작은 의자 위에 올려져 있어 체중 분산과 자세의 지지가 매우 현실적이고 안정적입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지정된 카센터 창고의 구조를 정확히 따랐으며, 인물의 외형과 지정된 의상을 완벽히 유지한 채 눈을 크게 뜬 표정과 붕대를 감은 다리의 설정을 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "공간과 인물의 얼굴은 잘 묘사되었으나, 캐릭터 참조 이미지에서 유지해야 할 의상 규정을 무시하고 임의의 옷을 렌더링하여 심각한 지침 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "인물의 시선은 화면 우측을 향하고 있으며, 두 눈을 크게 뜨고 주변을 살피는 듯한 뚜렷한 응시가 보입니다.",
        "built_space": "카센터 창고의 배경이 참조 사진과 매우 일치합니다. 침상, 작업대, 캠핑카가 올바른 위치에 있으며, 침상 앞의 작은 검은 의자가 붕대 감긴 다리를 받치는 용도로 적절히 활용되었습니다. 벽면과 좌측 구조물에 록밴드 포스터가 부착되어 있습니다.",
        "entities": "인물의 얼굴, 헤어스타일, 체형뿐만 아니라 참조 이미지에 명시된 낡은 어두운 후드티와 찢어진 청바지까지 완벽하게 일치합니다. 왼쪽 다리와 발에 붕대가 감겨 있으며, 작업대 위에는 플라이어가 놓여 있습니다.",
        "hard_violations": [],
        "physics": "침상 가장자리에 앉아 있으며, 오른쪽 발은 바닥을 향하고 왼쪽 다리는 작은 의자 위에 올려져 있어 체중 분산과 자세의 지지가 매우 현실적이고 안정적입니다."
       },
       {
        "label": "A",
        "direction": "인물의 시선은 천장을 향해 위쪽으로 쏠려 있습니다.",
        "built_space": "카센터 창고 내부의 구조(침상, 작업대, 뒤편의 캠핑카)가 참조 사진의 배치와 잘 맞습니다. 좌측 벽면에 록밴드 포스터들이 붙어 있습니다.",
        "entities": "인물의 얼굴과 머리 모양은 참조 이미지와 흡사하나, 의상이 짙은 회색 반팔 티셔츠와 밝은 반바지로 완전히 변경되었습니다. 왼쪽 종아리에 붕대가 감겨 있습니다.",
        "hard_violations": [
         "캐릭터 참조 지침 위반: 참조 이미지에서 반드시 유지해야 할 의상(후드티와 청바지)을 무시하고 완전히 다른 형태의 의상(반팔 티셔츠와 반바지)을 입혔습니다."
        ],
        "physics": "침상 가장자리에 안정적으로 걸터앉아 있으며, 양손을 침상 바닥과 무릎에 올려두어 몸을 지탱하고 있습니다. 물리적으로 어긋난 부분은 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지정된 카센터 창고의 구조를 정확히 따랐으며, 인물의 외형과 지정된 의상을 완벽히 유지한 채 눈을 크게 뜬 표정과 붕대를 감은 다리의 설정을 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "공간과 인물의 얼굴은 잘 묘사되었으나, 캐릭터 참조 이미지에서 유지해야 할 의상 규정을 무시하고 임의의 옷을 렌더링하여 심각한 지침 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "인물의 시선은 화면 우측을 향하고 있으며, 두 눈을 크게 뜨고 주변을 살피는 듯한 뚜렷한 응시가 보입니다.",
        "built_space": "카센터 창고의 배경이 참조 사진과 매우 일치합니다. 침상, 작업대, 캠핑카가 올바른 위치에 있으며, 침상 앞의 작은 검은 의자가 붕대 감긴 다리를 받치는 용도로 적절히 활용되었습니다. 벽면과 좌측 구조물에 록밴드 포스터가 부착되어 있습니다.",
        "entities": "인물의 얼굴, 헤어스타일, 체형뿐만 아니라 참조 이미지에 명시된 낡은 어두운 후드티와 찢어진 청바지까지 완벽하게 일치합니다. 왼쪽 다리와 발에 붕대가 감겨 있으며, 작업대 위에는 플라이어가 놓여 있습니다.",
        "hard_violations": [],
        "physics": "침상 가장자리에 앉아 있으며, 오른쪽 발은 바닥을 향하고 왼쪽 다리는 작은 의자 위에 올려져 있어 체중 분산과 자세의 지지가 매우 현실적이고 안정적입니다."
       },
       {
        "label": "A",
        "direction": "인물의 시선은 천장을 향해 위쪽으로 쏠려 있습니다.",
        "built_space": "카센터 창고 내부의 구조(침상, 작업대, 뒤편의 캠핑카)가 참조 사진의 배치와 잘 맞습니다. 좌측 벽면에 록밴드 포스터들이 붙어 있습니다.",
        "entities": "인물의 얼굴과 머리 모양은 참조 이미지와 흡사하나, 의상이 짙은 회색 반팔 티셔츠와 밝은 반바지로 완전히 변경되었습니다. 왼쪽 종아리에 붕대가 감겨 있습니다.",
        "hard_violations": [
         "캐릭터 참조 지침 위반: 참조 이미지에서 반드시 유지해야 할 의상(후드티와 청바지)을 무시하고 완전히 다른 형태의 의상(반팔 티셔츠와 반바지)을 입혔습니다."
        ],
        "physics": "침상 가장자리에 안정적으로 걸터앉아 있으며, 양손을 침상 바닥과 무릎에 올려두어 몸을 지탱하고 있습니다. 물리적으로 어긋난 부분은 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴·후드 의상과 침상 주변 배치는 비교적 충실하지만, 포스터의 읽히는 영문 제목이 무문자 조건을 위반하고 다리 전체까지 펼친 구도도 지정된 미디엄 숏보다 넓다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "침상 가장자리에 막 일어나 앉은 동작과 처치한 다리는 자연스럽지만, 차고 전경을 강조한 넓은 구도와 반팔·반바지는 지정 프레이밍 및 인물 참조 의상과 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개를 카메라 쪽으로 돌리고 렌즈 가까운 화면 밖을 바라본다. 시선의 구체적인 대상은 보이지 않으며 원문도 대상을 지정하지 않는다. 눈은 정상적인 형태로 떠 있지만 두 눈을 번쩍 뜬 순간의 놀람은 약하다. 붕대 다리는 화면 오른쪽 앞으로 뻗어 있다.",
        "built_space": "왼쪽 벽을 따라 침상 하나, 바로 오른쪽에 금속 작업대 하나, 작업대 아래 서랍장 하나, 뒤쪽 수납 선반 한 구역, 오른쪽에 캠핑카 일부가 보인다. 침상 옆 검은 스툴 하나가 뻗은 발 아래에 있다. 참조의 침상·작업대·후방 선반 관계는 대체로 유지된다. 다만 침상 상판과 차고 내부가 지정된 제한적 배경보다 많이 노출된다. 반사는 없다.",
        "entities": "실제 인물은 젊은 동아시아계 남성 한 명이며 검은 헝클어진 머리, 얼굴 윤곽, 마른 체격, 낡은 짙은 후드와 찢어진 청바지가 인물 참조에 가깝다. 정확한 나이와 국적은 외양만으로 확정할 수 없다. 한쪽 종아리와 발에 붕대가 있고 얼굴에도 작은 상처가 보인다. 작업대에는 큰 빨간 손잡이 펜치가 놓여 있다. 벽의 인쇄된 밴드 얼굴은 실제 추가 인물이 아니지만, 포스터 일부의 영문 제목은 읽을 수 있는 형태로 드러난다. 스피커와 신발 속 카드는 확인되지 않으며, 화면 밖 물건의 누락으로 판단하지 않는다.",
        "hard_violations": [
         "배경 포스터에 읽을 수 있는 영문 밴드 제목이 노출되어, 읽히는 글자와 특정 밴드명 도입을 금지한 조건을 위반한다."
        ],
        "physics": "엉덩이와 허벅지가 매트리스에 놓여 체중을 지탱한다. 앞으로 뻗은 붕대 다리의 발뒤꿈치 아래에는 검은 스툴 좌면이 보여 지지가 가능하다. 손은 무릎 부근에 놓여 있고 펜치는 작업대 상판에 놓여 있다. 명백하게 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "고개를 들고 화면 오른쪽 위의 프레임 밖을 바라본다. 보이는 시선 대상은 없지만 원문에 특정 목표가 없으므로 방향 자체의 모순은 없다. 벌어진 입과 열린 눈은 낯선 공간을 살피는 반응으로 읽히며, 눈을 번쩍 뜨는 강도는 다소 약하다. 두 다리는 침상 가장자리에서 아래쪽으로 내려온다.",
        "built_space": "왼쪽에 침상 하나, 그 옆 작업대 하나와 하부 서랍장 하나, 뒤편 수납 선반, 오른쪽에 보닛을 연 캠핑카 한 대가 있다. 작업대 앞 스툴 하나, 오른쪽 벽의 스피커 캐비닛 두 개와 공구함, 높은 창 두 곳도 보인다. 침상·작업대·차량의 상대적 배치는 참조와 대체로 맞는다. 천장에는 참조에서 확인되지 않는 매달린 조명이 보인다. 넓은 바닥과 차량 전체에 가까운 모습까지 포함하여 미디엄 숏보다는 차고를 소개하는 넓은 장면에 가깝다. 반사 오류는 보이지 않는다.",
        "entities": "실제 인물은 검은 헝클어진 머리의 젊은 동아시아계 남성 한 명이며 얼굴과 체격은 참조와 대체로 유사하다. 다만 짙은 후드·청바지 대신 회색 반팔과 밝은 반바지를 입었다. 한쪽 종아리에 붕대가 있고 얼굴에는 작은 상처 흔적이 보인다. 큰 빨간 손잡이 펜치는 손이 닿을 법한 작업대 위에 있다. 벽 곳곳에 공연자와 밴드 사진 포스터가 있으며, 추가 실제 인물은 없다. 열린 보닛은 차량 수리 상황을 보여 주지만 엔진·브레이크 손상의 정도는 확인할 수 없다. 신발 속 카드는 프레임에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "엉덩이는 매트리스 가장자리에 닿아 있고 화면 왼쪽 손은 침상을 눌러 상체를 지탱한다. 다른 손은 허벅지 위에 놓여 있다. 무릎을 굽힌 다리가 아래로 내려가므로 막 몸을 일으킨 착석 동작이 가능하다. 발의 바닥 접촉은 프레임 밖이지만 좌면과 손의 지지가 명확하다. 펜치와 용기는 작업대 위에 안정적으로 놓여 있으며 지지 없이 떠 있는 대상은 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "얼굴·후드 의상과 침상 주변 배치는 비교적 충실하지만, 포스터의 읽히는 영문 제목이 무문자 조건을 위반하고 다리 전체까지 펼친 구도도 지정된 미디엄 숏보다 넓다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "침상 가장자리에 막 일어나 앉은 동작과 처치한 다리는 자연스럽지만, 차고 전경을 강조한 넓은 구도와 반팔·반바지는 지정 프레이밍 및 인물 참조 의상과 다르다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "고개를 카메라 쪽으로 돌리고 렌즈 가까운 화면 밖을 바라본다. 시선의 구체적인 대상은 보이지 않으며 원문도 대상을 지정하지 않는다. 눈은 정상적인 형태로 떠 있지만 두 눈을 번쩍 뜬 순간의 놀람은 약하다. 붕대 다리는 화면 오른쪽 앞으로 뻗어 있다.",
        "built_space": "왼쪽 벽을 따라 침상 하나, 바로 오른쪽에 금속 작업대 하나, 작업대 아래 서랍장 하나, 뒤쪽 수납 선반 한 구역, 오른쪽에 캠핑카 일부가 보인다. 침상 옆 검은 스툴 하나가 뻗은 발 아래에 있다. 참조의 침상·작업대·후방 선반 관계는 대체로 유지된다. 다만 침상 상판과 차고 내부가 지정된 제한적 배경보다 많이 노출된다. 반사는 없다.",
        "entities": "실제 인물은 젊은 동아시아계 남성 한 명이며 검은 헝클어진 머리, 얼굴 윤곽, 마른 체격, 낡은 짙은 후드와 찢어진 청바지가 인물 참조에 가깝다. 정확한 나이와 국적은 외양만으로 확정할 수 없다. 한쪽 종아리와 발에 붕대가 있고 얼굴에도 작은 상처가 보인다. 작업대에는 큰 빨간 손잡이 펜치가 놓여 있다. 벽의 인쇄된 밴드 얼굴은 실제 추가 인물이 아니지만, 포스터 일부의 영문 제목은 읽을 수 있는 형태로 드러난다. 스피커와 신발 속 카드는 확인되지 않으며, 화면 밖 물건의 누락으로 판단하지 않는다.",
        "hard_violations": [
         "배경 포스터에 읽을 수 있는 영문 밴드 제목이 노출되어, 읽히는 글자와 특정 밴드명 도입을 금지한 조건을 위반한다."
        ],
        "physics": "엉덩이와 허벅지가 매트리스에 놓여 체중을 지탱한다. 앞으로 뻗은 붕대 다리의 발뒤꿈치 아래에는 검은 스툴 좌면이 보여 지지가 가능하다. 손은 무릎 부근에 놓여 있고 펜치는 작업대 상판에 놓여 있다. 명백하게 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "고개를 들고 화면 오른쪽 위의 프레임 밖을 바라본다. 보이는 시선 대상은 없지만 원문에 특정 목표가 없으므로 방향 자체의 모순은 없다. 벌어진 입과 열린 눈은 낯선 공간을 살피는 반응으로 읽히며, 눈을 번쩍 뜨는 강도는 다소 약하다. 두 다리는 침상 가장자리에서 아래쪽으로 내려온다.",
        "built_space": "왼쪽에 침상 하나, 그 옆 작업대 하나와 하부 서랍장 하나, 뒤편 수납 선반, 오른쪽에 보닛을 연 캠핑카 한 대가 있다. 작업대 앞 스툴 하나, 오른쪽 벽의 스피커 캐비닛 두 개와 공구함, 높은 창 두 곳도 보인다. 침상·작업대·차량의 상대적 배치는 참조와 대체로 맞는다. 천장에는 참조에서 확인되지 않는 매달린 조명이 보인다. 넓은 바닥과 차량 전체에 가까운 모습까지 포함하여 미디엄 숏보다는 차고를 소개하는 넓은 장면에 가깝다. 반사 오류는 보이지 않는다.",
        "entities": "실제 인물은 검은 헝클어진 머리의 젊은 동아시아계 남성 한 명이며 얼굴과 체격은 참조와 대체로 유사하다. 다만 짙은 후드·청바지 대신 회색 반팔과 밝은 반바지를 입었다. 한쪽 종아리에 붕대가 있고 얼굴에는 작은 상처 흔적이 보인다. 큰 빨간 손잡이 펜치는 손이 닿을 법한 작업대 위에 있다. 벽 곳곳에 공연자와 밴드 사진 포스터가 있으며, 추가 실제 인물은 없다. 열린 보닛은 차량 수리 상황을 보여 주지만 엔진·브레이크 손상의 정도는 확인할 수 없다. 신발 속 카드는 프레임에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "엉덩이는 매트리스 가장자리에 닿아 있고 화면 왼쪽 손은 침상을 눌러 상체를 지탱한다. 다른 손은 허벅지 위에 놓여 있다. 무릎을 굽힌 다리가 아래로 내려가므로 막 몸을 일으킨 착석 동작이 가능하다. 발의 바닥 접촉은 프레임 밖이지만 좌면과 손의 지지가 명확하다. 펜치와 용기는 작업대 위에 안정적으로 놓여 있으며 지지 없이 떠 있는 대상은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 캐릭터 참조 지침 위반: 참조 이미지에서 반드시 유지해야 할 의상(후드티와 청바지)을 무시하고 완전히 다른 형태의 의상(반팔 티셔츠와 반바지)을 입혔습니다."
    ],
    "B": [
     "[gpt-high] 배경 포스터에 읽을 수 있는 영문 밴드 제목이 노출되어, 읽히는 글자와 특정 밴드명 도입을 금지한 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1250,
   "A": 1083
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "지정된 카센터 창고의 구조를 정확히 따랐으며, 인물의 외형과 지정된 의상을 완벽히 유지한 채 눈을 크게 뜬 표정과 붕대를 감은 다리의 설정을 훌륭하게 구현했습니다.  ★위반: [gpt-high] 배경 포스터에 읽을 수 있는 영문 밴드 제목이 노출되어, 읽히는 글자와 특정 밴드명 도입을 금지한 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "공간과 인물의 얼굴은 잘 묘사되었으나, 캐릭터 참조 이미지에서 유지해야 할 의상 규정을 무시하고 임의의 옷을 렌더링하여 심각한 지침 위반이 발생했습니다.  ★위반: [gemini-pro] 캐릭터 참조 지침 위반: 참조 이미지에서 반드시 유지해야 할 의상(후드티와 청바지)을 무시하고 완전히 다른 형태의 의상(반팔 티셔츠와 반바지)을 입혔습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L67B01.png",
    "asset_id": "38f979f5-28cd-471a-a942-4cbe22909078",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdda-aad3-7e1c-8a92-5889e55d5eef",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S50sh1__bgfirst_bg.png",
   "bg_asset_id": "dbdd5fe6-62a1-4304-92a6-29ab6d4465e4",
   "bg_record_key": "S50sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S50sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:43:54.764374+00:00",
  "fingerprint": "482aaaba4fc9a63b90960953530cafb54fddf7432dc048e771bf29eb8b25a8b4",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S50sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S50sh1_sel.png",
  "source_sha256": "80ee69d122f5f5e97bdd5caddbaa325d8e489c25410a2bb2e9c367996c96d7c9",
  "file": "S50sh1_cine.png",
  "staged_sha256": "0cd960005863cc4f8a923946dd93bc370cb4eb86a57ac165f9113348d3be5491",
  "latency_ms": 11066
 },
 "S50sh8::signage": {
  "fp": "1a7562a5a3723fe3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S50sh8::bgfirst_bg": {
  "input_fingerprint": "2a7102d42267a582",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 태진이 캠핑카의 엔진 투입구 안으로 검은색 용액을 콸콸 쏟아 붓는 힘이 실린 자세.\n\nLOCATION (lock): At the damaged camper's engine service point inside the auto repair garage, in daytime workshop ambient light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camp​er engine opening (Receiving the repair solution) — The accessible opening is seen obliquely from above beside 태진's forearms; used as Anchors the destination of the pour at lower right; Repair-solution bottle (Tilted with black solution pouring out) — Its mouth points downward toward the vehicle opening; used as Connects her physical effort to the stream without obscuring her hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the daytime interior preserves facial detail and clear separation between her hands, the bottle, and the black solution.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 태진이 캠핑카의 엔진 투입구 안으로 검은색 용액을 콸콸 쏟아 붓는 힘이 실린 자세.\n\nLOCATION (lock): At the damaged camper's engine service point inside the auto repair garage, in daytime workshop ambient light.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camp​er engine opening (Receiving the repair solution) — The accessible opening is seen obliquely from above beside 태진's forearms; used as Anchors the destination of the pour at lower right; Repair-solution bottle (Tilted with black solution pouring out) — Its mouth points downward toward the vehicle opening; used as Connects her physical effort to the stream without obscuring her hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the daytime interior preserves facial detail and clear separation between her hands, the bottle, and the black solution.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S50sh8__bgfirst_bg.png",
  "asset_id": "e00b4baa-a2b1-4328-83bb-eaa26aa8af94",
  "input_asset_ids": [
   "7f8dfef6-34cc-4edc-9807-541725a38394",
   "38f979f5-28cd-471a-a942-4cbe22909078"
  ]
 },
 "S50sh8": {
  "input_fingerprint": "5b0734ef81b2e328",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 태진이 캠핑카의 엔진 투입구 안으로 검은색 용액을 콸콸 쏟아 붓는 힘이 실린 자세.\n\nLOCATION (lock): At the damaged camper's engine service point inside the auto repair garage, in daytime workshop ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camp​er engine opening (Receiving the repair solution) — The accessible opening is seen obliquely from above beside 태진's forearms; used as Anchors the destination of the pour at lower right; Repair-solution bottle (Tilted with black solution pouring out) — Its mouth points downward toward the vehicle opening; used as Connects her physical effort to the stream without obscuring her hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the daytime interior preserves facial detail and clear separation between her hands, the bottle, and the black solution.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains under repair with extensive engine and brake damage; repair solution is being poured into its inlet. Rock-band posters remain on the storage-room walls, and the large pliers have not been established as picked up. 태진: She stands at the camper holding and tipping the bottle of repair solution.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 태진 right now, so 태진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 태진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (30세 여성, 단발머리, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 태진이 캠핑카의 엔진 투입구 안으로 검은색 용액을 콸콸 쏟아 붓는 힘이 실린 자세.\n\nLOCATION (lock): At the damaged camper's engine service point inside the auto repair garage, in daytime workshop ambient light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camp​er engine opening (Receiving the repair solution) — The accessible opening is seen obliquely from above beside 태진's forearms; used as Anchors the destination of the pour at lower right; Repair-solution bottle (Tilted with black solution pouring out) — Its mouth points downward toward the vehicle opening; used as Connects her physical effort to the stream without obscuring her hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the daytime interior preserves facial detail and clear separation between her hands, the bottle, and the black solution.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains under repair with extensive engine and brake damage; repair solution is being poured into its inlet. Rock-band posters remain on the storage-room walls, and the large pliers have not been established as picked up. 태진: She stands at the camper holding and tipping the bottle of repair solution.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 태진 right now, so 태진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 태진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (30세 여성, 단발머리, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 태진이 캠핑카의 엔진 투입구 안으로 검은색 용액을 콸콸 쏟아 붓는 힘이 실린 자세.\n\nLOCATION (lock): At the damaged camper's engine service point inside the auto repair garage, in daytime workshop ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camp​er engine opening (Receiving the repair solution) — The accessible opening is seen obliquely from above beside 태진's forearms; used as Anchors the destination of the pour at lower right; Repair-solution bottle (Tilted with black solution pouring out) — Its mouth points downward toward the vehicle opening; used as Connects her physical effort to the stream without obscuring her hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the daytime interior preserves facial detail and clear separation between her hands, the bottle, and the black solution.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains under repair with extensive engine and brake damage; repair solution is being poured into its inlet. Rock-band posters remain on the storage-room walls, and the large pliers have not been established as picked up. 태진: She stands at the camper holding and tipping the bottle of repair solution.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 태진 right now, so 태진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 태진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (30세 여성, 단발머리, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S50sh8__bgfirst_bg.png",
     "asset_id": "e00b4baa-a2b1-4328-83bb-eaa26aa8af94",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S50sh8.png",
     "asset_id": "7f8dfef6-34cc-4edc-9807-541725a38394",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:942632>",
     "asset_id": "cb567190-df6f-4833-8024-26cec5633f22",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L67B01.png",
     "asset_id": "38f979f5-28cd-471a-a942-4cbe22909078",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:942632>",
     "asset_id": "cb567190-df6f-4833-8024-26cec5633f22",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선과 통 입구가 엔진 투입구를 향하며 검은 용액이 투입구로 정확히 떨어짐.",
    "built_space": "차고 내부의 작업대와 공구 배치는 레퍼런스와 일치하나, 전경의 엔진룸과 배경의 캠핑카가 전혀 연결되지 않고 분리된 채 묘사됨.",
    "entities": "태진의 복장(모자, 반다나), 작업대의 펜치, 벽의 록 밴드 포스터가 프롬프트 및 레퍼런스와 정확히 일치함.",
    "hard_violations": [
     "[gemini-pro] 발명된 객체 및 물리적으로 불가능한 무대 구조(전경의 엔진룸과 배경의 캠핑카가 분리되어 두 대의 차량처럼 묘사됨)"
    ],
    "physics": "두 손으로 통을 쥐고 용액을 붓는 자세와 중력에 따른 용액의 낙하가 물리적으로 자연스러움."
   },
   {
    "label": "B",
    "direction": "시선과 통 입구가 프롬프트에 없는 깔때기를 향하고 용액이 그곳으로 떨어짐.",
    "built_space": "차고 내부 배경은 유사하나, 전경의 엔진룸과 배경의 차체가 구조적으로 연결되지 않는 오류가 동일하게 존재함.",
    "entities": "녹색 작업복은 입었으나 필수 복장인 모자와 반다나가 누락되었고, 벽면의 포스터도 묘사되지 않음.",
    "hard_violations": [
     "[gemini-pro] 발명된 객체(프롬프트에 없는 깔때기)",
     "[gemini-pro] 물리적으로 불가능한 무대 구조(전경의 엔진룸과 배경의 캠핑카가 구조적으로 분리됨)",
     "[gpt-high] 지정되지 않은 큰 검은 깔때기를 핵심 전경 소품으로 추가하여, 용액을 직접 받는 엔진 투입구 대신 깔때기로 붓는 장면을 만들었습니다."
    ],
    "physics": "오른팔이 몸을 가로질러 통 바닥을 받치는 자세가 해부학적으로 매우 부자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "필수 복장(모자, 반다나)과 소품(포스터, 펜치)을 정확히 구현했으나, 차량 구조의 심각한 오류(엔진룸과 차체 분리)가 발생함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "필수 복장 누락, 불필요한 깔때기 추가, 부자연스러운 자세 및 차량 구조 오류로 인해 지시사항을 다수 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선과 통 입구가 엔진 투입구를 향하며 검은 용액이 투입구로 정확히 떨어짐.",
        "built_space": "차고 내부의 작업대와 공구 배치는 레퍼런스와 일치하나, 전경의 엔진룸과 배경의 캠핑카가 전혀 연결되지 않고 분리된 채 묘사됨.",
        "entities": "태진의 복장(모자, 반다나), 작업대의 펜치, 벽의 록 밴드 포스터가 프롬프트 및 레퍼런스와 정확히 일치함.",
        "hard_violations": [
         "발명된 객체 및 물리적으로 불가능한 무대 구조(전경의 엔진룸과 배경의 캠핑카가 분리되어 두 대의 차량처럼 묘사됨)"
        ],
        "physics": "두 손으로 통을 쥐고 용액을 붓는 자세와 중력에 따른 용액의 낙하가 물리적으로 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선과 통 입구가 프롬프트에 없는 깔때기를 향하고 용액이 그곳으로 떨어짐.",
        "built_space": "차고 내부 배경은 유사하나, 전경의 엔진룸과 배경의 차체가 구조적으로 연결되지 않는 오류가 동일하게 존재함.",
        "entities": "녹색 작업복은 입었으나 필수 복장인 모자와 반다나가 누락되었고, 벽면의 포스터도 묘사되지 않음.",
        "hard_violations": [
         "발명된 객체(프롬프트에 없는 깔때기)",
         "물리적으로 불가능한 무대 구조(전경의 엔진룸과 배경의 캠핑카가 구조적으로 분리됨)"
        ],
        "physics": "오른팔이 몸을 가로질러 통 바닥을 받치는 자세가 해부학적으로 매우 부자연스러움."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "필수 복장(모자, 반다나)과 소품(포스터, 펜치)을 정확히 구현했으나, 차량 구조의 심각한 오류(엔진룸과 차체 분리)가 발생함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "필수 복장 누락, 불필요한 깔때기 추가, 부자연스러운 자세 및 차량 구조 오류로 인해 지시사항을 다수 위반함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "시선과 통 입구가 엔진 투입구를 향하며 검은 용액이 투입구로 정확히 떨어짐.",
        "built_space": "차고 내부의 작업대와 공구 배치는 레퍼런스와 일치하나, 전경의 엔진룸과 배경의 캠핑카가 전혀 연결되지 않고 분리된 채 묘사됨.",
        "entities": "태진의 복장(모자, 반다나), 작업대의 펜치, 벽의 록 밴드 포스터가 프롬프트 및 레퍼런스와 정확히 일치함.",
        "hard_violations": [
         "발명된 객체 및 물리적으로 불가능한 무대 구조(전경의 엔진룸과 배경의 캠핑카가 분리되어 두 대의 차량처럼 묘사됨)"
        ],
        "physics": "두 손으로 통을 쥐고 용액을 붓는 자세와 중력에 따른 용액의 낙하가 물리적으로 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선과 통 입구가 프롬프트에 없는 깔때기를 향하고 용액이 그곳으로 떨어짐.",
        "built_space": "차고 내부 배경은 유사하나, 전경의 엔진룸과 배경의 차체가 구조적으로 연결되지 않는 오류가 동일하게 존재함.",
        "entities": "녹색 작업복은 입었으나 필수 복장인 모자와 반다나가 누락되었고, 벽면의 포스터도 묘사되지 않음.",
        "hard_violations": [
         "발명된 객체(프롬프트에 없는 깔때기)",
         "물리적으로 불가능한 무대 구조(전경의 엔진룸과 배경의 캠핑카가 구조적으로 분리됨)"
        ],
        "physics": "오른팔이 몸을 가로질러 통 바닥을 받치는 자세가 해부학적으로 매우 부자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "허리 중심의 미디엄 구도와 힘주어 붓는 동작은 좋지만, 지정되지 않은 큰 깔때기가 투입구를 대체하고 태진의 얼굴·모자·목수건이 참조와 다릅니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "프레임이 허벅지까지 다소 넓지만, 참조에 가까운 태진이 두 손으로 용기를 기울여 오른쪽 아래 엔진 투입구에 검은 용액을 직접 붓는 관계를 정확히 보여줍니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "태진의 눈은 손과 액체가 떨어지는 곳을 향합니다. 용기 입구는 아래쪽으로 기울어 있고 굵은 검은 액체가 전경 중앙의 깔때기 안으로 들어갑니다. 흐름은 수용부를 맞히지만, 지정된 오른쪽 아래 엔진 투입구 대신 중앙의 큰 깔때기가 시각적 목표가 됩니다.",
        "built_space": "오른쪽에 캠핑카 차체 한 대와 열린 보닛 하나, 아래에 엔진 공간 하나가 보입니다. 뒤쪽에는 금속 수납 선반, 왼쪽에는 작업대와 작은 서랍함, 벽과 천장에는 노출 구조재가 있습니다. 참조의 낡은 정비소 재료와 분위기는 유사합니다. 태진은 엔진 옆에 서 있지만 깔때기가 실제 투입구와 주변 연결부를 가립니다.",
        "entities": "보이는 사람은 단발의 성인 한 명뿐입니다. 참조보다 얼굴이 길고 각져 보이며 나이 인상도 더 높아, 지정된 30세 여성 태진의 동일성이 약합니다. 올리브색 작업복과 갈색 벨트는 맞지만 회색 모자와 파란 목수건이 없습니다. 큰 반투명 용기와 광택 있는 검은 용액은 존재합니다. 엔진 손상과 브레이크 손상의 정도는 확인하기 어렵습니다. 배경의 작은 표기들은 확실히 읽히지 않습니다.",
        "hard_violations": [
         "지정되지 않은 큰 검은 깔때기를 핵심 전경 소품으로 추가하여, 용액을 직접 받는 엔진 투입구 대신 깔때기로 붓는 장면을 만들었습니다."
        ],
        "physics": "한 손은 용기 손잡이를 잡고 다른 손은 바닥을 받칩니다. 들어 올린 팔꿈치와 앞으로 숙인 몸통이 무거운 액체를 힘주어 붓는 동작에 맞습니다. 액체는 용기 입구에서 아래로 연속해서 떨어지고, 깔때기 하단은 엔진 쪽에 꽂힌 것으로 보입니다. 하체는 잘려 있어 발 접촉은 확인할 수 없지만 몸이 공중에 떠 있다는 증거는 없습니다."
       },
       {
        "label": "B",
        "direction": "태진은 아래쪽 주입 지점을 바라봅니다. 용기 입구는 오른쪽 아래를 향하며 검은 액체 줄기가 오른쪽 아래의 열린 원형 투입구 안으로 정확히 들어갑니다. 두 손, 용기 입구, 액체와 수용부가 서로 가리지 않고 연결됩니다.",
        "built_space": "오른쪽에 캠핑카 차체와 열린 보닛 하나, 전경에 엔진 공간 하나가 있습니다. 왼쪽 벽에는 공구판과 창, 그 아래에는 작업대와 서랍함이 있고 뒤쪽에는 수납 선반과 이동식 공구함이 보입니다. 참조의 정비소와 주요 재료·설비 구성이 가깝습니다. 태진은 작업대와 엔진 사이에 서 있으며, 엔진 투입구는 팔 옆 오른쪽 아래에서 비스듬히 내려다보입니다. 화면은 머리부터 허벅지까지 포함해 요구된 미디엄보다 조금 넓습니다.",
        "entities": "단발의 젊은 성인 여성 한 명이 있으며 얼굴과 체형이 태진 참조에 가깝습니다. 회색 모자, 파란 무늬 목수건, 오염된 올리브색 작업복, 갈색 벨트가 일치합니다. 반투명 수리액 용기, 검은 용액, 열린 엔진 투입구가 모두 보입니다. 큰 펜치는 왼쪽 작업대에 놓여 있고 손에 들려 있지 않습니다. 뒤쪽에 밴드 포스터들이 있으나 문구는 선명하지 않습니다. 심각한 엔진·브레이크 손상은 이 화면만으로 확인하기 어렵습니다.",
        "hard_violations": [],
        "physics": "한 손으로 손잡이를 쥐고 다른 손으로 용기 밑면을 받쳐 무게를 지탱합니다. 팔을 앞으로 뻗고 몸을 약간 기울인 자세에서 액체가 중력 방향으로 흘러 투입구에 들어갑니다. 보닛에는 지지대가 보이고 투입구와 용기는 각각 차량과 손에 의해 지지됩니다. 발은 프레임 밖이지만 보이는 하체 자세에 부유나 불가능한 관절 꺾임은 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "허리 중심의 미디엄 구도와 힘주어 붓는 동작은 좋지만, 지정되지 않은 큰 깔때기가 투입구를 대체하고 태진의 얼굴·모자·목수건이 참조와 다릅니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프레임이 허벅지까지 다소 넓지만, 참조에 가까운 태진이 두 손으로 용기를 기울여 오른쪽 아래 엔진 투입구에 검은 용액을 직접 붓는 관계를 정확히 보여줍니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "태진의 눈은 손과 액체가 떨어지는 곳을 향합니다. 용기 입구는 아래쪽으로 기울어 있고 굵은 검은 액체가 전경 중앙의 깔때기 안으로 들어갑니다. 흐름은 수용부를 맞히지만, 지정된 오른쪽 아래 엔진 투입구 대신 중앙의 큰 깔때기가 시각적 목표가 됩니다.",
        "built_space": "오른쪽에 캠핑카 차체 한 대와 열린 보닛 하나, 아래에 엔진 공간 하나가 보입니다. 뒤쪽에는 금속 수납 선반, 왼쪽에는 작업대와 작은 서랍함, 벽과 천장에는 노출 구조재가 있습니다. 참조의 낡은 정비소 재료와 분위기는 유사합니다. 태진은 엔진 옆에 서 있지만 깔때기가 실제 투입구와 주변 연결부를 가립니다.",
        "entities": "보이는 사람은 단발의 성인 한 명뿐입니다. 참조보다 얼굴이 길고 각져 보이며 나이 인상도 더 높아, 지정된 30세 여성 태진의 동일성이 약합니다. 올리브색 작업복과 갈색 벨트는 맞지만 회색 모자와 파란 목수건이 없습니다. 큰 반투명 용기와 광택 있는 검은 용액은 존재합니다. 엔진 손상과 브레이크 손상의 정도는 확인하기 어렵습니다. 배경의 작은 표기들은 확실히 읽히지 않습니다.",
        "hard_violations": [
         "지정되지 않은 큰 검은 깔때기를 핵심 전경 소품으로 추가하여, 용액을 직접 받는 엔진 투입구 대신 깔때기로 붓는 장면을 만들었습니다."
        ],
        "physics": "한 손은 용기 손잡이를 잡고 다른 손은 바닥을 받칩니다. 들어 올린 팔꿈치와 앞으로 숙인 몸통이 무거운 액체를 힘주어 붓는 동작에 맞습니다. 액체는 용기 입구에서 아래로 연속해서 떨어지고, 깔때기 하단은 엔진 쪽에 꽂힌 것으로 보입니다. 하체는 잘려 있어 발 접촉은 확인할 수 없지만 몸이 공중에 떠 있다는 증거는 없습니다."
       },
       {
        "label": "A",
        "direction": "태진은 아래쪽 주입 지점을 바라봅니다. 용기 입구는 오른쪽 아래를 향하며 검은 액체 줄기가 오른쪽 아래의 열린 원형 투입구 안으로 정확히 들어갑니다. 두 손, 용기 입구, 액체와 수용부가 서로 가리지 않고 연결됩니다.",
        "built_space": "오른쪽에 캠핑카 차체와 열린 보닛 하나, 전경에 엔진 공간 하나가 있습니다. 왼쪽 벽에는 공구판과 창, 그 아래에는 작업대와 서랍함이 있고 뒤쪽에는 수납 선반과 이동식 공구함이 보입니다. 참조의 정비소와 주요 재료·설비 구성이 가깝습니다. 태진은 작업대와 엔진 사이에 서 있으며, 엔진 투입구는 팔 옆 오른쪽 아래에서 비스듬히 내려다보입니다. 화면은 머리부터 허벅지까지 포함해 요구된 미디엄보다 조금 넓습니다.",
        "entities": "단발의 젊은 성인 여성 한 명이 있으며 얼굴과 체형이 태진 참조에 가깝습니다. 회색 모자, 파란 무늬 목수건, 오염된 올리브색 작업복, 갈색 벨트가 일치합니다. 반투명 수리액 용기, 검은 용액, 열린 엔진 투입구가 모두 보입니다. 큰 펜치는 왼쪽 작업대에 놓여 있고 손에 들려 있지 않습니다. 뒤쪽에 밴드 포스터들이 있으나 문구는 선명하지 않습니다. 심각한 엔진·브레이크 손상은 이 화면만으로 확인하기 어렵습니다.",
        "hard_violations": [],
        "physics": "한 손으로 손잡이를 쥐고 다른 손으로 용기 밑면을 받쳐 무게를 지탱합니다. 팔을 앞으로 뻗고 몸을 약간 기울인 자세에서 액체가 중력 방향으로 흘러 투입구에 들어갑니다. 보닛에는 지지대가 보이고 투입구와 용기는 각각 차량과 손에 의해 지지됩니다. 발은 프레임 밖이지만 보이는 하체 자세에 부유나 불가능한 관절 꺾임은 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.1
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.85
   },
   "violations": {
    "A": [
     "[gemini-pro] 발명된 객체 및 물리적으로 불가능한 무대 구조(전경의 엔진룸과 배경의 캠핑카가 분리되어 두 대의 차량처럼 묘사됨)"
    ],
    "B": [
     "[gemini-pro] 발명된 객체(프롬프트에 없는 깔때기)",
     "[gemini-pro] 물리적으로 불가능한 무대 구조(전경의 엔진룸과 배경의 캠핑카가 구조적으로 분리됨)",
     "[gpt-high] 지정되지 않은 큰 검은 깔때기를 핵심 전경 소품으로 추가하여, 용액을 직접 받는 엔진 투입구 대신 깔때기로 붓는 장면을 만들었습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 850
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "필수 복장(모자, 반다나)과 소품(포스터, 펜치)을 정확히 구현했으나, 차량 구조의 심각한 오류(엔진룸과 차체 분리)가 발생함.  ★위반: [gemini-pro] 발명된 객체 및 물리적으로 불가능한 무대 구조(전경의 엔진룸과 배경의 캠핑카가 분리되어 두 대의 차량처럼 묘사됨)"
   },
   {
    "label": "B",
    "score": 850,
    "verdict_ko": "필수 복장 누락, 불필요한 깔때기 추가, 부자연스러운 자세 및 차량 구조 오류로 인해 지시사항을 다수 위반함.  ★위반: [gemini-pro] 발명된 객체(프롬프트에 없는 깔때기) / [gemini-pro] 물리적으로 불가능한 무대 구조(전경의 엔진룸과 배경의 캠핑카가 구조적으로 분리됨) / [gpt-high] 지정되지 않은 큰 검은 깔때기를 핵심 전경 소품으로 추가하여, 용액을 직접 받는 엔진 투입구 대신 깔때기로 붓는 장면을 만들었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L67B01.png",
    "asset_id": "38f979f5-28cd-471a-a942-4cbe22909078",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:942632>",
    "asset_id": "cb567190-df6f-4833-8024-26cec5633f22",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacddb-62b1-7b47-95fb-360d10be00b5",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S50sh8__bgfirst_bg.png",
   "bg_asset_id": "e00b4baa-a2b1-4328-83bb-eaa26aa8af94",
   "bg_record_key": "S50sh8::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S50sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:44:06.254744+00:00",
  "fingerprint": "5bedfc640b766d67155e1a3556c25b5e7952d9224d9e76e8f8d80562ba1c1ea7",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S50sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S50sh8_sel.png",
  "source_sha256": "35b0b75b9e8e524429093b86483b11c02c618bd2f5cda43d99608c4b3f2b6209",
  "file": "S50sh8_cine.png",
  "staged_sha256": "8810dbcceb7322f0260baa3bb23617ce98d482bb483d1b0122b747c7aab3aba2",
  "latency_ms": 12024
 },
 "S50sh10::signage": {
  "fp": "f0d60468ee86f7a4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S50sh10": {
  "input_fingerprint": "a8eec777980de407",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 펜치를 쥔 채 당황한 표정으로 은영과 태진이 있는 쪽을 바라보는 어리둥절한 얼굴.\n\nLOCATION (lock): In the garage's storage-room resting area, facing the adjoining repair bay in dim daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Large pliers (Gripped by 이현우) — Held obliquely at the lower frame edge, with the gripping hand visible; used as Retains the evidence of his defensive impulse beneath the facial reaction; Rock-band posters (Attached around the room) — Their printed band imagery faces into the room and is partially visible behind him; used as Provides subdued spatial continuity without competing with his expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime interior illumination consistent with the repair shot, allowing restrained facial contrast to reveal his uncertainty.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged camper remains in the repair area after repair solution has been poured into its inlet. The storage room retains its rock-band posters and large pliers; the screenplay does not establish that the pliers were picked up. 이현우: He remains awake with a bandaged leg and other visible treatment marks, and his contact card remains concealed in his shoe. His hands are not established as holding the pliers.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 펜치를 쥔 채 당황한 표정으로 은영과 태진이 있는 쪽을 바라보는 어리둥절한 얼굴.\n\nLOCATION (lock): In the garage's storage-room resting area, facing the adjoining repair bay in dim daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Large pliers (Gripped by 이현우) — Held obliquely at the lower frame edge, with the gripping hand visible; used as Retains the evidence of his defensive impulse beneath the facial reaction; Rock-band posters (Attached around the room) — Their printed band imagery faces into the room and is partially visible behind him; used as Provides subdued spatial continuity without competing with his expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime interior illumination consistent with the repair shot, allowing restrained facial contrast to reveal his uncertainty.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged camper remains in the repair area after repair solution has been poured into its inlet. The storage room retains its rock-band posters and large pliers; the screenplay does not establish that the pliers were picked up. 이현우: He remains awake with a bandaged leg and other visible treatment marks, and his contact card remains concealed in his shoe. His hands are not established as holding the pliers.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 펜치를 쥔 채 당황한 표정으로 은영과 태진이 있는 쪽을 바라보는 어리둥절한 얼굴.\n\nLOCATION (lock): In the garage's storage-room resting area, facing the adjoining repair bay in dim daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Large pliers (Gripped by 이현우) — Held obliquely at the lower frame edge, with the gripping hand visible; used as Retains the evidence of his defensive impulse beneath the facial reaction; Rock-band posters (Attached around the room) — Their printed band imagery faces into the room and is partially visible behind him; used as Provides subdued spatial continuity without competing with his expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime interior illumination consistent with the repair shot, allowing restrained facial contrast to reveal his uncertainty.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged camper remains in the repair area after repair solution has been poured into its inlet. The storage room retains its rock-band posters and large pliers; the screenplay does not establish that the pliers were picked up. 이현우: He remains awake with a bandaged leg and other visible treatment marks, and his contact card remains concealed in his shoe. His hands are not established as holding the pliers.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 우측 밖을 향하고 있으며, 펜치는 화면 하단에 위치함.",
    "built_space": "포스터가 가득 붙은 평면적인 벽만 보이며, 지시된 수리 공간(repair bay)을 향하는 카메라 각도가 완전히 무시됨.",
    "entities": "이현우의 외모는 일치하나, 펜치가 이전 샷의 소품과 완전히 다른 형태임.",
    "hard_violations": [
     "[gemini-pro] physically impossible object (펜치에 경첩이 없고 손잡이와 머리가 하나로 융합된 불가능한 형태)",
     "[gemini-pro] staging violation (지정된 수리 공간 방향이 아닌 잘못된 카메라 앵글)",
     "[gpt-high] 오른쪽 배경 포스터의 큰 영문 표제 일부가 읽을 수 있는 상태로 남아 있어 문자 비가독성 조건을 위반한다."
    ],
    "physics": "손이 펜치를 쥐고 있으나, 도구 자체가 물리적 관절이 없는 비정상적인 금속 덩어리로 변형되어 있음."
   },
   {
    "label": "B",
    "direction": "시선은 화면 좌측 전경에 있는 두 인물을 향하고 있음.",
    "built_space": "좌측의 포스터 벽, 침대, 그리고 우측 후경의 수리 공간까지 레퍼런스의 공간 구조와 방향을 정확히 구현함.",
    "entities": "이현우의 인상착의가 레퍼런스와 일치하며, 펜치 역시 이전 샷과 동일한 빨간색 손잡이 모델임. 전경에 두 명의 타 인물이 등장함.",
    "hard_violations": [
     "[gemini-pro] extra bodies (PEOPLE 지침에서 단독 등장을 명시했으나 전경에 금지된 두 인물이 노출됨)",
     "[gpt-high] 이현우만 보여야 하는 장면에 왼쪽 전경의 얼굴·어깨와 오른쪽 전경의 머리로 두 명의 추가 인물이 등장한다.",
     "[gpt-high] 왼쪽 아래 포스터 등에 식별 가능한 영문 표기가 남아 있어 읽을 수 있는 문자 금지 조건을 위반한다."
    ],
    "physics": "오른손으로 펜치의 손잡이를 자연스럽게 쥐고 있으며, 물체와 손의 접촉 및 지지 상태가 정상적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "배경 구조와 이전 샷의 소품(빨간 펜치)을 완벽히 유지했으나, 지침에서 엄격히 금지한 프레임 내 타 인물 노출로 인해 치명적 감점이 발생했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "카메라가 수리 공간이 아닌 벽면을 향해 공간 지침을 어겼고, 펜치가 물리적으로 불가능한 금속 덩어리로 묘사되었습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "시선은 화면 좌측 전경에 있는 두 인물을 향하고 있음.",
        "built_space": "좌측의 포스터 벽, 침대, 그리고 우측 후경의 수리 공간까지 레퍼런스의 공간 구조와 방향을 정확히 구현함.",
        "entities": "이현우의 인상착의가 레퍼런스와 일치하며, 펜치 역시 이전 샷과 동일한 빨간색 손잡이 모델임. 전경에 두 명의 타 인물이 등장함.",
        "hard_violations": [
         "extra bodies (PEOPLE 지침에서 단독 등장을 명시했으나 전경에 금지된 두 인물이 노출됨)"
        ],
        "physics": "오른손으로 펜치의 손잡이를 자연스럽게 쥐고 있으며, 물체와 손의 접촉 및 지지 상태가 정상적임."
       },
       {
        "label": "A",
        "direction": "시선은 화면 우측 밖을 향하고 있으며, 펜치는 화면 하단에 위치함.",
        "built_space": "포스터가 가득 붙은 평면적인 벽만 보이며, 지시된 수리 공간(repair bay)을 향하는 카메라 각도가 완전히 무시됨.",
        "entities": "이현우의 외모는 일치하나, 펜치가 이전 샷의 소품과 완전히 다른 형태임.",
        "hard_violations": [
         "physically impossible object (펜치에 경첩이 없고 손잡이와 머리가 하나로 융합된 불가능한 형태)",
         "staging violation (지정된 수리 공간 방향이 아닌 잘못된 카메라 앵글)"
        ],
        "physics": "손이 펜치를 쥐고 있으나, 도구 자체가 물리적 관절이 없는 비정상적인 금속 덩어리로 변형되어 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "배경 구조와 이전 샷의 소품(빨간 펜치)을 완벽히 유지했으나, 지침에서 엄격히 금지한 프레임 내 타 인물 노출로 인해 치명적 감점이 발생했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "카메라가 수리 공간이 아닌 벽면을 향해 공간 지침을 어겼고, 펜치가 물리적으로 불가능한 금속 덩어리로 묘사되었습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "시선은 화면 좌측 전경에 있는 두 인물을 향하고 있음.",
        "built_space": "좌측의 포스터 벽, 침대, 그리고 우측 후경의 수리 공간까지 레퍼런스의 공간 구조와 방향을 정확히 구현함.",
        "entities": "이현우의 인상착의가 레퍼런스와 일치하며, 펜치 역시 이전 샷과 동일한 빨간색 손잡이 모델임. 전경에 두 명의 타 인물이 등장함.",
        "hard_violations": [
         "extra bodies (PEOPLE 지침에서 단독 등장을 명시했으나 전경에 금지된 두 인물이 노출됨)"
        ],
        "physics": "오른손으로 펜치의 손잡이를 자연스럽게 쥐고 있으며, 물체와 손의 접촉 및 지지 상태가 정상적임."
       },
       {
        "label": "A",
        "direction": "시선은 화면 우측 밖을 향하고 있으며, 펜치는 화면 하단에 위치함.",
        "built_space": "포스터가 가득 붙은 평면적인 벽만 보이며, 지시된 수리 공간(repair bay)을 향하는 카메라 각도가 완전히 무시됨.",
        "entities": "이현우의 외모는 일치하나, 펜치가 이전 샷의 소품과 완전히 다른 형태임.",
        "hard_violations": [
         "physically impossible object (펜치에 경첩이 없고 손잡이와 머리가 하나로 융합된 불가능한 형태)",
         "staging violation (지정된 수리 공간 방향이 아닌 잘못된 카메라 앵글)"
        ],
        "physics": "손이 펜치를 쥐고 있으나, 도구 자체가 물리적 관절이 없는 비정상적인 금속 덩어리로 변형되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "이현우 외의 인물이 양쪽 전경에 추가되어 단독 얼굴 장면을 위반하며, 펜치도 하단에 비스듬히 놓이기보다 세워져 있다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "단독 클로즈업, 화면 밖을 향한 당황한 시선과 손에 쥔 비스듬한 펜치를 더 충실하게 구현하지만, 배경 포스터의 읽히는 문자는 금지 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 눈은 화면 왼쪽 전경 인물을 향한다. 오른쪽 전경에도 다른 인물의 머리 일부가 보인다. 은영과 태진 쪽을 보는 관계를 사람을 직접 추가하는 방식으로 표현했으나, 이 장면의 등장인물 제한에 맞지 않는다. 펜치 턱은 거의 수직으로 위를 향하며 특정 사람을 겨누지는 않는다.",
        "built_space": "뒤쪽에는 밴드 포스터가 겹쳐 붙은 벽, 침대 한 개와 베개 한 개, 밝은 개구부 한 곳, 오른쪽 끝의 작은 서랍 수납함 일부가 보인다. 낡은 휴식 공간의 재질과 낮의 약한 실내광은 참고와 유사하다. 이현우의 골반은 잘려 있어 정확한 착석 접점은 확인할 수 없다. 중복된 침대나 불가능한 반사는 보이지 않는다.",
        "entities": "중앙 인물은 참고와 유사한 젊은 한국계 남성의 얼굴, 헝클어진 검은 머리, 낡은 짙은 회색 후드를 갖췄다. 주름진 미간과 벌어진 입으로 당황한 반응이 보인다. 붉은 손잡이의 큰 펜치를 쥐었지만 턱 형태는 참고 도구와 다소 다르다. 양쪽 전경의 실제 사람 두 명은 허용되지 않은 추가 인물이다. 포스터에는 밴드 사진과 식별 가능한 영문 글자가 남아 있다. 다리 붕대와 신발 속 카드는 프레임 밖이다.",
        "hard_violations": [
         "이현우만 보여야 하는 장면에 왼쪽 전경의 얼굴·어깨와 오른쪽 전경의 머리로 두 명의 추가 인물이 등장한다.",
         "왼쪽 아래 포스터 등에 식별 가능한 영문 표기가 남아 있어 읽을 수 있는 문자 금지 조건을 위반한다."
        ],
        "physics": "펜치 손잡이는 이현우의 손가락과 손바닥에 잡혀 있고 손목은 후드 소매로 이어져 있어 지지가 성립한다. 팔을 굽혀 도구를 든 자세는 가능하다. 하체는 프레임 밖이며, 보이는 상체나 도구에서 근거 없는 부유는 없다."
       },
       {
        "label": "B",
        "direction": "이현우의 얼굴과 두 눈은 화면 오른쪽 바깥을 향한다. 은영과 태진은 보이지 않지만 그들이 있는 화면 밖 방향을 바라보는 연기로 읽힌다. 펜치는 손에서 오른쪽 위로 비스듬히 뻗으며, 상대를 직접 겨누는 동작은 아니다.",
        "built_space": "배경에는 참고의 얼룩진 밝은 벽과 여러 장의 겹친 밴드 포스터가 보인다. 큰 초상 포스터는 왼쪽, 선글라스를 쓴 인물 사진은 중앙 뒤, 단체 사진은 오른쪽에 놓여 참고 공간의 연속성이 높다. 침대·작업대·캠핑카는 이 클로즈업 밖이므로 그 생략은 결함이 아니다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "실제 등장인물은 이현우 한 명이며, 젊은 한국계 남성의 얼굴과 검은 흐트러진 머리, 때 묻고 해진 짙은 후드가 참고와 부합한다. 올라간 눈썹, 옆으로 향한 눈, 벌어진 입이 어리둥절한 반응을 전달한다. 손과 손목이 보이고 큰 금속 펜치를 잡고 있으나, 참고의 붉은 손잡이 도구와 외형은 다소 다르다. 배경의 밴드 인물들은 인쇄 사진이다. 오른쪽 포스터 하단의 큰 영문 표기는 완전히 흐려지지 않았다. 붕대와 신발은 프레임 밖이다.",
        "hard_violations": [
         "오른쪽 배경 포스터의 큰 영문 표제 일부가 읽을 수 있는 상태로 남아 있어 문자 비가독성 조건을 위반한다."
        ],
        "physics": "이현우의 손이 펜치 손잡이를 감싸고 있으며 손목과 굽힌 팔이 소매로 자연스럽게 이어진다. 비스듬히 들어 올린 도구의 무게를 손과 팔이 받치는 자세가 가능하다. 하체 지지면은 클로즈업 밖이고, 보이는 신체나 도구에 지지 없는 부유는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "이현우 외의 인물이 양쪽 전경에 추가되어 단독 얼굴 장면을 위반하며, 펜치도 하단에 비스듬히 놓이기보다 세워져 있다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "단독 클로즈업, 화면 밖을 향한 당황한 시선과 손에 쥔 비스듬한 펜치를 더 충실하게 구현하지만, 배경 포스터의 읽히는 문자는 금지 조건을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 눈은 화면 왼쪽 전경 인물을 향한다. 오른쪽 전경에도 다른 인물의 머리 일부가 보인다. 은영과 태진 쪽을 보는 관계를 사람을 직접 추가하는 방식으로 표현했으나, 이 장면의 등장인물 제한에 맞지 않는다. 펜치 턱은 거의 수직으로 위를 향하며 특정 사람을 겨누지는 않는다.",
        "built_space": "뒤쪽에는 밴드 포스터가 겹쳐 붙은 벽, 침대 한 개와 베개 한 개, 밝은 개구부 한 곳, 오른쪽 끝의 작은 서랍 수납함 일부가 보인다. 낡은 휴식 공간의 재질과 낮의 약한 실내광은 참고와 유사하다. 이현우의 골반은 잘려 있어 정확한 착석 접점은 확인할 수 없다. 중복된 침대나 불가능한 반사는 보이지 않는다.",
        "entities": "중앙 인물은 참고와 유사한 젊은 한국계 남성의 얼굴, 헝클어진 검은 머리, 낡은 짙은 회색 후드를 갖췄다. 주름진 미간과 벌어진 입으로 당황한 반응이 보인다. 붉은 손잡이의 큰 펜치를 쥐었지만 턱 형태는 참고 도구와 다소 다르다. 양쪽 전경의 실제 사람 두 명은 허용되지 않은 추가 인물이다. 포스터에는 밴드 사진과 식별 가능한 영문 글자가 남아 있다. 다리 붕대와 신발 속 카드는 프레임 밖이다.",
        "hard_violations": [
         "이현우만 보여야 하는 장면에 왼쪽 전경의 얼굴·어깨와 오른쪽 전경의 머리로 두 명의 추가 인물이 등장한다.",
         "왼쪽 아래 포스터 등에 식별 가능한 영문 표기가 남아 있어 읽을 수 있는 문자 금지 조건을 위반한다."
        ],
        "physics": "펜치 손잡이는 이현우의 손가락과 손바닥에 잡혀 있고 손목은 후드 소매로 이어져 있어 지지가 성립한다. 팔을 굽혀 도구를 든 자세는 가능하다. 하체는 프레임 밖이며, 보이는 상체나 도구에서 근거 없는 부유는 없다."
       },
       {
        "label": "A",
        "direction": "이현우의 얼굴과 두 눈은 화면 오른쪽 바깥을 향한다. 은영과 태진은 보이지 않지만 그들이 있는 화면 밖 방향을 바라보는 연기로 읽힌다. 펜치는 손에서 오른쪽 위로 비스듬히 뻗으며, 상대를 직접 겨누는 동작은 아니다.",
        "built_space": "배경에는 참고의 얼룩진 밝은 벽과 여러 장의 겹친 밴드 포스터가 보인다. 큰 초상 포스터는 왼쪽, 선글라스를 쓴 인물 사진은 중앙 뒤, 단체 사진은 오른쪽에 놓여 참고 공간의 연속성이 높다. 침대·작업대·캠핑카는 이 클로즈업 밖이므로 그 생략은 결함이 아니다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "실제 등장인물은 이현우 한 명이며, 젊은 한국계 남성의 얼굴과 검은 흐트러진 머리, 때 묻고 해진 짙은 후드가 참고와 부합한다. 올라간 눈썹, 옆으로 향한 눈, 벌어진 입이 어리둥절한 반응을 전달한다. 손과 손목이 보이고 큰 금속 펜치를 잡고 있으나, 참고의 붉은 손잡이 도구와 외형은 다소 다르다. 배경의 밴드 인물들은 인쇄 사진이다. 오른쪽 포스터 하단의 큰 영문 표기는 완전히 흐려지지 않았다. 붕대와 신발은 프레임 밖이다.",
        "hard_violations": [
         "오른쪽 배경 포스터의 큰 영문 표제 일부가 읽을 수 있는 상태로 남아 있어 문자 비가독성 조건을 위반한다."
        ],
        "physics": "이현우의 손이 펜치 손잡이를 감싸고 있으며 손목과 굽힌 팔이 소매로 자연스럽게 이어진다. 비스듬히 들어 올린 도구의 무게를 손과 팔이 받치는 자세가 가능하다. 하체 지지면은 클로즈업 밖이고, 보이는 신체나 도구에 지지 없는 부유는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.333
   },
   "adjusted": {
    "A": 1.35,
    "B": 1.083
   },
   "violations": {
    "B": [
     "[gemini-pro] extra bodies (PEOPLE 지침에서 단독 등장을 명시했으나 전경에 금지된 두 인물이 노출됨)",
     "[gpt-high] 이현우만 보여야 하는 장면에 왼쪽 전경의 얼굴·어깨와 오른쪽 전경의 머리로 두 명의 추가 인물이 등장한다.",
     "[gpt-high] 왼쪽 아래 포스터 등에 식별 가능한 영문 표기가 남아 있어 읽을 수 있는 문자 금지 조건을 위반한다."
    ],
    "A": [
     "[gemini-pro] physically impossible object (펜치에 경첩이 없고 손잡이와 머리가 하나로 융합된 불가능한 형태)",
     "[gemini-pro] staging violation (지정된 수리 공간 방향이 아닌 잘못된 카메라 앵글)",
     "[gpt-high] 오른쪽 배경 포스터의 큰 영문 표제 일부가 읽을 수 있는 상태로 남아 있어 문자 비가독성 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1083,
   "A": 1350
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1083,
    "verdict_ko": "배경 구조와 이전 샷의 소품(빨간 펜치)을 완벽히 유지했으나, 지침에서 엄격히 금지한 프레임 내 타 인물 노출로 인해 치명적 감점이 발생했습니다.  ★위반: [gemini-pro] extra bodies (PEOPLE 지침에서 단독 등장을 명시했으나 전경에 금지된 두 인물이 노출됨) / [gpt-high] 이현우만 보여야 하는 장면에 왼쪽 전경의 얼굴·어깨와 오른쪽 전경의 머리로 두 명의 추가 인물이 등장한다. / [gpt-high] 왼쪽 아래 포스터 등에 식별 가능한 영문 표기가 남아 있어 읽을 수 있는 문자 금지 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "카메라가 수리 공간이 아닌 벽면을 향해 공간 지침을 어겼고, 펜치가 물리적으로 불가능한 금속 덩어리로 묘사되었습니다.  ★위반: [gemini-pro] physically impossible object (펜치에 경첩이 없고 손잡이와 머리가 하나로 융합된 불가능한 형태) / [gemini-pro] staging violation (지정된 수리 공간 방향이 아닌 잘못된 카메라 앵글) / [gpt-high] 오른쪽 배경 포스터의 큰 영문 표제 일부가 읽을 수 있는 상태로 남아 있어 문자 비가독성 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S50sh1_sel.png",
    "asset_id": "41bdd7e5-7b6a-41da-91fa-1eed086540fd",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacddc-29cc-7543-a0a8-ea54bcb3181a",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S50sh1"
  }
 },
 "S50sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:44:18.622264+00:00",
  "fingerprint": "943278dd52eabfbc429b3b0cdac876045a6adf5d86635cbb5fdcf16750170ba3",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S50sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S50sh10_sel.png",
  "source_sha256": "bbda8eb3543e28c8ee38734a6444ae468cc333a327efa63466316a70f8223295",
  "file": "S50sh10_cine.png",
  "staged_sha256": "54a8a151fc6c21ef0bd5bc7d59525e251ac6941531a578c55ff05721174a9831",
  "latency_ms": 13477
 },
 "S51sh5::signage": {
  "fp": "c5cf87b0f8b1e449",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::058e63c9811e6c5e": {
  "subjects": [],
  "subject_text": "태진과 은영의 카센터 앞 공터\n외따로 선 정비소 건물 앞의 열린 공터. 창고 출입문과 옆쪽 컨테이너형 간이식당이 함께 보이는 배치다.",
  "identity": "canonical",
  "scope_id": "L69",
  "scope_role": "location_exterior",
  "scope_sha": "3c6b24db15c8947d"
 },
 "S51sh5::bgfirst_bg": {
  "input_fingerprint": "e896e8c6beba0afc",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 이현우의 허리를 두 팔로 와락 껴안은 구도.\n\nLOCATION (lock): Just inside the entrance to a container canteen beside the auto repair garage, with daylight entering through the open door.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining-space entrance door (Open following 이현우's entrance) — The open doorway is seen obliquely behind the pair; used as Locates the interrupted arrival without boxing the embrace into a frontal tableau; Container dining-space interior (In use as a temporary dining area) — The interior recedes to the side of the entrance; used as Leaves a small area of spatial context around the intimate two-person frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime interior light keeps the two faces and enclosing arms gently readable without changing the scene's natural tonal balance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 이현우의 허리를 두 팔로 와락 껴안은 구도.\n\nLOCATION (lock): Just inside the entrance to a container canteen beside the auto repair garage, with daylight entering through the open door.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining-space entrance door (Open following 이현우's entrance) — The open doorway is seen obliquely behind the pair; used as Locates the interrupted arrival without boxing the embrace into a frontal tableau; Container dining-space interior (In use as a temporary dining area) — The interior recedes to the side of the entrance; used as Leaves a small area of spatial context around the intimate two-person frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime interior light keeps the two faces and enclosing arms gently readable without changing the scene's natural tonal balance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S51sh5__bgfirst_bg.png",
  "asset_id": "d0cff22f-3817-4d20-9056-f8530bd9cfe5",
  "input_asset_ids": [
   "1823ba43-81fe-4b4c-a268-f10ad30775fb",
   "72d02da5-1e63-4c21-8b17-0c3762d667b5"
  ]
 },
 "S51sh5": {
  "input_fingerprint": "805aa343b708d3c2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 이현우의 허리를 두 팔로 와락 껴안은 구도.\n\nLOCATION (lock): Just inside the entrance to a container canteen beside the auto repair garage, with daylight entering through the open door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining-space entrance door (Open following 이현우's entrance) — The open doorway is seen obliquely behind the pair; used as Locates the interrupted arrival without boxing the embrace into a frontal tableau; Container dining-space interior (In use as a temporary dining area) — The interior recedes to the side of the entrance; used as Leaves a small area of spatial context around the intimate two-person frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime interior light keeps the two faces and enclosing arms gently readable without changing the scene's natural tonal balance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bowls of rice soup remain in the makeshift container dining area beside the car center. Charlie has been repaired and can rotate the previously injured arm without the earlier shoulder sparks. 이현우: He has entered the dining area with his leg bandaged and visible treatment marks still present. His contact card remains concealed in his shoe. 앰버: She has risen from her meal and rushed forward with her arms extended in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 이현우의 허리를 두 팔로 와락 껴안은 구도.\n\nLOCATION (lock): Just inside the entrance to a container canteen beside the auto repair garage, with daylight entering through the open door. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining-space entrance door (Open following 이현우's entrance) — The open doorway is seen obliquely behind the pair; used as Locates the interrupted arrival without boxing the embrace into a frontal tableau; Container dining-space interior (In use as a temporary dining area) — The interior recedes to the side of the entrance; used as Leaves a small area of spatial context around the intimate two-person frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime interior light keeps the two faces and enclosing arms gently readable without changing the scene's natural tonal balance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bowls of rice soup remain in the makeshift container dining area beside the car center. Charlie has been repaired and can rotate the previously injured arm without the earlier shoulder sparks. 이현우: He has entered the dining area with his leg bandaged and visible treatment marks still present. His contact card remains concealed in his shoe. 앰버: She has risen from her meal and rushed forward with her arms extended in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 이현우의 허리를 두 팔로 와락 껴안은 구도.\n\nLOCATION (lock): Just inside the entrance to a container canteen beside the auto repair garage, with daylight entering through the open door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining-space entrance door (Open following 이현우's entrance) — The open doorway is seen obliquely behind the pair; used as Locates the interrupted arrival without boxing the embrace into a frontal tableau; Container dining-space interior (In use as a temporary dining area) — The interior recedes to the side of the entrance; used as Leaves a small area of spatial context around the intimate two-person frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime interior light keeps the two faces and enclosing arms gently readable without changing the scene's natural tonal balance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bowls of rice soup remain in the makeshift container dining area beside the car center. Charlie has been repaired and can rotate the previously injured arm without the earlier shoulder sparks. 이현우: He has entered the dining area with his leg bandaged and visible treatment marks still present. His contact card remains concealed in his shoe. 앰버: She has risen from her meal and rushed forward with her arms extended in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S51sh5__bgfirst_bg.png",
     "asset_id": "d0cff22f-3817-4d20-9056-f8530bd9cfe5",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S51sh5.png",
     "asset_id": "1823ba43-81fe-4b4c-a268-f10ad30775fb",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L69B01.png",
     "asset_id": "72d02da5-1e63-4c21-8b17-0c3762d667b5",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "앰버의 시선은 이현우의 얼굴을 향하고 있으며, 두 팔은 그의 허리를 감싸 안고 있음. 이현우의 시선은 앰버를 향해 아래로 향함.",
    "built_space": "컨테이너 식당 내부. 왼쪽에 열린 문을 통해 정비소가 보이며, 오른쪽으로는 국수 그릇이 놓인 테이블과 의자들이 배치되어 있음.",
    "entities": "앰버(얼굴과 머리 모양 일치, 재킷 부재), 이현우(레퍼런스와 일치하는 의상과 외모, 찢어진 바지 틈으로 붕대 일부 보임). 추가 인물 없음.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥에 발을 딛고 서 있으며, 앰버의 팔은 이현우의 몸통에 의해 지지됨."
   },
   {
    "label": "B",
    "direction": "앰버는 이현우의 허리를 껴안고 얼굴을 그의 가슴에 묻고 있음. 이현우는 아래를 내려다봄.",
    "built_space": "컨테이너 식당 내부. 왼쪽에 열린 문이 있고 오른쪽 공간에 다수의 식사 테이블이 놓여 있음.",
    "entities": "앰버(레퍼런스 의상 일치), 이현우(다리에 붕대 감겨 있음). 지시사항에 없는 다수의 남성 인물들이 배경에서 식사 중임.",
    "hard_violations": [
     "[gemini-pro] invented people (프롬프트에 명시되지 않은 다수의 인물들이 배경에 추가됨)",
     "[gpt-high] 샷에 허용된 앰버와 이현우 외에 식사객 일곱 명을 추가했습니다."
    ],
    "physics": "두 인물 모두 바닥에 안정적으로 서 있으며, 앰버의 포옹 자세는 자연스럽게 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 두 인물만을 정확히 배치하고, 열린 문과 식당 내부의 공간적 구도를 훌륭하게 구현하여 프롬프트의 요구사항을 잘 충족함."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트에 명시되지 않은 다수의 배경 인물들을 임의로 추가하여 엄격한 등장인물 제한 규칙을 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선은 이현우의 얼굴을 향하고 있으며, 두 팔은 그의 허리를 감싸 안고 있음. 이현우의 시선은 앰버를 향해 아래로 향함.",
        "built_space": "컨테이너 식당 내부. 왼쪽에 열린 문을 통해 정비소가 보이며, 오른쪽으로는 국수 그릇이 놓인 테이블과 의자들이 배치되어 있음.",
        "entities": "앰버(얼굴과 머리 모양 일치, 재킷 부재), 이현우(레퍼런스와 일치하는 의상과 외모, 찢어진 바지 틈으로 붕대 일부 보임). 추가 인물 없음.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 발을 딛고 서 있으며, 앰버의 팔은 이현우의 몸통에 의해 지지됨."
       },
       {
        "label": "B",
        "direction": "앰버는 이현우의 허리를 껴안고 얼굴을 그의 가슴에 묻고 있음. 이현우는 아래를 내려다봄.",
        "built_space": "컨테이너 식당 내부. 왼쪽에 열린 문이 있고 오른쪽 공간에 다수의 식사 테이블이 놓여 있음.",
        "entities": "앰버(레퍼런스 의상 일치), 이현우(다리에 붕대 감겨 있음). 지시사항에 없는 다수의 남성 인물들이 배경에서 식사 중임.",
        "hard_violations": [
         "invented people (프롬프트에 명시되지 않은 다수의 인물들이 배경에 추가됨)"
        ],
        "physics": "두 인물 모두 바닥에 안정적으로 서 있으며, 앰버의 포옹 자세는 자연스럽게 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 두 인물만을 정확히 배치하고, 열린 문과 식당 내부의 공간적 구도를 훌륭하게 구현하여 프롬프트의 요구사항을 잘 충족함."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트에 명시되지 않은 다수의 배경 인물들을 임의로 추가하여 엄격한 등장인물 제한 규칙을 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선은 이현우의 얼굴을 향하고 있으며, 두 팔은 그의 허리를 감싸 안고 있음. 이현우의 시선은 앰버를 향해 아래로 향함.",
        "built_space": "컨테이너 식당 내부. 왼쪽에 열린 문을 통해 정비소가 보이며, 오른쪽으로는 국수 그릇이 놓인 테이블과 의자들이 배치되어 있음.",
        "entities": "앰버(얼굴과 머리 모양 일치, 재킷 부재), 이현우(레퍼런스와 일치하는 의상과 외모, 찢어진 바지 틈으로 붕대 일부 보임). 추가 인물 없음.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 발을 딛고 서 있으며, 앰버의 팔은 이현우의 몸통에 의해 지지됨."
       },
       {
        "label": "B",
        "direction": "앰버는 이현우의 허리를 껴안고 얼굴을 그의 가슴에 묻고 있음. 이현우는 아래를 내려다봄.",
        "built_space": "컨테이너 식당 내부. 왼쪽에 열린 문이 있고 오른쪽 공간에 다수의 식사 테이블이 놓여 있음.",
        "entities": "앰버(레퍼런스 의상 일치), 이현우(다리에 붕대 감겨 있음). 지시사항에 없는 다수의 남성 인물들이 배경에서 식사 중임.",
        "hard_violations": [
         "invented people (프롬프트에 명시되지 않은 다수의 인물들이 배경에 추가됨)"
        ],
        "physics": "두 인물 모두 바닥에 안정적으로 서 있으며, 앰버의 포옹 자세는 자연스럽게 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "허리를 끌어안는 동작은 보이지만, 허용되지 않은 식사객들을 추가했고 미디엄 숏을 전신 구도로 넓혀 핵심 조건을 위반했습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 사람만의 허리 포옹과 비스듬한 열린 출입구, 옆으로 이어지는 식당 공간을 충실히 구현했지만, 구도가 다소 넓고 앰버의 의상이 참조와 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 고개와 시선을 앰버에게 내리고, 앰버는 눈을 감은 채 그의 몸통에 얼굴을 붙입니다. 앰버의 팔은 이현우의 허리 위쪽을 둘러싸며, 보이는 손은 반대쪽 옆구리에 닿습니다. 카메라를 바라보는 주인공은 없습니다.",
        "built_space": "왼쪽에 열린 출입구 하나가 있고, 두 사람은 그 바로 안쪽 통로에 서 있습니다. 골강판 벽과 천장, 왼쪽 작은 창 하나와 오른쪽 큰 창 하나, 뒤쪽 입식 냉방기 하나와 흰 냉동고 하나가 보입니다. 왼쪽 식탁 구역 하나와 오른쪽 식탁 구역 세 곳이 드러나며 여러 의자가 배치되어 있습니다. 낮빛과 식당 용도는 맞지만, 전신과 넓은 홀을 보여 주어 요구된 친밀한 미디엄 숏보다 훨씬 넓습니다. 출입문 유리의 희미한 반사에 명백한 광학적 모순은 보이지 않습니다.",
        "entities": "금발의 어린 여자아이와 헝클어진 검은 머리의 젊은 한국계 남성이 보이며, 나이와 체격은 대체로 지정에 부합합니다. 앰버의 녹색 재킷·회색 원피스·갈색 부츠와 이현우의 낡은 후드·청바지·운동화는 참조와 가깝지만 앰버의 마스크는 없습니다. 이현우의 종아리 붕대와 상처, 식탁 위 국그릇이 보입니다. 그러나 두 주인공 외에 배경과 오른쪽 전경에 식사객 일곱 명이 추가되어 있습니다. 찰리나 노출된 연락처 카드는 없고, 뚜렷하게 판독되는 문구는 확인되지 않습니다.",
        "hard_violations": [
         "샷에 허용된 앰버와 이현우 외에 식사객 일곱 명을 추가했습니다."
        ],
        "physics": "두 사람 모두 신발이 바닥에 닿아 서 있으며, 앰버가 앞으로 기울인 몸은 자신의 다리와 이현우 몸통에 닿은 포옹으로 지지됩니다. 팔과 손의 접촉은 가능한 자세입니다. 식사객들은 의자에 앉아 있고 그릇은 식탁 위에 놓여 있어, 지지 없이 뜬 몸이나 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "앰버는 이현우의 얼굴을 올려다보고, 이현우는 앰버를 내려다봅니다. 앰버의 보이는 팔은 이현우의 배 앞을 지나 허리를 감싸고 손은 반대쪽 옆구리에 닿습니다. 다른 팔의 대부분은 몸 뒤에 가려져 있지만 두 팔로 허리를 감싸는 관계로 읽힙니다.",
        "built_space": "왼쪽 뒤에 열린 출입구 하나가 비스듬히 보이고 두 사람은 문턱 바로 안쪽에 있습니다. 식당은 오른쪽으로 이어지며, 뒤쪽 창 하나, 식탁 두 개, 냉장고 하나, 식기 선반 하나, 천장 형광등 두 개와 여러 원형 의자가 보입니다. 노란 골강판과 문밖 정비소가 장소를 연결합니다. 참조의 노란 출입문과 달리 회색 금속 문짝이며 내부 세부 배치는 참조만으로 확정하기 어렵습니다. 인물 중심 구도는 A보다 요구에 가깝지만 무릎 부근까지 보여 주어 미디엄 숏보다는 약간 넓습니다.",
        "entities": "등장인물은 앰버와 이현우 두 명뿐입니다. 앰버는 금발의 어린 여자아이로 보이지만 옆얼굴만으로 참조의 혼혈 얼굴과 정확히 일치하는지는 판단이 제한됩니다. 참조의 녹색 재킷과 마스크가 없고 회청색 얼룩무늬 반소매 원피스를 입었습니다. 이현우의 젊은 한국계 얼굴, 헝클어진 검은 머리, 낡은 회색 후드와 청바지는 대체로 맞으며 얼굴의 상처와 무릎 부근 치료 흔적이 보입니다. 식탁 위에는 국그릇이 남아 있고, 연락처 카드나 찰리는 보이지 않습니다. 판독 가능한 글자는 확인되지 않습니다.",
        "hard_violations": [],
        "physics": "두 인물의 다리는 화면 아래로 이어지고 발은 프레임 밖입니다. 몸통이 수직으로 서 있는 자세이며 공중에 떠 있다는 징후는 없습니다. 앰버의 상체는 이현우에게 기울어 있고 팔과 손이 허리에 밀착되어 실제 포옹으로 가능한 자세입니다. 그릇과 보온 용기는 식탁 위에, 식기는 선반 위에 놓여 있어 지지가 자연스럽습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "허리를 끌어안는 동작은 보이지만, 허용되지 않은 식사객들을 추가했고 미디엄 숏을 전신 구도로 넓혀 핵심 조건을 위반했습니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 사람만의 허리 포옹과 비스듬한 열린 출입구, 옆으로 이어지는 식당 공간을 충실히 구현했지만, 구도가 다소 넓고 앰버의 의상이 참조와 다릅니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 고개와 시선을 앰버에게 내리고, 앰버는 눈을 감은 채 그의 몸통에 얼굴을 붙입니다. 앰버의 팔은 이현우의 허리 위쪽을 둘러싸며, 보이는 손은 반대쪽 옆구리에 닿습니다. 카메라를 바라보는 주인공은 없습니다.",
        "built_space": "왼쪽에 열린 출입구 하나가 있고, 두 사람은 그 바로 안쪽 통로에 서 있습니다. 골강판 벽과 천장, 왼쪽 작은 창 하나와 오른쪽 큰 창 하나, 뒤쪽 입식 냉방기 하나와 흰 냉동고 하나가 보입니다. 왼쪽 식탁 구역 하나와 오른쪽 식탁 구역 세 곳이 드러나며 여러 의자가 배치되어 있습니다. 낮빛과 식당 용도는 맞지만, 전신과 넓은 홀을 보여 주어 요구된 친밀한 미디엄 숏보다 훨씬 넓습니다. 출입문 유리의 희미한 반사에 명백한 광학적 모순은 보이지 않습니다.",
        "entities": "금발의 어린 여자아이와 헝클어진 검은 머리의 젊은 한국계 남성이 보이며, 나이와 체격은 대체로 지정에 부합합니다. 앰버의 녹색 재킷·회색 원피스·갈색 부츠와 이현우의 낡은 후드·청바지·운동화는 참조와 가깝지만 앰버의 마스크는 없습니다. 이현우의 종아리 붕대와 상처, 식탁 위 국그릇이 보입니다. 그러나 두 주인공 외에 배경과 오른쪽 전경에 식사객 일곱 명이 추가되어 있습니다. 찰리나 노출된 연락처 카드는 없고, 뚜렷하게 판독되는 문구는 확인되지 않습니다.",
        "hard_violations": [
         "샷에 허용된 앰버와 이현우 외에 식사객 일곱 명을 추가했습니다."
        ],
        "physics": "두 사람 모두 신발이 바닥에 닿아 서 있으며, 앰버가 앞으로 기울인 몸은 자신의 다리와 이현우 몸통에 닿은 포옹으로 지지됩니다. 팔과 손의 접촉은 가능한 자세입니다. 식사객들은 의자에 앉아 있고 그릇은 식탁 위에 놓여 있어, 지지 없이 뜬 몸이나 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "앰버는 이현우의 얼굴을 올려다보고, 이현우는 앰버를 내려다봅니다. 앰버의 보이는 팔은 이현우의 배 앞을 지나 허리를 감싸고 손은 반대쪽 옆구리에 닿습니다. 다른 팔의 대부분은 몸 뒤에 가려져 있지만 두 팔로 허리를 감싸는 관계로 읽힙니다.",
        "built_space": "왼쪽 뒤에 열린 출입구 하나가 비스듬히 보이고 두 사람은 문턱 바로 안쪽에 있습니다. 식당은 오른쪽으로 이어지며, 뒤쪽 창 하나, 식탁 두 개, 냉장고 하나, 식기 선반 하나, 천장 형광등 두 개와 여러 원형 의자가 보입니다. 노란 골강판과 문밖 정비소가 장소를 연결합니다. 참조의 노란 출입문과 달리 회색 금속 문짝이며 내부 세부 배치는 참조만으로 확정하기 어렵습니다. 인물 중심 구도는 A보다 요구에 가깝지만 무릎 부근까지 보여 주어 미디엄 숏보다는 약간 넓습니다.",
        "entities": "등장인물은 앰버와 이현우 두 명뿐입니다. 앰버는 금발의 어린 여자아이로 보이지만 옆얼굴만으로 참조의 혼혈 얼굴과 정확히 일치하는지는 판단이 제한됩니다. 참조의 녹색 재킷과 마스크가 없고 회청색 얼룩무늬 반소매 원피스를 입었습니다. 이현우의 젊은 한국계 얼굴, 헝클어진 검은 머리, 낡은 회색 후드와 청바지는 대체로 맞으며 얼굴의 상처와 무릎 부근 치료 흔적이 보입니다. 식탁 위에는 국그릇이 남아 있고, 연락처 카드나 찰리는 보이지 않습니다. 판독 가능한 글자는 확인되지 않습니다.",
        "hard_violations": [],
        "physics": "두 인물의 다리는 화면 아래로 이어지고 발은 프레임 밖입니다. 몸통이 수직으로 서 있는 자세이며 공중에 떠 있다는 징후는 없습니다. 앰버의 상체는 이현우에게 기울어 있고 팔과 손이 허리에 밀착되어 실제 포옹으로 가능한 자세입니다. 그릇과 보온 용기는 식탁 위에, 식기는 선반 위에 놓여 있어 지지가 자연스럽습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.536
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.286
   },
   "violations": {
    "B": [
     "[gemini-pro] invented people (프롬프트에 명시되지 않은 다수의 인물들이 배경에 추가됨)",
     "[gpt-high] 샷에 허용된 앰버와 이현우 외에 식사객 일곱 명을 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 286
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 두 인물만을 정확히 배치하고, 열린 문과 식당 내부의 공간적 구도를 훌륭하게 구현하여 프롬프트의 요구사항을 잘 충족함."
   },
   {
    "label": "B",
    "score": 286,
    "verdict_ko": "프롬프트에 명시되지 않은 다수의 배경 인물들을 임의로 추가하여 엄격한 등장인물 제한 규칙을 위반함.  ★위반: [gemini-pro] invented people (프롬프트에 명시되지 않은 다수의 인물들이 배경에 추가됨) / [gpt-high] 샷에 허용된 앰버와 이현우 외에 식사객 일곱 명을 추가했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L69B01.png",
    "asset_id": "72d02da5-1e63-4c21-8b17-0c3762d667b5",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacddd-06b7-7a1c-a3ea-8f0911834bcc",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S51sh5__bgfirst_bg.png",
   "bg_asset_id": "d0cff22f-3817-4d20-9056-f8530bd9cfe5",
   "bg_record_key": "S51sh5::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S51sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:44:32.512333+00:00",
  "fingerprint": "22dc8c56d01000a2671c17959a6e4ab54aa0bfff76c225d66449fb5472e27718",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S51sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S51sh5_sel.png",
  "source_sha256": "c3f5d4117ef95db74f6c62f205c447e79c8962c1340f1ec8739743ef7931e0d8",
  "file": "S51sh5_cine.png",
  "staged_sha256": "3d6fe14eb8df17fa550cbec97c98fb5ebbe90cdb60a19c470664b35a2b62dcd2",
  "latency_ms": 13226
 },
 "S51sh19::signage": {
  "fp": "d9c5f3d42913c814",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S51sh19": {
  "input_fingerprint": "dace0f09a9bab12f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 종이 편지를 쥔 채 두 눈을 크게 치켜뜨고 상체를 벌떡 일으킨 역동적인 찰나.\n\nLOCATION (lock): At the sleeping place in the garage's temporary lodging area the following morning, in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 찰리's paper letter (Held in 이현우's hand after being given to him) — The written face angles toward 이현우 and away from the camera, so its message is not presented for reading; used as Keeps the cause of his alarm visible beneath his face without becoming a text insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use consistent ambient illumination for the next-day interior, preserving detail through the abrupt movement without introducing a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's paper farewell note remains, saying thanks, inviting a visit to Haenam and signing off as Amber's friend Charlie; the camper has been left at the car center. Charlie has departed alone, already wearing the large straw hat, oversized boots and colorful raincoat seen on the road. 이현우: He is being awakened the following morning and has been given the paper farewell note. His treated leg remains bandaged, and the contact card remains concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 종이 편지를 쥔 채 두 눈을 크게 치켜뜨고 상체를 벌떡 일으킨 역동적인 찰나.\n\nLOCATION (lock): At the sleeping place in the garage's temporary lodging area the following morning, in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 찰리's paper letter (Held in 이현우's hand after being given to him) — The written face angles toward 이현우 and away from the camera, so its message is not presented for reading; used as Keeps the cause of his alarm visible beneath his face without becoming a text insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use consistent ambient illumination for the next-day interior, preserving detail through the abrupt movement without introducing a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's paper farewell note remains, saying thanks, inviting a visit to Haenam and signing off as Amber's friend Charlie; the camper has been left at the car center. Charlie has departed alone, already wearing the large straw hat, oversized boots and colorful raincoat seen on the road. 이현우: He is being awakened the following morning and has been given the paper farewell note. His treated leg remains bandaged, and the contact card remains concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 종이 편지를 쥔 채 두 눈을 크게 치켜뜨고 상체를 벌떡 일으킨 역동적인 찰나.\n\nLOCATION (lock): At the sleeping place in the garage's temporary lodging area the following morning, in daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 찰리's paper letter (Held in 이현우's hand after being given to him) — The written face angles toward 이현우 and away from the camera, so its message is not presented for reading; used as Keeps the cause of his alarm visible beneath his face without becoming a text insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use consistent ambient illumination for the next-day interior, preserving detail through the abrupt movement without introducing a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's paper farewell note remains, saying thanks, inviting a visit to Haenam and signing off as Amber's friend Charlie; the camper has been left at the car center. Charlie has departed alone, already wearing the large straw hat, oversized boots and colorful raincoat seen on the road. 이현우: He is being awakened the following morning and has been given the paper farewell note. His treated leg remains bandaged, and the contact card remains concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선은 렌즈 너머 정면을 향하며, 오른손에 쥔 편지의 평평한 면이 카메라 쪽을 향하고 있음.",
    "built_space": "이전 샷과 동일한 차고 내부. 벽면의 포스터, 좌측의 철제 캐비닛, 우측의 작업대와 뒤편의 캠핑카가 정확한 위치에 유지됨.",
    "entities": "이현우의 인상착의(더러운 후드티, 찢어진 청바지, 다리의 붕대)와 종이 편지가 레퍼런스와 일치함.",
    "hard_violations": [
     "[gpt-high] 왼쪽 벽 포스터의 영문 밴드명이 읽혀, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
    ],
    "physics": "침대 위에 엉덩이를 대고 편안하게 앉아 있으며, 허공에 든 편지는 오른손이 안정적으로 쥐고 있음."
   },
   {
    "label": "B",
    "direction": "이현우의 시선은 정면을 향하고, 왼손에 든 편지는 내용면이 인물을 향하고 뒷면이 카메라를 향해 있음.",
    "built_space": "이전 샷의 차고 임시 숙소와 동일. 포스터, 캐비닛, 선반, 캠핑카 등의 고정 요소들이 일관되게 배치되어 있음.",
    "entities": "이현우의 외모, 착장, 붕대 감은 다리 및 종이 편지의 재질감이 레퍼런스의 특징을 정확히 반영함.",
    "hard_violations": [
     "[gpt-high] 왼쪽 벽 포스터의 영문 밴드명이 읽혀, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
    ],
    "physics": "오른손으로 침대의 금속 프레임을 강하게 짚어 상체를 앞으로 내민 역동적인 자세를 지탱하고 있으며, 왼손이 편지를 자연스럽게 쥐고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "상체를 벌떡 일으키는 역동적인 동작이 누락되어 이전 샷과 동일한 정적인 자세에 머물렀으며, 편지가 카메라를 향해 있어 방향 지시도 어긋납니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "침대 프레임을 짚고 상체를 벌떡 일으킨 역동적인 찰나를 잘 포착했으며, 편지의 내용면이 인물을 향하도록 한 프롬프트의 지시를 정확히 따랐습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 렌즈 너머 정면을 향하며, 오른손에 쥔 편지의 평평한 면이 카메라 쪽을 향하고 있음.",
        "built_space": "이전 샷과 동일한 차고 내부. 벽면의 포스터, 좌측의 철제 캐비닛, 우측의 작업대와 뒤편의 캠핑카가 정확한 위치에 유지됨.",
        "entities": "이현우의 인상착의(더러운 후드티, 찢어진 청바지, 다리의 붕대)와 종이 편지가 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "침대 위에 엉덩이를 대고 편안하게 앉아 있으며, 허공에 든 편지는 오른손이 안정적으로 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 정면을 향하고, 왼손에 든 편지는 내용면이 인물을 향하고 뒷면이 카메라를 향해 있음.",
        "built_space": "이전 샷의 차고 임시 숙소와 동일. 포스터, 캐비닛, 선반, 캠핑카 등의 고정 요소들이 일관되게 배치되어 있음.",
        "entities": "이현우의 외모, 착장, 붕대 감은 다리 및 종이 편지의 재질감이 레퍼런스의 특징을 정확히 반영함.",
        "hard_violations": [],
        "physics": "오른손으로 침대의 금속 프레임을 강하게 짚어 상체를 앞으로 내민 역동적인 자세를 지탱하고 있으며, 왼손이 편지를 자연스럽게 쥐고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "상체를 벌떡 일으키는 역동적인 동작이 누락되어 이전 샷과 동일한 정적인 자세에 머물렀으며, 편지가 카메라를 향해 있어 방향 지시도 어긋납니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "침대 프레임을 짚고 상체를 벌떡 일으킨 역동적인 찰나를 잘 포착했으며, 편지의 내용면이 인물을 향하도록 한 프롬프트의 지시를 정확히 따랐습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 렌즈 너머 정면을 향하며, 오른손에 쥔 편지의 평평한 면이 카메라 쪽을 향하고 있음.",
        "built_space": "이전 샷과 동일한 차고 내부. 벽면의 포스터, 좌측의 철제 캐비닛, 우측의 작업대와 뒤편의 캠핑카가 정확한 위치에 유지됨.",
        "entities": "이현우의 인상착의(더러운 후드티, 찢어진 청바지, 다리의 붕대)와 종이 편지가 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "침대 위에 엉덩이를 대고 편안하게 앉아 있으며, 허공에 든 편지는 오른손이 안정적으로 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 정면을 향하고, 왼손에 든 편지는 내용면이 인물을 향하고 뒷면이 카메라를 향해 있음.",
        "built_space": "이전 샷의 차고 임시 숙소와 동일. 포스터, 캐비닛, 선반, 캠핑카 등의 고정 요소들이 일관되게 배치되어 있음.",
        "entities": "이현우의 외모, 착장, 붕대 감은 다리 및 종이 편지의 재질감이 레퍼런스의 특징을 정확히 반영함.",
        "hard_violations": [],
        "physics": "오른손으로 침대의 금속 프레임을 강하게 짚어 상체를 앞으로 내민 역동적인 자세를 지탱하고 있으며, 왼손이 편지를 자연스럽게 쥐고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "침대 틀을 짚고 상체를 앞으로 일으키는 동작이 더 분명하지만, 지정된 미디엄 샷보다 넓고 배경 포스터의 읽히는 영문이 금지 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "크게 뜬 눈과 인물·장소의 연속성은 맞지만, 상체를 벌떡 일으키기보다 앉아서 돌아보는 모습이며 넓은 구도와 읽히는 포스터 문구도 문제다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈은 편지가 아닌 카메라 가까운 화면 밖 정면을 향한다. 놀라 편지에서 시선을 든 순간으로는 가능하다. 편지는 얼굴 아래 화면 오른쪽에 있으며, 카메라에는 글 없는 면이 보이고 반대 면은 인물 쪽으로 기울어 있다.",
        "built_space": "왼쪽 벽을 따라 침대 하나, 전경 금속 수납장 하나, 침대 오른쪽 작업대 하나, 뒤쪽 다단 선반 하나, 오른쪽 캠핑카 한 대가 보인다. 작업대 위 소형 서랍함 하나와 아래 금속 서랍의 전면 세 개가 보인다. 인물은 침대 가장자리에 앉아 전경의 둥근 침대 틀을 짚는다. 기존 차고의 배치와 낡은 재질, 낮의 실내광은 대체로 유지되지만 무릎과 붕대 감은 하퇴까지 포함해 미디엄 샷보다 넓다.",
        "entities": "실제 인물은 이현우 한 명이며, 한국계의 십 대 후반 남성으로 보이는 얼굴, 헝클어진 검은 머리, 낡은 짙은 후드와 찢어진 청바지가 참조와 부합한다. 다리 붕대도 보인다. 손에 든 것은 구김과 접힌 자국이 있는 미색 종이 한 장으로, 참조보다 가로 접힘이 두드러진다. 편지 내용은 노출되지 않으며 신발 속 카드는 프레임에서 확인할 수 없다. 왼쪽 벽 포스터에는 식별 가능한 영문이 남아 있다.",
        "hard_violations": [
         "왼쪽 벽 포스터의 영문 밴드명이 읽혀, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "골반은 매트리스에 놓여 있고 한 손은 금속 침대 틀을 잡아 상체를 일으킬 지지점을 만든다. 다른 손의 손가락은 편지 아래 가장자리를 실제로 쥐고 있다. 앞으로 이동한 몸통과 굽힌 팔은 급히 일어나는 동작으로 가능하다. 붕대 감은 다리의 끝과 지지면은 화면 밖이므로 공중에 떠 있다고 단정할 근거는 없다."
       },
       {
        "label": "B",
        "direction": "얼굴과 크게 뜬 두 눈은 화면 왼쪽 바깥을 향하며, 편지나 렌즈를 직접 보지는 않는다. 편지는 얼굴 아래 가슴 앞에 있고 글 없는 뒷면이 카메라를 향한다. 글이 있을 반대 면은 인물 쪽을 향해 소품 방향 조건에 부합한다.",
        "built_space": "침대 하나가 왼쪽 포스터 벽에 붙어 있고, 전경 수납장 하나, 중앙 오른쪽 작업대 하나, 후면 선반 하나, 오른쪽 캠핑카 한 대가 참조와 같은 관계로 놓여 있다. 작업대 위에는 소형 서랍함 하나, 아래에는 금속 서랍 전면 두 개가 뚜렷하게 보인다. 인물은 침대 가장자리에 앉고 뻗은 다리 아래에는 낮은 받침이 있다. 발끝과 넓은 차고 공간까지 담아 지정된 미디엄 샷보다 확연히 넓다.",
        "entities": "이현우 한 명만 등장하며 젊은 한국계 남성의 얼굴, 검은 흐트러진 머리, 낡은 후드와 청바지가 참조에 가깝다. 붕대 감은 다리와 드러난 발가락도 이어진다. 한 손에는 낡은 미색 종이 편지가 있고 내용은 읽히지 않는다. 숨겨진 연락처 카드는 확인 대상이 되는 위치가 보이지 않는다. 배경 포스터의 인쇄된 얼굴은 추가 등장인물이 아니지만, 영문 밴드명은 읽을 수 있다.",
        "hard_violations": [
         "왼쪽 벽 포스터의 영문 밴드명이 읽혀, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "골반은 침대 매트리스에 지지되고 뻗은 다리는 발목 아래 낮은 받침에 놓여 있다. 편지는 손가락으로 가장자리를 쥐고 있어 떠 있지 않다. 몸통을 돌려 왼쪽을 보는 자세는 물리적으로 자연스럽지만, 상체를 벌떡 일으키는 추진이나 체중 이동은 약하고 이미 앉아 있는 자세에 가깝다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "침대 틀을 짚고 상체를 앞으로 일으키는 동작이 더 분명하지만, 지정된 미디엄 샷보다 넓고 배경 포스터의 읽히는 영문이 금지 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "크게 뜬 눈과 인물·장소의 연속성은 맞지만, 상체를 벌떡 일으키기보다 앉아서 돌아보는 모습이며 넓은 구도와 읽히는 포스터 문구도 문제다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "두 눈은 편지가 아닌 카메라 가까운 화면 밖 정면을 향한다. 놀라 편지에서 시선을 든 순간으로는 가능하다. 편지는 얼굴 아래 화면 오른쪽에 있으며, 카메라에는 글 없는 면이 보이고 반대 면은 인물 쪽으로 기울어 있다.",
        "built_space": "왼쪽 벽을 따라 침대 하나, 전경 금속 수납장 하나, 침대 오른쪽 작업대 하나, 뒤쪽 다단 선반 하나, 오른쪽 캠핑카 한 대가 보인다. 작업대 위 소형 서랍함 하나와 아래 금속 서랍의 전면 세 개가 보인다. 인물은 침대 가장자리에 앉아 전경의 둥근 침대 틀을 짚는다. 기존 차고의 배치와 낡은 재질, 낮의 실내광은 대체로 유지되지만 무릎과 붕대 감은 하퇴까지 포함해 미디엄 샷보다 넓다.",
        "entities": "실제 인물은 이현우 한 명이며, 한국계의 십 대 후반 남성으로 보이는 얼굴, 헝클어진 검은 머리, 낡은 짙은 후드와 찢어진 청바지가 참조와 부합한다. 다리 붕대도 보인다. 손에 든 것은 구김과 접힌 자국이 있는 미색 종이 한 장으로, 참조보다 가로 접힘이 두드러진다. 편지 내용은 노출되지 않으며 신발 속 카드는 프레임에서 확인할 수 없다. 왼쪽 벽 포스터에는 식별 가능한 영문이 남아 있다.",
        "hard_violations": [
         "왼쪽 벽 포스터의 영문 밴드명이 읽혀, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "골반은 매트리스에 놓여 있고 한 손은 금속 침대 틀을 잡아 상체를 일으킬 지지점을 만든다. 다른 손의 손가락은 편지 아래 가장자리를 실제로 쥐고 있다. 앞으로 이동한 몸통과 굽힌 팔은 급히 일어나는 동작으로 가능하다. 붕대 감은 다리의 끝과 지지면은 화면 밖이므로 공중에 떠 있다고 단정할 근거는 없다."
       },
       {
        "label": "A",
        "direction": "얼굴과 크게 뜬 두 눈은 화면 왼쪽 바깥을 향하며, 편지나 렌즈를 직접 보지는 않는다. 편지는 얼굴 아래 가슴 앞에 있고 글 없는 뒷면이 카메라를 향한다. 글이 있을 반대 면은 인물 쪽을 향해 소품 방향 조건에 부합한다.",
        "built_space": "침대 하나가 왼쪽 포스터 벽에 붙어 있고, 전경 수납장 하나, 중앙 오른쪽 작업대 하나, 후면 선반 하나, 오른쪽 캠핑카 한 대가 참조와 같은 관계로 놓여 있다. 작업대 위에는 소형 서랍함 하나, 아래에는 금속 서랍 전면 두 개가 뚜렷하게 보인다. 인물은 침대 가장자리에 앉고 뻗은 다리 아래에는 낮은 받침이 있다. 발끝과 넓은 차고 공간까지 담아 지정된 미디엄 샷보다 확연히 넓다.",
        "entities": "이현우 한 명만 등장하며 젊은 한국계 남성의 얼굴, 검은 흐트러진 머리, 낡은 후드와 청바지가 참조에 가깝다. 붕대 감은 다리와 드러난 발가락도 이어진다. 한 손에는 낡은 미색 종이 편지가 있고 내용은 읽히지 않는다. 숨겨진 연락처 카드는 확인 대상이 되는 위치가 보이지 않는다. 배경 포스터의 인쇄된 얼굴은 추가 등장인물이 아니지만, 영문 밴드명은 읽을 수 있다.",
        "hard_violations": [
         "왼쪽 벽 포스터의 영문 밴드명이 읽혀, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "골반은 침대 매트리스에 지지되고 뻗은 다리는 발목 아래 낮은 받침에 놓여 있다. 편지는 손가락으로 가장자리를 쥐고 있어 떠 있지 않다. 몸통을 돌려 왼쪽을 보는 자세는 물리적으로 자연스럽지만, 상체를 벌떡 일으키는 추진이나 체중 이동은 약하고 이미 앉아 있는 자세에 가깝다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.306,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.056,
    "B": 1.75
   },
   "violations": {
    "B": [
     "[gpt-high] 왼쪽 벽 포스터의 영문 밴드명이 읽혀, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
    ],
    "A": [
     "[gpt-high] 왼쪽 벽 포스터의 영문 밴드명이 읽혀, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 1056,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1056,
    "verdict_ko": "상체를 벌떡 일으키는 역동적인 동작이 누락되어 이전 샷과 동일한 정적인 자세에 머물렀으며, 편지가 카메라를 향해 있어 방향 지시도 어긋납니다.  ★위반: [gpt-high] 왼쪽 벽 포스터의 영문 밴드명이 읽혀, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "침대 프레임을 짚고 상체를 벌떡 일으킨 역동적인 찰나를 잘 포착했으며, 편지의 내용면이 인물을 향하도록 한 프롬프트의 지시를 정확히 따랐습니다.  ★위반: [gpt-high] 왼쪽 벽 포스터의 영문 밴드명이 읽혀, 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S50sh1_sel.png",
    "asset_id": "41bdd7e5-7b6a-41da-91fa-1eed086540fd",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 찰리의 종이 편지: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1274588>",
    "asset_id": "14316ae9-c57a-4bfa-89bd-efc9ecbc861f",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacddd-e11b-7089-8197-6ad0c1517c10",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S50sh1"
  }
 },
 "S51sh19::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:44:46.079018+00:00",
  "fingerprint": "1419b16373d96494abcd25ec29a755c057791c4578679fea664d22765a5d7ead",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S51sh19_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S51sh19_sel.png",
  "source_sha256": "615e678185a0e60926e192eefcbf41a4d81a919f65c443edccc1e5735cf463e9",
  "file": "S51sh19_cine.png",
  "staged_sha256": "110125659f309074ef52e656385f216d41663c2e5e517cf23622f6ede4b81a1c",
  "latency_ms": 10683
 },
 "S51sh23::signage": {
  "fp": "93ddf3f152c68e15",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S51sh23": {
  "input_fingerprint": "caf4c5bc3885837b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 흙먼지가 바퀴 뒤로 뿜어져 나오는 가운데 카센터 공터 밖을 향해 맹렬히 뻗어나가는 캠핑카의 후면 외관.\n\nLOCATION (lock): Across the auto repair garage's outdoor forecourt, at the driveway leading back to the road. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Forecourt departure route in the upper-right of the frame, background; Dust trail behind the camper in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Departing camper (Accelerating out of the car-center forecourt) — Rear and one side remain visible as the vehicle travels away; used as Provides the diminishing mechanical subject against the fixed surroundings; Car-center forecourt (Being crossed by the departing camper) — The departure route recedes diagonally toward the upper right; used as Supplies negative space and a stable reference for the vehicle's retreat; Dust behind the rear wheels (Thrown backward during acceleration); used as Traces the departure path while leaving the rear wheels readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime exterior illumination keeps the departing rear body, wheels, and trailing dust distinguishable without adding a new light source or color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is operable again and departs with luggage, added food and a renewed supply of Amber's medicine. Charlie remains absent from the vehicle, traveling separately in the straw hat, oversized boots and colorful raincoat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 흙먼지가 바퀴 뒤로 뿜어져 나오는 가운데 카센터 공터 밖을 향해 맹렬히 뻗어나가는 캠핑카의 후면 외관.\n\nLOCATION (lock): Across the auto repair garage's outdoor forecourt, at the driveway leading back to the road. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Forecourt departure route in the upper-right of the frame, background; Dust trail behind the camper in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Departing camper (Accelerating out of the car-center forecourt) — Rear and one side remain visible as the vehicle travels away; used as Provides the diminishing mechanical subject against the fixed surroundings; Car-center forecourt (Being crossed by the departing camper) — The departure route recedes diagonally toward the upper right; used as Supplies negative space and a stable reference for the vehicle's retreat; Dust behind the rear wheels (Thrown backward during acceleration); used as Traces the departure path while leaving the rear wheels readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime exterior illumination keeps the departing rear body, wheels, and trailing dust distinguishable without adding a new light source or color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is operable again and departs with luggage, added food and a renewed supply of Amber's medicine. Charlie remains absent from the vehicle, traveling separately in the straw hat, oversized boots and colorful raincoat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 흙먼지가 바퀴 뒤로 뿜어져 나오는 가운데 카센터 공터 밖을 향해 맹렬히 뻗어나가는 캠핑카의 후면 외관.\n\nLOCATION (lock): Across the auto repair garage's outdoor forecourt, at the driveway leading back to the road. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Forecourt departure route in the upper-right of the frame, background; Dust trail behind the camper in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Departing camper (Accelerating out of the car-center forecourt) — Rear and one side remain visible as the vehicle travels away; used as Provides the diminishing mechanical subject against the fixed surroundings; Car-center forecourt (Being crossed by the departing camper) — The departure route recedes diagonally toward the upper right; used as Supplies negative space and a stable reference for the vehicle's retreat; Dust behind the rear wheels (Thrown backward during acceleration); used as Traces the departure path while leaving the rear wheels readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime exterior illumination keeps the departing rear body, wheels, and trailing dust distinguishable without adding a new light source or color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is operable again and departs with luggage, added food and a renewed supply of Amber's medicine. Charlie remains absent from the vehicle, traveling separately in the straw hat, oversized boots and colorful raincoat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "캠핑카가 화면 우측 후방을 향해 이동 중이며, 차량 뒤쪽으로 먼지구름이 화면 하단 중앙을 향해 길게 뻗어 나옴.",
    "built_space": "화면 좌측에 검은색 외벽과 롤업 도어 4개가 있는 카센터 건물이 위치하며, 프레임 내에 구조물 중복 없이 자연스럽게 배치됨.",
    "entities": "후면과 측면이 보이는 캠핑카는 레퍼런스와 형태 및 비율이 일치하며, 지시문대로 인물은 등장하지 않음. 건물 간판 등 텍스트가 모두 지워짐.",
    "hard_violations": [
     "[gpt-high] 캠핑카 후면의 영문 상표가 선명하게 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반합니다."
    ],
    "physics": "캠핑카 바퀴가 자갈 바닥에 안정적으로 닿아 있으며, 가속으로 인해 뒤로 뿜어지는 흙먼지의 물리적 궤적이 자연스러움."
   },
   {
    "label": "B",
    "direction": "캠핑카가 화면 중앙에서 약간 우측 후방으로 주행하며, 좌측 하단 방향으로 흙먼지가 뿜어짐.",
    "built_space": "카센터 내부에서 바깥 공터를 바라보는 시점이나, 공터 건너편에 레퍼런스의 카센터 건물(식당 컨테이너 포함)이 또 하나 존재함.",
    "entities": "캠핑카의 형태는 레퍼런스와 일치하나, 건물 외벽 컨테이너에 '문왕이네 식당' 등 읽을 수 있는 한글 텍스트가 선명하게 렌더링됨.",
    "hard_violations": [
     "[gemini-pro] 텍스트 노출 금지 지시 위반 (건물 간판에 식별 가능한 한글 생성)",
     "[gemini-pro] 구조물 중복 (카센터 내부 시점임에도 외부 공터에 동일한 카센터 건물이 통째로 중복 생성됨)",
     "[gpt-high] 캠핑카 후면의 영문 상표와 부속동의 한글 간판이 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반합니다."
    ],
    "physics": "차량이 지면에 닿아 이동 중이며 흙먼지가 일어나는 모습이 확인됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요청된 프레이밍(우측 후경으로 향하는 경로, 하단 중앙의 먼지)을 정확히 구현했으며, 읽을 수 있는 텍스트를 배제하라는 지시를 충실히 따름."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "텍스트 노출 금지 지시를 어기고 간판 글자를 명확히 렌더링했으며, 카센터 내부 시점임에도 바깥에 동일한 건물을 중복 생성하는 치명적인 위반이 있음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카가 화면 우측 후방을 향해 이동 중이며, 차량 뒤쪽으로 먼지구름이 화면 하단 중앙을 향해 길게 뻗어 나옴.",
        "built_space": "화면 좌측에 검은색 외벽과 롤업 도어 4개가 있는 카센터 건물이 위치하며, 프레임 내에 구조물 중복 없이 자연스럽게 배치됨.",
        "entities": "후면과 측면이 보이는 캠핑카는 레퍼런스와 형태 및 비율이 일치하며, 지시문대로 인물은 등장하지 않음. 건물 간판 등 텍스트가 모두 지워짐.",
        "hard_violations": [],
        "physics": "캠핑카 바퀴가 자갈 바닥에 안정적으로 닿아 있으며, 가속으로 인해 뒤로 뿜어지는 흙먼지의 물리적 궤적이 자연스러움."
       },
       {
        "label": "B",
        "direction": "캠핑카가 화면 중앙에서 약간 우측 후방으로 주행하며, 좌측 하단 방향으로 흙먼지가 뿜어짐.",
        "built_space": "카센터 내부에서 바깥 공터를 바라보는 시점이나, 공터 건너편에 레퍼런스의 카센터 건물(식당 컨테이너 포함)이 또 하나 존재함.",
        "entities": "캠핑카의 형태는 레퍼런스와 일치하나, 건물 외벽 컨테이너에 '문왕이네 식당' 등 읽을 수 있는 한글 텍스트가 선명하게 렌더링됨.",
        "hard_violations": [
         "텍스트 노출 금지 지시 위반 (건물 간판에 식별 가능한 한글 생성)",
         "구조물 중복 (카센터 내부 시점임에도 외부 공터에 동일한 카센터 건물이 통째로 중복 생성됨)"
        ],
        "physics": "차량이 지면에 닿아 이동 중이며 흙먼지가 일어나는 모습이 확인됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요청된 프레이밍(우측 후경으로 향하는 경로, 하단 중앙의 먼지)을 정확히 구현했으며, 읽을 수 있는 텍스트를 배제하라는 지시를 충실히 따름."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "텍스트 노출 금지 지시를 어기고 간판 글자를 명확히 렌더링했으며, 카센터 내부 시점임에도 바깥에 동일한 건물을 중복 생성하는 치명적인 위반이 있음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카가 화면 우측 후방을 향해 이동 중이며, 차량 뒤쪽으로 먼지구름이 화면 하단 중앙을 향해 길게 뻗어 나옴.",
        "built_space": "화면 좌측에 검은색 외벽과 롤업 도어 4개가 있는 카센터 건물이 위치하며, 프레임 내에 구조물 중복 없이 자연스럽게 배치됨.",
        "entities": "후면과 측면이 보이는 캠핑카는 레퍼런스와 형태 및 비율이 일치하며, 지시문대로 인물은 등장하지 않음. 건물 간판 등 텍스트가 모두 지워짐.",
        "hard_violations": [],
        "physics": "캠핑카 바퀴가 자갈 바닥에 안정적으로 닿아 있으며, 가속으로 인해 뒤로 뿜어지는 흙먼지의 물리적 궤적이 자연스러움."
       },
       {
        "label": "B",
        "direction": "캠핑카가 화면 중앙에서 약간 우측 후방으로 주행하며, 좌측 하단 방향으로 흙먼지가 뿜어짐.",
        "built_space": "카센터 내부에서 바깥 공터를 바라보는 시점이나, 공터 건너편에 레퍼런스의 카센터 건물(식당 컨테이너 포함)이 또 하나 존재함.",
        "entities": "캠핑카의 형태는 레퍼런스와 일치하나, 건물 외벽 컨테이너에 '문왕이네 식당' 등 읽을 수 있는 한글 텍스트가 선명하게 렌더링됨.",
        "hard_violations": [
         "텍스트 노출 금지 지시 위반 (건물 간판에 식별 가능한 한글 생성)",
         "구조물 중복 (카센터 내부 시점임에도 외부 공터에 동일한 카센터 건물이 통째로 중복 생성됨)"
        ],
        "physics": "차량이 지면에 닿아 이동 중이며 흙먼지가 일어나는 모습이 확인됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "읽히는 차량 로고와 건물 간판이 금지 조건을 위반하며, 캠핑카가 중앙에 크게 놓여 우상단으로 멀어지는 배경 피사체라는 지정 구도에도 덜 맞습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "우상단으로 떠나는 후면·측면과 하단 중앙의 먼지 궤적은 더 충실하지만, 읽히는 차량 로고 때문에 최종 사용 조건은 충족하지 못합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카는 후면과 오른쪽 측면을 보이며 화면 오른쪽 안쪽으로 진행합니다. 도로 출구는 우상단에 보이지만 차량은 아직 중앙에 있고, 현재 차체 축이 출구에 직접 정렬되지는 않았습니다. 먼지는 뒷바퀴에서 카메라 쪽 하단 중앙과 왼쪽으로 퍼집니다. 사람이나 시선, 무기는 없습니다.",
        "built_space": "카메라는 큰 정비고 개구부 안쪽에 있으며 양쪽 문기둥, 왼쪽 공구류, 오른쪽 받침대가 보입니다. 공터 맞은편에는 짙은 금속 외벽의 정비소와 오른쪽 낮은 부속동이 다시 배치되어 있습니다. 맞은편 셔터는 차량에 가려 일부만 보여 기준 사진의 다섯 개를 모두 확인할 수 없습니다. 구조물 외형은 참조를 따르지만, 실내 정비고와 맞은편 정비소를 함께 둔 공간 관계는 장소 사진에서 확인되지 않습니다.",
        "entities": "흰색 캠핑카 한 대의 후면, 세로형 후미등, 측면 창문과 출입문, 지붕 장비가 보이며 참조 차량과 대체로 일치합니다. 자갈 공터, 타이어 더미, 드럼통, 수목과 주간 조명이 있습니다. 인물은 없고 차량 내부의 짐·식량·약은 보이지 않으므로 확인 대상에서 제외합니다. 차량의 영문 상표와 부속동의 한글 간판은 읽힙니다.",
        "hard_violations": [
         "캠핑카 후면의 영문 상표와 부속동의 한글 간판이 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반합니다."
        ],
        "physics": "보이는 앞뒤 타이어는 자갈 지면에 닿아 차체를 지지합니다. 먼지 일부가 뒷바퀴를 가리지만 접지 관계는 확인됩니다. 바퀴 부근에서 시작해 뒤쪽으로 부푸는 먼지는 가속으로 흙이 튀는 동작과 양립하며, 지지 없이 떠 있는 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "캠핑카의 전방은 화면 우상단의 공터 통로를 향하고 후면과 오른쪽 측면이 카메라에 보입니다. 먼지 궤적은 뒷바퀴에서 하단 중앙 전경까지 대각선으로 이어져 차량의 퇴장 방향과 일치합니다. 통로가 도로와 만나는 지점 자체는 뚜렷하지 않습니다. 사람이나 시선, 무기는 없습니다.",
        "built_space": "카메라는 실외 자갈 공터에 있습니다. 왼쪽 정비소에는 다섯 개 셔터, 왼쪽 보행자 문 하나, 오른쪽 높은 창 구획과 낮은 부속동이 보여 구조 참조의 주요 구성을 유지합니다. 오른쪽 배경에는 흰 컨테이너와 수목이 있고 그 사이 통로가 뒤로 이어집니다. 공터의 빈 공간과 우상단 퇴장 경로가 읽히며, 불가능한 반사나 중복된 고정 설비는 보이지 않습니다.",
        "entities": "흰색 캠핑카 한 대의 후면과 측면, 세로형 후미등, 측면 문과 창, 지붕 장비가 참조 차량에 부합합니다. 정비소, 자갈, 흙먼지와 주간 자연광이 보입니다. 인물은 추가되지 않았으며 내부 짐·식량·약은 외관 구도로 확인할 수 없습니다. 건물 간판은 비워져 있지만 차량 후면의 영문 상표는 선명하게 읽힙니다.",
        "hard_violations": [
         "캠핑카 후면의 영문 상표가 선명하게 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반합니다."
        ],
        "physics": "보이는 앞뒤 바퀴가 자갈 지면에 접촉해 차체를 지지합니다. 뒷바퀴 뒤의 먼지와 작은 자갈 입자는 타이어가 지면을 밀며 튀긴 것으로 해석 가능한 위치에 있습니다. 뒤로 길게 남은 먼지 띠와 접지된 차량은 가속 출발 동작에 부합하며, 근거 없이 공중에 뜬 물체는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "읽히는 차량 로고와 건물 간판이 금지 조건을 위반하며, 캠핑카가 중앙에 크게 놓여 우상단으로 멀어지는 배경 피사체라는 지정 구도에도 덜 맞습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "우상단으로 떠나는 후면·측면과 하단 중앙의 먼지 궤적은 더 충실하지만, 읽히는 차량 로고 때문에 최종 사용 조건은 충족하지 못합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카는 후면과 오른쪽 측면을 보이며 화면 오른쪽 안쪽으로 진행합니다. 도로 출구는 우상단에 보이지만 차량은 아직 중앙에 있고, 현재 차체 축이 출구에 직접 정렬되지는 않았습니다. 먼지는 뒷바퀴에서 카메라 쪽 하단 중앙과 왼쪽으로 퍼집니다. 사람이나 시선, 무기는 없습니다.",
        "built_space": "카메라는 큰 정비고 개구부 안쪽에 있으며 양쪽 문기둥, 왼쪽 공구류, 오른쪽 받침대가 보입니다. 공터 맞은편에는 짙은 금속 외벽의 정비소와 오른쪽 낮은 부속동이 다시 배치되어 있습니다. 맞은편 셔터는 차량에 가려 일부만 보여 기준 사진의 다섯 개를 모두 확인할 수 없습니다. 구조물 외형은 참조를 따르지만, 실내 정비고와 맞은편 정비소를 함께 둔 공간 관계는 장소 사진에서 확인되지 않습니다.",
        "entities": "흰색 캠핑카 한 대의 후면, 세로형 후미등, 측면 창문과 출입문, 지붕 장비가 보이며 참조 차량과 대체로 일치합니다. 자갈 공터, 타이어 더미, 드럼통, 수목과 주간 조명이 있습니다. 인물은 없고 차량 내부의 짐·식량·약은 보이지 않으므로 확인 대상에서 제외합니다. 차량의 영문 상표와 부속동의 한글 간판은 읽힙니다.",
        "hard_violations": [
         "캠핑카 후면의 영문 상표와 부속동의 한글 간판이 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반합니다."
        ],
        "physics": "보이는 앞뒤 타이어는 자갈 지면에 닿아 차체를 지지합니다. 먼지 일부가 뒷바퀴를 가리지만 접지 관계는 확인됩니다. 바퀴 부근에서 시작해 뒤쪽으로 부푸는 먼지는 가속으로 흙이 튀는 동작과 양립하며, 지지 없이 떠 있는 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "캠핑카의 전방은 화면 우상단의 공터 통로를 향하고 후면과 오른쪽 측면이 카메라에 보입니다. 먼지 궤적은 뒷바퀴에서 하단 중앙 전경까지 대각선으로 이어져 차량의 퇴장 방향과 일치합니다. 통로가 도로와 만나는 지점 자체는 뚜렷하지 않습니다. 사람이나 시선, 무기는 없습니다.",
        "built_space": "카메라는 실외 자갈 공터에 있습니다. 왼쪽 정비소에는 다섯 개 셔터, 왼쪽 보행자 문 하나, 오른쪽 높은 창 구획과 낮은 부속동이 보여 구조 참조의 주요 구성을 유지합니다. 오른쪽 배경에는 흰 컨테이너와 수목이 있고 그 사이 통로가 뒤로 이어집니다. 공터의 빈 공간과 우상단 퇴장 경로가 읽히며, 불가능한 반사나 중복된 고정 설비는 보이지 않습니다.",
        "entities": "흰색 캠핑카 한 대의 후면과 측면, 세로형 후미등, 측면 문과 창, 지붕 장비가 참조 차량에 부합합니다. 정비소, 자갈, 흙먼지와 주간 자연광이 보입니다. 인물은 추가되지 않았으며 내부 짐·식량·약은 외관 구도로 확인할 수 없습니다. 건물 간판은 비워져 있지만 차량 후면의 영문 상표는 선명하게 읽힙니다.",
        "hard_violations": [
         "캠핑카 후면의 영문 상표가 선명하게 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반합니다."
        ],
        "physics": "보이는 앞뒤 바퀴가 자갈 지면에 접촉해 차체를 지지합니다. 뒷바퀴 뒤의 먼지와 작은 자갈 입자는 타이어가 지면을 밀며 튀긴 것으로 해석 가능한 위치에 있습니다. 뒤로 길게 남은 먼지 띠와 접지된 차량은 가속 출발 동작에 부합하며, 근거 없이 공중에 뜬 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.179
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.929
   },
   "violations": {
    "B": [
     "[gemini-pro] 텍스트 노출 금지 지시 위반 (건물 간판에 식별 가능한 한글 생성)",
     "[gemini-pro] 구조물 중복 (카센터 내부 시점임에도 외부 공터에 동일한 카센터 건물이 통째로 중복 생성됨)",
     "[gpt-high] 캠핑카 후면의 영문 상표와 부속동의 한글 간판이 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반합니다."
    ],
    "A": [
     "[gpt-high] 캠핑카 후면의 영문 상표가 선명하게 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 929
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "요청된 프레이밍(우측 후경으로 향하는 경로, 하단 중앙의 먼지)을 정확히 구현했으며, 읽을 수 있는 텍스트를 배제하라는 지시를 충실히 따름.  ★위반: [gpt-high] 캠핑카 후면의 영문 상표가 선명하게 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반합니다."
   },
   {
    "label": "B",
    "score": 929,
    "verdict_ko": "텍스트 노출 금지 지시를 어기고 간판 글자를 명확히 렌더링했으며, 카센터 내부 시점임에도 바깥에 동일한 건물을 중복 생성하는 치명적인 위반이 있음.  ★위반: [gemini-pro] 텍스트 노출 금지 지시 위반 (건물 간판에 식별 가능한 한글 생성) / [gemini-pro] 구조물 중복 (카센터 내부 시점임에도 외부 공터에 동일한 카센터 건물이 통째로 중복 생성됨) / [gpt-high] 캠핑카 후면의 영문 상표와 부속동의 한글 간판이 읽혀, 읽을 수 있는 글자와 로고를 전면 금지한 조건을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L69B02.png",
    "asset_id": "ac71a4a8-e8a1-4575-83dc-c7f136c6a0b1",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_auto_repair_complex_sel.png",
    "asset_id": "e518cee8-6c42-4dbf-88e3-6e2ae23e7ad0",
    "role": "structure_seed_look"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdde-9159-7f3c-9112-c5e356956071",
  "ref_mode": "플레이트+seed만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S51sh23::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:44:57.100354+00:00",
  "fingerprint": "dff282e9fead38747c56f0a6460a410833827c434b195d41a9bae15a16d27b82",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S51sh23_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S51sh23_sel.png",
  "source_sha256": "e82d97ec9a16b35247714c9ed229a742d413cf4bccd37652652540da78a296ba",
  "file": "S51sh23_cine.png",
  "staged_sha256": "c0171030f7efe953267e7ed61f85aebc2e23af95bece74c6823f0a913e340a72",
  "latency_ms": 12048
 },
 "S52sh1::signage": {
  "fp": "d66d24474eca6ad4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S52sh1": {
  "input_fingerprint": "771fc3589b3fb593",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 한적한 지방 아스팔트 도로 위, 커다란 밀짚모자와 알록달록한 비닐 우비를 뒤집어쓴 찰리가 우뚝 선 전신.\n\nLOCATION (lock): On an empty stretch of provincial asphalt road bordered by open countryside in daylight. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Quiet provincial asphalt road (Traversed by 찰리) — Its length continues across the frame in his direction of travel; used as Makes his solitary journey readable and leaves room for the coming parallel camera movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light preserves the raincoat's stated vivid colors while keeping the surrounding tonal treatment restrained.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie walks along the provincial road wearing a large straw hat, oversized boots and a colorful raincoat. The previously damaged shoulder has been repaired.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 한적한 지방 아스팔트 도로 위, 커다란 밀짚모자와 알록달록한 비닐 우비를 뒤집어쓴 찰리가 우뚝 선 전신.\n\nLOCATION (lock): On an empty stretch of provincial asphalt road bordered by open countryside in daylight. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Quiet provincial asphalt road (Traversed by 찰리) — Its length continues across the frame in his direction of travel; used as Makes his solitary journey readable and leaves room for the coming parallel camera movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light preserves the raincoat's stated vivid colors while keeping the surrounding tonal treatment restrained.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie walks along the provincial road wearing a large straw hat, oversized boots and a colorful raincoat. The previously damaged shoulder has been repaired.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 한적한 지방 아스팔트 도로 위, 커다란 밀짚모자와 알록달록한 비닐 우비를 뒤집어쓴 찰리가 우뚝 선 전신.\n\nLOCATION (lock): On an empty stretch of provincial asphalt road bordered by open countryside in daylight. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Quiet provincial asphalt road (Traversed by 찰리) — Its length continues across the frame in his direction of travel; used as Makes his solitary journey readable and leaves room for the coming parallel camera movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light preserves the raincoat's stated vivid colors while keeping the surrounding tonal treatment restrained.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie walks along the provincial road wearing a large straw hat, oversized boots and a colorful raincoat. The previously damaged shoulder has been repaired.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라를 정면으로 바라보고 있으며, 도로는 캐릭터 뒤로 이어짐.",
    "built_space": "아스팔트 도로와 주변 들판, 우측 배경에 시골 주택들이 배치됨.",
    "entities": "찰리의 기계 머리와 밀짚모자는 일치하나, 알록달록한 우비의 질감이 비닐이 아닌 천처럼 보임.",
    "hard_violations": [
     "[gpt-high] 장소를 지정된 빈 지방도로와 개방된 전원으로만 구성하라는 제한에도 불구하고, 배경에 여러 주택과 부속 구조물로 된 주택군을 새로 추가했다."
    ],
    "physics": "두 발로 지면에 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "찰리는 화면 좌측의 도로가 뻗어 있는 방향을 응시하고 있음.",
    "built_space": "시골의 빈 아스팔트 도로가 프레임 좌측 깊숙이 뻗어 있으며 주변은 탁 트인 들판임.",
    "entities": "고릴라형 기계 머리(흰 얼굴판, 점 눈, 선 입), 밀짚모자, 알록달록한 비닐 우비, 큰 장화를 착용한 찰리가 일치하게 묘사됨.",
    "hard_violations": [],
    "physics": "두 발로 도로 위에 자연스럽게 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "광택이 나는 비닐 우비의 재질을 완벽하게 구현했으며, 도로가 뻗은 방향으로 선 자세가 샷 텍스트의 카메라 무빙 연출 의도와 정확히 일치합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "우비의 재질이 비닐보다는 천에 가깝게 표현되었으며, 캐릭터가 정면을 향해 서 있어 도로를 따라가는 공간 연출 의도가 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 화면 좌측의 도로가 뻗어 있는 방향을 응시하고 있음.",
        "built_space": "시골의 빈 아스팔트 도로가 프레임 좌측 깊숙이 뻗어 있으며 주변은 탁 트인 들판임.",
        "entities": "고릴라형 기계 머리(흰 얼굴판, 점 눈, 선 입), 밀짚모자, 알록달록한 비닐 우비, 큰 장화를 착용한 찰리가 일치하게 묘사됨.",
        "hard_violations": [],
        "physics": "두 발로 도로 위에 자연스럽게 서 있음."
       },
       {
        "label": "A",
        "direction": "카메라를 정면으로 바라보고 있으며, 도로는 캐릭터 뒤로 이어짐.",
        "built_space": "아스팔트 도로와 주변 들판, 우측 배경에 시골 주택들이 배치됨.",
        "entities": "찰리의 기계 머리와 밀짚모자는 일치하나, 알록달록한 우비의 질감이 비닐이 아닌 천처럼 보임.",
        "hard_violations": [],
        "physics": "두 발로 지면에 안정적으로 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "광택이 나는 비닐 우비의 재질을 완벽하게 구현했으며, 도로가 뻗은 방향으로 선 자세가 샷 텍스트의 카메라 무빙 연출 의도와 정확히 일치합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "우비의 재질이 비닐보다는 천에 가깝게 표현되었으며, 캐릭터가 정면을 향해 서 있어 도로를 따라가는 공간 연출 의도가 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 화면 좌측의 도로가 뻗어 있는 방향을 응시하고 있음.",
        "built_space": "시골의 빈 아스팔트 도로가 프레임 좌측 깊숙이 뻗어 있으며 주변은 탁 트인 들판임.",
        "entities": "고릴라형 기계 머리(흰 얼굴판, 점 눈, 선 입), 밀짚모자, 알록달록한 비닐 우비, 큰 장화를 착용한 찰리가 일치하게 묘사됨.",
        "hard_violations": [],
        "physics": "두 발로 도로 위에 자연스럽게 서 있음."
       },
       {
        "label": "A",
        "direction": "카메라를 정면으로 바라보고 있으며, 도로는 캐릭터 뒤로 이어짐.",
        "built_space": "아스팔트 도로와 주변 들판, 우측 배경에 시골 주택들이 배치됨.",
        "entities": "찰리의 기계 머리와 밀짚모자는 일치하나, 알록달록한 우비의 질감이 비닐이 아닌 천처럼 보임.",
        "hard_violations": [],
        "physics": "두 발로 지면에 안정적으로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "빈 지방도로에 우뚝 선 전신과 이동할 여백을 살렸으며, 큰 밀짚모자·선명한 비닐 우비·큰 장화 및 기계 얼굴이 지시와 잘 맞는다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전신 정지 자세와 찰리의 얼굴은 맞지만, 허용되지 않은 주택군을 추가했고 우비의 색도 요구된 선명함보다 탁하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 장화 앞코가 화면 왼쪽을 향하고, 시선도 왼쪽 도로 공간으로 향한다. 도로는 전경에서 왼쪽 먼 들판으로 이어져 여행할 공간이 읽힌다. 특정 응시 대상이나 겨눌 물체는 지시되지 않았으며, 무기나 지시 물체도 없다.",
        "built_space": "아스팔트 도로 한 줄기가 들판 사이로 이어진다. 희미한 중앙선과 양쪽 도로 가장자리가 보이며, 찰리는 화면 오른쪽의 아스팔트 위에 서 있다. 전경의 건물이나 별도 시설물 없이 개방된 전원과 낮은 산이 배경을 이룬다. 반사상이나 중복된 구조물은 없다.",
        "entities": "찰리 한 명만 보인다. 각진 어두운 기계 머리, 흰 고릴라형 얼굴판, 점 형태의 눈과 선 형태의 입이 모두 있다. 큰 밀짚모자, 원색 조각으로 된 광택 있는 비닐 우비, 큼직한 검은 장화가 보인다. 어깨에는 은색 보수 조각처럼 보이는 부분이 있으나 그 아래 상태는 명확하지 않다. 지정된 기계 외형이므로 인간의 나이·성별·민족성은 판별할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 장화의 밑창이 아스팔트에 닿아 몸을 지탱하고, 다리는 곧게 서 있다. 모자는 머리에 얹혀 있으며 우비는 어깨에서 자연스럽게 내려온다. 두 손은 팔 끝에 연결되어 옆으로 늘어져 있다. 도로에 드리운 그림자도 접지와 일치한다. 공중에 떠 있는 신체나 물건은 없고, 우뚝 멈춰 선 순간으로 성립한다."
       },
       {
        "label": "B",
        "direction": "얼굴과 몸통은 거의 카메라 정면을 향하고, 시선은 카메라보다 약간 왼쪽으로 향한다. 장화 앞코도 주로 전경을 향한다. 도로는 뒤에서 화면 왼쪽으로 굽어 이어지므로 여행 경로는 보이지만, 인물의 방향과 도로의 연결은 A보다 덜 분명하다. 특정 응시 대상이나 조준 대상은 없다.",
        "built_space": "아스팔트 도로 한 줄기와 중앙선, 양쪽 가장자리가 보이며 찰리는 도로 위에 서 있다. 들판 외에 화면 오른쪽 배경에 최소 세 동의 주택성 건물과 부속 구조물이 추가되어 있다. 이는 빈 지방도로와 개방된 전원만으로 구성하라는 장소 제한을 넘어선다. 건물들의 원근 크기는 자연스럽고 불가능한 반사는 없다.",
        "entities": "찰리 한 명의 전신이 보인다. 각진 기계 머리, 흰 고릴라형 얼굴판, 점 눈과 선 입, 큰 밀짚모자는 일치한다. 우비는 광택과 여러 색을 지녔지만 상당히 바래고 탁하며, 앞이 열려 내부의 누더기 작업복이 크게 드러난다. 큰 장화는 있으나 A보다 일반적인 고무장화 비율에 가깝다. 어깨에 벌어진 손상은 보이지 않는다. 기계 외형에서 인간의 나이·성별·민족성은 판별되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장소를 지정된 빈 지방도로와 개방된 전원으로만 구성하라는 제한에도 불구하고, 배경에 여러 주택과 부속 구조물로 된 주택군을 새로 추가했다."
        ],
        "physics": "두 장화가 도로에 닿아 벌어진 다리와 몸통을 지탱한다. 모자는 머리에 얹혀 있고 우비와 내부 옷은 어깨 및 몸에 걸려 있다. 양손은 팔 끝에 자연스럽게 내려와 있으며, 열린 우비 자락도 착용부에서 이어진다. 긴 지면 그림자가 발과 연결되어 있고 지지 없이 떠 있는 물체는 없다. 우뚝 선 정지 자세는 물리적으로 가능하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "빈 지방도로에 우뚝 선 전신과 이동할 여백을 살렸으며, 큰 밀짚모자·선명한 비닐 우비·큰 장화 및 기계 얼굴이 지시와 잘 맞는다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "전신 정지 자세와 찰리의 얼굴은 맞지만, 허용되지 않은 주택군을 추가했고 우비의 색도 요구된 선명함보다 탁하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 장화 앞코가 화면 왼쪽을 향하고, 시선도 왼쪽 도로 공간으로 향한다. 도로는 전경에서 왼쪽 먼 들판으로 이어져 여행할 공간이 읽힌다. 특정 응시 대상이나 겨눌 물체는 지시되지 않았으며, 무기나 지시 물체도 없다.",
        "built_space": "아스팔트 도로 한 줄기가 들판 사이로 이어진다. 희미한 중앙선과 양쪽 도로 가장자리가 보이며, 찰리는 화면 오른쪽의 아스팔트 위에 서 있다. 전경의 건물이나 별도 시설물 없이 개방된 전원과 낮은 산이 배경을 이룬다. 반사상이나 중복된 구조물은 없다.",
        "entities": "찰리 한 명만 보인다. 각진 어두운 기계 머리, 흰 고릴라형 얼굴판, 점 형태의 눈과 선 형태의 입이 모두 있다. 큰 밀짚모자, 원색 조각으로 된 광택 있는 비닐 우비, 큼직한 검은 장화가 보인다. 어깨에는 은색 보수 조각처럼 보이는 부분이 있으나 그 아래 상태는 명확하지 않다. 지정된 기계 외형이므로 인간의 나이·성별·민족성은 판별할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 장화의 밑창이 아스팔트에 닿아 몸을 지탱하고, 다리는 곧게 서 있다. 모자는 머리에 얹혀 있으며 우비는 어깨에서 자연스럽게 내려온다. 두 손은 팔 끝에 연결되어 옆으로 늘어져 있다. 도로에 드리운 그림자도 접지와 일치한다. 공중에 떠 있는 신체나 물건은 없고, 우뚝 멈춰 선 순간으로 성립한다."
       },
       {
        "label": "A",
        "direction": "얼굴과 몸통은 거의 카메라 정면을 향하고, 시선은 카메라보다 약간 왼쪽으로 향한다. 장화 앞코도 주로 전경을 향한다. 도로는 뒤에서 화면 왼쪽으로 굽어 이어지므로 여행 경로는 보이지만, 인물의 방향과 도로의 연결은 A보다 덜 분명하다. 특정 응시 대상이나 조준 대상은 없다.",
        "built_space": "아스팔트 도로 한 줄기와 중앙선, 양쪽 가장자리가 보이며 찰리는 도로 위에 서 있다. 들판 외에 화면 오른쪽 배경에 최소 세 동의 주택성 건물과 부속 구조물이 추가되어 있다. 이는 빈 지방도로와 개방된 전원만으로 구성하라는 장소 제한을 넘어선다. 건물들의 원근 크기는 자연스럽고 불가능한 반사는 없다.",
        "entities": "찰리 한 명의 전신이 보인다. 각진 기계 머리, 흰 고릴라형 얼굴판, 점 눈과 선 입, 큰 밀짚모자는 일치한다. 우비는 광택과 여러 색을 지녔지만 상당히 바래고 탁하며, 앞이 열려 내부의 누더기 작업복이 크게 드러난다. 큰 장화는 있으나 A보다 일반적인 고무장화 비율에 가깝다. 어깨에 벌어진 손상은 보이지 않는다. 기계 외형에서 인간의 나이·성별·민족성은 판별되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장소를 지정된 빈 지방도로와 개방된 전원으로만 구성하라는 제한에도 불구하고, 배경에 여러 주택과 부속 구조물로 된 주택군을 새로 추가했다."
        ],
        "physics": "두 장화가 도로에 닿아 벌어진 다리와 몸통을 지탱한다. 모자는 머리에 얹혀 있고 우비와 내부 옷은 어깨 및 몸에 걸려 있다. 양손은 팔 끝에 자연스럽게 내려와 있으며, 열린 우비 자락도 착용부에서 이어진다. 긴 지면 그림자가 발과 연결되어 있고 지지 없이 떠 있는 물체는 없다. 우뚝 선 정지 자세는 물리적으로 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.27,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.02,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gpt-high] 장소를 지정된 빈 지방도로와 개방된 전원으로만 구성하라는 제한에도 불구하고, 배경에 여러 주택과 부속 구조물로 된 주택군을 새로 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1020
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "광택이 나는 비닐 우비의 재질을 완벽하게 구현했으며, 도로가 뻗은 방향으로 선 자세가 샷 텍스트의 카메라 무빙 연출 의도와 정확히 일치합니다."
   },
   {
    "label": "A",
    "score": 1020,
    "verdict_ko": "우비의 재질이 비닐보다는 천에 가깝게 표현되었으며, 캐릭터가 정면을 향해 서 있어 도로를 따라가는 공간 연출 의도가 부족합니다.  ★위반: [gpt-high] 장소를 지정된 빈 지방도로와 개방된 전원으로만 구성하라는 제한에도 불구하고, 배경에 여러 주택과 부속 구조물로 된 주택군을 새로 추가했다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacddf-5788-7f32-ba0b-5886b61b488b",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S52sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:45:09.478486+00:00",
  "fingerprint": "257d1ba58096fcc54d298d3bb0b9195021976ada25c8dd9239202d418b22b2f0",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S52sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S52sh1_sel.png",
  "source_sha256": "dcb09bb9f8a656f96059992ecf052b4627ddc017f6b83be6096fd4f518dd64a5",
  "file": "S52sh1_cine.png",
  "staged_sha256": "eb2d8f693349349298075f15512f5834f4f774a79acb866af7158c779afb7dba",
  "latency_ms": 12888
 },
 "S53sh1::signage": {
  "fp": "5771105178ee99af",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "한반도 지도"
   }
  ],
  "dropped": []
 },
 "S53sh1": {
  "input_fingerprint": "26b4fc3df70c1b4e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두운 민병대 사무실, 커다란 한반도 지도가 쫙 펼쳐진 책상 위로 상체를 숙인 박철진의 구도.\n\nLOCATION (lock): At the map-covered desk inside a dim militia office in the metropolitan area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Map of the Korean Peninsula (Spread across the desk) — The printed geographic face lies upward and is visible obliquely beneath 박철진; used as Establishes the object of his search and the destination of the following camera descent; Desk (Supporting the spread map) — The near corner and adjacent edges are seen from above; used as Defines the camera's outside position and preserves the side of the action axis.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office dim as described, with sufficient tonal separation to read his bent posture and the map without specifying a visible fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A map of the Korean Peninsula lies spread across the militia-office desk. 박철진: He has darts in use while considering directions on the map.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두운 민병대 사무실, 커다란 한반도 지도가 쫙 펼쳐진 책상 위로 상체를 숙인 박철진의 구도.\n\nLOCATION (lock): At the map-covered desk inside a dim militia office in the metropolitan area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Map of the Korean Peninsula (Spread across the desk) — The printed geographic face lies upward and is visible obliquely beneath 박철진; used as Establishes the object of his search and the destination of the following camera descent; Desk (Supporting the spread map) — The near corner and adjacent edges are seen from above; used as Defines the camera's outside position and preserves the side of the action axis.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office dim as described, with sufficient tonal separation to read his bent posture and the map without specifying a visible fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A map of the Korean Peninsula lies spread across the militia-office desk. 박철진: He has darts in use while considering directions on the map.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두운 민병대 사무실, 커다란 한반도 지도가 쫙 펼쳐진 책상 위로 상체를 숙인 박철진의 구도.\n\nLOCATION (lock): At the map-covered desk inside a dim militia office in the metropolitan area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Map of the Korean Peninsula (Spread across the desk) — The printed geographic face lies upward and is visible obliquely beneath 박철진; used as Establishes the object of his search and the destination of the following camera descent; Desk (Supporting the spread map) — The near corner and adjacent edges are seen from above; used as Defines the camera's outside position and preserves the side of the action axis.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office dim as described, with sufficient tonal separation to read his bent posture and the map without specifying a visible fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A map of the Korean Peninsula lies spread across the militia-office desk. 박철진: He has darts in use while considering directions on the map.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 책상 위에 펼쳐진 지도를 향하고 있음.",
    "built_space": "어두운 사무실 내부. 책상, 전화기, 나무 판벽 등 기준 이미지의 공간 요소들이 적절한 위치에 배치되어 있음.",
    "entities": "인물의 얼굴과 머리는 기준과 일치하나, MP 완장이 왼쪽이 아닌 오른팔에 있음. 지도는 한반도의 형태가 아니며, 다트는 형태가 불분명한 파편처럼 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 지지대 없이 공중에 떠 있는 물체 (지도 오른쪽 위의 얇은 조각)",
     "[gpt-high] 지도에 영문 지명이 판독 가능하게 남아 있고 완장에도 읽을 수 있는 글자가 보여, 판독 가능한 문자 금지 조건을 위반한다."
    ],
    "physics": "양손을 책상 모서리에 짚고 상체를 지탱하고 있으나, 지도 위 오른쪽 끝에 있는 조각은 아무런 지지나 접촉 없이 허공에 떠 있음."
   },
   {
    "label": "B",
    "direction": "시선은 손에 든 다트와 지도의 특정 지점을 향하고 있음.",
    "built_space": "어두운 사무실 내부. 창문, 책상, 나무 판벽 등 기준 이미지의 고정 요소와 분위기가 잘 유지됨.",
    "entities": "인물의 얼굴, 머리, 군복이 기준과 일치하며 완장도 올바르게 왼팔에 위치함. 지도는 한반도의 형태를 갖추었고 붉은색 다트 형태가 명확함.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (다트 기둥이 오른손 엄지손가락 살을 관통함)",
     "[gpt-high] 완장에 큰 영문 ‘MP’가 선명하게 읽혀, 화면 어디에도 판독 가능한 문자를 두지 말라는 조건을 위반한다."
    ],
    "physics": "상체를 숙여 자세를 지탱하고 있으나, 오른손에 쥔 다트가 엄지손가락을 물리적으로 관통하고 있어 쥐는 동작이 불가능한 형태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지도의 형태와 완장의 위치 등 구성 요소는 기준에 부합하나, 오른손의 다트가 엄지손가락을 관통하는 치명적인 물리적 오류가 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지도가 한반도의 형태가 아니고 완장이 반대 팔에 잘못 위치했으며, 지도 위에 지지대 없이 떠 있는 물체가 있어 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 책상 위에 펼쳐진 지도를 향하고 있음.",
        "built_space": "어두운 사무실 내부. 책상, 전화기, 나무 판벽 등 기준 이미지의 공간 요소들이 적절한 위치에 배치되어 있음.",
        "entities": "인물의 얼굴과 머리는 기준과 일치하나, MP 완장이 왼쪽이 아닌 오른팔에 있음. 지도는 한반도의 형태가 아니며, 다트는 형태가 불분명한 파편처럼 묘사됨.",
        "hard_violations": [
         "지지대 없이 공중에 떠 있는 물체 (지도 오른쪽 위의 얇은 조각)"
        ],
        "physics": "양손을 책상 모서리에 짚고 상체를 지탱하고 있으나, 지도 위 오른쪽 끝에 있는 조각은 아무런 지지나 접촉 없이 허공에 떠 있음."
       },
       {
        "label": "B",
        "direction": "시선은 손에 든 다트와 지도의 특정 지점을 향하고 있음.",
        "built_space": "어두운 사무실 내부. 창문, 책상, 나무 판벽 등 기준 이미지의 고정 요소와 분위기가 잘 유지됨.",
        "entities": "인물의 얼굴, 머리, 군복이 기준과 일치하며 완장도 올바르게 왼팔에 위치함. 지도는 한반도의 형태를 갖추었고 붉은색 다트 형태가 명확함.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (다트 기둥이 오른손 엄지손가락 살을 관통함)"
        ],
        "physics": "상체를 숙여 자세를 지탱하고 있으나, 오른손에 쥔 다트가 엄지손가락을 물리적으로 관통하고 있어 쥐는 동작이 불가능한 형태임."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지도의 형태와 완장의 위치 등 구성 요소는 기준에 부합하나, 오른손의 다트가 엄지손가락을 관통하는 치명적인 물리적 오류가 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지도가 한반도의 형태가 아니고 완장이 반대 팔에 잘못 위치했으며, 지도 위에 지지대 없이 떠 있는 물체가 있어 감점되었습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 책상 위에 펼쳐진 지도를 향하고 있음.",
        "built_space": "어두운 사무실 내부. 책상, 전화기, 나무 판벽 등 기준 이미지의 공간 요소들이 적절한 위치에 배치되어 있음.",
        "entities": "인물의 얼굴과 머리는 기준과 일치하나, MP 완장이 왼쪽이 아닌 오른팔에 있음. 지도는 한반도의 형태가 아니며, 다트는 형태가 불분명한 파편처럼 묘사됨.",
        "hard_violations": [
         "지지대 없이 공중에 떠 있는 물체 (지도 오른쪽 위의 얇은 조각)"
        ],
        "physics": "양손을 책상 모서리에 짚고 상체를 지탱하고 있으나, 지도 위 오른쪽 끝에 있는 조각은 아무런 지지나 접촉 없이 허공에 떠 있음."
       },
       {
        "label": "B",
        "direction": "시선은 손에 든 다트와 지도의 특정 지점을 향하고 있음.",
        "built_space": "어두운 사무실 내부. 창문, 책상, 나무 판벽 등 기준 이미지의 고정 요소와 분위기가 잘 유지됨.",
        "entities": "인물의 얼굴, 머리, 군복이 기준과 일치하며 완장도 올바르게 왼팔에 위치함. 지도는 한반도의 형태를 갖추었고 붉은색 다트 형태가 명확함.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (다트 기둥이 오른손 엄지손가락 살을 관통함)"
        ],
        "physics": "상체를 숙여 자세를 지탱하고 있으나, 오른손에 쥔 다트가 엄지손가락을 물리적으로 관통하고 있어 쥐는 동작이 불가능한 형태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "미디엄 숏에서 숙인 상체, 위로 펼쳐진 지도, 책상의 가까운 모서리와 양쪽 가장자리를 잘 담았지만, 완장에 선명한 ‘MP’가 보여 문자 금지 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지도에 체중을 실어 숙인 자세와 사무실 연속성은 좋지만, 가까운 책상 모서리가 잘리고 지도와 완장의 글자가 판독되어 필수 조건을 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 얼굴과 눈은 아래쪽 지도 및 손으로 잡은 다트 부근을 향한다. 잡은 빨간 다트의 촉은 지도 위 지점을 향해 내려가 있어 지도를 검토하는 행동과 맞는다. 뒤쪽에 세워진 다트들도 촉이 종이 쪽을 향한다.",
        "built_space": "전경 지도 책상 하나, 왼쪽 뒤 책상 하나, 오른쪽 뒤와 화면 끝의 책상 두 개가 보인다. 녹색 의자는 세 개가 식별되며, 왼쪽과 중앙의 창, 낡은 회벽, 오른쪽 목재 벽이 이전 장소의 재료를 이어간다. 인물은 지도 책상 건너편에 서 있고 카메라는 바깥에서 가까운 모서리와 그 양쪽 가장자리를 내려다본다. 창밖은 어두워 낮이라는 단서는 약하다.",
        "entities": "중년 동아시아계 남성 한 명만 있으며, 짧은 회색 머리와 수염, 올리브색 군복은 이전 숏의 박철진과 대체로 일치한다. 모자를 쓰지 않은 것도 이전 숏과 맞는다. 다만 검은 완장의 표기가 큰 영문 ‘MP’로 바뀌었다. 책상에는 한반도 해안선과 도로가 인쇄된 큰 종이 지도가 앞면을 위로 하고 펼쳐져 있으며, 접힘과 옅은 색조는 소품 참조와 유사하다. 손에 든 빨간 다트 하나와 뒤쪽의 추가 다트들이 보인다.",
        "hard_violations": [
         "완장에 큰 영문 ‘MP’가 선명하게 읽혀, 화면 어디에도 판독 가능한 문자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "상체는 책상 쪽으로 기울어 있지만 하체가 아래로 이어져 서 있는 자세로 읽힌다. 발은 프레임 밖이므로 접지는 확인할 수 없다. 오른손은 다트를 실제로 잡고 있고 촉은 지도에 닿는다. 뒤쪽 다트는 촉이 지도 또는 책상에 꽂힌 상태로 보이며, 누운 다트와 서류는 상판 위에 놓여 있다. 지도도 책상이 받치고 있어 명백한 무지지 부유는 없다."
       },
       {
        "label": "B",
        "direction": "박철진은 고개를 숙여 책상 중앙과 오른쪽의 지도 영역을 내려다본다. 지도에 꽂힌 여러 다트의 촉은 각각 지도 표면을 향한다. 양손은 다트를 조준하기보다 책상 양쪽을 짚고 있어 이미 표시한 지점을 검토하는 순간으로 읽힌다.",
        "built_space": "전경 지도 책상과 좌우 뒤쪽 책상 네 개가 보이며, 녹색 의자 다섯 개가 부분적으로 드러난다. 뒤 창, 열린 회색 문, 오른쪽 목재 벽, 전화기가 놓인 책상 배치는 이전 사무실과 잘 이어진다. 인물은 지도 책상 왼쪽에 서서 몸을 뻗는다. 카메라가 상판을 위에서 보지만 가까운 책상 모서리는 화면 아래로 잘려, 요구된 모서리와 인접한 두 가장자리의 연결이 A보다 불명확하다. 낮을 확인할 외부광 단서는 약하다.",
        "entities": "짧은 회색 머리와 수염을 가진 중년 동아시아계 남성 한 명이 올리브색 군복과 검은 완장을 착용해 이전 숏의 박철진을 잘 이어간다. 앞면이 위로 향한 한반도 종이 지도에는 참조와 유사한 도로망, 색조, 접힘이 있다. 지도 위 다트는 여섯 개가 보인다. 다만 지도 왼쪽의 영문 지명과 완장의 일부 글자가 읽혀 무문자 조건과 충돌한다.",
        "hard_violations": [
         "지도에 영문 지명이 판독 가능하게 남아 있고 완장에도 읽을 수 있는 글자가 보여, 판독 가능한 문자 금지 조건을 위반한다."
        ],
        "physics": "양손바닥이 책상 가장자리와 지도 위에 닿고 두 팔이 숙인 상체를 지탱한다. 하체는 화면 왼쪽 아래로 이어지며 발은 프레임 밖이다. 다트들은 수직 또는 비스듬하게 촉이 종이에 꽂힌 것으로 보이고, 앞쪽의 기울어진 다트도 촉이 지도 표면까지 이어진다. 종이 지도와 전화기, 서류는 각각 책상에 받쳐져 있으며 명백히 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "미디엄 숏에서 숙인 상체, 위로 펼쳐진 지도, 책상의 가까운 모서리와 양쪽 가장자리를 잘 담았지만, 완장에 선명한 ‘MP’가 보여 문자 금지 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지도에 체중을 실어 숙인 자세와 사무실 연속성은 좋지만, 가까운 책상 모서리가 잘리고 지도와 완장의 글자가 판독되어 필수 조건을 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박철진의 얼굴과 눈은 아래쪽 지도 및 손으로 잡은 다트 부근을 향한다. 잡은 빨간 다트의 촉은 지도 위 지점을 향해 내려가 있어 지도를 검토하는 행동과 맞는다. 뒤쪽에 세워진 다트들도 촉이 종이 쪽을 향한다.",
        "built_space": "전경 지도 책상 하나, 왼쪽 뒤 책상 하나, 오른쪽 뒤와 화면 끝의 책상 두 개가 보인다. 녹색 의자는 세 개가 식별되며, 왼쪽과 중앙의 창, 낡은 회벽, 오른쪽 목재 벽이 이전 장소의 재료를 이어간다. 인물은 지도 책상 건너편에 서 있고 카메라는 바깥에서 가까운 모서리와 그 양쪽 가장자리를 내려다본다. 창밖은 어두워 낮이라는 단서는 약하다.",
        "entities": "중년 동아시아계 남성 한 명만 있으며, 짧은 회색 머리와 수염, 올리브색 군복은 이전 숏의 박철진과 대체로 일치한다. 모자를 쓰지 않은 것도 이전 숏과 맞는다. 다만 검은 완장의 표기가 큰 영문 ‘MP’로 바뀌었다. 책상에는 한반도 해안선과 도로가 인쇄된 큰 종이 지도가 앞면을 위로 하고 펼쳐져 있으며, 접힘과 옅은 색조는 소품 참조와 유사하다. 손에 든 빨간 다트 하나와 뒤쪽의 추가 다트들이 보인다.",
        "hard_violations": [
         "완장에 큰 영문 ‘MP’가 선명하게 읽혀, 화면 어디에도 판독 가능한 문자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "상체는 책상 쪽으로 기울어 있지만 하체가 아래로 이어져 서 있는 자세로 읽힌다. 발은 프레임 밖이므로 접지는 확인할 수 없다. 오른손은 다트를 실제로 잡고 있고 촉은 지도에 닿는다. 뒤쪽 다트는 촉이 지도 또는 책상에 꽂힌 상태로 보이며, 누운 다트와 서류는 상판 위에 놓여 있다. 지도도 책상이 받치고 있어 명백한 무지지 부유는 없다."
       },
       {
        "label": "A",
        "direction": "박철진은 고개를 숙여 책상 중앙과 오른쪽의 지도 영역을 내려다본다. 지도에 꽂힌 여러 다트의 촉은 각각 지도 표면을 향한다. 양손은 다트를 조준하기보다 책상 양쪽을 짚고 있어 이미 표시한 지점을 검토하는 순간으로 읽힌다.",
        "built_space": "전경 지도 책상과 좌우 뒤쪽 책상 네 개가 보이며, 녹색 의자 다섯 개가 부분적으로 드러난다. 뒤 창, 열린 회색 문, 오른쪽 목재 벽, 전화기가 놓인 책상 배치는 이전 사무실과 잘 이어진다. 인물은 지도 책상 왼쪽에 서서 몸을 뻗는다. 카메라가 상판을 위에서 보지만 가까운 책상 모서리는 화면 아래로 잘려, 요구된 모서리와 인접한 두 가장자리의 연결이 A보다 불명확하다. 낮을 확인할 외부광 단서는 약하다.",
        "entities": "짧은 회색 머리와 수염을 가진 중년 동아시아계 남성 한 명이 올리브색 군복과 검은 완장을 착용해 이전 숏의 박철진을 잘 이어간다. 앞면이 위로 향한 한반도 종이 지도에는 참조와 유사한 도로망, 색조, 접힘이 있다. 지도 위 다트는 여섯 개가 보인다. 다만 지도 왼쪽의 영문 지명과 완장의 일부 글자가 읽혀 무문자 조건과 충돌한다.",
        "hard_violations": [
         "지도에 영문 지명이 판독 가능하게 남아 있고 완장에도 읽을 수 있는 글자가 보여, 판독 가능한 문자 금지 조건을 위반한다."
        ],
        "physics": "양손바닥이 책상 가장자리와 지도 위에 닿고 두 팔이 숙인 상체를 지탱한다. 하체는 화면 왼쪽 아래로 이어지며 발은 프레임 밖이다. 다트들은 수직 또는 비스듬하게 촉이 종이에 꽂힌 것으로 보이고, 앞쪽의 기울어진 다트도 촉이 지도 표면까지 이어진다. 종이 지도와 전화기, 서류는 각각 책상에 받쳐져 있으며 명백히 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.55,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.3,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 지지대 없이 공중에 떠 있는 물체 (지도 오른쪽 위의 얇은 조각)",
     "[gpt-high] 지도에 영문 지명이 판독 가능하게 남아 있고 완장에도 읽을 수 있는 글자가 보여, 판독 가능한 문자 금지 조건을 위반한다."
    ],
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (다트 기둥이 오른손 엄지손가락 살을 관통함)",
     "[gpt-high] 완장에 큰 영문 ‘MP’가 선명하게 읽혀, 화면 어디에도 판독 가능한 문자를 두지 말라는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 1300
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지도의 형태와 완장의 위치 등 구성 요소는 기준에 부합하나, 오른손의 다트가 엄지손가락을 관통하는 치명적인 물리적 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학적 구조 (다트 기둥이 오른손 엄지손가락 살을 관통함) / [gpt-high] 완장에 큰 영문 ‘MP’가 선명하게 읽혀, 화면 어디에도 판독 가능한 문자를 두지 말라는 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 1300,
    "verdict_ko": "지도가 한반도의 형태가 아니고 완장이 반대 팔에 잘못 위치했으며, 지도 위에 지지대 없이 떠 있는 물체가 있어 감점되었습니다.  ★위반: [gemini-pro] 지지대 없이 공중에 떠 있는 물체 (지도 오른쪽 위의 얇은 조각) / [gpt-high] 지도에 영문 지명이 판독 가능하게 남아 있고 완장에도 읽을 수 있는 글자가 보여, 판독 가능한 문자 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S43sh10_sel.png",
    "asset_id": "c14f4d84-0f17-41a4-bb7f-dd76fd3475ae",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 종이 지도: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1338462>",
    "asset_id": "8c50d23a-a6d7-4581-99af-57beaa5d211d",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde0-2b63-7747-aa7e-c2ab871876ac",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S43sh10"
  }
 },
 "S53sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:45:22.725719+00:00",
  "fingerprint": "053131777d1646ceed22d7551656e22566e6e0eb301908e5d1501cb244db7786",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S53sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S53sh1_sel.png",
  "source_sha256": "23f0f9ac1d439e639eb9f16cfc658cfc55d293eeb34ee791fc74e9f86b5188a1",
  "file": "S53sh1_cine.png",
  "staged_sha256": "b2a9502d716df2d4af5613d32e9dedb7d292d109573459901d220abd609d3e02",
  "latency_ms": 13256
 },
 "S53sh5::signage": {
  "fp": "c46d24a2826c4a93",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S53sh5": {
  "input_fingerprint": "328fab06f1276686",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진이 다트 끝을 탁자 위 지도를 향해 세차게 내리찍는 타격의 순간.\n\nLOCATION (lock): Over the map spread across the desk inside the dim militia command office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Dart (Tip striking the map) — Held point-down beneath 박철진's gripping hand; used as Marks the precise contact point without exaggerated foreground scale; Map of the Korean Peninsula (Spread on the desk beneath the striking dart) — A portion of its printed geographic face is visible at an oblique downward angle; used as Provides the meaningful target and immediate context for the impact; Desk edge (Visible beside the map) — The adjacent edge runs across the lower boundary of the composition; used as Maintains spatial continuity with the establishing desk-corner view.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established dim office illumination, using consistent contrast to separate the gripping hand and dart tip from the map.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the desk, spread-out peninsula map, and subdued office finishes and lighting. EXCLUDE any extra darts or new punctures not caused by the current action.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Korean Peninsula map remains spread across the desk; no puncture or embedded dart is established by the screenplay. 박철진: He remains engaged with the darts during the office discussion.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진이 다트 끝을 탁자 위 지도를 향해 세차게 내리찍는 타격의 순간.\n\nLOCATION (lock): Over the map spread across the desk inside the dim militia command office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Dart (Tip striking the map) — Held point-down beneath 박철진's gripping hand; used as Marks the precise contact point without exaggerated foreground scale; Map of the Korean Peninsula (Spread on the desk beneath the striking dart) — A portion of its printed geographic face is visible at an oblique downward angle; used as Provides the meaningful target and immediate context for the impact; Desk edge (Visible beside the map) — The adjacent edge runs across the lower boundary of the composition; used as Maintains spatial continuity with the establishing desk-corner view.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established dim office illumination, using consistent contrast to separate the gripping hand and dart tip from the map.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the desk, spread-out peninsula map, and subdued office finishes and lighting. EXCLUDE any extra darts or new punctures not caused by the current action.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Korean Peninsula map remains spread across the desk; no puncture or embedded dart is established by the screenplay. 박철진: He remains engaged with the darts during the office discussion.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진이 다트 끝을 탁자 위 지도를 향해 세차게 내리찍는 타격의 순간.\n\nLOCATION (lock): Over the map spread across the desk inside the dim militia command office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Dart (Tip striking the map) — Held point-down beneath 박철진's gripping hand; used as Marks the precise contact point without exaggerated foreground scale; Map of the Korean Peninsula (Spread on the desk beneath the striking dart) — A portion of its printed geographic face is visible at an oblique downward angle; used as Provides the meaningful target and immediate context for the impact; Desk edge (Visible beside the map) — The adjacent edge runs across the lower boundary of the composition; used as Maintains spatial continuity with the establishing desk-corner view.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established dim office illumination, using consistent contrast to separate the gripping hand and dart tip from the map.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the desk, spread-out peninsula map, and subdued office finishes and lighting. EXCLUDE any extra darts or new punctures not caused by the current action.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Korean Peninsula map remains spread across the desk; no puncture or embedded dart is established by the screenplay. 박철진: He remains engaged with the darts during the office discussion.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "다트 끝이 탁자 위 지도를 향해 똑바로 겨냥되어 있음.",
    "built_space": "책상 가장자리가 지시대로 화면 하단 경계를 가로지르며, 지도가 그 위에 넓게 펼쳐져 있음.",
    "entities": "박철진의 손과 소매, 빨간색 다트, 한반도 지도가 식별되나, 제외하라고 명시된 여분의 다트가 배경에 등장함.",
    "hard_violations": [
     "[gemini-pro] 손 안에서 다트의 상단 날개와 하단 금속 팁의 각도가 전혀 맞지 않아 축이 꺾인 것처럼 보이는 물리적 오류.",
     "[gemini-pro] 프롬프트에서 명시적으로 제외를 지시한 여분의 다트(배경의 붉은색과 푸른색 물체)가 화면에 포함됨.",
     "[gpt-high] 현재 손에 쥔 다트 외에 배경의 붉은색·파란색 추가 다트들을 남겨, 추가 다트를 제외하라는 명시적 조건을 위반한다.",
     "[gpt-high] 지도에 부산의 영문 지명 등 읽을 수 있는 인쇄 글자가 남아 있어 무가독 문자 조건을 위반한다."
    ],
    "physics": "주먹을 꽉 쥐고 다트를 강하게 내리찍는 힘이 묘사되어 손이 행동을 잘 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "다트 끝이 지도의 서울 부근을 향해 똑바로 닿아 있음.",
    "built_space": "책상 가장자리가 프롬프트가 요구한 하단이 아닌 화면 왼쪽 대각선에 잘못 배치됨.",
    "entities": "박철진의 손과 소매, 빨간색 다트, 한반도 지도가 등장하며 배경에 여분의 다트로 추정되는 붉은 형태가 희미하게 있음.",
    "hard_violations": [
     "[gpt-high] 지도에 서울과 평양의 영문 지명이 판독 가능하게 노출되어, 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반한다."
    ],
    "physics": "손가락 끝으로 다트를 가볍게 쥐고 있어 지시된 '세차게 내리찍는' 동작의 힘이나 속도감을 만들어낼 수 없는 자세임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지시된 화면 하단의 책상 가장자리 구도를 정확히 지켰고 '세차게 내리찍는' 주먹의 타격감을 잘 살렸으나, 다트 축이 어긋난 물리적 오류와 금지된 여분 다트가 배경에 생성된 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "책상 가장자리가 화면 하단이 아닌 왼쪽에 위치해 구도 지시를 어겼으며, 다트를 꼬집듯 가볍게 쥐고 있어 프롬프트가 요구한 강렬한 타격의 행동을 전혀 표현하지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "다트 끝이 탁자 위 지도를 향해 똑바로 겨냥되어 있음.",
        "built_space": "책상 가장자리가 지시대로 화면 하단 경계를 가로지르며, 지도가 그 위에 넓게 펼쳐져 있음.",
        "entities": "박철진의 손과 소매, 빨간색 다트, 한반도 지도가 식별되나, 제외하라고 명시된 여분의 다트가 배경에 등장함.",
        "hard_violations": [
         "손 안에서 다트의 상단 날개와 하단 금속 팁의 각도가 전혀 맞지 않아 축이 꺾인 것처럼 보이는 물리적 오류.",
         "프롬프트에서 명시적으로 제외를 지시한 여분의 다트(배경의 붉은색과 푸른색 물체)가 화면에 포함됨."
        ],
        "physics": "주먹을 꽉 쥐고 다트를 강하게 내리찍는 힘이 묘사되어 손이 행동을 잘 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "다트 끝이 지도의 서울 부근을 향해 똑바로 닿아 있음.",
        "built_space": "책상 가장자리가 프롬프트가 요구한 하단이 아닌 화면 왼쪽 대각선에 잘못 배치됨.",
        "entities": "박철진의 손과 소매, 빨간색 다트, 한반도 지도가 등장하며 배경에 여분의 다트로 추정되는 붉은 형태가 희미하게 있음.",
        "hard_violations": [],
        "physics": "손가락 끝으로 다트를 가볍게 쥐고 있어 지시된 '세차게 내리찍는' 동작의 힘이나 속도감을 만들어낼 수 없는 자세임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지시된 화면 하단의 책상 가장자리 구도를 정확히 지켰고 '세차게 내리찍는' 주먹의 타격감을 잘 살렸으나, 다트 축이 어긋난 물리적 오류와 금지된 여분 다트가 배경에 생성된 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "책상 가장자리가 화면 하단이 아닌 왼쪽에 위치해 구도 지시를 어겼으며, 다트를 꼬집듯 가볍게 쥐고 있어 프롬프트가 요구한 강렬한 타격의 행동을 전혀 표현하지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "다트 끝이 탁자 위 지도를 향해 똑바로 겨냥되어 있음.",
        "built_space": "책상 가장자리가 지시대로 화면 하단 경계를 가로지르며, 지도가 그 위에 넓게 펼쳐져 있음.",
        "entities": "박철진의 손과 소매, 빨간색 다트, 한반도 지도가 식별되나, 제외하라고 명시된 여분의 다트가 배경에 등장함.",
        "hard_violations": [
         "손 안에서 다트의 상단 날개와 하단 금속 팁의 각도가 전혀 맞지 않아 축이 꺾인 것처럼 보이는 물리적 오류.",
         "프롬프트에서 명시적으로 제외를 지시한 여분의 다트(배경의 붉은색과 푸른색 물체)가 화면에 포함됨."
        ],
        "physics": "주먹을 꽉 쥐고 다트를 강하게 내리찍는 힘이 묘사되어 손이 행동을 잘 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "다트 끝이 지도의 서울 부근을 향해 똑바로 닿아 있음.",
        "built_space": "책상 가장자리가 프롬프트가 요구한 하단이 아닌 화면 왼쪽 대각선에 잘못 배치됨.",
        "entities": "박철진의 손과 소매, 빨간색 다트, 한반도 지도가 등장하며 배경에 여분의 다트로 추정되는 붉은 형태가 희미하게 있음.",
        "hard_violations": [],
        "physics": "손가락 끝으로 다트를 가볍게 쥐고 있어 지시된 '세차게 내리찍는' 동작의 힘이나 속도감을 만들어낼 수 없는 자세임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "다트 끝과 지도 접촉점의 클로즈업은 맞지만, 읽히는 지명이 금지 조건을 위반하고 손끝으로 누르는 듯한 파지가 세찬 타격을 약화한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "주먹으로 내리찍는 동작과 하단 책상 테두리는 더 정확하지만, 배경의 추가 다트들과 읽히는 지도 글자로 필수 제외 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "손에 잡힌 다트의 금속 끝이 아래로 향해 지도에서 서울 표기 왼쪽의 지점에 닿아 있다. 목표가 지도라는 관계는 정확하다. 얼굴과 눈은 프레임 밖이라 시선은 보이지 않는다.",
        "built_space": "전경에는 지도 한 장을 펼친 책상 하나가 있고, 갈색 상판과 검은 테두리가 왼쪽에서 하단으로 비스듬히 이어진다. 뒤에는 흐릿한 보조 책상, 서류 더미, 창틀과 벽이 보여 이전 사무실의 재질과 조도를 대체로 유지한다. 손과 소매 중심의 클로즈업이며 반사는 없다.",
        "entities": "주름진 성인 남성의 손 하나와 올리브색 군복 소매가 보이며 참고 인물의 노출 부위와 대체로 부합한다. 얼굴이 없어 정확한 나이와 인물 동일성은 확인할 수 없다. 붉은 날개와 금속 몸통을 가진 다트, 청색 수계와 녹색 지형이 인쇄된 종이 한반도 지도가 있다. 지도상의 서울과 평양 영문 지명이 읽힌다. 뒤쪽 붉은 물체들은 흐려서 추가 다트인지 확정하기 어렵다.",
        "hard_violations": [
         "지도에 서울과 평양의 영문 지명이 판독 가능하게 노출되어, 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반한다."
        ],
        "physics": "다트는 손가락들이 실제로 잡고 있고 끝은 책상에 받쳐진 지도에 닿아 있다. 접촉점의 작은 종이 들뜸은 현재 타격으로 생긴 것으로 해석할 수 있다. 손목과 팔도 소매로 연결되어 지지 없는 물체나 신체는 없다. 다만 손끝으로 집는 파지는 강하게 내리찍는 순간보다 조심스럽게 누르는 동작에 가깝다."
       },
       {
        "label": "B",
        "direction": "주먹 아래로 나온 다트 끝이 아래쪽으로 향해 지도 남동쪽 해안의 한 지점에 닿는다. 내려찍는 축이 지도라는 목표에 도달한다. 얼굴과 시선은 프레임 밖이다.",
        "built_space": "지도 한 장이 전경 책상 상판 위에 펼쳐져 있고, 갈색 상판 가장자리와 검은 테두리가 화면 하단을 가로지른다. 뒤쪽에는 흐릿한 서류 더미와 사무실 가구 일부가 있다. 중앙의 손, 다트 끝, 비스듬히 내려다본 지도를 함께 담은 클로즈업으로 요청한 공간 관계에 잘 맞는다. 거울이나 반사는 없다.",
        "entities": "성인 남성의 거친 손 하나가 다트를 주먹으로 쥐고 있으며 올리브색 군복 소매는 참고와 부합한다. 얼굴이 없어 정확한 인물 동일성은 확인할 수 없다. 중심 소품은 붉은 날개의 금속 다트이고 지도는 참고와 유사한 접힘, 색상, 해안선을 가진 종이 지도다. 지도 하단 왼쪽의 부산 영문 표기가 읽힌다. 배경에는 별도의 붉은 다트와 파란 다트가 서 있고, 그 옆에도 붉은 다트 형태가 보인다.",
        "hard_violations": [
         "현재 손에 쥔 다트 외에 배경의 붉은색·파란색 추가 다트들을 남겨, 추가 다트를 제외하라는 명시적 조건을 위반한다.",
         "지도에 부산의 영문 지명 등 읽을 수 있는 인쇄 글자가 남아 있어 무가독 문자 조건을 위반한다."
        ],
        "physics": "주먹이 다트 축을 감싸 잡고, 금속 끝은 책상에 놓인 지도에 닿아 있어 힘이 전달되는 경로가 보인다. 손목과 전완의 정렬도 내리찍는 동작으로 가능하다. 배경 다트들은 끝을 지도 쪽에 대고 서 있는 형태이며, 지지 없이 공중에 떠 있다고 볼 근거는 없다. 다만 이 추가 다트의 존재 자체가 제외 지시와 충돌한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "다트 끝과 지도 접촉점의 클로즈업은 맞지만, 읽히는 지명이 금지 조건을 위반하고 손끝으로 누르는 듯한 파지가 세찬 타격을 약화한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "주먹으로 내리찍는 동작과 하단 책상 테두리는 더 정확하지만, 배경의 추가 다트들과 읽히는 지도 글자로 필수 제외 조건을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "손에 잡힌 다트의 금속 끝이 아래로 향해 지도에서 서울 표기 왼쪽의 지점에 닿아 있다. 목표가 지도라는 관계는 정확하다. 얼굴과 눈은 프레임 밖이라 시선은 보이지 않는다.",
        "built_space": "전경에는 지도 한 장을 펼친 책상 하나가 있고, 갈색 상판과 검은 테두리가 왼쪽에서 하단으로 비스듬히 이어진다. 뒤에는 흐릿한 보조 책상, 서류 더미, 창틀과 벽이 보여 이전 사무실의 재질과 조도를 대체로 유지한다. 손과 소매 중심의 클로즈업이며 반사는 없다.",
        "entities": "주름진 성인 남성의 손 하나와 올리브색 군복 소매가 보이며 참고 인물의 노출 부위와 대체로 부합한다. 얼굴이 없어 정확한 나이와 인물 동일성은 확인할 수 없다. 붉은 날개와 금속 몸통을 가진 다트, 청색 수계와 녹색 지형이 인쇄된 종이 한반도 지도가 있다. 지도상의 서울과 평양 영문 지명이 읽힌다. 뒤쪽 붉은 물체들은 흐려서 추가 다트인지 확정하기 어렵다.",
        "hard_violations": [
         "지도에 서울과 평양의 영문 지명이 판독 가능하게 노출되어, 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반한다."
        ],
        "physics": "다트는 손가락들이 실제로 잡고 있고 끝은 책상에 받쳐진 지도에 닿아 있다. 접촉점의 작은 종이 들뜸은 현재 타격으로 생긴 것으로 해석할 수 있다. 손목과 팔도 소매로 연결되어 지지 없는 물체나 신체는 없다. 다만 손끝으로 집는 파지는 강하게 내리찍는 순간보다 조심스럽게 누르는 동작에 가깝다."
       },
       {
        "label": "A",
        "direction": "주먹 아래로 나온 다트 끝이 아래쪽으로 향해 지도 남동쪽 해안의 한 지점에 닿는다. 내려찍는 축이 지도라는 목표에 도달한다. 얼굴과 시선은 프레임 밖이다.",
        "built_space": "지도 한 장이 전경 책상 상판 위에 펼쳐져 있고, 갈색 상판 가장자리와 검은 테두리가 화면 하단을 가로지른다. 뒤쪽에는 흐릿한 서류 더미와 사무실 가구 일부가 있다. 중앙의 손, 다트 끝, 비스듬히 내려다본 지도를 함께 담은 클로즈업으로 요청한 공간 관계에 잘 맞는다. 거울이나 반사는 없다.",
        "entities": "성인 남성의 거친 손 하나가 다트를 주먹으로 쥐고 있으며 올리브색 군복 소매는 참고와 부합한다. 얼굴이 없어 정확한 인물 동일성은 확인할 수 없다. 중심 소품은 붉은 날개의 금속 다트이고 지도는 참고와 유사한 접힘, 색상, 해안선을 가진 종이 지도다. 지도 하단 왼쪽의 부산 영문 표기가 읽힌다. 배경에는 별도의 붉은 다트와 파란 다트가 서 있고, 그 옆에도 붉은 다트 형태가 보인다.",
        "hard_violations": [
         "현재 손에 쥔 다트 외에 배경의 붉은색·파란색 추가 다트들을 남겨, 추가 다트를 제외하라는 명시적 조건을 위반한다.",
         "지도에 부산의 영문 지명 등 읽을 수 있는 인쇄 글자가 남아 있어 무가독 문자 조건을 위반한다."
        ],
        "physics": "주먹이 다트 축을 감싸 잡고, 금속 끝은 책상에 놓인 지도에 닿아 있어 힘이 전달되는 경로가 보인다. 손목과 전완의 정렬도 내리찍는 동작으로 가능하다. 배경 다트들은 끝을 지도 쪽에 대고 서 있는 형태이며, 지지 없이 공중에 떠 있다고 볼 근거는 없다. 다만 이 추가 다트의 존재 자체가 제외 지시와 충돌한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 손 안에서 다트의 상단 날개와 하단 금속 팁의 각도가 전혀 맞지 않아 축이 꺾인 것처럼 보이는 물리적 오류.",
     "[gemini-pro] 프롬프트에서 명시적으로 제외를 지시한 여분의 다트(배경의 붉은색과 푸른색 물체)가 화면에 포함됨.",
     "[gpt-high] 현재 손에 쥔 다트 외에 배경의 붉은색·파란색 추가 다트들을 남겨, 추가 다트를 제외하라는 명시적 조건을 위반한다.",
     "[gpt-high] 지도에 부산의 영문 지명 등 읽을 수 있는 인쇄 글자가 남아 있어 무가독 문자 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 지도에 서울과 평양의 영문 지명이 판독 가능하게 노출되어, 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1417,
   "B": 1417
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "지시된 화면 하단의 책상 가장자리 구도를 정확히 지켰고 '세차게 내리찍는' 주먹의 타격감을 잘 살렸으나, 다트 축이 어긋난 물리적 오류와 금지된 여분 다트가 배경에 생성된 점이 감점 요인입니다.  ★위반: [gemini-pro] 손 안에서 다트의 상단 날개와 하단 금속 팁의 각도가 전혀 맞지 않아 축이 꺾인 것처럼 보이는 물리적 오류. / [gemini-pro] 프롬프트에서 명시적으로 제외를 지시한 여분의 다트(배경의 붉은색과 푸른색 물체)가 화면에 포함됨. / [gpt-high] 현재 손에 쥔 다트 외에 배경의 붉은색·파란색 추가 다트들을 남겨, 추가 다트를 제외하라는 명시적 조건을 위반한다. / [gpt-high] 지도에 부산의 영문 지명 등 읽을 수 있는 인쇄 글자가 남아 있어 무가독 문자 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "책상 가장자리가 화면 하단이 아닌 왼쪽에 위치해 구도 지시를 어겼으며, 다트를 꼬집듯 가볍게 쥐고 있어 프롬프트가 요구한 강렬한 타격의 행동을 전혀 표현하지 못했습니다.  ★위반: [gpt-high] 지도에 서울과 평양의 영문 지명이 판독 가능하게 노출되어, 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S53sh1_sel.png",
    "asset_id": "9c3faec1-b0f8-4602-8343-90601892c0cc",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 종이 지도: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1338462>",
    "asset_id": "8c50d23a-a6d7-4581-99af-57beaa5d211d",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde1-052f-7d97-b061-53f241098627",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S53sh1"
  }
 },
 "S53sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:45:36.335713+00:00",
  "fingerprint": "8799bedbe0ba3a0c1de9679ce4b2f7834fe04c06918c7087e996cf6d69fcf2f8",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S53sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S53sh5_sel.png",
  "source_sha256": "dc9389a1fa3d0f3d454746cff0ef57c7dbe15d166e254879530c64315e7c04d3",
  "file": "S53sh5_cine.png",
  "staged_sha256": "ba9df09c7336123255af0639593db9e3a74b9e39cbb4efa48a299f6c86041578",
  "latency_ms": 11096
 },
 "S54sh5::signage": {
  "fp": "9618590e5194412c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S54sh5": {
  "input_fingerprint": "7ccf1d2f4c7c5d94",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 국방장관이 두 눈을 동그랗게 뜬 채 입을 살짝 벌린 당황한 구도.\n\nLOCATION (lock): In the visitor-conversation area of the defense minister's private office, in daytime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the daytime office preserves the small changes around his eyes and mouth without an expressive lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 국방장관 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 국방장관이 두 눈을 동그랗게 뜬 채 입을 살짝 벌린 당황한 구도.\n\nLOCATION (lock): In the visitor-conversation area of the defense minister's private office, in daytime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the daytime office preserves the small changes around his eyes and mouth without an expressive lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 국방장관 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 국방장관이 두 눈을 동그랗게 뜬 채 입을 살짝 벌린 당황한 구도.\n\nLOCATION (lock): In the visitor-conversation area of the defense minister's private office, in daytime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the daytime office preserves the small changes around his eyes and mouth without an expressive lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 국방장관 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 카메라 우측 상단 화면 밖을 향하고 있음.",
    "built_space": "주간의 사무실. 창문이 인물의 왼쪽에 위치하고, 두 깃발(태극기와 국방부기)이 인물 바로 뒤에 모여 있어 참조된 이전 샷의 공간 배치와 일치하지 않음.",
    "entities": "국방장관: 참조 이미지의 얼굴과 복장(짙은 회색 정장, 넥타이)이 일치함. 두 눈을 크게 뜨고 입을 살짝 벌린 표정.",
    "hard_violations": [
     "[gemini-pro] 이전 샷에 고정된 공간 구조 위반 (창문과 깃발의 상대적 위치 왜곡)"
    ],
    "physics": "서 있는 상태로 바닥에 지탱됨."
   },
   {
    "label": "B",
    "direction": "시선은 카메라 우측 화면 밖의 특정 대상을 향하고 있음.",
    "built_space": "주간의 사무실. 인물 뒤로 넓은 창문이 있고 왼쪽에 태극기, 오른쪽에 국방부기와 화분이 배치되어 이전 샷의 공간 구조를 정확히 유지함.",
    "entities": "국방장관: 참조 이미지의 인물 외모 및 복장과 일치함. 두 눈을 동그랗게 뜨고 입을 살짝 벌려 당황한 표정을 명확히 보여줌.",
    "hard_violations": [],
    "physics": "검은색 가죽 의자에 앉아 체중이 의자에 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지시된 표정은 잘 연출되었으나, 이전 샷에서 고정된 공간 구조(창문과 깃발의 위치)를 임의로 왜곡하여 연속성을 상실함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이전 샷의 사무실 공간 구조와 가구를 정확히 유지하면서, 눈을 둥글게 뜨고 입을 벌린 당황한 표정을 클로즈업으로 훌륭하게 구현함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 카메라 우측 상단 화면 밖을 향하고 있음.",
        "built_space": "주간의 사무실. 창문이 인물의 왼쪽에 위치하고, 두 깃발(태극기와 국방부기)이 인물 바로 뒤에 모여 있어 참조된 이전 샷의 공간 배치와 일치하지 않음.",
        "entities": "국방장관: 참조 이미지의 얼굴과 복장(짙은 회색 정장, 넥타이)이 일치함. 두 눈을 크게 뜨고 입을 살짝 벌린 표정.",
        "hard_violations": [
         "이전 샷에 고정된 공간 구조 위반 (창문과 깃발의 상대적 위치 왜곡)"
        ],
        "physics": "서 있는 상태로 바닥에 지탱됨."
       },
       {
        "label": "B",
        "direction": "시선은 카메라 우측 화면 밖의 특정 대상을 향하고 있음.",
        "built_space": "주간의 사무실. 인물 뒤로 넓은 창문이 있고 왼쪽에 태극기, 오른쪽에 국방부기와 화분이 배치되어 이전 샷의 공간 구조를 정확히 유지함.",
        "entities": "국방장관: 참조 이미지의 인물 외모 및 복장과 일치함. 두 눈을 동그랗게 뜨고 입을 살짝 벌려 당황한 표정을 명확히 보여줌.",
        "hard_violations": [],
        "physics": "검은색 가죽 의자에 앉아 체중이 의자에 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지시된 표정은 잘 연출되었으나, 이전 샷에서 고정된 공간 구조(창문과 깃발의 위치)를 임의로 왜곡하여 연속성을 상실함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이전 샷의 사무실 공간 구조와 가구를 정확히 유지하면서, 눈을 둥글게 뜨고 입을 벌린 당황한 표정을 클로즈업으로 훌륭하게 구현함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 카메라 우측 상단 화면 밖을 향하고 있음.",
        "built_space": "주간의 사무실. 창문이 인물의 왼쪽에 위치하고, 두 깃발(태극기와 국방부기)이 인물 바로 뒤에 모여 있어 참조된 이전 샷의 공간 배치와 일치하지 않음.",
        "entities": "국방장관: 참조 이미지의 얼굴과 복장(짙은 회색 정장, 넥타이)이 일치함. 두 눈을 크게 뜨고 입을 살짝 벌린 표정.",
        "hard_violations": [
         "이전 샷에 고정된 공간 구조 위반 (창문과 깃발의 상대적 위치 왜곡)"
        ],
        "physics": "서 있는 상태로 바닥에 지탱됨."
       },
       {
        "label": "B",
        "direction": "시선은 카메라 우측 화면 밖의 특정 대상을 향하고 있음.",
        "built_space": "주간의 사무실. 인물 뒤로 넓은 창문이 있고 왼쪽에 태극기, 오른쪽에 국방부기와 화분이 배치되어 이전 샷의 공간 구조를 정확히 유지함.",
        "entities": "국방장관: 참조 이미지의 인물 외모 및 복장과 일치함. 두 눈을 동그랗게 뜨고 입을 살짝 벌려 당황한 표정을 명확히 보여줌.",
        "hard_violations": [],
        "physics": "검은색 가죽 의자에 앉아 체중이 의자에 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "얼굴 중심의 근접 구도에서 동그랗게 뜬 두 눈과 살짝 열린 입을 명확히 구현하며, 이전 장면의 의자·창·양쪽 깃발 배치도 더 잘 유지한다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "당황한 표정과 인물 외형은 충실하지만, 몸통과 주변 공간을 더 많이 보여 주어 요구한 클로즈업보다 느슨하고 왼쪽 창의 공간 연속성도 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 정면이고 두 눈은 렌즈 부근을 응시한다. 화면 안에 시선의 대상은 없으며, 지문도 특정 대상을 지정하지 않는다. 눈을 크게 뜨고 입을 조금 벌려 당황한 순간을 표현한다. 무기나 방향을 검사할 휴대 물체는 없다.",
        "built_space": "인물 바로 뒤에 검은 가죽 등받이 하나, 뒤쪽 벽에 세로 프레임으로 나뉜 창, 왼쪽 태극기 하나와 오른쪽 청색 깃발 하나가 보인다. 창 아래 목재 수납부와 작은 원형 장식 하나, 오른쪽 가장자리 식물 일부도 보인다. 이전 장면의 주요 배치와 재료를 유지하며 의자와 인물의 방향이 맞는다. 불가능한 반사나 중복된 집기는 보이지 않는다.",
        "entities": "성인 남성 한 명만 등장한다. 짙은 뒤로 넘긴 머리, 갈색 피부, 짧은 수염과 얼굴 윤곽이 참고 인물에 가깝다. 특정 민족은 지문에 명시되지 않았다. 이전 장면처럼 모자 없이 짙은 회색 정장과 흰 셔츠, 회색 넥타이를 착용하나 넥타이는 조금 더 밝다. 눈은 정상적인 홍채와 동공을 유지한다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "등과 어깨 뒤에 의자 등받이가 있어 앉은 상체를 지지하는 관계가 자연스럽다. 엉덩이와 발은 프레임 밖이므로 접촉 상태를 단정할 수 없지만 공중에 떠 있다는 징후는 없다. 깃발은 봉과 바닥 받침으로 지지되고 원형 장식은 창 아래 가구 위에 놓여 있다. 손에 든 물체나 움직이는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "얼굴과 시선은 카메라 왼쪽의 화면 밖을 향한다. 구체적인 대상은 보이지 않지만 지문에 지정된 시선 대상도 없다. 두 눈을 크게 뜨고 입술을 조금 벌린 당황한 연기가 보인다. 무기나 지향성 휴대 물체는 없다.",
        "built_space": "인물 뒤에 검은 의자 등받이 하나, 부분적으로 가려진 태극기 하나, 오른쪽 청색 깃발 하나가 있다. 왼쪽 벽에는 액자 하나가 있고, 화면 왼쪽과 뒤쪽에 창이 보인다. 목재와 밝은 벽 패널은 참고 장소와 유사하지만 왼쪽의 큰 창은 이전 장면의 액자 중심 측벽과 연결이 덜 분명하다. 의자와 인물 사이의 정확한 거리는 몸통에 가려져 확인하기 어렵다. 중복 집기나 불가능한 반사는 없다.",
        "entities": "참고 인물과 유사한 성인 남성 한 명으로, 피부색과 얼굴형, 짧은 수염, 뒤로 넘긴 검은 머리를 유지한다. 짙은 회색 더블브레스트 정장, 흰 셔츠, 회색 넥타이가 보이며 이전 장면처럼 모자는 없다. 정상적인 홍채와 동공이 보인다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "상체는 수직으로 자연스럽게 유지되며 뒤에 의자 등받이가 일부 보인다. 좌면·골반·발이 잘려 있어 앉았는지 서 있는지와 하중 접촉점은 확정할 수 없지만, 부유하거나 불가능하게 꺾인 자세는 아니다. 깃발은 봉에 매달려 있고 액자는 벽에 부착되어 있다. 손에 든 물체나 공중에 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "얼굴 중심의 근접 구도에서 동그랗게 뜬 두 눈과 살짝 열린 입을 명확히 구현하며, 이전 장면의 의자·창·양쪽 깃발 배치도 더 잘 유지한다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "당황한 표정과 인물 외형은 충실하지만, 몸통과 주변 공간을 더 많이 보여 주어 요구한 클로즈업보다 느슨하고 왼쪽 창의 공간 연속성도 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 정면이고 두 눈은 렌즈 부근을 응시한다. 화면 안에 시선의 대상은 없으며, 지문도 특정 대상을 지정하지 않는다. 눈을 크게 뜨고 입을 조금 벌려 당황한 순간을 표현한다. 무기나 방향을 검사할 휴대 물체는 없다.",
        "built_space": "인물 바로 뒤에 검은 가죽 등받이 하나, 뒤쪽 벽에 세로 프레임으로 나뉜 창, 왼쪽 태극기 하나와 오른쪽 청색 깃발 하나가 보인다. 창 아래 목재 수납부와 작은 원형 장식 하나, 오른쪽 가장자리 식물 일부도 보인다. 이전 장면의 주요 배치와 재료를 유지하며 의자와 인물의 방향이 맞는다. 불가능한 반사나 중복된 집기는 보이지 않는다.",
        "entities": "성인 남성 한 명만 등장한다. 짙은 뒤로 넘긴 머리, 갈색 피부, 짧은 수염과 얼굴 윤곽이 참고 인물에 가깝다. 특정 민족은 지문에 명시되지 않았다. 이전 장면처럼 모자 없이 짙은 회색 정장과 흰 셔츠, 회색 넥타이를 착용하나 넥타이는 조금 더 밝다. 눈은 정상적인 홍채와 동공을 유지한다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "등과 어깨 뒤에 의자 등받이가 있어 앉은 상체를 지지하는 관계가 자연스럽다. 엉덩이와 발은 프레임 밖이므로 접촉 상태를 단정할 수 없지만 공중에 떠 있다는 징후는 없다. 깃발은 봉과 바닥 받침으로 지지되고 원형 장식은 창 아래 가구 위에 놓여 있다. 손에 든 물체나 움직이는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "얼굴과 시선은 카메라 왼쪽의 화면 밖을 향한다. 구체적인 대상은 보이지 않지만 지문에 지정된 시선 대상도 없다. 두 눈을 크게 뜨고 입술을 조금 벌린 당황한 연기가 보인다. 무기나 지향성 휴대 물체는 없다.",
        "built_space": "인물 뒤에 검은 의자 등받이 하나, 부분적으로 가려진 태극기 하나, 오른쪽 청색 깃발 하나가 있다. 왼쪽 벽에는 액자 하나가 있고, 화면 왼쪽과 뒤쪽에 창이 보인다. 목재와 밝은 벽 패널은 참고 장소와 유사하지만 왼쪽의 큰 창은 이전 장면의 액자 중심 측벽과 연결이 덜 분명하다. 의자와 인물 사이의 정확한 거리는 몸통에 가려져 확인하기 어렵다. 중복 집기나 불가능한 반사는 없다.",
        "entities": "참고 인물과 유사한 성인 남성 한 명으로, 피부색과 얼굴형, 짧은 수염, 뒤로 넘긴 검은 머리를 유지한다. 짙은 회색 더블브레스트 정장, 흰 셔츠, 회색 넥타이가 보이며 이전 장면처럼 모자는 없다. 정상적인 홍채와 동공이 보인다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "상체는 수직으로 자연스럽게 유지되며 뒤에 의자 등받이가 일부 보인다. 좌면·골반·발이 잘려 있어 앉았는지 서 있는지와 하중 접촉점은 확정할 수 없지만, 부유하거나 불가능하게 꺾인 자세는 아니다. 깃발은 봉에 매달려 있고 액자는 벽에 부착되어 있다. 손에 든 물체나 공중에 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.349,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.099,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 샷에 고정된 공간 구조 위반 (창문과 깃발의 상대적 위치 왜곡)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 1099,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1099,
    "verdict_ko": "지시된 표정은 잘 연출되었으나, 이전 샷에서 고정된 공간 구조(창문과 깃발의 위치)를 임의로 왜곡하여 연속성을 상실함.  ★위반: [gemini-pro] 이전 샷에 고정된 공간 구조 위반 (창문과 깃발의 상대적 위치 왜곡)"
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "이전 샷의 사무실 공간 구조와 가구를 정확히 유지하면서, 눈을 둥글게 뜨고 입을 벌린 당황한 표정을 클로즈업으로 훌륭하게 구현함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 국방장관 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S43sh1_sel.png",
    "asset_id": "7865ba3d-28be-4eb0-a417-2d8e3608c202",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:684284>",
    "asset_id": "b8636b60-6a4d-4eb8-bb28-08231e30fcd5",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde1-bc60-7568-9093-75fc57f0dbd3",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S43sh1"
  }
 },
 "S54sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:45:47.782742+00:00",
  "fingerprint": "db3a7a1f5b770acd541d2dd82c4cd78c3a31792e0dc185aa3dde837d61062953",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S54sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S54sh5_sel.png",
  "source_sha256": "e19a3382c9496c048a4283cf5657f7a5e01bdff526345812a06756767793447c",
  "file": "S54sh5_cine.png",
  "staged_sha256": "922ea0adb20608629e5fbf678a723ca288c86842946108ade98ede6a0fc7d766",
  "latency_ms": 10102
 },
 "S54sh6::signage": {
  "fp": "aa772d19061169f5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S54sh6": {
  "input_fingerprint": "83d62e1768e4ed45",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬이 턱을 살짝 치켜든 채 비릿하고 거만한 미소를 띤 찰나.\n\nLOCATION (lock): Opposite the minister inside his private office, under daytime office ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding office illumination and tonal balance, letting the low viewing angle and restrained smile convey confidence rather than a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬이 턱을 살짝 치켜든 채 비릿하고 거만한 미소를 띤 찰나.\n\nLOCATION (lock): Opposite the minister inside his private office, under daytime office ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding office illumination and tonal balance, letting the low viewing angle and restrained smile convey confidence rather than a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬이 턱을 살짝 치켜든 채 비릿하고 거만한 미소를 띤 찰나.\n\nLOCATION (lock): Opposite the minister inside his private office, under daytime office ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding office illumination and tonal balance, letting the low viewing angle and restrained smile convey confidence rather than a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라 정면을 향해 턱을 살짝 들고 거만한 미소를 띠고 있음.",
    "built_space": "사무실 내부의 맞은편 좌석, 어두운 가죽 의자와 나무 패널 벽면, 왼쪽에 태극기가 일부 보임.",
    "entities": "윤성찬 레퍼런스와 완벽하게 일치 (백발 노인 남성, 네이비 3피스 정장, 흰 셔츠, 파란색 넥타이).",
    "hard_violations": [],
    "physics": "의자에 앉아 체중이 자연스럽게 실려 있음."
   },
   {
    "label": "B",
    "direction": "약간 우측을 응시하며 턱을 들고 미소를 띠고 있음.",
    "built_space": "사무실 내부, 이전 샷과 동일한 창문 배경, 책상, 왼쪽 태극기, 오른쪽 유엔기(UN flag).",
    "entities": "윤성찬의 얼굴은 일치하나 복장 지침 위반 (조끼와 넥타이 누락, 단추가 풀린 흰 셔츠).",
    "hard_violations": [
     "[gpt-high] 장관 맞은편이라는 지정과 달리, 이전 장관 뒤의 창·양쪽 깃발·검은 등받이를 그대로 뒤에 두어 윤성찬을 장관 좌석에 배치했다."
    ],
    "physics": "책상 앞 의자에 기대어 앉아 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 클로즈업 프레이밍과 캐릭터의 복장(넥타이, 조끼)을 정확히 준수했으며, 요구된 거만한 표정과 장소의 논리(맞은편 벽면)를 훌륭하게 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터 레퍼런스의 핵심 복장(넥타이, 조끼)을 누락했고, 프레이밍이 지정된 샷보다 넓으며, 우측 깃발이 유엔기로 잘못 변형됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라 정면을 향해 턱을 살짝 들고 거만한 미소를 띠고 있음.",
        "built_space": "사무실 내부의 맞은편 좌석, 어두운 가죽 의자와 나무 패널 벽면, 왼쪽에 태극기가 일부 보임.",
        "entities": "윤성찬 레퍼런스와 완벽하게 일치 (백발 노인 남성, 네이비 3피스 정장, 흰 셔츠, 파란색 넥타이).",
        "hard_violations": [],
        "physics": "의자에 앉아 체중이 자연스럽게 실려 있음."
       },
       {
        "label": "B",
        "direction": "약간 우측을 응시하며 턱을 들고 미소를 띠고 있음.",
        "built_space": "사무실 내부, 이전 샷과 동일한 창문 배경, 책상, 왼쪽 태극기, 오른쪽 유엔기(UN flag).",
        "entities": "윤성찬의 얼굴은 일치하나 복장 지침 위반 (조끼와 넥타이 누락, 단추가 풀린 흰 셔츠).",
        "hard_violations": [],
        "physics": "책상 앞 의자에 기대어 앉아 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 클로즈업 프레이밍과 캐릭터의 복장(넥타이, 조끼)을 정확히 준수했으며, 요구된 거만한 표정과 장소의 논리(맞은편 벽면)를 훌륭하게 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터 레퍼런스의 핵심 복장(넥타이, 조끼)을 누락했고, 프레이밍이 지정된 샷보다 넓으며, 우측 깃발이 유엔기로 잘못 변형됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라 정면을 향해 턱을 살짝 들고 거만한 미소를 띠고 있음.",
        "built_space": "사무실 내부의 맞은편 좌석, 어두운 가죽 의자와 나무 패널 벽면, 왼쪽에 태극기가 일부 보임.",
        "entities": "윤성찬 레퍼런스와 완벽하게 일치 (백발 노인 남성, 네이비 3피스 정장, 흰 셔츠, 파란색 넥타이).",
        "hard_violations": [],
        "physics": "의자에 앉아 체중이 자연스럽게 실려 있음."
       },
       {
        "label": "B",
        "direction": "약간 우측을 응시하며 턱을 들고 미소를 띠고 있음.",
        "built_space": "사무실 내부, 이전 샷과 동일한 창문 배경, 책상, 왼쪽 태극기, 오른쪽 유엔기(UN flag).",
        "entities": "윤성찬의 얼굴은 일치하나 복장 지침 위반 (조끼와 넥타이 누락, 단추가 풀린 흰 셔츠).",
        "hard_violations": [],
        "physics": "책상 앞 의자에 기대어 앉아 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "장관 맞은편이 아니라 이전 장관의 창가 좌석을 차지한 구도이며, 클로즈업보다 넓고 참고 의상의 넥타이도 빠져 있다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "얼굴 중심 클로즈업과 절제된 거만한 미소, 남색 넥타이가 더 충실하지만, 턱을 치켜든 동작과 낮은 시점은 약하고 장소의 정확한 연속성은 불확실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 거의 렌즈 정면을 향하고 턱은 조금 들려 있다. 화면 밖 정면의 장관을 보는 연기로 해석할 수 있으나 장관 자체는 보이지 않는다. 입꼬리는 살짝 올라가 있으며, 겨누거나 이동하는 물체는 없다.",
        "built_space": "뒤에 가로로 긴 분할 창 하나, 왼쪽 태극기 하나, 오른쪽 파란 깃발 하나, 검은 의자 등받이 하나가 보인다. 창 아래 목재 가구 위에는 작은 시계와 전화기 형태의 물체가 있고, 전경에는 넓은 목재 상판이 있다. 창밖 산과 양쪽 깃발 및 의자의 배열이 이전 장관 좌석의 배치를 재현하므로 윤성찬이 장관 맞은편이 아니라 장관 자리에 앉은 것으로 읽힌다. 오른쪽 깃발의 색과 문양도 참고와 다르다.",
        "entities": "고령의 회색 머리 남성 한 명만 등장한다. 나이대와 성별은 부합하지만 참고보다 얼굴이 넓고 머리 윗부분이 풍성해 정확한 동일 인물 재현에는 차이가 있다. 남색 재킷과 흰 셔츠는 맞지만 참고의 남색 넥타이가 없고 셔츠 목이 열려 있다. 이전 장관의 얼굴은 복제하지 않았다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "장관 맞은편이라는 지정과 달리, 이전 장관 뒤의 창·양쪽 깃발·검은 등받이를 그대로 뒤에 두어 윤성찬을 장관 좌석에 배치했다."
        ],
        "physics": "상체 뒤에 의자 등받이가 있어 앉아 있는 자세로 자연스럽게 읽힌다. 엉덩이와 발은 상판 및 화면 밖에 가려져 접촉을 직접 확인할 수 없지만 부유를 나타내는 징후는 없다. 머리는 목으로 지지되고, 배경 물체들은 가구 위에 놓여 있다. 전경 상판의 흐린 반사는 물리적으로 무리하지 않다."
       },
       {
        "label": "B",
        "direction": "눈은 거의 렌즈 정면의 상대를 바라본다. 화면 밖 장관을 향한 시선으로 읽을 수 있다. 입술을 다문 작은 미소와 약간 좁힌 눈은 자신만만한 표정에 가깝지만, 턱은 뚜렷하게 들리지 않았고 시점도 낮다기보다 눈높이에 가깝다. 방향을 확인할 무기나 이동 물체는 없다.",
        "built_space": "얼굴 뒤에 갈색 가죽 의자 등받이 하나, 목재 벽면, 왼쪽 가장자리의 태극기 일부 하나가 보인다. 창과 다른 깃발은 이 좁은 프레임에 나오지 않으므로 누락 자체는 문제가 아니다. 목재 마감과 주간의 부드러운 빛은 사무실과 어울리지만, 참고의 검은 의자와 다른 등받이 및 창 없는 배경 때문에 같은 방의 정확한 좌석 위치는 확인하기 어렵다. 중복된 시설이나 불가능한 반사는 보이지 않는다.",
        "entities": "고령의 회색 머리 남성 한 명이며, 남색 정장과 흰 셔츠, 남색 넥타이가 참고 의상에 부합한다. 참고보다 머리가 풍성하고 얼굴 윤곽이 다소 넓지만, 연령감과 회색 머리 등 주요 특징은 유지된다. 조끼와 하반신은 클로즈업 밖이므로 판단 대상이 아니다. 추가 인물이나 읽을 수 있는 문구는 없다.",
        "hard_violations": [],
        "physics": "상체가 의자 등받이 앞에 자연스럽게 놓여 있고 목과 어깨가 머리를 정상적으로 지지한다. 좌판과 발은 화면 밖이라 직접 확인할 수 없지만, 떠 있거나 지지 없이 매달린 신체는 보이지 않는다. 미소와 눈의 형태도 실제 배우가 만들 수 있는 정상적인 표정이다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "장관 맞은편이 아니라 이전 장관의 창가 좌석을 차지한 구도이며, 클로즈업보다 넓고 참고 의상의 넥타이도 빠져 있다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "얼굴 중심 클로즈업과 절제된 거만한 미소, 남색 넥타이가 더 충실하지만, 턱을 치켜든 동작과 낮은 시점은 약하고 장소의 정확한 연속성은 불확실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "시선은 거의 렌즈 정면을 향하고 턱은 조금 들려 있다. 화면 밖 정면의 장관을 보는 연기로 해석할 수 있으나 장관 자체는 보이지 않는다. 입꼬리는 살짝 올라가 있으며, 겨누거나 이동하는 물체는 없다.",
        "built_space": "뒤에 가로로 긴 분할 창 하나, 왼쪽 태극기 하나, 오른쪽 파란 깃발 하나, 검은 의자 등받이 하나가 보인다. 창 아래 목재 가구 위에는 작은 시계와 전화기 형태의 물체가 있고, 전경에는 넓은 목재 상판이 있다. 창밖 산과 양쪽 깃발 및 의자의 배열이 이전 장관 좌석의 배치를 재현하므로 윤성찬이 장관 맞은편이 아니라 장관 자리에 앉은 것으로 읽힌다. 오른쪽 깃발의 색과 문양도 참고와 다르다.",
        "entities": "고령의 회색 머리 남성 한 명만 등장한다. 나이대와 성별은 부합하지만 참고보다 얼굴이 넓고 머리 윗부분이 풍성해 정확한 동일 인물 재현에는 차이가 있다. 남색 재킷과 흰 셔츠는 맞지만 참고의 남색 넥타이가 없고 셔츠 목이 열려 있다. 이전 장관의 얼굴은 복제하지 않았다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "장관 맞은편이라는 지정과 달리, 이전 장관 뒤의 창·양쪽 깃발·검은 등받이를 그대로 뒤에 두어 윤성찬을 장관 좌석에 배치했다."
        ],
        "physics": "상체 뒤에 의자 등받이가 있어 앉아 있는 자세로 자연스럽게 읽힌다. 엉덩이와 발은 상판 및 화면 밖에 가려져 접촉을 직접 확인할 수 없지만 부유를 나타내는 징후는 없다. 머리는 목으로 지지되고, 배경 물체들은 가구 위에 놓여 있다. 전경 상판의 흐린 반사는 물리적으로 무리하지 않다."
       },
       {
        "label": "A",
        "direction": "눈은 거의 렌즈 정면의 상대를 바라본다. 화면 밖 장관을 향한 시선으로 읽을 수 있다. 입술을 다문 작은 미소와 약간 좁힌 눈은 자신만만한 표정에 가깝지만, 턱은 뚜렷하게 들리지 않았고 시점도 낮다기보다 눈높이에 가깝다. 방향을 확인할 무기나 이동 물체는 없다.",
        "built_space": "얼굴 뒤에 갈색 가죽 의자 등받이 하나, 목재 벽면, 왼쪽 가장자리의 태극기 일부 하나가 보인다. 창과 다른 깃발은 이 좁은 프레임에 나오지 않으므로 누락 자체는 문제가 아니다. 목재 마감과 주간의 부드러운 빛은 사무실과 어울리지만, 참고의 검은 의자와 다른 등받이 및 창 없는 배경 때문에 같은 방의 정확한 좌석 위치는 확인하기 어렵다. 중복된 시설이나 불가능한 반사는 보이지 않는다.",
        "entities": "고령의 회색 머리 남성 한 명이며, 남색 정장과 흰 셔츠, 남색 넥타이가 참고 의상에 부합한다. 참고보다 머리가 풍성하고 얼굴 윤곽이 다소 넓지만, 연령감과 회색 머리 등 주요 특징은 유지된다. 조끼와 하반신은 클로즈업 밖이므로 판단 대상이 아니다. 추가 인물이나 읽을 수 있는 문구는 없다.",
        "hard_violations": [],
        "physics": "상체가 의자 등받이 앞에 자연스럽게 놓여 있고 목과 어깨가 머리를 정상적으로 지지한다. 좌판과 발은 화면 밖이라 직접 확인할 수 없지만, 떠 있거나 지지 없이 매달린 신체는 보이지 않는다. 미소와 눈의 형태도 실제 배우가 만들 수 있는 정상적인 표정이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.857
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.607
   },
   "violations": {
    "B": [
     "[gpt-high] 장관 맞은편이라는 지정과 달리, 이전 장관 뒤의 창·양쪽 깃발·검은 등받이를 그대로 뒤에 두어 윤성찬을 장관 좌석에 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 607
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 클로즈업 프레이밍과 캐릭터의 복장(넥타이, 조끼)을 정확히 준수했으며, 요구된 거만한 표정과 장소의 논리(맞은편 벽면)를 훌륭하게 구현함."
   },
   {
    "label": "B",
    "score": 607,
    "verdict_ko": "캐릭터 레퍼런스의 핵심 복장(넥타이, 조끼)을 누락했고, 프레이밍이 지정된 샷보다 넓으며, 우측 깃발이 유엔기로 잘못 변형됨.  ★위반: [gpt-high] 장관 맞은편이라는 지정과 달리, 이전 장관 뒤의 창·양쪽 깃발·검은 등받이를 그대로 뒤에 두어 윤성찬을 장관 좌석에 배치했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S54sh5_sel.png",
    "asset_id": "0482f8d1-c434-44d4-99b2-552a79f895d4",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:781789>",
    "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde2-6381-7fea-9f4b-469095bac61f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S54sh5"
  }
 },
 "S54sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:45:58.228367+00:00",
  "fingerprint": "c38758d535ac2bfda1aa64661a7d88e62c219933954929813ad5575e7de00482",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S54sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S54sh6_sel.png",
  "source_sha256": "a12dbc652d36f0a7c37ea329c55402c4b4ec469fb11681d4469662d2f9658f69",
  "file": "S54sh6_cine.png",
  "staged_sha256": "137d26c4c606cb9f51b65b159ec94e10cb03093d43eea18599fde90abc784b6a",
  "latency_ms": 11106
 },
 "S55sh1::signage": {
  "fp": "8f9be8b184f9e69a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::7bfa8c95975a769d": {
  "subjects": [],
  "subject_text": "익산 늪지대와 캠핑카가 빠진 진흙 공터\n숲 가장자리에 펼쳐진 황폐한 늪과 진흙 공터. 출입 금지 팻말과 표식이 희미해진 녹슨 방사능 경고판이 서 있다.",
  "identity": "canonical",
  "scope_id": "L72",
  "scope_role": "location_exterior",
  "scope_sha": "ef93357fcf325506"
 },
 "groupbg::camper_living": {
  "input_fingerprint": "3cdb0e3bd6130f43",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "camper_living",
    "tags": [
     "S44sh6",
     "S45sh14",
     "S45sh8",
     "S55sh1"
    ]
   },
   "context_sig": "6fc7ea6386e00160"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: Inside the camper's rear seating area while traveling along a provincial road, with post-shower daylight through the side windows.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우 일행의 캠핑카 내부: 침대와 좌석이 빽빽하게 들어찬 생활용 차량으로, 지붕 틈새로 빗물이 뚝뚝 떨어지는 낡은 내부. (특징: 좁은 뒷좌석과 침대 공간; 천장에서 새어 들어오는 빗물과 이를 막으려는 옷가지들; 어두운 실내를 비추는 작은 전등) / 익산 늪지대와 캠핑카가 빠진 진흙 공터: 부식되어 글씨를 알아보기 힘든 방사능 경고 팻말이 꽂혀 있고, 바퀴가 푹 빠진 깊은 진흙 웅덩이가 있는 황량한 늪. (특징: 차량 바퀴가 절반쯤 잠긴 깊은 진흙 웅덩이; 녹슬고 부식된 '방사능 구역' 출입 금지 팻말) / 캠핑카가 보관된 폐공장 옆 창고: 위로 들어 올리는 셔터 문이 있고 내부에 어둠에 싸인 소형 캠핑카가 숨겨져 있는 낡은 차고. (특징: 먼지 쌓인 철제 롤업 셔터; 어둠 속에서 모습을 드러내는 소형 캠핑 트럭)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 네 사람이 거뜬히 탈 만한, 침대도 있고... 꽤나 그럴싸한 비주얼이다.\n- 뒤에서는 앰버와 찰리가 앉아있다.\n- 보아하니 태진이 고치고 있는 건 현우의 캠핑카다.\n- 55. 캠핑차 안 + 지방도로 -D\n\nTIME OF DAY (lock): day, after a rain shower.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: Inside the camper's rear seating area while traveling along a provincial road, with post-shower daylight through the side windows.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우 일행의 캠핑카 내부: 침대와 좌석이 빽빽하게 들어찬 생활용 차량으로, 지붕 틈새로 빗물이 뚝뚝 떨어지는 낡은 내부. (특징: 좁은 뒷좌석과 침대 공간; 천장에서 새어 들어오는 빗물과 이를 막으려는 옷가지들; 어두운 실내를 비추는 작은 전등) / 익산 늪지대와 캠핑카가 빠진 진흙 공터: 부식되어 글씨를 알아보기 힘든 방사능 경고 팻말이 꽂혀 있고, 바퀴가 푹 빠진 깊은 진흙 웅덩이가 있는 황량한 늪. (특징: 차량 바퀴가 절반쯤 잠긴 깊은 진흙 웅덩이; 녹슬고 부식된 '방사능 구역' 출입 금지 팻말) / 캠핑카가 보관된 폐공장 옆 창고: 위로 들어 올리는 셔터 문이 있고 내부에 어둠에 싸인 소형 캠핑카가 숨겨져 있는 낡은 차고. (특징: 먼지 쌓인 철제 롤업 셔터; 어둠 속에서 모습을 드러내는 소형 캠핑 트럭)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 네 사람이 거뜬히 탈 만한, 침대도 있고... 꽤나 그럴싸한 비주얼이다.\n- 뒤에서는 앰버와 찰리가 앉아있다.\n- 보아하니 태진이 고치고 있는 건 현우의 캠핑카다.\n- 55. 캠핑차 안 + 지방도로 -D\n\nTIME OF DAY (lock): day, after a rain shower.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_camper_living_809ab7.png",
  "asset_id": "0a075c81-0701-4544-9b02-0334a12fab36",
  "input_asset_ids": [
   "45256585-5451-4c0b-85a7-77766ce17dca"
  ],
  "origin_tag": "S55sh1",
  "place_text": "Inside the camper's rear seating area while traveling along a provincial road, with post-shower daylight through the side windows.",
  "origin_inputs": {
   "place_text": "Inside the camper's rear seating area while traveling along a provincial road, with post-shower daylight through the side windows.",
   "time_of_day_en": "day, after a rain shower",
   "conti_asset_id": "45256585-5451-4c0b-85a7-77766ce17dca"
  }
 },
 "S55sh1::bgfirst_bg": {
  "input_fingerprint": "62bcc122e0ca50ed",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 낡은 캠핑카 뒷좌석에서 앰버와 라울이 젖은 머리카락을 한 채 창밖을 내다보는 찰나.\n\nLOCATION (lock): Inside the camper's rear seating area while traveling along a provincial road, with post-shower daylight through the side windows.\n\nTIME OF DAY (lock): day, after a rain shower.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rear seats (Occupied by the exhausted passengers inside the old camper) — Seat backs appear obliquely beneath the passengers; used as Low foreground boundary establishing the camera behind the passengers; Side windows (Passengers search outside through them) — Seen obliquely from the aisle, with the roadside visible beyond; used as Carry the passengers' outward attention without becoming a reflected view; Camper aisle (Connects the rear seating area to the driving position) — Recedes from the lower frame toward the front of the camper; used as Establishes the continuing camera route.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light through the windows gently separates their tired faces from the interior without romanticizing their distress.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 낡은 캠핑카 뒷좌석에서 앰버와 라울이 젖은 머리카락을 한 채 창밖을 내다보는 찰나.\n\nLOCATION (lock): Inside the camper's rear seating area while traveling along a provincial road, with post-shower daylight through the side windows.\n\nTIME OF DAY (lock): day, after a rain shower.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rear seats (Occupied by the exhausted passengers inside the old camper) — Seat backs appear obliquely beneath the passengers; used as Low foreground boundary establishing the camera behind the passengers; Side windows (Passengers search outside through them) — Seen obliquely from the aisle, with the roadside visible beyond; used as Carry the passengers' outward attention without becoming a reflected view; Camper aisle (Connects the rear seating area to the driving position) — Recedes from the lower frame toward the front of the camper; used as Establishes the continuing camera route.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light through the windows gently separates their tired faces from the interior without romanticizing their distress.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S55sh1__bgfirst_bg.png",
  "asset_id": "f43f0854-763c-4679-8ff3-b631473172d6",
  "input_asset_ids": [
   "45256585-5451-4c0b-85a7-77766ce17dca",
   "0a075c81-0701-4544-9b02-0334a12fab36"
  ]
 },
 "S55sh1": {
  "input_fingerprint": "63b6cff8ab821dd0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day, after a rain shower.\n\nSHOT TEXT (authoritative, Korean): 낡은 캠핑카 뒷좌석에서 앰버와 라울이 젖은 머리카락을 한 채 창밖을 내다보는 찰나.\n\nLOCATION (lock): Inside the camper's rear seating area while traveling along a provincial road, with post-shower daylight through the side windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rear seats (Occupied by the exhausted passengers inside the old camper) — Seat backs appear obliquely beneath the passengers; used as Low foreground boundary establishing the camera behind the passengers; Side windows (Passengers search outside through them) — Seen obliquely from the aisle, with the roadside visible beyond; used as Carry the passengers' outward attention without becoming a reflected view; Camper aisle (Connects the rear seating area to the driving position) — Recedes from the lower frame toward the front of the camper; used as Establishes the continuing camera route.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light through the windows gently separates their tired faces from the interior without romanticizing their distress.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The repaired camper carries the luggage, food and medicine supplied at the car center. A rain shower has just passed. 앰버: She rides in the camper looking outside, with wet hair and clothing and an exhausted appearance. 라울: He rides in the camper looking outside, soaked and tired, with his earlier injury treated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day, after a rain shower.\n\nSHOT TEXT (authoritative, Korean): 낡은 캠핑카 뒷좌석에서 앰버와 라울이 젖은 머리카락을 한 채 창밖을 내다보는 찰나.\n\nLOCATION (lock): Inside the camper's rear seating area while traveling along a provincial road, with post-shower daylight through the side windows. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rear seats (Occupied by the exhausted passengers inside the old camper) — Seat backs appear obliquely beneath the passengers; used as Low foreground boundary establishing the camera behind the passengers; Side windows (Passengers search outside through them) — Seen obliquely from the aisle, with the roadside visible beyond; used as Carry the passengers' outward attention without becoming a reflected view; Camper aisle (Connects the rear seating area to the driving position) — Recedes from the lower frame toward the front of the camper; used as Establishes the continuing camera route.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light through the windows gently separates their tired faces from the interior without romanticizing their distress.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The repaired camper carries the luggage, food and medicine supplied at the car center. A rain shower has just passed. 앰버: She rides in the camper looking outside, with wet hair and clothing and an exhausted appearance. 라울: He rides in the camper looking outside, soaked and tired, with his earlier injury treated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day, after a rain shower.\n\nSHOT TEXT (authoritative, Korean): 낡은 캠핑카 뒷좌석에서 앰버와 라울이 젖은 머리카락을 한 채 창밖을 내다보는 찰나.\n\nLOCATION (lock): Inside the camper's rear seating area while traveling along a provincial road, with post-shower daylight through the side windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rear seats (Occupied by the exhausted passengers inside the old camper) — Seat backs appear obliquely beneath the passengers; used as Low foreground boundary establishing the camera behind the passengers; Side windows (Passengers search outside through them) — Seen obliquely from the aisle, with the roadside visible beyond; used as Carry the passengers' outward attention without becoming a reflected view; Camper aisle (Connects the rear seating area to the driving position) — Recedes from the lower frame toward the front of the camper; used as Establishes the continuing camera route.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light through the windows gently separates their tired faces from the interior without romanticizing their distress.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The repaired camper carries the luggage, food and medicine supplied at the car center. A rain shower has just passed. 앰버: She rides in the camper looking outside, with wet hair and clothing and an exhausted appearance. 라울: He rides in the camper looking outside, soaked and tired, with his earlier injury treated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S55sh1__bgfirst_bg.png",
     "asset_id": "f43f0854-763c-4679-8ff3-b631473172d6",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S55sh1.png",
     "asset_id": "45256585-5451-4c0b-85a7-77766ce17dca",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_camper_living_809ab7.png",
     "asset_id": "0a075c81-0701-4544-9b02-0334a12fab36",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "앰버는 왼쪽 창밖을, 라울은 오른쪽 창밖을 응시함.",
    "built_space": "캠핑카 내부. 전경에 좌석 등받이가 있고 중앙 통로가 앞쪽으로 이어지며 양쪽 좌석에 인물이 자연스럽게 배치됨.",
    "entities": "앰버(금발, 젖은 머리), 라울(꽁지머리, 젖은 머리, 팔에 반창고) 모두 지친 기색이 뚜렷하며 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "앰버는 손으로 턱을 괴고 기댔으며, 라울도 팔을 등받이에 걸치고 안정적으로 좌석에 앉아 있음."
   },
   {
    "label": "B",
    "direction": "앰버와 라울 모두 오른쪽 창밖을 응시함.",
    "built_space": "캠핑카 내부. 통로와 전경 좌석이 있으나, 두 인물이 오른쪽 좌석에만 몰려 앉아 배치와 공간감이 다소 답답함.",
    "entities": "앰버와 라울의 외형은 레퍼런스를 따르며, 라울의 팔에 붕대가 감겨 있음.",
    "hard_violations": [],
    "physics": "앰버가 좌석 끝에 걸터앉아 다리를 통로로 내린 자세가 주변 공간 및 라울과의 크기 비례상 다소 부자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "양쪽 좌석과 중앙 통로를 활용한 균형 잡힌 구도 속에서 인물들의 지친 상태와 젖은 모습을 매우 사실적으로 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물들이 한쪽 좌석에 몰려 있어 공간감이 떨어지며, 앰버의 앉은 자세와 신체 비례가 공간에 비해 다소 어색함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 왼쪽 창밖을, 라울은 오른쪽 창밖을 응시함.",
        "built_space": "캠핑카 내부. 전경에 좌석 등받이가 있고 중앙 통로가 앞쪽으로 이어지며 양쪽 좌석에 인물이 자연스럽게 배치됨.",
        "entities": "앰버(금발, 젖은 머리), 라울(꽁지머리, 젖은 머리, 팔에 반창고) 모두 지친 기색이 뚜렷하며 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "앰버는 손으로 턱을 괴고 기댔으며, 라울도 팔을 등받이에 걸치고 안정적으로 좌석에 앉아 있음."
       },
       {
        "label": "B",
        "direction": "앰버와 라울 모두 오른쪽 창밖을 응시함.",
        "built_space": "캠핑카 내부. 통로와 전경 좌석이 있으나, 두 인물이 오른쪽 좌석에만 몰려 앉아 배치와 공간감이 다소 답답함.",
        "entities": "앰버와 라울의 외형은 레퍼런스를 따르며, 라울의 팔에 붕대가 감겨 있음.",
        "hard_violations": [],
        "physics": "앰버가 좌석 끝에 걸터앉아 다리를 통로로 내린 자세가 주변 공간 및 라울과의 크기 비례상 다소 부자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "양쪽 좌석과 중앙 통로를 활용한 균형 잡힌 구도 속에서 인물들의 지친 상태와 젖은 모습을 매우 사실적으로 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물들이 한쪽 좌석에 몰려 있어 공간감이 떨어지며, 앰버의 앉은 자세와 신체 비례가 공간에 비해 다소 어색함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 왼쪽 창밖을, 라울은 오른쪽 창밖을 응시함.",
        "built_space": "캠핑카 내부. 전경에 좌석 등받이가 있고 중앙 통로가 앞쪽으로 이어지며 양쪽 좌석에 인물이 자연스럽게 배치됨.",
        "entities": "앰버(금발, 젖은 머리), 라울(꽁지머리, 젖은 머리, 팔에 반창고) 모두 지친 기색이 뚜렷하며 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "앰버는 손으로 턱을 괴고 기댔으며, 라울도 팔을 등받이에 걸치고 안정적으로 좌석에 앉아 있음."
       },
       {
        "label": "B",
        "direction": "앰버와 라울 모두 오른쪽 창밖을 응시함.",
        "built_space": "캠핑카 내부. 통로와 전경 좌석이 있으나, 두 인물이 오른쪽 좌석에만 몰려 앉아 배치와 공간감이 다소 답답함.",
        "entities": "앰버와 라울의 외형은 레퍼런스를 따르며, 라울의 팔에 붕대가 감겨 있음.",
        "hard_violations": [],
        "physics": "앰버가 좌석 끝에 걸터앉아 다리를 통로로 내린 자세가 주변 공간 및 라울과의 크기 비례상 다소 부자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "창밖을 보는 행동과 장소는 맞지만, 인물을 오른쪽에 작게 몰아놓고 다리까지 보여 주는 넓은 구도가 지정된 미디엄 숏에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 아이의 상반신과 각 측면 창을 향한 시선, 전방으로 이어지는 통로가 지정 구도에 더 충실하지만, 전경 등받이가 다소 높고 앰버의 참조 마스크는 없습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버와 라울 모두 고개와 눈을 화면 오른쪽으로 돌려 오른쪽 측면 창 너머 도로변을 봅니다. 서로나 카메라를 보는 것은 아니며, 창밖을 내다보라는 행동은 맞습니다. 정지 화면만으로 차량의 이동은 확인하기 어렵습니다.",
        "built_space": "중앙 통로 양쪽으로 전경 등받이 두 개와 그 너머 등받이 두 개, 창 아래 측면 쿠션이 보이고, 전방에는 운전석과 조수석 두 개가 있습니다. 좌우 큰 측면 창, 오른쪽 앞의 작은 창, 목재 상부장, 커튼, 천창과 운전실 위 침상이 참조 장소에 대응합니다. 두 아이는 오른쪽 좌석에 함께 있으며 왼쪽 좌석은 비어 있습니다. 통로는 화면 아래에서 운전실로 이어지지만, 실내 전체가 크게 드러나고 아이들은 상대적으로 작아 미디엄 숏보다 넓습니다. 창밖은 직접 보이는 풍경이며 부자연스러운 반사는 없습니다.",
        "entities": "등장인물은 어린 여자아이와 남자아이 두 명뿐입니다. 앰버의 밝은 금발, 밝은 피부, 큰 눈, 녹색 재킷과 회색 원피스는 참조와 대체로 맞지만 흰 마스크는 없습니다. 라울은 짙은 피부와 어린 얼굴, 뒤로 묶은 곱슬머리, 낡은 올리브색 티셔츠로 참조에 대체로 부합합니다. 지정된 혼혈 배경 자체를 외모만으로 확정할 수는 없습니다. 두 사람의 젖은 머리와 지친 표정이 보이고 라울의 위팔 붕대는 치료 상태를 나타냅니다. 가방, 수납 상자와 금속 용기가 보이나 식품과 약품 내용물은 식별되지 않습니다. 빗방울 맺힌 창과 젖은 지방도로, 낮빛이 맞으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 아이의 몸은 오른쪽 좌석 쿠션에 지지됩니다. 앰버는 통로 쪽 가장자리에 앉아 한쪽 다리를 아래로 내리고 다른 쪽 다리를 굽힌 모습이며, 라울의 하체는 전경 등받이에 가려집니다. 밀집된 자세이지만 몸이 공중에 뜨거나 좌석을 관통한다고 단정할 근거는 없습니다. 가방과 용기들은 수납장 및 받침면 위에 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "앰버는 화면 왼쪽 측면 창 너머 가드레일과 풍경을 보고, 라울은 오른쪽 측면 창 너머 도로변을 봅니다. 두 시선 모두 실제 창밖을 향하며 카메라나 반사상으로 향하지 않습니다. 차량 이동 자체는 정지 화면에서 뚜렷하게 드러나지 않습니다.",
        "built_space": "중앙 통로 양쪽에 전경 등받이 두 개와 그 너머 등받이 두 개, 창 아래 측면 쿠션이 보이고, 전방 운전석과 조수석도 두 개입니다. 좌우 큰 창, 오른쪽 앞 작은 창, 커튼, 목재 상부장, 천창과 운전실 위 침상이 참조의 배치와 재질을 유지합니다. 앰버는 왼쪽, 라울은 오른쪽 뒷좌석에 각각 앉아 창 쪽으로 몸을 돌립니다. 통로가 아래 중앙에서 운전실로 이어지고 두 아이가 상반신 중심의 크기로 잡혀 A보다 미디엄 숏에 가깝습니다. 다만 전경 등받이가 화면 하단을 상당히 차지해 요구한 낮은 경계보다는 높습니다. 불가능한 반사는 보이지 않습니다.",
        "entities": "어린 여자아이 앰버와 어린 남자아이 라울만 보입니다. 앰버의 젖은 금발, 큰 눈, 밝은 피부, 녹색 재킷과 회색 안쪽 옷은 참조와 대체로 맞지만 흰 마스크는 빠졌습니다. 라울의 짙은 피부, 어린 얼굴 윤곽, 뒤로 묶은 곱슬머리와 낡은 올리브색 티셔츠도 참조에 대체로 맞습니다. 외모만으로 지정된 혼혈 배경을 확정할 수는 없습니다. 두 아이의 피곤한 표정과 축축한 머리가 보이고 라울의 위팔 반창고는 치료 상태를 표현합니다. 가방과 상자, 금속 보관 용기는 있으나 식품·약품의 내용물은 확인되지 않습니다. 비가 그친 낮의 젖은 도로와 창의 물방울이 보이며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 아이의 하체는 대부분 가려져 있지만 몸통은 각 좌석 위에 자연스럽게 자리합니다. 앰버는 좌석 측면에 팔꿈치를 대고 손으로 머리를 받치며, 라울은 창 아래 쿠션에 팔을 걸쳐 지지합니다. 창밖을 보려고 몸을 돌린 자세로 성립하며 공중에 뜬 신체는 없습니다. 짐과 금속 용기도 받침면에 놓여 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "창밖을 보는 행동과 장소는 맞지만, 인물을 오른쪽에 작게 몰아놓고 다리까지 보여 주는 넓은 구도가 지정된 미디엄 숏에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 아이의 상반신과 각 측면 창을 향한 시선, 전방으로 이어지는 통로가 지정 구도에 더 충실하지만, 전경 등받이가 다소 높고 앰버의 참조 마스크는 없습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버와 라울 모두 고개와 눈을 화면 오른쪽으로 돌려 오른쪽 측면 창 너머 도로변을 봅니다. 서로나 카메라를 보는 것은 아니며, 창밖을 내다보라는 행동은 맞습니다. 정지 화면만으로 차량의 이동은 확인하기 어렵습니다.",
        "built_space": "중앙 통로 양쪽으로 전경 등받이 두 개와 그 너머 등받이 두 개, 창 아래 측면 쿠션이 보이고, 전방에는 운전석과 조수석 두 개가 있습니다. 좌우 큰 측면 창, 오른쪽 앞의 작은 창, 목재 상부장, 커튼, 천창과 운전실 위 침상이 참조 장소에 대응합니다. 두 아이는 오른쪽 좌석에 함께 있으며 왼쪽 좌석은 비어 있습니다. 통로는 화면 아래에서 운전실로 이어지지만, 실내 전체가 크게 드러나고 아이들은 상대적으로 작아 미디엄 숏보다 넓습니다. 창밖은 직접 보이는 풍경이며 부자연스러운 반사는 없습니다.",
        "entities": "등장인물은 어린 여자아이와 남자아이 두 명뿐입니다. 앰버의 밝은 금발, 밝은 피부, 큰 눈, 녹색 재킷과 회색 원피스는 참조와 대체로 맞지만 흰 마스크는 없습니다. 라울은 짙은 피부와 어린 얼굴, 뒤로 묶은 곱슬머리, 낡은 올리브색 티셔츠로 참조에 대체로 부합합니다. 지정된 혼혈 배경 자체를 외모만으로 확정할 수는 없습니다. 두 사람의 젖은 머리와 지친 표정이 보이고 라울의 위팔 붕대는 치료 상태를 나타냅니다. 가방, 수납 상자와 금속 용기가 보이나 식품과 약품 내용물은 식별되지 않습니다. 빗방울 맺힌 창과 젖은 지방도로, 낮빛이 맞으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 아이의 몸은 오른쪽 좌석 쿠션에 지지됩니다. 앰버는 통로 쪽 가장자리에 앉아 한쪽 다리를 아래로 내리고 다른 쪽 다리를 굽힌 모습이며, 라울의 하체는 전경 등받이에 가려집니다. 밀집된 자세이지만 몸이 공중에 뜨거나 좌석을 관통한다고 단정할 근거는 없습니다. 가방과 용기들은 수납장 및 받침면 위에 놓여 있습니다."
       },
       {
        "label": "A",
        "direction": "앰버는 화면 왼쪽 측면 창 너머 가드레일과 풍경을 보고, 라울은 오른쪽 측면 창 너머 도로변을 봅니다. 두 시선 모두 실제 창밖을 향하며 카메라나 반사상으로 향하지 않습니다. 차량 이동 자체는 정지 화면에서 뚜렷하게 드러나지 않습니다.",
        "built_space": "중앙 통로 양쪽에 전경 등받이 두 개와 그 너머 등받이 두 개, 창 아래 측면 쿠션이 보이고, 전방 운전석과 조수석도 두 개입니다. 좌우 큰 창, 오른쪽 앞 작은 창, 커튼, 목재 상부장, 천창과 운전실 위 침상이 참조의 배치와 재질을 유지합니다. 앰버는 왼쪽, 라울은 오른쪽 뒷좌석에 각각 앉아 창 쪽으로 몸을 돌립니다. 통로가 아래 중앙에서 운전실로 이어지고 두 아이가 상반신 중심의 크기로 잡혀 A보다 미디엄 숏에 가깝습니다. 다만 전경 등받이가 화면 하단을 상당히 차지해 요구한 낮은 경계보다는 높습니다. 불가능한 반사는 보이지 않습니다.",
        "entities": "어린 여자아이 앰버와 어린 남자아이 라울만 보입니다. 앰버의 젖은 금발, 큰 눈, 밝은 피부, 녹색 재킷과 회색 안쪽 옷은 참조와 대체로 맞지만 흰 마스크는 빠졌습니다. 라울의 짙은 피부, 어린 얼굴 윤곽, 뒤로 묶은 곱슬머리와 낡은 올리브색 티셔츠도 참조에 대체로 맞습니다. 외모만으로 지정된 혼혈 배경을 확정할 수는 없습니다. 두 아이의 피곤한 표정과 축축한 머리가 보이고 라울의 위팔 반창고는 치료 상태를 표현합니다. 가방과 상자, 금속 보관 용기는 있으나 식품·약품의 내용물은 확인되지 않습니다. 비가 그친 낮의 젖은 도로와 창의 물방울이 보이며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 아이의 하체는 대부분 가려져 있지만 몸통은 각 좌석 위에 자연스럽게 자리합니다. 앰버는 좌석 측면에 팔꿈치를 대고 손으로 머리를 받치며, 라울은 창 아래 쿠션에 팔을 걸쳐 지지합니다. 창밖을 보려고 몸을 돌린 자세로 성립하며 공중에 뜬 신체는 없습니다. 짐과 금속 용기도 받침면에 놓여 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.25
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.25
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "양쪽 좌석과 중앙 통로를 활용한 균형 잡힌 구도 속에서 인물들의 지친 상태와 젖은 모습을 매우 사실적으로 구현함."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "인물들이 한쪽 좌석에 몰려 있어 공간감이 떨어지며, 앰버의 앉은 자세와 신체 비례가 공간에 비해 다소 어색함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_camper_living_809ab7.png",
    "asset_id": "0a075c81-0701-4544-9b02-0334a12fab36",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:943979>",
    "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde3-1aa9-7a2c-9874-0e8d2f4b01b7",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S55sh1__bgfirst_bg.png",
   "bg_asset_id": "f43f0854-763c-4679-8ff3-b631473172d6",
   "bg_record_key": "S55sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "camper_living",
   "groupbg_asset_id": "0a075c81-0701-4544-9b02-0334a12fab36"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S55sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:46:09.824827+00:00",
  "fingerprint": "808c5eea3048df3fb38c078de6b1c79dd3dafd99cce12ed75600929e8af2ccab",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S55sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S55sh1_sel.png",
  "source_sha256": "d3cf58ef98e6690a3d232ca5dd494053203a5de75fb4f9547201969ddbd63444",
  "file": "S55sh1_cine.png",
  "staged_sha256": "1af40111d5beb297f2062a7a098fded402e368a0c223252c7d3ea157d8288fe6",
  "latency_ms": 11001
 },
 "S55sh6::confined_fp_apt": {
  "applies": true,
  "reason_ko": "캠핑카 운전석 내부라는 제한되고 조작 장치가 중요한 공간에서 진행되며, 운전대를 잡고 있는 주인공의 위치와 차량 조작부의 정확한 배치가 샷의 설득력에 필수적이기 때문입니다.",
  "input_fingerprint": "b32ba3fc90c35830"
 },
 "S55sh6::signage": {
  "fp": "7724202a0b823c45",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::dd456e722f78": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/confinedfp_base_dd456e722f78.png",
  "place_text": "Inside the camper's compact driving cab, with the steering wheel and controls close to the driver's seat and daylight through the windshield.",
  "input_fingerprint": "d71f67c099392947"
 },
 "S55sh6::confined_fp": {
  "reads": {
   "controls": "A steering wheel is located in front of the driver's seat on the left.",
   "mirrors": "No mirrors are present in the diagram.",
   "camera": "The camera is positioned behind the center gap between the two front seats, facing straight forward toward the windshield.",
   "occupants": "이현우 occupies the driver's seat on the left."
  },
  "mismatches": [],
  "scene_description_en": "The camera is positioned in the center rear of the cab, looking straight forward. In the left midground, the back and right side of the driver's seat are visible, occupied by 이현우 facing away from the camera. Directly in front of him on the left sits the steering wheel. In the right midground is the back of the unoccupied passenger seat. A wide windshield spans the entire background across the top of the frame. No mirrors are present in this view.",
  "fixed": false,
  "input_fingerprint": "99cd23dc91053751"
 },
 "S55sh6": {
  "input_fingerprint": "21facabfdb7531d4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day, after a rain shower.\n\nSHOT TEXT (authoritative, Korean): 이현우가 운전대를 두 손으로 꽉 쥔 채 옆으로 거칠게 꺾은 상태로 멈춘 mid-action 구도.\n\nLOCATION (lock): Inside the camper's compact driving cab, with the steering wheel and controls close to the driver's seat and daylight through the windshield. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Steering wheel (Sharply turned and tightly gripped with both hands) — Its driver-facing surface is seen obliquely from above and behind; used as Makes the decisive change of direction legible beside his torso; Driver's seat (Occupied by 이현우 during the turn) — A partial rear and passenger-side view borders his body; used as Maintains scale and continuity with the aisle approach.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established daytime window light and restrained contrast across his tense face and gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper still carries the luggage, food and medicine as it turns toward the forest after the rain shower. 이현우: He remains at the wheel, wet and exhausted, with his treated leg bandaged. The contact card remains concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned in the center rear of the cab, looking straight forward. In the left midground, the back and right side of the driver's seat are visible, occupied by 이현우 facing away from the camera. Directly in front of him on the left sits the steering wheel. In the right midground is the back of the unoccupied passenger seat. A wide windshield spans the entire background across the top of the frame. No mirrors are present in this view.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day, after a rain shower.\n\nSHOT TEXT (authoritative, Korean): 이현우가 운전대를 두 손으로 꽉 쥔 채 옆으로 거칠게 꺾은 상태로 멈춘 mid-action 구도.\n\nLOCATION (lock): Inside the camper's compact driving cab, with the steering wheel and controls close to the driver's seat and daylight through the windshield. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established daytime window light and restrained contrast across his tense face and gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper still carries the luggage, food and medicine as it turns toward the forest after the rain shower. 이현우: He remains at the wheel, wet and exhausted, with his treated leg bandaged. The contact card remains concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned in the center rear of the cab, looking straight forward. In the left midground, the back and right side of the driver's seat are visible, occupied by 이현우 facing away from the camera. Directly in front of him on the left sits the steering wheel. In the right midground is the back of the unoccupied passenger seat. A wide windshield spans the entire background across the top of the frame. No mirrors are present in this view.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day, after a rain shower.\n\nSHOT TEXT (authoritative, Korean): 이현우가 운전대를 두 손으로 꽉 쥔 채 옆으로 거칠게 꺾은 상태로 멈춘 mid-action 구도.\n\nLOCATION (lock): Inside the camper's compact driving cab, with the steering wheel and controls close to the driver's seat and daylight through the windshield. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established daytime window light and restrained contrast across his tense face and gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper still carries the luggage, food and medicine as it turns toward the forest after the rain shower. 이현우: He remains at the wheel, wet and exhausted, with his treated leg bandaged. The contact card remains concealed in his shoe.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S55sh6_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S55sh6_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "머리와 시선은 앞유리 너머 젖은 도로를 향한다. 왼손은 운전대 상단, 오른손은 오른쪽 테두리를 잡고 있다. 운전대의 운전자 쪽 면이 뒤쪽 위에서 비스듬히 보이지만, 손과 바퀴살의 배치에서 급격한 회전은 명확하지 않다.",
    "built_space": "앞유리 하나, 앞좌석 두 개와 각각의 머리받침, 왼쪽 운전대 하나, 중앙 조작부와 변속 레버가 보인다. 이현우는 왼쪽 운전석에 있고 오른쪽 좌석은 비어 있다. 카메라는 두 좌석 뒤 통로 쪽에 있어 배치도와 맞으며, 운전석의 뒤와 조수석 쪽 측면이 인물을 둘러싼다. 다만 조수석과 도로가 상당한 면적을 차지하여 인물 중심의 미디엄 샷보다 다소 넓게 읽힌다.",
    "entities": "사람은 한 명뿐이다. 검고 젖은 듯 헝클어진 머리, 마른 체격, 해진 짙은 후드와 청바지가 인물 참조와 맞는다. 얼굴 대부분이 가려져 한국계 얼굴의 세부 동일성과 정확한 나이는 확인하기 어렵다. 운전대와 가까운 조작부는 확인되며 창의 물방울, 젖은 도로와 숲, 낮빛은 비 온 뒤 낮 설정에 부합한다. 붕대와 신발 속 카드, 짐·식량·약은 이 구도에서 확인되지 않는다. 읽을 수 있는 문구는 식별되지 않는다.",
    "hard_violations": [],
    "physics": "몸은 운전석 좌판에 앉아 지지되고 등은 등받이 앞에 있다. 두 손의 손가락이 실제 운전대 테두리를 감싸며 팔도 몸통에 자연스럽게 연결된다. 운전대는 조향축에 연결되어 있고 공중에 지지 없이 떠 있는 물체는 없다. 가능한 운전 자세이지만 거친 조향 중 정지한 몸의 비틀림과 힘의 전달은 약하다."
   },
   {
    "label": "A",
    "direction": "머리는 앞유리 쪽을 향하고 약간 아래로 숙여져 있다. 왼손은 운전대 상단을 잡지만 오른손은 중앙 송풍구 위에서 빈 주먹을 쥐고 있으며 운전대에 닿지 않는다. 따라서 양손이 같은 운전대를 꺾는 방향 관계가 성립하지 않는다.",
    "built_space": "앞유리 하나, 머리받침이 있는 앞좌석 두 개, 왼쪽 운전대 하나, 중앙 조작부와 변속 레버가 보인다. 왼쪽 운전석에 한 명이 앉고 조수석은 비어 있으며 후방 통로 시점도 배치와 맞는다. 인물 상체와 팔을 더 가깝게 담고 운전석 뒤쪽이 화면을 경계 짓는다. 운전대의 운전자 쪽 면은 맞는 방향으로 보인다.",
    "entities": "한 명의 젊은 남성으로 보이며 검은 헝클어진 머리와 해진 짙은 후드, 청바지는 참조와 대체로 맞는다. 뒷모습 중심이어서 얼굴 동일성이나 정확한 나이는 검증하기 어렵다. 무릎 부근에는 붕대가 보인다. 창의 물방울과 낮빛은 있으나 바깥은 흐린 건물과 식생으로 보여 숲으로 향하는 상황은 A보다 불명확하다. 카드와 적재 물자는 보이지 않으며 읽을 수 있는 문구는 식별되지 않는다.",
    "hard_violations": [],
    "physics": "몸과 다리는 운전석에 의해 지지되며 왼손은 조향축에 연결된 운전대를 잡는다. 오른손은 팔꿈치를 굽힌 팔로 지지되므로 부유하거나 해부학적으로 불가능한 것은 아니다. 그러나 그 손에는 잡은 물체가 없고 운전대와 간격이 있어, 두 손으로 힘껏 운전대를 꺾는 동작으로는 성립하지 않는다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "후방 사선 시점과 양손의 실제 운전대 접촉은 충실하지만, 운전대를 거칠게 꺾어 멈춘 순간의 회전량과 긴장감은 약하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물에 가까운 구도는 적절하지만 오른손이 운전대에서 완전히 떨어져 있어, 양손으로 꽉 쥔 급조향이라는 핵심 동작을 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리와 시선은 앞유리 너머 젖은 도로를 향한다. 왼손은 운전대 상단, 오른손은 오른쪽 테두리를 잡고 있다. 운전대의 운전자 쪽 면이 뒤쪽 위에서 비스듬히 보이지만, 손과 바퀴살의 배치에서 급격한 회전은 명확하지 않다.",
        "built_space": "앞유리 하나, 앞좌석 두 개와 각각의 머리받침, 왼쪽 운전대 하나, 중앙 조작부와 변속 레버가 보인다. 이현우는 왼쪽 운전석에 있고 오른쪽 좌석은 비어 있다. 카메라는 두 좌석 뒤 통로 쪽에 있어 배치도와 맞으며, 운전석의 뒤와 조수석 쪽 측면이 인물을 둘러싼다. 다만 조수석과 도로가 상당한 면적을 차지하여 인물 중심의 미디엄 샷보다 다소 넓게 읽힌다.",
        "entities": "사람은 한 명뿐이다. 검고 젖은 듯 헝클어진 머리, 마른 체격, 해진 짙은 후드와 청바지가 인물 참조와 맞는다. 얼굴 대부분이 가려져 한국계 얼굴의 세부 동일성과 정확한 나이는 확인하기 어렵다. 운전대와 가까운 조작부는 확인되며 창의 물방울, 젖은 도로와 숲, 낮빛은 비 온 뒤 낮 설정에 부합한다. 붕대와 신발 속 카드, 짐·식량·약은 이 구도에서 확인되지 않는다. 읽을 수 있는 문구는 식별되지 않는다.",
        "hard_violations": [],
        "physics": "몸은 운전석 좌판에 앉아 지지되고 등은 등받이 앞에 있다. 두 손의 손가락이 실제 운전대 테두리를 감싸며 팔도 몸통에 자연스럽게 연결된다. 운전대는 조향축에 연결되어 있고 공중에 지지 없이 떠 있는 물체는 없다. 가능한 운전 자세이지만 거친 조향 중 정지한 몸의 비틀림과 힘의 전달은 약하다."
       },
       {
        "label": "B",
        "direction": "머리는 앞유리 쪽을 향하고 약간 아래로 숙여져 있다. 왼손은 운전대 상단을 잡지만 오른손은 중앙 송풍구 위에서 빈 주먹을 쥐고 있으며 운전대에 닿지 않는다. 따라서 양손이 같은 운전대를 꺾는 방향 관계가 성립하지 않는다.",
        "built_space": "앞유리 하나, 머리받침이 있는 앞좌석 두 개, 왼쪽 운전대 하나, 중앙 조작부와 변속 레버가 보인다. 왼쪽 운전석에 한 명이 앉고 조수석은 비어 있으며 후방 통로 시점도 배치와 맞는다. 인물 상체와 팔을 더 가깝게 담고 운전석 뒤쪽이 화면을 경계 짓는다. 운전대의 운전자 쪽 면은 맞는 방향으로 보인다.",
        "entities": "한 명의 젊은 남성으로 보이며 검은 헝클어진 머리와 해진 짙은 후드, 청바지는 참조와 대체로 맞는다. 뒷모습 중심이어서 얼굴 동일성이나 정확한 나이는 검증하기 어렵다. 무릎 부근에는 붕대가 보인다. 창의 물방울과 낮빛은 있으나 바깥은 흐린 건물과 식생으로 보여 숲으로 향하는 상황은 A보다 불명확하다. 카드와 적재 물자는 보이지 않으며 읽을 수 있는 문구는 식별되지 않는다.",
        "hard_violations": [],
        "physics": "몸과 다리는 운전석에 의해 지지되며 왼손은 조향축에 연결된 운전대를 잡는다. 오른손은 팔꿈치를 굽힌 팔로 지지되므로 부유하거나 해부학적으로 불가능한 것은 아니다. 그러나 그 손에는 잡은 물체가 없고 운전대와 간격이 있어, 두 손으로 힘껏 운전대를 꺾는 동작으로는 성립하지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "후방 사선 시점과 양손의 실제 운전대 접촉은 충실하지만, 운전대를 거칠게 꺾어 멈춘 순간의 회전량과 긴장감은 약하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물에 가까운 구도는 적절하지만 오른손이 운전대에서 완전히 떨어져 있어, 양손으로 꽉 쥔 급조향이라는 핵심 동작을 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "머리와 시선은 앞유리 너머 젖은 도로를 향한다. 왼손은 운전대 상단, 오른손은 오른쪽 테두리를 잡고 있다. 운전대의 운전자 쪽 면이 뒤쪽 위에서 비스듬히 보이지만, 손과 바퀴살의 배치에서 급격한 회전은 명확하지 않다.",
        "built_space": "앞유리 하나, 앞좌석 두 개와 각각의 머리받침, 왼쪽 운전대 하나, 중앙 조작부와 변속 레버가 보인다. 이현우는 왼쪽 운전석에 있고 오른쪽 좌석은 비어 있다. 카메라는 두 좌석 뒤 통로 쪽에 있어 배치도와 맞으며, 운전석의 뒤와 조수석 쪽 측면이 인물을 둘러싼다. 다만 조수석과 도로가 상당한 면적을 차지하여 인물 중심의 미디엄 샷보다 다소 넓게 읽힌다.",
        "entities": "사람은 한 명뿐이다. 검고 젖은 듯 헝클어진 머리, 마른 체격, 해진 짙은 후드와 청바지가 인물 참조와 맞는다. 얼굴 대부분이 가려져 한국계 얼굴의 세부 동일성과 정확한 나이는 확인하기 어렵다. 운전대와 가까운 조작부는 확인되며 창의 물방울, 젖은 도로와 숲, 낮빛은 비 온 뒤 낮 설정에 부합한다. 붕대와 신발 속 카드, 짐·식량·약은 이 구도에서 확인되지 않는다. 읽을 수 있는 문구는 식별되지 않는다.",
        "hard_violations": [],
        "physics": "몸은 운전석 좌판에 앉아 지지되고 등은 등받이 앞에 있다. 두 손의 손가락이 실제 운전대 테두리를 감싸며 팔도 몸통에 자연스럽게 연결된다. 운전대는 조향축에 연결되어 있고 공중에 지지 없이 떠 있는 물체는 없다. 가능한 운전 자세이지만 거친 조향 중 정지한 몸의 비틀림과 힘의 전달은 약하다."
       },
       {
        "label": "A",
        "direction": "머리는 앞유리 쪽을 향하고 약간 아래로 숙여져 있다. 왼손은 운전대 상단을 잡지만 오른손은 중앙 송풍구 위에서 빈 주먹을 쥐고 있으며 운전대에 닿지 않는다. 따라서 양손이 같은 운전대를 꺾는 방향 관계가 성립하지 않는다.",
        "built_space": "앞유리 하나, 머리받침이 있는 앞좌석 두 개, 왼쪽 운전대 하나, 중앙 조작부와 변속 레버가 보인다. 왼쪽 운전석에 한 명이 앉고 조수석은 비어 있으며 후방 통로 시점도 배치와 맞는다. 인물 상체와 팔을 더 가깝게 담고 운전석 뒤쪽이 화면을 경계 짓는다. 운전대의 운전자 쪽 면은 맞는 방향으로 보인다.",
        "entities": "한 명의 젊은 남성으로 보이며 검은 헝클어진 머리와 해진 짙은 후드, 청바지는 참조와 대체로 맞는다. 뒷모습 중심이어서 얼굴 동일성이나 정확한 나이는 검증하기 어렵다. 무릎 부근에는 붕대가 보인다. 창의 물방울과 낮빛은 있으나 바깥은 흐린 건물과 식생으로 보여 숲으로 향하는 상황은 A보다 불명확하다. 카드와 적재 물자는 보이지 않으며 읽을 수 있는 문구는 식별되지 않는다.",
        "hard_violations": [],
        "physics": "몸과 다리는 운전석에 의해 지지되며 왼손은 조향축에 연결된 운전대를 잡는다. 오른손은 팔꿈치를 굽힌 팔로 지지되므로 부유하거나 해부학적으로 불가능한 것은 아니다. 그러나 그 손에는 잡은 물체가 없고 운전대와 간격이 있어, 두 손으로 힘껏 운전대를 꺾는 동작으로는 성립하지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "후방 사선 시점과 양손의 실제 운전대 접촉은 충실하지만, 운전대를 거칠게 꺾어 멈춘 순간의 회전량과 긴장감은 약하다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "인물에 가까운 구도는 적절하지만 오른손이 운전대에서 완전히 떨어져 있어, 양손으로 꽉 쥔 급조향이라는 핵심 동작을 놓쳤다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S55sh6_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "이현우",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde3-d278-7705-9cda-ecbfb9009510",
  "confined_fp": {
   "base_key": "confinedfp::dd456e722f78",
   "apt_reason": "캠핑카 운전석 내부라는 제한되고 조작 장치가 중요한 공간에서 진행되며, 운전대를 잡고 있는 주인공의 위치와 차량 조작부의 정확한 배치가 샷의 설득력에 필수적이기 때문입니다.",
   "fixed": false,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S55sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:46:21.235075+00:00",
  "fingerprint": "4cbb55819909d6605afb97b18d3b5f8bc19be3bf38cec780d3feb84194442c23",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S55sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S55sh6_sel.png",
  "source_sha256": "027f3c7c86adb22e3579316bd8f63f2e98066daaef85ae20353bc6cc9b520ba9",
  "file": "S55sh6_cine.png",
  "staged_sha256": "c4a2c6ef10bd0ca93d44979d80f2f7b185a39b388757e55f7aab80482b670ce6",
  "latency_ms": 11197
 },
 "S55sh8::signage": {
  "fp": "b7b82368bcf496d9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S55sh8": {
  "input_fingerprint": "51eb7768eb7c9527",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day, after a rain shower.\n\nSHOT TEXT (authoritative, Korean): 캠핑카의 앞바퀴가 질퍽한 늪지대 진흙 속에 푹 빠진 상태에서, 흙탕물이 공중으로 튀어 올라 멈춰 있는 타격의 순간.\n\nLOCATION (lock): At the muddy end of a forest track where the camper's front wheels sink into a marsh. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper front wheel (Deeply sunk in soft mud at the instant the vehicle stops) — Outer side and ground-contact region are visible from ahead of the axle; used as Primary action detail occupying roughly a quarter of the frame; Adjacent camper body (Stopped above the trapped wheel) — Only a narrow oblique portion of the lower side is visible; used as Provides realistic vehicle scale without overpowering the ground; Mud and splashing muddy water (Ground displaced around the sunken wheel, with muddy water still airborne); used as Spreads the impact outward through the lower and central frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight preserves clear separation between the wheel, muddy ground, and airborne muddy water for a blunt mechanical impact.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper has reached the rain-soaked marsh and is becoming stuck in the mud, with its luggage, food and medicine still aboard. No recovery has yet occurred.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day, after a rain shower.\n\nSHOT TEXT (authoritative, Korean): 캠핑카의 앞바퀴가 질퍽한 늪지대 진흙 속에 푹 빠진 상태에서, 흙탕물이 공중으로 튀어 올라 멈춰 있는 타격의 순간.\n\nLOCATION (lock): At the muddy end of a forest track where the camper's front wheels sink into a marsh. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper front wheel (Deeply sunk in soft mud at the instant the vehicle stops) — Outer side and ground-contact region are visible from ahead of the axle; used as Primary action detail occupying roughly a quarter of the frame; Adjacent camper body (Stopped above the trapped wheel) — Only a narrow oblique portion of the lower side is visible; used as Provides realistic vehicle scale without overpowering the ground; Mud and splashing muddy water (Ground displaced around the sunken wheel, with muddy water still airborne); used as Spreads the impact outward through the lower and central frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight preserves clear separation between the wheel, muddy ground, and airborne muddy water for a blunt mechanical impact.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper has reached the rain-soaked marsh and is becoming stuck in the mud, with its luggage, food and medicine still aboard. No recovery has yet occurred.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day, after a rain shower.\n\nSHOT TEXT (authoritative, Korean): 캠핑카의 앞바퀴가 질퍽한 늪지대 진흙 속에 푹 빠진 상태에서, 흙탕물이 공중으로 튀어 올라 멈춰 있는 타격의 순간.\n\nLOCATION (lock): At the muddy end of a forest track where the camper's front wheels sink into a marsh. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper front wheel (Deeply sunk in soft mud at the instant the vehicle stops) — Outer side and ground-contact region are visible from ahead of the axle; used as Primary action detail occupying roughly a quarter of the frame; Adjacent camper body (Stopped above the trapped wheel) — Only a narrow oblique portion of the lower side is visible; used as Provides realistic vehicle scale without overpowering the ground; Mud and splashing muddy water (Ground displaced around the sunken wheel, with muddy water still airborne); used as Spreads the impact outward through the lower and central frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight preserves clear separation between the wheel, muddy ground, and airborne muddy water for a blunt mechanical impact.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper has reached the rain-soaked marsh and is becoming stuck in the mud, with its luggage, food and medicine still aboard. No recovery has yet occurred.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 차량 앞바퀴의 전방에서 후방을 향해 측면을 바라보고 있음.",
    "built_space": "비 온 뒤의 숲길 끝자락에 위치한 깊은 진흙탕.",
    "entities": "깊게 빠진 타이어, 튀어오르는 흙탕물, 차량 하단 패널 모두 지시사항과 일치함.",
    "hard_violations": [],
    "physics": "바퀴의 충격으로 인해 진흙과 물이 공중으로 튀어 오르며, 그 순간이 물리적으로 설득력 있게 멈춰 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 차량 앞쪽에서 측면을 따라 시선을 향함.",
    "built_space": "나무가 늘어선 진흙투성이의 숲길.",
    "entities": "바퀴와 캠핑카 차체가 보이나, 차체의 노출 범위가 지시사항보다 훨씬 넓음. 타이어 측면에 글자가 일부 보임.",
    "hard_violations": [],
    "physics": "공중으로 튀어 오른 진흙의 형태가 바퀴의 충격이라기보다는 다소 인위적인 덩어리처럼 보임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 프레이밍(바퀴가 화면의 1/4을 차지하고 차체 하단만 좁게 보임)을 완벽하게 준수하였으며, 튀어오르는 진흙의 물리적 표현이 자연스럽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "차체 하단만 좁게 보여야 한다는 프레이밍 지시를 어기고 범퍼와 차체 측면 전체를 너무 많이 노출하여 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 차량 앞바퀴의 전방에서 후방을 향해 측면을 바라보고 있음.",
        "built_space": "비 온 뒤의 숲길 끝자락에 위치한 깊은 진흙탕.",
        "entities": "깊게 빠진 타이어, 튀어오르는 흙탕물, 차량 하단 패널 모두 지시사항과 일치함.",
        "hard_violations": [],
        "physics": "바퀴의 충격으로 인해 진흙과 물이 공중으로 튀어 오르며, 그 순간이 물리적으로 설득력 있게 멈춰 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 차량 앞쪽에서 측면을 따라 시선을 향함.",
        "built_space": "나무가 늘어선 진흙투성이의 숲길.",
        "entities": "바퀴와 캠핑카 차체가 보이나, 차체의 노출 범위가 지시사항보다 훨씬 넓음. 타이어 측면에 글자가 일부 보임.",
        "hard_violations": [],
        "physics": "공중으로 튀어 오른 진흙의 형태가 바퀴의 충격이라기보다는 다소 인위적인 덩어리처럼 보임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 프레이밍(바퀴가 화면의 1/4을 차지하고 차체 하단만 좁게 보임)을 완벽하게 준수하였으며, 튀어오르는 진흙의 물리적 표현이 자연스럽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "차체 하단만 좁게 보여야 한다는 프레이밍 지시를 어기고 범퍼와 차체 측면 전체를 너무 많이 노출하여 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 차량 앞바퀴의 전방에서 후방을 향해 측면을 바라보고 있음.",
        "built_space": "비 온 뒤의 숲길 끝자락에 위치한 깊은 진흙탕.",
        "entities": "깊게 빠진 타이어, 튀어오르는 흙탕물, 차량 하단 패널 모두 지시사항과 일치함.",
        "hard_violations": [],
        "physics": "바퀴의 충격으로 인해 진흙과 물이 공중으로 튀어 오르며, 그 순간이 물리적으로 설득력 있게 멈춰 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 차량 앞쪽에서 측면을 따라 시선을 향함.",
        "built_space": "나무가 늘어선 진흙투성이의 숲길.",
        "entities": "바퀴와 캠핑카 차체가 보이나, 차체의 노출 범위가 지시사항보다 훨씬 넓음. 타이어 측면에 글자가 일부 보임.",
        "hard_violations": [],
        "physics": "공중으로 튀어 오른 진흙의 형태가 바퀴의 충격이라기보다는 다소 인위적인 덩어리처럼 보임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "앞바퀴의 매몰과 공중에 멈춘 흙탕물은 잘 구현했지만, 차체 측면과 뒷바퀴까지 크게 보여 하부 측면만 좁게 담으라는 구도에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "차체를 하부 일부로 제한하고 깊이 잠긴 앞바퀴와 중앙·하단으로 퍼지는 흙탕물에 집중해 지정된 충격 순간과 구도에 더 충실합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "차량 앞부분은 화면 왼쪽에 있고 측면은 오른쪽 뒤로 이어집니다. 앞바퀴 바깥면과 접지부를 차축 앞쪽의 낮은 사선에서 봅니다. 흙탕물은 바퀴 주변에서 위쪽과 화면 중앙·오른쪽으로 튀며, 발생 지점이 매몰된 바퀴와 연결됩니다. 사람의 시선이나 조준 물체는 없습니다.",
        "built_space": "앞바퀴와 휠 아치가 각각 하나씩 크게 보이고, 오른쪽 뒤에는 뒷바퀴와 그 휠 아치가 하나씩 보입니다. 앞 범퍼 일부와 여러 측면 패널을 포함한 상당한 차체가 상단을 차지합니다. 숲속 진흙길이라는 장소는 맞지만, 차체 하부의 좁은 일부만 보이라는 요구보다 차량을 많이 드러냅니다. 중복된 구조물이나 불가능한 반사는 없습니다.",
        "entities": "캠핑카 한 대, 진흙에 빠진 앞바퀴, 젖은 숲길, 고인 흙탕물과 공중의 물방울이 보입니다. 앞바퀴 하부는 진흙 둔덕에 묻혀 있지만 바퀴 대부분은 노출되어 있습니다. 사람이나 복구 장비는 없고 읽을 수 있는 글자도 보이지 않습니다. 짐·식량·의약품은 실내에 있어 이 구도로는 확인할 수 없습니다. 젖은 표면과 자연광은 비가 그친 낮에 부합합니다.",
        "hard_violations": [],
        "physics": "차량은 진흙에 박힌 앞바퀴와 뒤쪽 지면에 닿은 뒷바퀴로 지지됩니다. 흙탕물 막은 바퀴 옆 웅덩이에서 솟고, 떨어져 나온 물방울은 타이어가 진흙을 밀어낸 충격으로 설명됩니다. 무지지 상태의 차량이나 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "차량 앞쪽은 화면 왼쪽이며 차체 측면은 오른쪽으로 이어집니다. 낮은 사선 구도에서 앞바퀴 바깥면과 진흙에 잠긴 접지부가 보입니다. 흙탕물은 바퀴 앞쪽에서 왼쪽 위로 솟고 바퀴 옆에서는 오른쪽과 전경으로 퍼져, 충격의 중심이 바퀴 주변으로 읽힙니다. 시선이나 조준 방향을 확인할 사람은 없습니다.",
        "built_space": "앞바퀴 하나와 휠 아치 하나, 작은 측면 표시등 하나 및 차체 하부 패널 일부가 보입니다. 다른 바퀴나 차체 상부를 펼쳐 보이지 않아 차량보다 진흙 지면이 중심이 됩니다. 뒤에는 나무와 물이 고인 숲길이 이어집니다. 별도의 시설, 중복 부품 또는 불가능한 반사는 보이지 않습니다.",
        "entities": "캠핑카의 앞바퀴로 읽히는 바퀴 하나가 화면 중앙 오른쪽에 있고, 하부가 부드러운 진흙과 흙탕물에 깊이 잠겨 있습니다. 차량 전체 형식은 제한된 크롭 때문에 A보다 덜 명확하지만, 보이는 차체는 요구와 모순되지 않습니다. 진흙 덩어리와 공중의 흙탕물이 분명하며 사람·복구 장비·읽을 수 있는 문자는 없습니다. 실내 적재물은 프레임 밖입니다. 숲과 젖은 지면의 낮 자연광도 요구에 맞습니다.",
        "hard_violations": [],
        "physics": "차체는 휠 아치 아래 바퀴와 연결되어 있고, 바퀴는 진흙에 잠긴 채 지면의 지지를 받습니다. 물막과 물방울은 타이어 둘레의 웅덩이에서 튀어나와 충격에 따른 비산으로 설명됩니다. 타이어에는 약한 회전 흐림이 남아 완전히 정지한 상태의 선명함은 덜하지만, 막 멈추는 충격 순간으로 가능한 모습입니다. 근거 없이 공중에 뜬 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "앞바퀴의 매몰과 공중에 멈춘 흙탕물은 잘 구현했지만, 차체 측면과 뒷바퀴까지 크게 보여 하부 측면만 좁게 담으라는 구도에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "차체를 하부 일부로 제한하고 깊이 잠긴 앞바퀴와 중앙·하단으로 퍼지는 흙탕물에 집중해 지정된 충격 순간과 구도에 더 충실합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "차량 앞부분은 화면 왼쪽에 있고 측면은 오른쪽 뒤로 이어집니다. 앞바퀴 바깥면과 접지부를 차축 앞쪽의 낮은 사선에서 봅니다. 흙탕물은 바퀴 주변에서 위쪽과 화면 중앙·오른쪽으로 튀며, 발생 지점이 매몰된 바퀴와 연결됩니다. 사람의 시선이나 조준 물체는 없습니다.",
        "built_space": "앞바퀴와 휠 아치가 각각 하나씩 크게 보이고, 오른쪽 뒤에는 뒷바퀴와 그 휠 아치가 하나씩 보입니다. 앞 범퍼 일부와 여러 측면 패널을 포함한 상당한 차체가 상단을 차지합니다. 숲속 진흙길이라는 장소는 맞지만, 차체 하부의 좁은 일부만 보이라는 요구보다 차량을 많이 드러냅니다. 중복된 구조물이나 불가능한 반사는 없습니다.",
        "entities": "캠핑카 한 대, 진흙에 빠진 앞바퀴, 젖은 숲길, 고인 흙탕물과 공중의 물방울이 보입니다. 앞바퀴 하부는 진흙 둔덕에 묻혀 있지만 바퀴 대부분은 노출되어 있습니다. 사람이나 복구 장비는 없고 읽을 수 있는 글자도 보이지 않습니다. 짐·식량·의약품은 실내에 있어 이 구도로는 확인할 수 없습니다. 젖은 표면과 자연광은 비가 그친 낮에 부합합니다.",
        "hard_violations": [],
        "physics": "차량은 진흙에 박힌 앞바퀴와 뒤쪽 지면에 닿은 뒷바퀴로 지지됩니다. 흙탕물 막은 바퀴 옆 웅덩이에서 솟고, 떨어져 나온 물방울은 타이어가 진흙을 밀어낸 충격으로 설명됩니다. 무지지 상태의 차량이나 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "차량 앞쪽은 화면 왼쪽이며 차체 측면은 오른쪽으로 이어집니다. 낮은 사선 구도에서 앞바퀴 바깥면과 진흙에 잠긴 접지부가 보입니다. 흙탕물은 바퀴 앞쪽에서 왼쪽 위로 솟고 바퀴 옆에서는 오른쪽과 전경으로 퍼져, 충격의 중심이 바퀴 주변으로 읽힙니다. 시선이나 조준 방향을 확인할 사람은 없습니다.",
        "built_space": "앞바퀴 하나와 휠 아치 하나, 작은 측면 표시등 하나 및 차체 하부 패널 일부가 보입니다. 다른 바퀴나 차체 상부를 펼쳐 보이지 않아 차량보다 진흙 지면이 중심이 됩니다. 뒤에는 나무와 물이 고인 숲길이 이어집니다. 별도의 시설, 중복 부품 또는 불가능한 반사는 보이지 않습니다.",
        "entities": "캠핑카의 앞바퀴로 읽히는 바퀴 하나가 화면 중앙 오른쪽에 있고, 하부가 부드러운 진흙과 흙탕물에 깊이 잠겨 있습니다. 차량 전체 형식은 제한된 크롭 때문에 A보다 덜 명확하지만, 보이는 차체는 요구와 모순되지 않습니다. 진흙 덩어리와 공중의 흙탕물이 분명하며 사람·복구 장비·읽을 수 있는 문자는 없습니다. 실내 적재물은 프레임 밖입니다. 숲과 젖은 지면의 낮 자연광도 요구에 맞습니다.",
        "hard_violations": [],
        "physics": "차체는 휠 아치 아래 바퀴와 연결되어 있고, 바퀴는 진흙에 잠긴 채 지면의 지지를 받습니다. 물막과 물방울은 타이어 둘레의 웅덩이에서 튀어나와 충격에 따른 비산으로 설명됩니다. 타이어에는 약한 회전 흐림이 남아 완전히 정지한 상태의 선명함은 덜하지만, 막 멈추는 충격 순간으로 가능한 모습입니다. 근거 없이 공중에 뜬 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.46
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.46
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1460
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 프레이밍(바퀴가 화면의 1/4을 차지하고 차체 하단만 좁게 보임)을 완벽하게 준수하였으며, 튀어오르는 진흙의 물리적 표현이 자연스럽습니다."
   },
   {
    "label": "B",
    "score": 1460,
    "verdict_ko": "차체 하단만 좁게 보여야 한다는 프레이밍 지시를 어기고 범퍼와 차체 측면 전체를 너무 많이 노출하여 감점되었습니다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde4-8c2c-7226-8dd5-0e9a31eb4a52",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S55sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:46:32.769824+00:00",
  "fingerprint": "b682874a5f6a2ff8c5f7b29c36a799bbfb1df3d011404cbcf45e25c6da62e1a1",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S55sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S55sh8_sel.png",
  "source_sha256": "3196e6e423b1d647238bcded802f6917e0bde7aa6f3c662fce28df44bc864ece",
  "file": "S55sh8_cine.png",
  "staged_sha256": "76d3609d64966d6f0852db0fd1ea64be60b440d1ed5ba81e19153c02ce0a6e0b",
  "latency_ms": 10581
 },
 "S56sh12::signage": {
  "fp": "ab524ba9adbe709c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S56sh12": {
  "input_fingerprint": "b925efd0109767c6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 몸통 위로 두꺼운 밧줄 그물망이 맹렬히 덮쳐오는 공중의 순간.\n\nLOCATION (lock): On a wooded path near the marsh, beside flowering vegetation where a rope net drops onto the robot. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Thick rope net (Descending toward 찰리's upper body before settling around him) — Its open mesh crosses the view diagonally from above; used as Foreground interruption and visual transition, with 찰리 visible through its openings; Forest trees (Surround the forest-path location where 찰리 has been exploring); used as Peripheral environmental continuity behind the body and open mesh.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the metal body and thick rope distinct, allowing the threatening interruption to emerge from an otherwise gentle moment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains stuck in the marsh; nearby are large bootprints, a no-entry sign and a corroded, barely legible radiation warning. Elsewhere on the forest path, a capture net descends over Charlie, whose straw hat, colorful raincoat and oversized boots remain on.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 몸통 위로 두꺼운 밧줄 그물망이 맹렬히 덮쳐오는 공중의 순간.\n\nLOCATION (lock): On a wooded path near the marsh, beside flowering vegetation where a rope net drops onto the robot. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Thick rope net (Descending toward 찰리's upper body before settling around him) — Its open mesh crosses the view diagonally from above; used as Foreground interruption and visual transition, with 찰리 visible through its openings; Forest trees (Surround the forest-path location where 찰리 has been exploring); used as Peripheral environmental continuity behind the body and open mesh.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the metal body and thick rope distinct, allowing the threatening interruption to emerge from an otherwise gentle moment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains stuck in the marsh; nearby are large bootprints, a no-entry sign and a corroded, barely legible radiation warning. Elsewhere on the forest path, a capture net descends over Charlie, whose straw hat, colorful raincoat and oversized boots remain on.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 몸통 위로 두꺼운 밧줄 그물망이 맹렬히 덮쳐오는 공중의 순간.\n\nLOCATION (lock): On a wooded path near the marsh, beside flowering vegetation where a rope net drops onto the robot. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Thick rope net (Descending toward 찰리's upper body before settling around him) — Its open mesh crosses the view diagonally from above; used as Foreground interruption and visual transition, with 찰리 visible through its openings; Forest trees (Surround the forest-path location where 찰리 has been exploring); used as Peripheral environmental continuity behind the body and open mesh.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the metal body and thick rope distinct, allowing the threatening interruption to emerge from an otherwise gentle moment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains stuck in the marsh; nearby are large bootprints, a no-entry sign and a corroded, barely legible radiation warning. Elsewhere on the forest path, a capture net descends over Charlie, whose straw hat, colorful raincoat and oversized boots remain on.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리는 약간 우측 정면을 바라보고 있으며, 밧줄 그물이 대각선 방향으로 상체 위로 떨어지고 있습니다.",
    "built_space": "숲속 오솔길과 우측에 늪지대가 보이나, 찰리의 뒤편으로 프롬프트에 없는 거대한 나무 기둥/구조물이 세워져 있습니다.",
    "entities": "찰리는 둥근 눈과 선 형태의 입을 가진 로봇이지만, 요구된 복장(짚모자, 우비, 장화)을 입고 있지 않으며 대신 프롬프트에 없는 배낭을 메고 있습니다.",
    "hard_violations": [
     "[gemini-pro] 지시되지 않은 사물 추가 (등에 멘 배낭, 배경의 거대한 나무 기둥 구조물)",
     "[gpt-high] 장소 설명에 없는 굵은 목재 기둥과 사선 부재로 이루어진 인공 구조물을 추가했다.",
     "[gpt-high] 지정되지 않은 커다란 배낭을 찰리의 등에 추가했다."
    ],
    "physics": "찰리는 걷는 자세로 땅을 딛고 있으며, 밧줄 그물은 공중에서 떨어지는 중첩된 형태를 보여줍니다."
   },
   {
    "label": "B",
    "direction": "찰리의 시선은 정면을 향해 있으며, 그물이 대각선으로 그의 상체를 덮치며 떨어지고 있습니다.",
    "built_space": "숲속 오솔길과 주변의 식물들이 보이며, 지정된 공간 구성과 일치합니다.",
    "entities": "찰리는 짚모자, 다채로운 색의 우비, 장화를 착용한 로봇으로 정확히 묘사되었으나, 우측 가장자리에 체크무늬 셔츠를 입은 알 수 없는 인물의 어깨와 팔이 등장합니다.",
    "hard_violations": [
     "[gemini-pro] 지시되지 않은 인물 추가 (우측 가장자리의 체크무늬 셔츠를 입은 팔과 어깨)"
    ],
    "physics": "찰리의 발은 땅을 딛고 자세를 유지 중이며, 그물은 위에서 아래로 떨어지는 중이므로 물리적으로 타당합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "찰리의 기계 외형과 짚모자, 우비 등 명시된 복장 지침을 훌륭하게 구현했으나, 화면 우측 가장자리에 지시되지 않은 인물이 등장하여 심각한 감점 요인이 되었습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트가 요구한 복장(우비, 짚모자, 장화)을 완전히 누락했으며, 지시되지 않은 배낭과 나무 구조물을 임의로 추가하여 프롬프트 충실도가 낮습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 시선은 정면을 향해 있으며, 그물이 대각선으로 그의 상체를 덮치며 떨어지고 있습니다.",
        "built_space": "숲속 오솔길과 주변의 식물들이 보이며, 지정된 공간 구성과 일치합니다.",
        "entities": "찰리는 짚모자, 다채로운 색의 우비, 장화를 착용한 로봇으로 정확히 묘사되었으나, 우측 가장자리에 체크무늬 셔츠를 입은 알 수 없는 인물의 어깨와 팔이 등장합니다.",
        "hard_violations": [
         "지시되지 않은 인물 추가 (우측 가장자리의 체크무늬 셔츠를 입은 팔과 어깨)"
        ],
        "physics": "찰리의 발은 땅을 딛고 자세를 유지 중이며, 그물은 위에서 아래로 떨어지는 중이므로 물리적으로 타당합니다."
       },
       {
        "label": "A",
        "direction": "찰리는 약간 우측 정면을 바라보고 있으며, 밧줄 그물이 대각선 방향으로 상체 위로 떨어지고 있습니다.",
        "built_space": "숲속 오솔길과 우측에 늪지대가 보이나, 찰리의 뒤편으로 프롬프트에 없는 거대한 나무 기둥/구조물이 세워져 있습니다.",
        "entities": "찰리는 둥근 눈과 선 형태의 입을 가진 로봇이지만, 요구된 복장(짚모자, 우비, 장화)을 입고 있지 않으며 대신 프롬프트에 없는 배낭을 메고 있습니다.",
        "hard_violations": [
         "지시되지 않은 사물 추가 (등에 멘 배낭, 배경의 거대한 나무 기둥 구조물)"
        ],
        "physics": "찰리는 걷는 자세로 땅을 딛고 있으며, 밧줄 그물은 공중에서 떨어지는 중첩된 형태를 보여줍니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "찰리의 기계 외형과 짚모자, 우비 등 명시된 복장 지침을 훌륭하게 구현했으나, 화면 우측 가장자리에 지시되지 않은 인물이 등장하여 심각한 감점 요인이 되었습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트가 요구한 복장(우비, 짚모자, 장화)을 완전히 누락했으며, 지시되지 않은 배낭과 나무 구조물을 임의로 추가하여 프롬프트 충실도가 낮습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 시선은 정면을 향해 있으며, 그물이 대각선으로 그의 상체를 덮치며 떨어지고 있습니다.",
        "built_space": "숲속 오솔길과 주변의 식물들이 보이며, 지정된 공간 구성과 일치합니다.",
        "entities": "찰리는 짚모자, 다채로운 색의 우비, 장화를 착용한 로봇으로 정확히 묘사되었으나, 우측 가장자리에 체크무늬 셔츠를 입은 알 수 없는 인물의 어깨와 팔이 등장합니다.",
        "hard_violations": [
         "지시되지 않은 인물 추가 (우측 가장자리의 체크무늬 셔츠를 입은 팔과 어깨)"
        ],
        "physics": "찰리의 발은 땅을 딛고 자세를 유지 중이며, 그물은 위에서 아래로 떨어지는 중이므로 물리적으로 타당합니다."
       },
       {
        "label": "A",
        "direction": "찰리는 약간 우측 정면을 바라보고 있으며, 밧줄 그물이 대각선 방향으로 상체 위로 떨어지고 있습니다.",
        "built_space": "숲속 오솔길과 우측에 늪지대가 보이나, 찰리의 뒤편으로 프롬프트에 없는 거대한 나무 기둥/구조물이 세워져 있습니다.",
        "entities": "찰리는 둥근 눈과 선 형태의 입을 가진 로봇이지만, 요구된 복장(짚모자, 우비, 장화)을 입고 있지 않으며 대신 프롬프트에 없는 배낭을 메고 있습니다.",
        "hard_violations": [
         "지시되지 않은 사물 추가 (등에 멘 배낭, 배경의 거대한 나무 기둥 구조물)"
        ],
        "physics": "찰리는 걷는 자세로 땅을 딛고 있으며, 밧줄 그물은 공중에서 떨어지는 중첩된 형태를 보여줍니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "상반신 중심의 미디엄 숏과 대각선 전경 그물, 밀짚모자·색색의 우비를 충실히 구현하지만, 그물이 이미 몸에 닿아 맹렬한 낙하 순간은 다소 약하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "그물이 찰리의 몸통을 향하지만, 허용되지 않은 목재 구조물과 배낭을 추가하고 화면에 보여야 할 밀짚모자·우비를 누락했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴은 화면 오른쪽 아래를 향한다. 두꺼운 그물은 화면 위와 오른쪽에서 왼쪽 아래로 대각선을 이루며 찰리의 어깨와 가슴 앞을 가로지른다. 그물의 도달 대상은 분명 찰리의 상체지만, 일부 밧줄이 이미 모자와 몸에 닿아 낙하 직전보다는 덮이기 시작한 순간으로 보인다.",
        "built_space": "꽃 핀 식생 사이의 흙길과 주변 나무들이 보이며 인공 구조물이나 고정 설비는 없다. 찰리가 화면 중앙을 크게 차지하고 머리부터 허벅지 윗부분까지 잡혀 상반신 중심의 미디엄 숏에 가깝다. 습지는 이 구도에서 직접 확인되지 않는다.",
        "entities": "각진 금속 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입을 가진 거대한 로봇 한 대가 보인다. 밀짚모자와 여러 색이 섞인 우비를 착용했고 우비 사이로 금속 몸통이 드러난다. 장화는 프레임 밖이다. 굵은 밧줄 그물, 숲, 꽃은 모두 확인되며 읽을 수 있는 글자는 없다. 오른쪽 가장자리에는 정체를 확정하기 어려운 흐릿한 색 띠 형태가 있다.",
        "hard_violations": [],
        "physics": "찰리는 몸을 앞으로 숙이고 있으며 하체가 화면 아래로 이어져 공중에 떠 있는 자세는 아니다. 모자는 머리에, 우비는 어깨와 팔에 지지된다. 그물은 상단 프레임 밖으로 이어지고 일부는 모자와 상체에 접촉한다. 중력에 따라 내려오며 몸에 걸리기 시작하는 형태는 가능하지만, 팽팽한 부분이 많아 맹렬히 덮쳐오는 운동감은 제한적이다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴은 화면 오른쪽 앞을 향한다. 그물은 위쪽과 오른쪽에서 대각선으로 내려와 머리 앞과 가슴을 가로지르며 찰리의 상체를 대상으로 한다. 다만 하단 밧줄이 허리 주변에 이미 모여 있어 공중에서 막 덮쳐오는 순간보다 포획이 진행된 상태에 가깝다.",
        "built_space": "왼쪽 뒤로 이어지는 숲길, 오른쪽 습지와 갈대, 길가의 꽃이 보인다. 여기에 왼쪽의 수직 목재 기둥 하나와 화면 뒤를 가로지르는 굵은 사선 목재 부재 두 개 이상이 추가되어 있다. 장소 지시에는 이런 구조물이 없다. 찰리는 머리부터 무릎 부근까지 보여 A보다 넓은 구도다.",
        "entities": "흰 얼굴판과 점 형태의 눈, 선 형태의 입을 가진 고릴라형 금속 로봇은 확인된다. 머리 외장은 요구된 각진 형태보다 둥글다. 머리와 어깨가 노출되어 있는데 밀짚모자와 색색의 우비가 없고, 대신 커다란 갈색 배낭을 메고 있다. 장화는 화면 밖이다. 굵은 그물, 나무, 꽃, 습지는 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장소 설명에 없는 굵은 목재 기둥과 사선 부재로 이루어진 인공 구조물을 추가했다.",
         "지정되지 않은 커다란 배낭을 찰리의 등에 추가했다."
        ],
        "physics": "찰리의 굽힌 다리는 화면 아래로 이어지고 몸통과 관절의 연결도 보이므로 떠 있는 몸으로 보이지 않는다. 배낭은 어깨끈으로 몸에 지지된다. 그물은 상단 화면 밖으로 이어지며 어깨와 몸통에 걸려 있고, 허리 쪽에는 밧줄이 겹쳐져 있다. 물리적으로 가능한 포획 자세지만 자유롭게 낙하 중인 그물보다 이미 몸을 감싸는 그물로 읽힌다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "상반신 중심의 미디엄 숏과 대각선 전경 그물, 밀짚모자·색색의 우비를 충실히 구현하지만, 그물이 이미 몸에 닿아 맹렬한 낙하 순간은 다소 약하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "그물이 찰리의 몸통을 향하지만, 허용되지 않은 목재 구조물과 배낭을 추가하고 화면에 보여야 할 밀짚모자·우비를 누락했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴은 화면 오른쪽 아래를 향한다. 두꺼운 그물은 화면 위와 오른쪽에서 왼쪽 아래로 대각선을 이루며 찰리의 어깨와 가슴 앞을 가로지른다. 그물의 도달 대상은 분명 찰리의 상체지만, 일부 밧줄이 이미 모자와 몸에 닿아 낙하 직전보다는 덮이기 시작한 순간으로 보인다.",
        "built_space": "꽃 핀 식생 사이의 흙길과 주변 나무들이 보이며 인공 구조물이나 고정 설비는 없다. 찰리가 화면 중앙을 크게 차지하고 머리부터 허벅지 윗부분까지 잡혀 상반신 중심의 미디엄 숏에 가깝다. 습지는 이 구도에서 직접 확인되지 않는다.",
        "entities": "각진 금속 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입을 가진 거대한 로봇 한 대가 보인다. 밀짚모자와 여러 색이 섞인 우비를 착용했고 우비 사이로 금속 몸통이 드러난다. 장화는 프레임 밖이다. 굵은 밧줄 그물, 숲, 꽃은 모두 확인되며 읽을 수 있는 글자는 없다. 오른쪽 가장자리에는 정체를 확정하기 어려운 흐릿한 색 띠 형태가 있다.",
        "hard_violations": [],
        "physics": "찰리는 몸을 앞으로 숙이고 있으며 하체가 화면 아래로 이어져 공중에 떠 있는 자세는 아니다. 모자는 머리에, 우비는 어깨와 팔에 지지된다. 그물은 상단 프레임 밖으로 이어지고 일부는 모자와 상체에 접촉한다. 중력에 따라 내려오며 몸에 걸리기 시작하는 형태는 가능하지만, 팽팽한 부분이 많아 맹렬히 덮쳐오는 운동감은 제한적이다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴은 화면 오른쪽 앞을 향한다. 그물은 위쪽과 오른쪽에서 대각선으로 내려와 머리 앞과 가슴을 가로지르며 찰리의 상체를 대상으로 한다. 다만 하단 밧줄이 허리 주변에 이미 모여 있어 공중에서 막 덮쳐오는 순간보다 포획이 진행된 상태에 가깝다.",
        "built_space": "왼쪽 뒤로 이어지는 숲길, 오른쪽 습지와 갈대, 길가의 꽃이 보인다. 여기에 왼쪽의 수직 목재 기둥 하나와 화면 뒤를 가로지르는 굵은 사선 목재 부재 두 개 이상이 추가되어 있다. 장소 지시에는 이런 구조물이 없다. 찰리는 머리부터 무릎 부근까지 보여 A보다 넓은 구도다.",
        "entities": "흰 얼굴판과 점 형태의 눈, 선 형태의 입을 가진 고릴라형 금속 로봇은 확인된다. 머리 외장은 요구된 각진 형태보다 둥글다. 머리와 어깨가 노출되어 있는데 밀짚모자와 색색의 우비가 없고, 대신 커다란 갈색 배낭을 메고 있다. 장화는 화면 밖이다. 굵은 그물, 나무, 꽃, 습지는 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장소 설명에 없는 굵은 목재 기둥과 사선 부재로 이루어진 인공 구조물을 추가했다.",
         "지정되지 않은 커다란 배낭을 찰리의 등에 추가했다."
        ],
        "physics": "찰리의 굽힌 다리는 화면 아래로 이어지고 몸통과 관절의 연결도 보이므로 떠 있는 몸으로 보이지 않는다. 배낭은 어깨끈으로 몸에 지지된다. 그물은 상단 화면 밖으로 이어지며 어깨와 몸통에 걸려 있고, 허리 쪽에는 밧줄이 겹쳐져 있다. 물리적으로 가능한 포획 자세지만 자유롭게 낙하 중인 그물보다 이미 몸을 감싸는 그물로 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.875,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.625,
    "B": 1.75
   },
   "violations": {
    "B": [
     "[gemini-pro] 지시되지 않은 인물 추가 (우측 가장자리의 체크무늬 셔츠를 입은 팔과 어깨)"
    ],
    "A": [
     "[gemini-pro] 지시되지 않은 사물 추가 (등에 멘 배낭, 배경의 거대한 나무 기둥 구조물)",
     "[gpt-high] 장소 설명에 없는 굵은 목재 기둥과 사선 부재로 이루어진 인공 구조물을 추가했다.",
     "[gpt-high] 지정되지 않은 커다란 배낭을 찰리의 등에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 625
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "찰리의 기계 외형과 짚모자, 우비 등 명시된 복장 지침을 훌륭하게 구현했으나, 화면 우측 가장자리에 지시되지 않은 인물이 등장하여 심각한 감점 요인이 되었습니다.  ★위반: [gemini-pro] 지시되지 않은 인물 추가 (우측 가장자리의 체크무늬 셔츠를 입은 팔과 어깨)"
   },
   {
    "label": "A",
    "score": 625,
    "verdict_ko": "프롬프트가 요구한 복장(우비, 짚모자, 장화)을 완전히 누락했으며, 지시되지 않은 배낭과 나무 구조물을 임의로 추가하여 프롬프트 충실도가 낮습니다.  ★위반: [gemini-pro] 지시되지 않은 사물 추가 (등에 멘 배낭, 배경의 거대한 나무 기둥 구조물) / [gpt-high] 장소 설명에 없는 굵은 목재 기둥과 사선 부재로 이루어진 인공 구조물을 추가했다. / [gpt-high] 지정되지 않은 커다란 배낭을 찰리의 등에 추가했다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde5-3af2-7452-8b8f-f496ef126711",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S56sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:46:43.694179+00:00",
  "fingerprint": "1fa2aebd2ff68442607a1cfa80327ec591a45f676b5cd260611f8fa26161b977",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S56sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S56sh12_sel.png",
  "source_sha256": "ac7f5f22afd9ffe5d29cd9eb0b27ce472c42b54a5de7c2d7c6d578b18b53ce7e",
  "file": "S56sh12_cine.png",
  "staged_sha256": "3ed324071fbfc83e876e062876b3798012d2d6f1f718b81820d3b78e8b18ace1",
  "latency_ms": 11161
 },
 "S56sh20::signage": {
  "fp": "933e8bf95a282e87",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::4f9b0ad8599ad581": {
  "subjects": [],
  "subject_text": "익산 숲속 함정 구덩이 내부\n숲 지면 아래로 깊게 파인 함정 구덩이. 가파른 흙벽과 바닥이 둘러싸고, 위쪽 개구부로 자연광이 들어온다.",
  "identity": "canonical",
  "scope_id": "L74",
  "scope_role": "location_exterior",
  "scope_sha": "481c6e056ad261ba"
 },
 "groupbg::forest_trap_pit": {
  "input_fingerprint": "fc7e44d62261530f",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "forest_trap_pit",
    "tags": [
     "S56sh20",
     "S56sh22"
    ]
   },
   "context_sig": "f303efbd5a8299da"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the bottom of an open-topped trap pit in the marshland forest, looking up toward armed figures at its rim.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 숲속 함정 구덩이 내부: 땅이 꺼지며 떨어지는 깊은 흙구덩이로, 하회탈을 쓴 궁사들이 위에서 창과 활을 겨누고 내려다보는 시점. (특징: 가파른 구덩이 흙벽; 위를 올려다볼 때 보이는 화려한 하회탈과 무기들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 순간 트랩에 빠지는 현우와 앰버, 아아아아악!\n- 구덩이에 빠진 현우와 앰버를 내려다본다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the bottom of an open-topped trap pit in the marshland forest, looking up toward armed figures at its rim.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 숲속 함정 구덩이 내부: 땅이 꺼지며 떨어지는 깊은 흙구덩이로, 하회탈을 쓴 궁사들이 위에서 창과 활을 겨누고 내려다보는 시점. (특징: 가파른 구덩이 흙벽; 위를 올려다볼 때 보이는 화려한 하회탈과 무기들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 순간 트랩에 빠지는 현우와 앰버, 아아아아악!\n- 구덩이에 빠진 현우와 앰버를 내려다본다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_forest_trap_pit_2a4ec2.png",
  "asset_id": "a0433856-9364-4997-85eb-675ef0f445ce",
  "input_asset_ids": [
   "045169e4-516e-4bb6-8a52-de69de8240d5"
  ],
  "origin_tag": "S56sh20",
  "place_text": "At the bottom of an open-topped trap pit in the marshland forest, looking up toward armed figures at its rim.",
  "origin_inputs": {
   "place_text": "At the bottom of an open-topped trap pit in the marshland forest, looking up toward armed figures at its rim.",
   "time_of_day_en": "day",
   "conti_asset_id": "045169e4-516e-4bb6-8a52-de69de8240d5"
  }
 },
 "S56sh20::bgfirst_bg": {
  "input_fingerprint": "6dedadb235256ade",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 구덩이 아래에서 올려다본 하늘을 배경으로, 형형색색의 하회탈 가죽 마스크를 쓴 아시아계 무장 괴한 여럿이 창과 활을 든 채 내려다보는 위압적인 구도.\n\nLOCATION (lock): At the bottom of an open-topped trap pit in the marshland forest, looking up toward armed figures at its rim.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pit walls and rim (Enclose the trapped captives below the armed men) — Interior walls rise along the frame edges toward the opening overhead; used as Creates an enclosing frame and establishes the men above the camera; Spears and bows (Held by the men watching the captives) — Weapon shafts and bow profiles enter at differing angles from the rim; used as Interrupt the sky without forming a mechanically repeated pattern; Sky above the pit (Visible through the opening between the men); used as Central negative space emphasizing the captives' confinement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight against the sky gives the masked men imposing contrast while preserving the distinct colors and leather detail of their masks.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 구덩이 아래에서 올려다본 하늘을 배경으로, 형형색색의 하회탈 가죽 마스크를 쓴 아시아계 무장 괴한 여럿이 창과 활을 든 채 내려다보는 위압적인 구도.\n\nLOCATION (lock): At the bottom of an open-topped trap pit in the marshland forest, looking up toward armed figures at its rim.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pit walls and rim (Enclose the trapped captives below the armed men) — Interior walls rise along the frame edges toward the opening overhead; used as Creates an enclosing frame and establishes the men above the camera; Spears and bows (Held by the men watching the captives) — Weapon shafts and bow profiles enter at differing angles from the rim; used as Interrupt the sky without forming a mechanically repeated pattern; Sky above the pit (Visible through the opening between the men); used as Central negative space emphasizing the captives' confinement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight against the sky gives the masked men imposing contrast while preserving the distinct colors and leather detail of their masks.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S56sh20__bgfirst_bg.png",
  "asset_id": "9a13b370-6148-4e95-ae7c-d218762a3a6a",
  "input_asset_ids": [
   "045169e4-516e-4bb6-8a52-de69de8240d5",
   "a0433856-9364-4997-85eb-675ef0f445ce"
  ]
 },
 "S56sh20": {
  "input_fingerprint": "7f45b597e527df35",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 구덩이 아래에서 올려다본 하늘을 배경으로, 형형색색의 하회탈 가죽 마스크를 쓴 아시아계 무장 괴한 여럿이 창과 활을 든 채 내려다보는 위압적인 구도.\n\nLOCATION (lock): At the bottom of an open-topped trap pit in the marshland forest, looking up toward armed figures at its rim. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pit walls and rim (Enclose the trapped captives below the armed men) — Interior walls rise along the frame edges toward the opening overhead; used as Creates an enclosing frame and establishes the men above the camera; Spears and bows (Held by the men watching the captives) — Weapon shafts and bow profiles enter at differing angles from the rim; used as Interrupt the sky without forming a mechanically repeated pattern; Sky above the pit (Visible through the opening between the men); used as Central negative space emphasizing the captives' confinement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight against the sky gives the masked men imposing contrast while preserving the distinct colors and leather detail of their masks.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains trapped in the marsh, and an arrow remains embedded in a tree at the forest capture site beside the open pit trap. Charlie has already been captured in the net elsewhere in the forest.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 아시아계 무장 괴한 여럿 right now, so 아시아계 무장 괴한 여럿's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 아시아계 무장 괴한 여럿: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 구덩이 아래에서 올려다본 하늘을 배경으로, 형형색색의 하회탈 가죽 마스크를 쓴 아시아계 무장 괴한 여럿이 창과 활을 든 채 내려다보는 위압적인 구도.\n\nLOCATION (lock): At the bottom of an open-topped trap pit in the marshland forest, looking up toward armed figures at its rim. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pit walls and rim (Enclose the trapped captives below the armed men) — Interior walls rise along the frame edges toward the opening overhead; used as Creates an enclosing frame and establishes the men above the camera; Spears and bows (Held by the men watching the captives) — Weapon shafts and bow profiles enter at differing angles from the rim; used as Interrupt the sky without forming a mechanically repeated pattern; Sky above the pit (Visible through the opening between the men); used as Central negative space emphasizing the captives' confinement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight against the sky gives the masked men imposing contrast while preserving the distinct colors and leather detail of their masks.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains trapped in the marsh, and an arrow remains embedded in a tree at the forest capture site beside the open pit trap. Charlie has already been captured in the net elsewhere in the forest.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 아시아계 무장 괴한 여럿 right now, so 아시아계 무장 괴한 여럿's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 아시아계 무장 괴한 여럿: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 구덩이 아래에서 올려다본 하늘을 배경으로, 형형색색의 하회탈 가죽 마스크를 쓴 아시아계 무장 괴한 여럿이 창과 활을 든 채 내려다보는 위압적인 구도.\n\nLOCATION (lock): At the bottom of an open-topped trap pit in the marshland forest, looking up toward armed figures at its rim. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pit walls and rim (Enclose the trapped captives below the armed men) — Interior walls rise along the frame edges toward the opening overhead; used as Creates an enclosing frame and establishes the men above the camera; Spears and bows (Held by the men watching the captives) — Weapon shafts and bow profiles enter at differing angles from the rim; used as Interrupt the sky without forming a mechanically repeated pattern; Sky above the pit (Visible through the opening between the men); used as Central negative space emphasizing the captives' confinement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight against the sky gives the masked men imposing contrast while preserving the distinct colors and leather detail of their masks.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains trapped in the marsh, and an arrow remains embedded in a tree at the forest capture site beside the open pit trap. Charlie has already been captured in the net elsewhere in the forest.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 아시아계 무장 괴한 여럿 right now, so 아시아계 무장 괴한 여럿's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 아시아계 무장 괴한 여럿: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S56sh20__bgfirst_bg.png",
     "asset_id": "9a13b370-6148-4e95-ae7c-d218762a3a6a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S56sh20.png",
     "asset_id": "045169e4-516e-4bb6-8a52-de69de8240d5",
     "role": "conti_light"
    },
    {
     "label": "PROP REFERENCE — 창: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:962114>",
     "asset_id": "c2369c07-dcb1-40c9-93a0-aafad7bacf3a",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_forest_trap_pit_2a4ec2.png",
     "asset_id": "a0433856-9364-4997-85eb-675ef0f445ce",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "PROP REFERENCE — 창: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:962114>",
     "asset_id": "c2369c07-dcb1-40c9-93a0-aafad7bacf3a",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물들은 아래를 내려다보고 있으며, 무기들은 프레임 중앙을 향하거나 제각각의 방향을 가리킴.",
    "built_space": "위로 올려다보는 구도가 아닌 수평적 참호처럼 보이며, 인물들이 구덩이 가장자리가 아닌 수직 벽면에 위치함.",
    "entities": "하회탈을 쓴 아시아계 인물들. 마스크 색상이 다양하지 않으며, 창의 형태가 레퍼런스와 다름.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 스테이징 (인물들이 중력을 거스르고 수직 벽면에 서 있음)"
    ],
    "physics": "좌우 인물들이 가파른 수직 벽면에 평지에 서 있듯 서 있거나 쪼그려 앉아 있어 중력에 위배됨."
   },
   {
    "label": "B",
    "direction": "인물들이 카메라를 똑바로 내려다보며, 다수의 창끝이 정확히 프레임 중앙을 겨냥함.",
    "built_space": "구덩이 바닥에서 하늘을 올려다보는 시점을 정확히 구현했으며, 인물들이 위쪽 가장자리에 알맞게 배치됨.",
    "entities": "형형색색의 하회탈을 쓴 인물들. 레퍼런스와 정확히 일치하는 형태의 창과 활을 들고 있음.",
    "hard_violations": [
     "[gemini-pro] 지지대 없는 물체 (손에 쥐어지지 않고 허공에 떠 있는 다수의 창)"
    ],
    "physics": "프레임 중앙과 상단에 있는 여러 개의 창들이 손에 들려있지 않고 창대가 하늘에서 끊어진 채 허공에 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "올려다보는 구도와 형형색색의 탈, 소품 디테일은 훌륭하나, 허공에 떠 있는 다수의 창들로 인해 실격 처리됨."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "구덩이의 수직 벽면에 인물들이 중력을 거스르고 서 있는 물리적으로 불가능한 구도이므로 실격 처리됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들은 아래를 내려다보고 있으며, 무기들은 프레임 중앙을 향하거나 제각각의 방향을 가리킴.",
        "built_space": "위로 올려다보는 구도가 아닌 수평적 참호처럼 보이며, 인물들이 구덩이 가장자리가 아닌 수직 벽면에 위치함.",
        "entities": "하회탈을 쓴 아시아계 인물들. 마스크 색상이 다양하지 않으며, 창의 형태가 레퍼런스와 다름.",
        "hard_violations": [
         "물리적으로 불가능한 스테이징 (인물들이 중력을 거스르고 수직 벽면에 서 있음)"
        ],
        "physics": "좌우 인물들이 가파른 수직 벽면에 평지에 서 있듯 서 있거나 쪼그려 앉아 있어 중력에 위배됨."
       },
       {
        "label": "B",
        "direction": "인물들이 카메라를 똑바로 내려다보며, 다수의 창끝이 정확히 프레임 중앙을 겨냥함.",
        "built_space": "구덩이 바닥에서 하늘을 올려다보는 시점을 정확히 구현했으며, 인물들이 위쪽 가장자리에 알맞게 배치됨.",
        "entities": "형형색색의 하회탈을 쓴 인물들. 레퍼런스와 정확히 일치하는 형태의 창과 활을 들고 있음.",
        "hard_violations": [
         "지지대 없는 물체 (손에 쥐어지지 않고 허공에 떠 있는 다수의 창)"
        ],
        "physics": "프레임 중앙과 상단에 있는 여러 개의 창들이 손에 들려있지 않고 창대가 하늘에서 끊어진 채 허공에 떠 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "올려다보는 구도와 형형색색의 탈, 소품 디테일은 훌륭하나, 허공에 떠 있는 다수의 창들로 인해 실격 처리됨."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "구덩이의 수직 벽면에 인물들이 중력을 거스르고 서 있는 물리적으로 불가능한 구도이므로 실격 처리됨."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "인물들은 아래를 내려다보고 있으며, 무기들은 프레임 중앙을 향하거나 제각각의 방향을 가리킴.",
        "built_space": "위로 올려다보는 구도가 아닌 수평적 참호처럼 보이며, 인물들이 구덩이 가장자리가 아닌 수직 벽면에 위치함.",
        "entities": "하회탈을 쓴 아시아계 인물들. 마스크 색상이 다양하지 않으며, 창의 형태가 레퍼런스와 다름.",
        "hard_violations": [
         "물리적으로 불가능한 스테이징 (인물들이 중력을 거스르고 수직 벽면에 서 있음)"
        ],
        "physics": "좌우 인물들이 가파른 수직 벽면에 평지에 서 있듯 서 있거나 쪼그려 앉아 있어 중력에 위배됨."
       },
       {
        "label": "B",
        "direction": "인물들이 카메라를 똑바로 내려다보며, 다수의 창끝이 정확히 프레임 중앙을 겨냥함.",
        "built_space": "구덩이 바닥에서 하늘을 올려다보는 시점을 정확히 구현했으며, 인물들이 위쪽 가장자리에 알맞게 배치됨.",
        "entities": "형형색색의 하회탈을 쓴 인물들. 레퍼런스와 정확히 일치하는 형태의 창과 활을 들고 있음.",
        "hard_violations": [
         "지지대 없는 물체 (손에 쥐어지지 않고 허공에 떠 있는 다수의 창)"
        ],
        "physics": "프레임 중앙과 상단에 있는 여러 개의 창들이 손에 들려있지 않고 창대가 하늘에서 끊어진 채 허공에 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "구덩이 바닥의 광각 시점, 중앙의 넓은 하늘, 가장자리에서 내려다보는 인물과 불규칙한 무기 배치가 더 정확하지만, 마스크가 주황·갈색으로 편중되어 형형색색 조건은 부족하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "참조의 흙벽과 다채로운 탈은 잘 살렸지만, 창끝들이 중앙으로 방사형 집결해 기계적 반복을 피하라는 구도 지시와 하늘의 여백을 약화한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴이 보이는 다섯 명은 탈의 정면을 구덩이 아래 카메라 쪽으로 기울인다. 위쪽에서 들어오는 창 세 자루는 개구부 중앙 쪽으로 내려오고, 아래·좌우의 창들은 반대로 중앙 하늘 쪽으로 올라가며 창끝들이 원형으로 모인다. 오른쪽 녹색 탈 인물은 활을 구덩이 안쪽으로 내밀지만 화살의 정확한 조준선은 명확하지 않다.",
        "built_space": "하나의 열린 흙구덩이이며 양쪽의 높은 흙벽, 노출된 뿌리, 중앙 뒤쪽 벽과 숲 가장자리가 참조와 매우 가깝다. 왼쪽에 탈을 쓴 두 명, 뒤쪽 중앙에 한 명, 오른쪽에 두 명이 있고, 위쪽에는 별도 인물의 잘린 몸통과 팔이 들어온다. 카메라는 벽 아래에 있으며 인물들은 모두 테두리 너머에 있다. 중앙 하늘은 남지만 여러 창끝이 여백을 집중적으로 분할한다.",
        "entities": "다섯 개의 탈은 황금색·황토색·적갈색·녹색 등으로 구분되며 하회탈 계열의 과장된 얼굴 형태와 표면 질감이 보인다. 인물들은 성인 남성으로 읽히지만 얼굴이 가려져 아시아계 여부를 확정하기 어렵다. 활 한 자루와 창날 일곱 개, 날이 보이지 않는 긴 자루 하나가 보인다. 일부 은색 창날은 참조의 넓고 긴 형태에 가깝지만 검은 소형 창날은 다르다. 화면 밖의 포로, 나무에 박힌 화살, 찰리는 이 프레이밍에서 확인할 수 없으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물들의 하체는 주로 흙 테두리 뒤에 가려져 있고 상체가 그 위로 기울어 있어 가장자리 지면에 지지된 자세로 읽힌다. 가까운 인물들의 손은 자루와 활을 쥐고 있다. 일부 창의 손잡이 쪽은 화면 밖이나 테두리 뒤로 이어지므로 손 접촉을 전부 확인할 수는 없지만, 공중에 독립적으로 떠 있는 물체라고 단정할 근거는 없다."
       },
       {
        "label": "B",
        "direction": "여섯 인물 모두 얼굴을 구덩이 아래로 향한다. 왼쪽 아래와 오른쪽 뒤쪽 궁수의 화살은 구덩이 내부를 향해 비스듬히 내려가며, 오른쪽 위 궁수도 활시위를 당겨 아래쪽을 겨눈다. 창 세 자루의 날은 하늘을 향해 올라가 있다. 창으로 아래를 찌르는 모습은 아니지만, 창을 든 채 내려다본다는 장면 지시에는 맞는다.",
        "built_space": "하나의 열린 구덩이 양옆으로 뿌리가 드러난 흙벽이 올라가고 뒤쪽 낮은 테두리 너머로 낮 하늘이 넓게 열린다. 왼쪽 세 명과 오른쪽 세 명이 각각 테두리에 서거나 웅크리고 있어 카메라보다 높은 위치가 분명하다. 참조의 흙·뿌리·수목 구성은 유지되지만 테두리 윤곽은 다소 다르다. 중앙 하늘이 뚜렷한 여백으로 남고 창과 활은 서로 다른 위치와 각도로 진입한다.",
        "entities": "성인 남성으로 보이는 여섯 명이 하회탈 형태의 마스크를 쓰고 창 세 자루와 활 세 자루를 든다. 얼굴이 가려져 민족적 정체성을 확정하기는 어렵다. 탈은 대부분 주황·갈색 계열이라 형형색색이라는 요구에 못 미치며, 가죽보다는 단단하게 조각된 표면처럼 보이는 부분도 있다. 창은 목제 자루와 금속 날을 갖추지만 일부 날의 톱니와 형태가 참조와 다르다. 궁수들의 화살과 화살통이 보이고 읽을 수 있는 글자는 없다. 화면 밖의 포획 상태는 평가할 수 없다.",
        "hard_violations": [],
        "physics": "서 있는 인물의 다리와 웅크린 인물의 굽힌 무릎이 가장자리 지면으로 이어져 체중 지지가 자연스럽다. 오른쪽 앞 인물은 한 손으로 흙 테두리를 짚는다. 창은 손으로 자루를 잡고 있으며 궁수들은 활을 잡는 손과 시위를 당기는 손을 나누어 사용한다. 지지 없이 떠 있는 인물이나 무기는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "구덩이 바닥의 광각 시점, 중앙의 넓은 하늘, 가장자리에서 내려다보는 인물과 불규칙한 무기 배치가 더 정확하지만, 마스크가 주황·갈색으로 편중되어 형형색색 조건은 부족하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "참조의 흙벽과 다채로운 탈은 잘 살렸지만, 창끝들이 중앙으로 방사형 집결해 기계적 반복을 피하라는 구도 지시와 하늘의 여백을 약화한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴이 보이는 다섯 명은 탈의 정면을 구덩이 아래 카메라 쪽으로 기울인다. 위쪽에서 들어오는 창 세 자루는 개구부 중앙 쪽으로 내려오고, 아래·좌우의 창들은 반대로 중앙 하늘 쪽으로 올라가며 창끝들이 원형으로 모인다. 오른쪽 녹색 탈 인물은 활을 구덩이 안쪽으로 내밀지만 화살의 정확한 조준선은 명확하지 않다.",
        "built_space": "하나의 열린 흙구덩이이며 양쪽의 높은 흙벽, 노출된 뿌리, 중앙 뒤쪽 벽과 숲 가장자리가 참조와 매우 가깝다. 왼쪽에 탈을 쓴 두 명, 뒤쪽 중앙에 한 명, 오른쪽에 두 명이 있고, 위쪽에는 별도 인물의 잘린 몸통과 팔이 들어온다. 카메라는 벽 아래에 있으며 인물들은 모두 테두리 너머에 있다. 중앙 하늘은 남지만 여러 창끝이 여백을 집중적으로 분할한다.",
        "entities": "다섯 개의 탈은 황금색·황토색·적갈색·녹색 등으로 구분되며 하회탈 계열의 과장된 얼굴 형태와 표면 질감이 보인다. 인물들은 성인 남성으로 읽히지만 얼굴이 가려져 아시아계 여부를 확정하기 어렵다. 활 한 자루와 창날 일곱 개, 날이 보이지 않는 긴 자루 하나가 보인다. 일부 은색 창날은 참조의 넓고 긴 형태에 가깝지만 검은 소형 창날은 다르다. 화면 밖의 포로, 나무에 박힌 화살, 찰리는 이 프레이밍에서 확인할 수 없으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물들의 하체는 주로 흙 테두리 뒤에 가려져 있고 상체가 그 위로 기울어 있어 가장자리 지면에 지지된 자세로 읽힌다. 가까운 인물들의 손은 자루와 활을 쥐고 있다. 일부 창의 손잡이 쪽은 화면 밖이나 테두리 뒤로 이어지므로 손 접촉을 전부 확인할 수는 없지만, 공중에 독립적으로 떠 있는 물체라고 단정할 근거는 없다."
       },
       {
        "label": "A",
        "direction": "여섯 인물 모두 얼굴을 구덩이 아래로 향한다. 왼쪽 아래와 오른쪽 뒤쪽 궁수의 화살은 구덩이 내부를 향해 비스듬히 내려가며, 오른쪽 위 궁수도 활시위를 당겨 아래쪽을 겨눈다. 창 세 자루의 날은 하늘을 향해 올라가 있다. 창으로 아래를 찌르는 모습은 아니지만, 창을 든 채 내려다본다는 장면 지시에는 맞는다.",
        "built_space": "하나의 열린 구덩이 양옆으로 뿌리가 드러난 흙벽이 올라가고 뒤쪽 낮은 테두리 너머로 낮 하늘이 넓게 열린다. 왼쪽 세 명과 오른쪽 세 명이 각각 테두리에 서거나 웅크리고 있어 카메라보다 높은 위치가 분명하다. 참조의 흙·뿌리·수목 구성은 유지되지만 테두리 윤곽은 다소 다르다. 중앙 하늘이 뚜렷한 여백으로 남고 창과 활은 서로 다른 위치와 각도로 진입한다.",
        "entities": "성인 남성으로 보이는 여섯 명이 하회탈 형태의 마스크를 쓰고 창 세 자루와 활 세 자루를 든다. 얼굴이 가려져 민족적 정체성을 확정하기는 어렵다. 탈은 대부분 주황·갈색 계열이라 형형색색이라는 요구에 못 미치며, 가죽보다는 단단하게 조각된 표면처럼 보이는 부분도 있다. 창은 목제 자루와 금속 날을 갖추지만 일부 날의 톱니와 형태가 참조와 다르다. 궁수들의 화살과 화살통이 보이고 읽을 수 있는 글자는 없다. 화면 밖의 포획 상태는 평가할 수 없다.",
        "hard_violations": [],
        "physics": "서 있는 인물의 다리와 웅크린 인물의 굽힌 무릎이 가장자리 지면으로 이어져 체중 지지가 자연스럽다. 오른쪽 앞 인물은 한 손으로 흙 테두리를 짚는다. 창은 손으로 자루를 잡고 있으며 궁수들은 활을 잡는 손과 시위를 당기는 손을 나누어 사용한다. 지지 없이 떠 있는 인물이나 무기는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.625
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 스테이징 (인물들이 중력을 거스르고 수직 벽면에 서 있음)"
    ],
    "B": [
     "[gemini-pro] 지지대 없는 물체 (손에 쥐어지지 않고 허공에 떠 있는 다수의 창)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1625,
   "A": 1500
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "올려다보는 구도와 형형색색의 탈, 소품 디테일은 훌륭하나, 허공에 떠 있는 다수의 창들로 인해 실격 처리됨.  ★위반: [gemini-pro] 지지대 없는 물체 (손에 쥐어지지 않고 허공에 떠 있는 다수의 창)"
   },
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "구덩이의 수직 벽면에 인물들이 중력을 거스르고 서 있는 물리적으로 불가능한 구도이므로 실격 처리됨.  ★위반: [gemini-pro] 물리적으로 불가능한 스테이징 (인물들이 중력을 거스르고 수직 벽면에 서 있음)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_forest_trap_pit_2a4ec2.png",
    "asset_id": "a0433856-9364-4997-85eb-675ef0f445ce",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "PROP REFERENCE — 창: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:962114>",
    "asset_id": "c2369c07-dcb1-40c9-93a0-aafad7bacf3a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde5-f2fd-709e-825d-3a9783b1ddb9",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S56sh20__bgfirst_bg.png",
   "bg_asset_id": "9a13b370-6148-4e95-ae7c-d218762a3a6a",
   "bg_record_key": "S56sh20::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "forest_trap_pit",
   "groupbg_asset_id": "a0433856-9364-4997-85eb-675ef0f445ce"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S56sh20::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:46:55.345391+00:00",
  "fingerprint": "fd96368f47bfefe7ceb2373788007b4e4a0a372eb14b7737be44b9830e90706f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S56sh20_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S56sh20_sel.png",
  "source_sha256": "e6526a6c765df21a5d4028facde52bc2dc1dc1dd6a0498b85613dd66e8f32f47",
  "file": "S56sh20_cine.png",
  "staged_sha256": "f7ee860596d0703c13857373e0fc50d9927e5ec6d61072417e061da94269c85b",
  "latency_ms": 11052
 },
 "S56sh22::signage": {
  "fp": "992eb25e53e2059c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S56sh22": {
  "input_fingerprint": "2edbf4bcffc13235",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 목에 마취 화살이 꽂힌 이현우의 두 눈이 반쯤 감긴 채 상체가 바닥을 향해 기울어진 위태로운 찰나.\n\nLOCATION (lock): On the earthen bottom of the open forest trap pit beneath the armed captors. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Tranquilizing dart (Embedded in 이현우's neck) — Projects visibly from the near side of his neck; used as Small causal detail linking the collapse to the attack; Pit floor (Immediately beneath his falling upper body); used as Open lower-frame landing space that makes the direction of collapse legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain objective daytime illumination within the pit, reserving the scripted fade-out for the transition rather than introducing subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is collapsing toward the bottom of the pit under sedation, with his upper body tipping downward, his eyes half closed, and a tranquilizer dart embedded in his neck. The positions of his head relative to his torso, his arms and legs, and any ground contact are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The pit remains open, the arrow remains embedded in the nearby tree, and the camper is still stuck in the marsh. Charlie remains captured elsewhere. 이현우: He is in the pit losing consciousness after a tranquilizer-dart hit; a tranquilizer dart is embedded in his neck. His treated leg remains bandaged and the contact card remains concealed in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 목에 마취 화살이 꽂힌 이현우의 두 눈이 반쯤 감긴 채 상체가 바닥을 향해 기울어진 위태로운 찰나.\n\nLOCATION (lock): On the earthen bottom of the open forest trap pit beneath the armed captors. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Tranquilizing dart (Embedded in 이현우's neck) — Projects visibly from the near side of his neck; used as Small causal detail linking the collapse to the attack; Pit floor (Immediately beneath his falling upper body); used as Open lower-frame landing space that makes the direction of collapse legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain objective daytime illumination within the pit, reserving the scripted fade-out for the transition rather than introducing subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is collapsing toward the bottom of the pit under sedation, with his upper body tipping downward, his eyes half closed, and a tranquilizer dart embedded in his neck. The positions of his head relative to his torso, his arms and legs, and any ground contact are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The pit remains open, the arrow remains embedded in the nearby tree, and the camper is still stuck in the marsh. Charlie remains captured elsewhere. 이현우: He is in the pit losing consciousness after a tranquilizer-dart hit; a tranquilizer dart is embedded in his neck. His treated leg remains bandaged and the contact card remains concealed in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 목에 마취 화살이 꽂힌 이현우의 두 눈이 반쯤 감긴 채 상체가 바닥을 향해 기울어진 위태로운 찰나.\n\nLOCATION (lock): On the earthen bottom of the open forest trap pit beneath the armed captors. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Tranquilizing dart (Embedded in 이현우's neck) — Projects visibly from the near side of his neck; used as Small causal detail linking the collapse to the attack; Pit floor (Immediately beneath his falling upper body); used as Open lower-frame landing space that makes the direction of collapse legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain objective daytime illumination within the pit, reserving the scripted fade-out for the transition rather than introducing subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is collapsing toward the bottom of the pit under sedation, with his upper body tipping downward, his eyes half closed, and a tranquilizer dart embedded in his neck. The positions of his head relative to his torso, his arms and legs, and any ground contact are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The pit remains open, the arrow remains embedded in the nearby tree, and the camper is still stuck in the marsh. Charlie remains captured elsewhere. 이현우: He is in the pit losing consciousness after a tranquilizer-dart hit; a tranquilizer dart is embedded in his neck. His treated leg remains bandaged and the contact card remains concealed in his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "반쯤 감긴 눈은 아래쪽 바닥을 향하고 있으며, 마취 화살은 목 측면에서 바깥을 향해 꽂혀 있음.",
    "built_space": "흙으로 이루어진 구덩이 내부 벽면과 바닥이 보이며, 프레임 좌측 하단으로 추락할 수 있는 여유 공간이 확보됨.",
    "entities": "이현우(레퍼런스와 일치하는 얼굴, 헝클어진 머리, 낡은 후드티), 마취 화살(나무 깃과 대).",
    "hard_violations": [],
    "physics": "하체(프레임 밖)가 지탱하는 가운데 상체가 중력에 의해 바닥으로 기울어지고 있어 의식을 잃어가는 순간의 물리적 상태가 자연스러움."
   },
   {
    "label": "B",
    "direction": "시선은 약간 앞쪽 바닥을 향하며 반쯤 감긴 눈을 함. 화살은 목에 꽂혀 있음.",
    "built_space": "좁고 긴 흙 구덩이 바닥에 위치하며 상단에는 숲의 나무와 하늘이 보임.",
    "entities": "이현우(얼굴, 후드티 일치), 마취 화살, 다리에 감긴 붕대 일부 노출됨.",
    "hard_violations": [],
    "physics": "양손과 몸이 이미 바닥에 엎드려 지탱하고 있으나, 머리와 상체 윗부분이 바닥에서 다소 들려 있어 완전히 의식을 잃고 추락하는 상태와 다름."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 클로즈업 프레이밍에 비교적 가깝게 피사체를 담아냈으며, 프레임 하단에 바닥 공간을 비워두어 상체가 기울어지며 추락하는 찰나의 방향성과 위태로움을 성공적으로 연출했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "클로즈업 지정에도 불구하고 피사체 전신이 보이는 풀샷에 가깝게 렌더링되었으며, 이미 상체가 바닥에 닿아 있어 텍스트가 요구한 '기울어지는 위태로운 찰나'와 하단 여백의 의도를 살리지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "반쯤 감긴 눈은 아래쪽 바닥을 향하고 있으며, 마취 화살은 목 측면에서 바깥을 향해 꽂혀 있음.",
        "built_space": "흙으로 이루어진 구덩이 내부 벽면과 바닥이 보이며, 프레임 좌측 하단으로 추락할 수 있는 여유 공간이 확보됨.",
        "entities": "이현우(레퍼런스와 일치하는 얼굴, 헝클어진 머리, 낡은 후드티), 마취 화살(나무 깃과 대).",
        "hard_violations": [],
        "physics": "하체(프레임 밖)가 지탱하는 가운데 상체가 중력에 의해 바닥으로 기울어지고 있어 의식을 잃어가는 순간의 물리적 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선은 약간 앞쪽 바닥을 향하며 반쯤 감긴 눈을 함. 화살은 목에 꽂혀 있음.",
        "built_space": "좁고 긴 흙 구덩이 바닥에 위치하며 상단에는 숲의 나무와 하늘이 보임.",
        "entities": "이현우(얼굴, 후드티 일치), 마취 화살, 다리에 감긴 붕대 일부 노출됨.",
        "hard_violations": [],
        "physics": "양손과 몸이 이미 바닥에 엎드려 지탱하고 있으나, 머리와 상체 윗부분이 바닥에서 다소 들려 있어 완전히 의식을 잃고 추락하는 상태와 다름."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 클로즈업 프레이밍에 비교적 가깝게 피사체를 담아냈으며, 프레임 하단에 바닥 공간을 비워두어 상체가 기울어지며 추락하는 찰나의 방향성과 위태로움을 성공적으로 연출했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "클로즈업 지정에도 불구하고 피사체 전신이 보이는 풀샷에 가깝게 렌더링되었으며, 이미 상체가 바닥에 닿아 있어 텍스트가 요구한 '기울어지는 위태로운 찰나'와 하단 여백의 의도를 살리지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "반쯤 감긴 눈은 아래쪽 바닥을 향하고 있으며, 마취 화살은 목 측면에서 바깥을 향해 꽂혀 있음.",
        "built_space": "흙으로 이루어진 구덩이 내부 벽면과 바닥이 보이며, 프레임 좌측 하단으로 추락할 수 있는 여유 공간이 확보됨.",
        "entities": "이현우(레퍼런스와 일치하는 얼굴, 헝클어진 머리, 낡은 후드티), 마취 화살(나무 깃과 대).",
        "hard_violations": [],
        "physics": "하체(프레임 밖)가 지탱하는 가운데 상체가 중력에 의해 바닥으로 기울어지고 있어 의식을 잃어가는 순간의 물리적 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선은 약간 앞쪽 바닥을 향하며 반쯤 감긴 눈을 함. 화살은 목에 꽂혀 있음.",
        "built_space": "좁고 긴 흙 구덩이 바닥에 위치하며 상단에는 숲의 나무와 하늘이 보임.",
        "entities": "이현우(얼굴, 후드티 일치), 마취 화살, 다리에 감긴 붕대 일부 노출됨.",
        "hard_violations": [],
        "physics": "양손과 몸이 이미 바닥에 엎드려 지탱하고 있으나, 머리와 상체 윗부분이 바닥에서 다소 들려 있어 완전히 의식을 잃고 추락하는 상태와 다름."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "목의 화살과 반쯤 감긴 눈, 흙바닥은 맞지만, 요구한 클로즈업보다 넓고 고개가 카메라 쪽으로 들려 있어 아래로 무너지는 찰나가 덜 선명하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "얼굴과 목 중심의 클로즈업 안에서 반쯤 감긴 눈, 아래로 처지는 상체, 목에 박힌 화살과 바로 아래 착지 공간을 가장 충실하게 연결한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈은 가늘게 열려 있고 시선은 바닥보다 카메라 앞쪽을 향한다. 머리는 화면 왼쪽으로 기울었으나 얼굴은 비교적 정면으로 들려 있다. 몸통은 앞쪽 흙바닥으로 낮아져 있다. 화살의 촉 쪽은 가까운 목 옆에 박혀 있고 깃 달린 꼬리는 화면 오른쪽으로 돌출된다.",
        "built_space": "좌우 흙벽과 뒤쪽 흙벽, 위쪽의 열린 숲과 아래 흙바닥이 보인다. 노출된 뿌리와 흙의 질감, 자연 주광은 장소 참고와 부합한다. 인물은 구덩이 바닥에 있으며 앞쪽 착지 공간이 넓게 확보되어 있다. 별도의 인공 고정 설비나 반사는 없다. 다만 주변 공간과 양손까지 넓게 담아 얼굴 중심 클로즈업보다 넓다.",
        "entities": "인물은 한 명뿐이며 젊은 한국계 남성으로 보이는 얼굴, 헝클어진 검은 머리, 얼굴의 상처와 낡고 구멍 난 짙은 회색 후드가 참고와 대체로 일치한다. 목에는 밝은 나무 축과 갈색 깃을 가진 작은 화살 한 개가 보인다. 참고 화살의 재질과 깃 모양을 따르지만 축은 훨씬 짧으며 촉은 삽입되어 확인할 수 없다. 다른 인물이나 읽을 수 있는 글자는 없다. 신발 속 카드와 구덩이 밖의 지속 소품은 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "화면 왼쪽 아래의 굽힌 팔과 팔꿈치 부근, 오른쪽 손이 흙바닥에 닿아 상체를 받친다. 따라서 몸이 지지 없이 떠 있는 것은 아니다. 화살은 목에 박힌 접점으로 고정되어 있다. 머리와 어깨를 아직 버티는 느낌이 있어 완전히 힘이 풀려 내려앉는 동작은 약하지만, 의식을 잃어가는 중간 순간으로는 물리적으로 가능하다."
       },
       {
        "label": "B",
        "direction": "반쯤 감긴 눈은 얼굴 바로 아래의 흙바닥을 향한다. 머리는 화면 왼쪽 아래로 처지고 몸통도 그 방향으로 기울어, 낙하 목표가 하단 왼쪽의 빈 바닥으로 읽힌다. 화살은 카메라에 가까운 목 옆에 박혀 있으며 깃 달린 꼬리가 오른쪽 위로 돌출된다.",
        "built_space": "좌우와 뒤쪽의 뿌리 드러난 흙벽, 자갈과 흙이 섞인 바닥이 보인다. 얼굴과 상체가 화면을 크게 차지하면서도 머리 아래와 왼쪽 하단에 착지할 바닥을 남긴다. 구덩이 내부의 낮은 시점이며 참고의 흙벽 재질과 주간 조명을 유지한다. 별도의 고정 설비나 반사는 없다. 구덩이 입구 전체가 잘린 것은 클로즈업에 맞는 크롭이다.",
        "entities": "인물은 한 명이며 참고와 닮은 젊은 한국계 남성의 얼굴, 검은 헝클어진 머리, 얼굴 상처, 해진 회색 후드와 일부 청바지가 보인다. 가까운 목에 작은 화살 한 개가 명확하게 박혀 있다. 밝은 나무 축과 줄무늬 갈색 깃은 소품 참고를 따르지만 전체 축 길이는 짧게 표현되었다. 눈은 정상적인 사람 눈으로 표현되고 글자나 추가 인물은 없다. 다리 붕대와 신발 속 카드는 크롭 밖이라 평가할 수 없다.",
        "hard_violations": [],
        "physics": "몸 아래로 접힌 팔의 팔꿈치·전완 부근이 흙바닥에 내려와 있고, 화면 오른쪽 아래의 골반과 허벅지도 바닥 쪽으로 낮아져 있다. 상체는 이 지지점 위에서 기울며 머리가 아래로 처져 있어 무지지 공중 부양으로 보이지 않는다. 화살은 목의 삽입 부위에 고정되고 후드 끈은 아래로 늘어진다. 완전히 바닥에 누운 이후가 아니라 지지가 무너지며 내려앉는 순간으로 자연스럽게 읽힌다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "목의 화살과 반쯤 감긴 눈, 흙바닥은 맞지만, 요구한 클로즈업보다 넓고 고개가 카메라 쪽으로 들려 있어 아래로 무너지는 찰나가 덜 선명하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "얼굴과 목 중심의 클로즈업 안에서 반쯤 감긴 눈, 아래로 처지는 상체, 목에 박힌 화살과 바로 아래 착지 공간을 가장 충실하게 연결한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 눈은 가늘게 열려 있고 시선은 바닥보다 카메라 앞쪽을 향한다. 머리는 화면 왼쪽으로 기울었으나 얼굴은 비교적 정면으로 들려 있다. 몸통은 앞쪽 흙바닥으로 낮아져 있다. 화살의 촉 쪽은 가까운 목 옆에 박혀 있고 깃 달린 꼬리는 화면 오른쪽으로 돌출된다.",
        "built_space": "좌우 흙벽과 뒤쪽 흙벽, 위쪽의 열린 숲과 아래 흙바닥이 보인다. 노출된 뿌리와 흙의 질감, 자연 주광은 장소 참고와 부합한다. 인물은 구덩이 바닥에 있으며 앞쪽 착지 공간이 넓게 확보되어 있다. 별도의 인공 고정 설비나 반사는 없다. 다만 주변 공간과 양손까지 넓게 담아 얼굴 중심 클로즈업보다 넓다.",
        "entities": "인물은 한 명뿐이며 젊은 한국계 남성으로 보이는 얼굴, 헝클어진 검은 머리, 얼굴의 상처와 낡고 구멍 난 짙은 회색 후드가 참고와 대체로 일치한다. 목에는 밝은 나무 축과 갈색 깃을 가진 작은 화살 한 개가 보인다. 참고 화살의 재질과 깃 모양을 따르지만 축은 훨씬 짧으며 촉은 삽입되어 확인할 수 없다. 다른 인물이나 읽을 수 있는 글자는 없다. 신발 속 카드와 구덩이 밖의 지속 소품은 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "화면 왼쪽 아래의 굽힌 팔과 팔꿈치 부근, 오른쪽 손이 흙바닥에 닿아 상체를 받친다. 따라서 몸이 지지 없이 떠 있는 것은 아니다. 화살은 목에 박힌 접점으로 고정되어 있다. 머리와 어깨를 아직 버티는 느낌이 있어 완전히 힘이 풀려 내려앉는 동작은 약하지만, 의식을 잃어가는 중간 순간으로는 물리적으로 가능하다."
       },
       {
        "label": "A",
        "direction": "반쯤 감긴 눈은 얼굴 바로 아래의 흙바닥을 향한다. 머리는 화면 왼쪽 아래로 처지고 몸통도 그 방향으로 기울어, 낙하 목표가 하단 왼쪽의 빈 바닥으로 읽힌다. 화살은 카메라에 가까운 목 옆에 박혀 있으며 깃 달린 꼬리가 오른쪽 위로 돌출된다.",
        "built_space": "좌우와 뒤쪽의 뿌리 드러난 흙벽, 자갈과 흙이 섞인 바닥이 보인다. 얼굴과 상체가 화면을 크게 차지하면서도 머리 아래와 왼쪽 하단에 착지할 바닥을 남긴다. 구덩이 내부의 낮은 시점이며 참고의 흙벽 재질과 주간 조명을 유지한다. 별도의 고정 설비나 반사는 없다. 구덩이 입구 전체가 잘린 것은 클로즈업에 맞는 크롭이다.",
        "entities": "인물은 한 명이며 참고와 닮은 젊은 한국계 남성의 얼굴, 검은 헝클어진 머리, 얼굴 상처, 해진 회색 후드와 일부 청바지가 보인다. 가까운 목에 작은 화살 한 개가 명확하게 박혀 있다. 밝은 나무 축과 줄무늬 갈색 깃은 소품 참고를 따르지만 전체 축 길이는 짧게 표현되었다. 눈은 정상적인 사람 눈으로 표현되고 글자나 추가 인물은 없다. 다리 붕대와 신발 속 카드는 크롭 밖이라 평가할 수 없다.",
        "hard_violations": [],
        "physics": "몸 아래로 접힌 팔의 팔꿈치·전완 부근이 흙바닥에 내려와 있고, 화면 오른쪽 아래의 골반과 허벅지도 바닥 쪽으로 낮아져 있다. 상체는 이 지지점 위에서 기울며 머리가 아래로 처져 있어 무지지 공중 부양으로 보이지 않는다. 화살은 목의 삽입 부위에 고정되고 후드 끈은 아래로 늘어진다. 완전히 바닥에 누운 이후가 아니라 지지가 무너지며 내려앉는 순간으로 자연스럽게 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.349
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.349
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1349
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 클로즈업 프레이밍에 비교적 가깝게 피사체를 담아냈으며, 프레임 하단에 바닥 공간을 비워두어 상체가 기울어지며 추락하는 찰나의 방향성과 위태로움을 성공적으로 연출했습니다."
   },
   {
    "label": "B",
    "score": 1349,
    "verdict_ko": "클로즈업 지정에도 불구하고 피사체 전신이 보이는 풀샷에 가깝게 렌더링되었으며, 이미 상체가 바닥에 닿아 있어 텍스트가 요구한 '기울어지는 위태로운 찰나'와 하단 여백의 의도를 살리지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S56sh20_sel.png",
    "asset_id": "0ed3fcd8-01e3-4e9f-8c0b-a7ec7fdfd967",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1218943>",
    "asset_id": "717460f5-31cc-4f17-a49b-efdad9614ee1",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 화살: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:883226>",
    "asset_id": "2134b4ba-9e55-444f-adb9-a28ecdc7f924",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde6-abaf-720e-bbe8-1b45bf4a50d2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S56sh20"
  }
 },
 "S56sh22::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:47:06.743183+00:00",
  "fingerprint": "60d8964760267118e7d7d84092c5c3d28c89059ceb77bc175e32e599efad9855",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S56sh22_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S56sh22_sel.png",
  "source_sha256": "d0708ef4ff7479198b2e67a947a9f634841209f2e3f82e0c1664079836f7cfa5",
  "file": "S56sh22_cine.png",
  "staged_sha256": "cacb53efbf3a3fcc0b065542f6ffa4bc00dd09581d571fc603773f5acfe91ccf",
  "latency_ms": 12591
 },
 "S57sh4::signage": {
  "fp": "f356e547fff510f9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::69a6a467d69c6a3d": {
  "subjects": [
   {
    "subject_native": "한국 전통 탈",
    "search_terms_native": [
     "한국 전통 탈",
     "하회탈 양반탈",
     "탈춤 가면"
    ],
    "language_lock_native": "반드시 한국어로만 검색하고 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "텍스트 모델은 한국 고유의 해학적인 전통 탈 대신 일본 가면이나 서양식 가면으로 왜곡할 위험이 높습니다."
   }
  ],
  "subject_text": "익산 마을 성당 앞\n성당 정면 출입구 앞의 열린 공간. 두 팔을 벌린 종교 인물 조각상 얼굴에 전통 탈 모양의 가면이 씌워져 있다.",
  "identity": "canonical",
  "scope_id": "L77",
  "scope_role": "location_exterior",
  "scope_sha": "f6cc34af3071a0f3"
 },
 "era_fail::c309b85d2d306d2b": {
  "stage": "research",
  "subject": "한국 전통 탈",
  "terms": [
   "한국 전통 탈",
   "하회탈 양반탈",
   "탈춤 가면"
  ],
  "status": "no_usable",
  "queries": [
   [
    "하회탈 양반탈 특징 국가유산"
   ]
  ],
  "candidate_urls": [
   "https://asset.m-gs.kr/prod/34855497/1/550",
   "https://www.antiquealive.com/upfile/8_2%2811%29.jpg",
   "https://shopofkorea.com/data/goods/1/2009/11/951_tmp_9e613e45dfe9390cf1a8f5f94953285a1892view.jpg",
   "https://i.etsystatic.com/24083336/r/il/b64e13/6996791191/il_1140xN.6996791191_lbai.jpg"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 100
   },
   {
    "index": 2,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 100
   },
   {
    "index": 3,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 100
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 100
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 2
 },
 "S57sh4::bgfirst_bg": {
  "input_fingerprint": "33f2bbb1dee75199",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 철창 밖으로 창과 몽둥이를 치켜든 하회탈 마스크의 아시아계 군중이 입을 크게 벌려 함성을 지르는 역동적인 전경.\n\nLOCATION (lock): Along the village street outside a prisoner cage carried on a truck, amid a crowd carrying spears and clubs.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Cage side bars (Confine 이현우 and 앰버 on the moving truck) — Seen from inside, narrowly bordering the outward view; used as Edge framing that establishes confinement without obscuring crowd gestures; Raised spears and clubs (Held aloft by the shouting crowd) — Different shaft angles and grip heights interrupt the upper frame; used as Extend the crowd's asynchronous threatening movement above their heads.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight renders the masked crowd clearly, using controlled contrast to retain readable gestures beyond the cage bars.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 철창 밖으로 창과 몽둥이를 치켜든 하회탈 마스크의 아시아계 군중이 입을 크게 벌려 함성을 지르는 역동적인 전경.\n\nLOCATION (lock): Along the village street outside a prisoner cage carried on a truck, amid a crowd carrying spears and clubs.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Cage side bars (Confine 이현우 and 앰버 on the moving truck) — Seen from inside, narrowly bordering the outward view; used as Edge framing that establishes confinement without obscuring crowd gestures; Raised spears and clubs (Held aloft by the shouting crowd) — Different shaft angles and grip heights interrupt the upper frame; used as Extend the crowd's asynchronous threatening movement above their heads.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight renders the masked crowd clearly, using controlled contrast to retain readable gestures beyond the cage bars.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S57sh4__bgfirst_bg.png",
  "asset_id": "8c30267c-6b8b-46f6-aad3-22656daef1db",
  "input_asset_ids": [
   "8a7d8f3c-a885-4e37-b8d7-1871e803ac45",
   "af710781-9127-4fcc-98c0-94ae3943bb3d"
  ]
 },
 "S57sh4": {
  "input_fingerprint": "07312ec688eea405",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 철창 밖으로 창과 몽둥이를 치켜든 하회탈 마스크의 아시아계 군중이 입을 크게 벌려 함성을 지르는 역동적인 전경.\n\nLOCATION (lock): Along the village street outside a prisoner cage carried on a truck, amid a crowd carrying spears and clubs. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Cage side bars (Confine 이현우 and 앰버 on the moving truck) — Seen from inside, narrowly bordering the outward view; used as Edge framing that establishes confinement without obscuring crowd gestures; Raised spears and clubs (Held aloft by the shouting crowd) — Different shaft angles and grip heights interrupt the upper frame; used as Extend the crowd's asynchronous threatening movement above their heads.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight renders the masked crowd clearly, using controlled contrast to retain readable gestures beyond the cage bars.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A barred cage is mounted on the moving truck, and a tall water-tank tower stands in the village. Water bottles are being distributed along the route.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 하회탈 마스크의 아시아계 군중 right now, so 하회탈 마스크의 아시아계 군중's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 하회탈 마스크의 아시아계 군중: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 철창 밖으로 창과 몽둥이를 치켜든 하회탈 마스크의 아시아계 군중이 입을 크게 벌려 함성을 지르는 역동적인 전경.\n\nLOCATION (lock): Along the village street outside a prisoner cage carried on a truck, amid a crowd carrying spears and clubs. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Cage side bars (Confine 이현우 and 앰버 on the moving truck) — Seen from inside, narrowly bordering the outward view; used as Edge framing that establishes confinement without obscuring crowd gestures; Raised spears and clubs (Held aloft by the shouting crowd) — Different shaft angles and grip heights interrupt the upper frame; used as Extend the crowd's asynchronous threatening movement above their heads.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight renders the masked crowd clearly, using controlled contrast to retain readable gestures beyond the cage bars.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A barred cage is mounted on the moving truck, and a tall water-tank tower stands in the village. Water bottles are being distributed along the route.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 하회탈 마스크의 아시아계 군중 right now, so 하회탈 마스크의 아시아계 군중's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 하회탈 마스크의 아시아계 군중: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 철창 밖으로 창과 몽둥이를 치켜든 하회탈 마스크의 아시아계 군중이 입을 크게 벌려 함성을 지르는 역동적인 전경.\n\nLOCATION (lock): Along the village street outside a prisoner cage carried on a truck, amid a crowd carrying spears and clubs. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Cage side bars (Confine 이현우 and 앰버 on the moving truck) — Seen from inside, narrowly bordering the outward view; used as Edge framing that establishes confinement without obscuring crowd gestures; Raised spears and clubs (Held aloft by the shouting crowd) — Different shaft angles and grip heights interrupt the upper frame; used as Extend the crowd's asynchronous threatening movement above their heads.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight renders the masked crowd clearly, using controlled contrast to retain readable gestures beyond the cage bars.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A barred cage is mounted on the moving truck, and a tall water-tank tower stands in the village. Water bottles are being distributed along the route.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 하회탈 마스크의 아시아계 군중 right now, so 하회탈 마스크의 아시아계 군중's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 하회탈 마스크의 아시아계 군중: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S57sh4__bgfirst_bg.png",
     "asset_id": "8c30267c-6b8b-46f6-aad3-22656daef1db",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S57sh4.png",
     "asset_id": "8a7d8f3c-a885-4e37-b8d7-1871e803ac45",
     "role": "conti_light"
    },
    {
     "label": "PROP REFERENCE — 창: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:962114>",
     "asset_id": "c2369c07-dcb1-40c9-93a0-aafad7bacf3a",
     "role": "prop_ref"
    },
    {
     "label": "PROP REFERENCE — 철창 케이지: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1148866>",
     "asset_id": "3e096b98-d7fb-48f0-b57f-d01d3f21144a",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L77B01.png",
     "asset_id": "af710781-9127-4fcc-98c0-94ae3943bb3d",
     "role": "location_plate"
    },
    {
     "label": "PROP REFERENCE — 창: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:962114>",
     "asset_id": "c2369c07-dcb1-40c9-93a0-aafad7bacf3a",
     "role": "prop_ref"
    },
    {
     "label": "PROP REFERENCE — 철창 케이지: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1148866>",
     "asset_id": "3e096b98-d7fb-48f0-b57f-d01d3f21144a",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "군중의 시선이 카메라를 향하며 무기들은 위를 가리킴.",
    "built_space": "전경에 나무 창살이 배치되었고, 배경에 레퍼런스에 없는 망루가 추가됨.",
    "entities": "군중의 얼굴 피부가 탈처럼 변형됨. 창의 형태가 레퍼런스와 다름.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 없는 구조물(망루) 추가",
     "[gemini-pro] 인체 구조 변형(피부가 탈 형태로 기괴하게 변형됨)",
     "[gemini-pro] 지정된 철창 대신 나무 창살 사용"
    ],
    "physics": "인물들의 손이 무기를 단단히 쥐고 있음."
   },
   {
    "label": "B",
    "direction": "군중이 철창 안쪽을 향해 시선을 두고 소리치며 무기를 치켜듦.",
    "built_space": "전경의 철창과 배경의 교회 건물 및 석재 질감이 레퍼런스와 정확히 일치함.",
    "entities": "하회탈을 쓴 아시아계 군중, 레퍼런스와 일치하는 창과 철창, 바닥에 놓인 생수병, 철창 안 인물들의 뒷모습.",
    "hard_violations": [],
    "physics": "손으로 무기와 철창을 자연스럽게 쥐고 있으며, 바닥에 생수병이 안정적으로 놓여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "레퍼런스와 다른 장소와 소품을 사용했으며, 인물의 얼굴이 탈과 혼합되어 기괴하게 변형되는 등 주요 지침을 위반했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 앵글, 레퍼런스와 완벽히 일치하는 장소와 소품(철창, 창), 탈 착용 상태 및 생수병 배치까지 모든 지침을 훌륭하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "군중의 시선이 카메라를 향하며 무기들은 위를 가리킴.",
        "built_space": "전경에 나무 창살이 배치되었고, 배경에 레퍼런스에 없는 망루가 추가됨.",
        "entities": "군중의 얼굴 피부가 탈처럼 변형됨. 창의 형태가 레퍼런스와 다름.",
        "hard_violations": [
         "레퍼런스에 없는 구조물(망루) 추가",
         "인체 구조 변형(피부가 탈 형태로 기괴하게 변형됨)",
         "지정된 철창 대신 나무 창살 사용"
        ],
        "physics": "인물들의 손이 무기를 단단히 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "군중이 철창 안쪽을 향해 시선을 두고 소리치며 무기를 치켜듦.",
        "built_space": "전경의 철창과 배경의 교회 건물 및 석재 질감이 레퍼런스와 정확히 일치함.",
        "entities": "하회탈을 쓴 아시아계 군중, 레퍼런스와 일치하는 창과 철창, 바닥에 놓인 생수병, 철창 안 인물들의 뒷모습.",
        "hard_violations": [],
        "physics": "손으로 무기와 철창을 자연스럽게 쥐고 있으며, 바닥에 생수병이 안정적으로 놓여 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "레퍼런스와 다른 장소와 소품을 사용했으며, 인물의 얼굴이 탈과 혼합되어 기괴하게 변형되는 등 주요 지침을 위반했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 앵글, 레퍼런스와 완벽히 일치하는 장소와 소품(철창, 창), 탈 착용 상태 및 생수병 배치까지 모든 지침을 훌륭하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "군중의 시선이 카메라를 향하며 무기들은 위를 가리킴.",
        "built_space": "전경에 나무 창살이 배치되었고, 배경에 레퍼런스에 없는 망루가 추가됨.",
        "entities": "군중의 얼굴 피부가 탈처럼 변형됨. 창의 형태가 레퍼런스와 다름.",
        "hard_violations": [
         "레퍼런스에 없는 구조물(망루) 추가",
         "인체 구조 변형(피부가 탈 형태로 기괴하게 변형됨)",
         "지정된 철창 대신 나무 창살 사용"
        ],
        "physics": "인물들의 손이 무기를 단단히 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "군중이 철창 안쪽을 향해 시선을 두고 소리치며 무기를 치켜듦.",
        "built_space": "전경의 철창과 배경의 교회 건물 및 석재 질감이 레퍼런스와 정확히 일치함.",
        "entities": "하회탈을 쓴 아시아계 군중, 레퍼런스와 일치하는 창과 철창, 바닥에 놓인 생수병, 철창 안 인물들의 뒷모습.",
        "hard_violations": [],
        "physics": "손으로 무기와 철창을 자연스럽게 쥐고 있으며, 바닥에 생수병이 안정적으로 놓여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "참조 성당의 장소 재현은 더 정확하지만, 전경 수감자와 화면 안쪽까지 들어온 철창이 시야를 차지해 군중 중심의 미디엄 숏과 좁은 가장자리 프레이밍에서 벗어난다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "철창 안에서 내다보는 군중 중심 미디엄 숏, 크게 벌린 입과 높이·각도가 다른 무기 동작을 더 충실히 구현하지만 성당 구조와 창 세부는 참조와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앞줄 군중은 철창 안의 두 전경 인물과 카메라 쪽을 바라보며 입을 벌린다. 창끝은 대체로 화면 오른쪽 위를 향하고 몽둥이도 머리 위로 올라가 있어, 수감자를 향해 찌르는 순간이 아니라 무기를 치켜들고 위협하는 순간이다. 가장 가까운 남자는 한 손으로 철창을 잡는다.",
        "built_space": "중앙의 열린 첨두아치 출입구 하나, 양쪽으로 열린 목문 두 짝, 상부의 작은 박공창 세 개, 왼쪽 석상과 오른쪽 낮은 부속 건물이 보여 참조 장소와 잘 맞는다. 가까운 케이지는 좌우에 각각 여러 세로봉과 약 다섯 높이의 가로대를 드러내며, 굵은 세로 기둥 두 개가 화면 안쪽까지 들어온다. 내부 인물 두 명의 머리와 어깨도 양쪽 전경을 크게 차지하므로 철창이 좁은 가장자리만 이루지는 않는다.",
        "entities": "성인 아시아계로 보이는 남녀 군중이 있고, 앞줄에는 흰색과 갈색 하회탈 형태의 마스크가 뚜렷하다. 뒤쪽에는 마스크를 쓰지 않은 듯한 얼굴도 보인다. 창은 갈색 목제 자루와 금속 날을 갖추지만 참조의 날 뿌리 구멍과 금속 띠 배치는 그대로 확인되지 않는다. 나무 몽둥이와 금속 케이지가 보이며, 케이지는 참조보다 격자 간격이 넓다. 내부의 두 인물은 뒷모습뿐이라 이현우와 앰버의 신원은 판별할 수 없다. 바닥에 물병 몇 개가 있으나 배급 행동은 없고, 트럭과 물탱크 탑은 화면에 나타나지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞줄 창 자루와 몽둥이는 손으로 잡혀 있고, 가까운 남자의 손은 세로 철봉을 감싼다. 뒤쪽 일부 인물의 발은 지면에 닿으며 앞줄 하체는 화면 아래로 잘려 있다. 상체의 기울기와 팔을 들어 올린 자세는 서서 함성을 지르는 동작으로 가능하다. 지지 없이 떠 있는 몸이나 무기는 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "군중의 얼굴과 열린 입이 케이지 내부의 카메라 쪽을 향한다. 앞줄 왼쪽 창은 왼쪽 위로 뻗어 끝이 프레임 밖으로 나가고, 중앙 창들은 거의 수직 또는 비스듬히 위를 향한다. 오른쪽 몽둥이는 오른쪽 위로 치켜들려 있다. 서로 다른 무기 각도와 손 높이가 수감자를 향한 집단적 위협과 함성을 만든다.",
        "built_space": "가까운 철창은 좌우의 굵은 세로봉 두 개와 하단 가로 구조로 주로 가장자리를 이루어 중앙 군중의 얼굴과 팔을 가리지 않는다. 군중은 케이지 밖에 밀집하고 앞줄은 허리 부근에서 잘리는 미디엄 숏이다. 뒤에는 석조 성당 출입구 하나와 열린 문, 왼쪽 석상 하나, 오른쪽 낮은 건물과 목제 탑 하나가 보인다. 다만 성당의 높은 단일 박공과 작은 첨탑은 참조의 세 박공창을 가진 낮고 넓은 정면과 달라 동일 장소의 건축 재현은 약하다.",
        "entities": "성인 아시아계로 보이는 군중이 하회탈을 닮은 갈색 얼굴 마스크와 전통식 겹옷·머리띠를 착용한다. 앞줄과 뒤쪽 모두 크게 벌어진 입이 반복되며, 일부 마스크는 실제 피부와의 경계가 덜 분명하다. 금속 날과 목제 자루의 창, 나무 몽둥이, 마모된 금속 케이지가 보인다. 창날의 윤곽과 결합부는 참조의 긴 홈과 구멍을 정확히 따르지 않는다. 오른쪽 뒤 탑은 있으나 덮개 때문에 물탱크 자체는 확인되지 않는다. 수감자, 트럭 차체와 물병 배급은 이 구도에 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞줄 왼쪽 인물의 손이 창 자루를 감싸고, 중앙 인물도 창을 움켜쥔다. 오른쪽 인물은 들어 올린 손으로 몽둥이 손잡이를 잡고 반대 손은 가슴 앞에서 쥔다. 팔꿈치와 어깨의 연결 및 앞으로 기운 몸통은 지면에 서서 무기를 드는 동작으로 가능하다. 발은 대체로 프레임 밖이거나 가려져 있지만 공중에 뜬 자세는 아니며, 명백히 지지 없이 떠 있는 무기도 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "참조 성당의 장소 재현은 더 정확하지만, 전경 수감자와 화면 안쪽까지 들어온 철창이 시야를 차지해 군중 중심의 미디엄 숏과 좁은 가장자리 프레이밍에서 벗어난다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "철창 안에서 내다보는 군중 중심 미디엄 숏, 크게 벌린 입과 높이·각도가 다른 무기 동작을 더 충실히 구현하지만 성당 구조와 창 세부는 참조와 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앞줄 군중은 철창 안의 두 전경 인물과 카메라 쪽을 바라보며 입을 벌린다. 창끝은 대체로 화면 오른쪽 위를 향하고 몽둥이도 머리 위로 올라가 있어, 수감자를 향해 찌르는 순간이 아니라 무기를 치켜들고 위협하는 순간이다. 가장 가까운 남자는 한 손으로 철창을 잡는다.",
        "built_space": "중앙의 열린 첨두아치 출입구 하나, 양쪽으로 열린 목문 두 짝, 상부의 작은 박공창 세 개, 왼쪽 석상과 오른쪽 낮은 부속 건물이 보여 참조 장소와 잘 맞는다. 가까운 케이지는 좌우에 각각 여러 세로봉과 약 다섯 높이의 가로대를 드러내며, 굵은 세로 기둥 두 개가 화면 안쪽까지 들어온다. 내부 인물 두 명의 머리와 어깨도 양쪽 전경을 크게 차지하므로 철창이 좁은 가장자리만 이루지는 않는다.",
        "entities": "성인 아시아계로 보이는 남녀 군중이 있고, 앞줄에는 흰색과 갈색 하회탈 형태의 마스크가 뚜렷하다. 뒤쪽에는 마스크를 쓰지 않은 듯한 얼굴도 보인다. 창은 갈색 목제 자루와 금속 날을 갖추지만 참조의 날 뿌리 구멍과 금속 띠 배치는 그대로 확인되지 않는다. 나무 몽둥이와 금속 케이지가 보이며, 케이지는 참조보다 격자 간격이 넓다. 내부의 두 인물은 뒷모습뿐이라 이현우와 앰버의 신원은 판별할 수 없다. 바닥에 물병 몇 개가 있으나 배급 행동은 없고, 트럭과 물탱크 탑은 화면에 나타나지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞줄 창 자루와 몽둥이는 손으로 잡혀 있고, 가까운 남자의 손은 세로 철봉을 감싼다. 뒤쪽 일부 인물의 발은 지면에 닿으며 앞줄 하체는 화면 아래로 잘려 있다. 상체의 기울기와 팔을 들어 올린 자세는 서서 함성을 지르는 동작으로 가능하다. 지지 없이 떠 있는 몸이나 무기는 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "군중의 얼굴과 열린 입이 케이지 내부의 카메라 쪽을 향한다. 앞줄 왼쪽 창은 왼쪽 위로 뻗어 끝이 프레임 밖으로 나가고, 중앙 창들은 거의 수직 또는 비스듬히 위를 향한다. 오른쪽 몽둥이는 오른쪽 위로 치켜들려 있다. 서로 다른 무기 각도와 손 높이가 수감자를 향한 집단적 위협과 함성을 만든다.",
        "built_space": "가까운 철창은 좌우의 굵은 세로봉 두 개와 하단 가로 구조로 주로 가장자리를 이루어 중앙 군중의 얼굴과 팔을 가리지 않는다. 군중은 케이지 밖에 밀집하고 앞줄은 허리 부근에서 잘리는 미디엄 숏이다. 뒤에는 석조 성당 출입구 하나와 열린 문, 왼쪽 석상 하나, 오른쪽 낮은 건물과 목제 탑 하나가 보인다. 다만 성당의 높은 단일 박공과 작은 첨탑은 참조의 세 박공창을 가진 낮고 넓은 정면과 달라 동일 장소의 건축 재현은 약하다.",
        "entities": "성인 아시아계로 보이는 군중이 하회탈을 닮은 갈색 얼굴 마스크와 전통식 겹옷·머리띠를 착용한다. 앞줄과 뒤쪽 모두 크게 벌어진 입이 반복되며, 일부 마스크는 실제 피부와의 경계가 덜 분명하다. 금속 날과 목제 자루의 창, 나무 몽둥이, 마모된 금속 케이지가 보인다. 창날의 윤곽과 결합부는 참조의 긴 홈과 구멍을 정확히 따르지 않는다. 오른쪽 뒤 탑은 있으나 덮개 때문에 물탱크 자체는 확인되지 않는다. 수감자, 트럭 차체와 물병 배급은 이 구도에 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞줄 왼쪽 인물의 손이 창 자루를 감싸고, 중앙 인물도 창을 움켜쥔다. 오른쪽 인물은 들어 올린 손으로 몽둥이 손잡이를 잡고 반대 손은 가슴 앞에서 쥔다. 팔꿈치와 어깨의 연결 및 앞으로 기운 몸통은 지면에 서서 무기를 드는 동작으로 가능하다. 발은 대체로 프레임 밖이거나 가려져 있지만 공중에 뜬 자세는 아니며, 명백히 지지 없이 떠 있는 무기도 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.875
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스에 없는 구조물(망루) 추가",
     "[gemini-pro] 인체 구조 변형(피부가 탈 형태로 기괴하게 변형됨)",
     "[gemini-pro] 지정된 철창 대신 나무 창살 사용"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "A": 1179,
   "B": 1875
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "레퍼런스와 다른 장소와 소품을 사용했으며, 인물의 얼굴이 탈과 혼합되어 기괴하게 변형되는 등 주요 지침을 위반했습니다.  ★위반: [gemini-pro] 레퍼런스에 없는 구조물(망루) 추가 / [gemini-pro] 인체 구조 변형(피부가 탈 형태로 기괴하게 변형됨) / [gemini-pro] 지정된 철창 대신 나무 창살 사용"
   },
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "지정된 앵글, 레퍼런스와 완벽히 일치하는 장소와 소품(철창, 창), 탈 착용 상태 및 생수병 배치까지 모든 지침을 훌륭하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L77B01.png",
    "asset_id": "af710781-9127-4fcc-98c0-94ae3943bb3d",
    "role": "location_plate"
   },
   {
    "label": "PROP REFERENCE — 창: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:962114>",
    "asset_id": "c2369c07-dcb1-40c9-93a0-aafad7bacf3a",
    "role": "prop_ref"
   },
   {
    "label": "PROP REFERENCE — 철창 케이지: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1148866>",
    "asset_id": "3e096b98-d7fb-48f0-b57f-d01d3f21144a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacde7-7aaa-751e-b034-1b19136a193a",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S57sh4__bgfirst_bg.png",
   "bg_asset_id": "8c30267c-6b8b-46f6-aad3-22656daef1db",
   "bg_record_key": "S57sh4::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S57sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:47:53.417530+00:00",
  "fingerprint": "7895c8066692125cfba6810c1a100385360826f4a5e385f3931e337265cc0dc4",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S57sh4_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S57sh4_sel.png",
  "source_sha256": "9a509a6e38536abe8886f1c56dce5a9925b0a3a217d20b70663b92d58efa6866",
  "file": "S57sh4_cine.png",
  "staged_sha256": "cf326a250aa222e160f229aab144539454bc3281937723054e60609d23b5e3d2",
  "latency_ms": 10748
 },
 "S57sh10::signage": {
  "fp": "5c4c29220f4acbac",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S57sh10": {
  "input_fingerprint": "30f360484375be9d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 팔을 벌린 거대한 예수 동상 얼굴 위에 기괴한 하회탈 마스크가 씌워진 상태의 클로즈업.\n\nLOCATION (lock): At the masked religious statue outside the village church's front entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Masked Jesus statue (A Hahoe mask covers the face of the statue whose arms are outstretched) — The masked face is seen from below at an oblique three-quarter angle; used as Isolated focal detail revealing the altered religious image; Church-like building (The stopped truck has arrived outside it) — An oblique exterior portion remains behind the statue detail; used as Provides location context and architectural scale; Sky (Visible beside the elevated statue); used as Negative space around the tightly selected face detail.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and restrained contrast make the imposed Hahoe mask readable without adding a supernatural glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Jesus statue outside the church-like building has outstretched arms and a Hahoe mask covering its face. The transport truck has stopped outside with its barred cage, and the tall water-tank tower remains in the village.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 팔을 벌린 거대한 예수 동상 얼굴 위에 기괴한 하회탈 마스크가 씌워진 상태의 클로즈업.\n\nLOCATION (lock): At the masked religious statue outside the village church's front entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Masked Jesus statue (A Hahoe mask covers the face of the statue whose arms are outstretched) — The masked face is seen from below at an oblique three-quarter angle; used as Isolated focal detail revealing the altered religious image; Church-like building (The stopped truck has arrived outside it) — An oblique exterior portion remains behind the statue detail; used as Provides location context and architectural scale; Sky (Visible beside the elevated statue); used as Negative space around the tightly selected face detail.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and restrained contrast make the imposed Hahoe mask readable without adding a supernatural glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Jesus statue outside the church-like building has outstretched arms and a Hahoe mask covering its face. The transport truck has stopped outside with its barred cage, and the tall water-tank tower remains in the village.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 팔을 벌린 거대한 예수 동상 얼굴 위에 기괴한 하회탈 마스크가 씌워진 상태의 클로즈업.\n\nLOCATION (lock): At the masked religious statue outside the village church's front entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Masked Jesus statue (A Hahoe mask covers the face of the statue whose arms are outstretched) — The masked face is seen from below at an oblique three-quarter angle; used as Isolated focal detail revealing the altered religious image; Church-like building (The stopped truck has arrived outside it) — An oblique exterior portion remains behind the statue detail; used as Provides location context and architectural scale; Sky (Visible beside the elevated statue); used as Negative space around the tightly selected face detail.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and restrained contrast make the imposed Hahoe mask readable without adding a supernatural glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Jesus statue outside the church-like building has outstretched arms and a Hahoe mask covering its face. The transport truck has stopped outside with its barred cage, and the tall water-tank tower remains in the village.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 마스크를 쓴 조각상의 얼굴을 아래에서 위로 비스듬히 바라보고 있으며, 조각상은 화면 우측을 향함.",
    "built_space": "배경에 아치형 입구를 가진 석조 교회 건물이 자리 잡고 있음.",
    "entities": "두 팔을 벌린 돌 재질의 예수 조각상, 천 끈으로 묶인 하회탈, 석조 교회 배경.",
    "hard_violations": [],
    "physics": "하회탈은 굵은 천 끈에 의해 조각상의 머리에 안정적으로 고정되어 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 조각상의 얼굴과 상반신을 올려다보고 있으며, 조각상은 화면 좌측을 향함.",
    "built_space": "화면 우측에 석조 교회가, 좌측 원경에 콘크리트 급수탑이 위치함.",
    "entities": "팔을 벌린 예수 조각상, 밧줄로 묶인 하회탈, 급수탑, 석조 교회.",
    "hard_violations": [],
    "physics": "하회탈이 밧줄로 머리에 묶여 지탱되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 '클로즈업' 프레이밍과 마스크 디테일에 집중하라는 지시를 가장 정확하게 준수했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "급수탑 등 더 많은 배경 요소를 보여주기 위해 프레임을 넓히면서 엄격한 클로즈업 지시를 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 마스크를 쓴 조각상의 얼굴을 아래에서 위로 비스듬히 바라보고 있으며, 조각상은 화면 우측을 향함.",
        "built_space": "배경에 아치형 입구를 가진 석조 교회 건물이 자리 잡고 있음.",
        "entities": "두 팔을 벌린 돌 재질의 예수 조각상, 천 끈으로 묶인 하회탈, 석조 교회 배경.",
        "hard_violations": [],
        "physics": "하회탈은 굵은 천 끈에 의해 조각상의 머리에 안정적으로 고정되어 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 조각상의 얼굴과 상반신을 올려다보고 있으며, 조각상은 화면 좌측을 향함.",
        "built_space": "화면 우측에 석조 교회가, 좌측 원경에 콘크리트 급수탑이 위치함.",
        "entities": "팔을 벌린 예수 조각상, 밧줄로 묶인 하회탈, 급수탑, 석조 교회.",
        "hard_violations": [],
        "physics": "하회탈이 밧줄로 머리에 묶여 지탱되고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 '클로즈업' 프레이밍과 마스크 디테일에 집중하라는 지시를 가장 정확하게 준수했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "급수탑 등 더 많은 배경 요소를 보여주기 위해 프레임을 넓히면서 엄격한 클로즈업 지시를 어겼습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 마스크를 쓴 조각상의 얼굴을 아래에서 위로 비스듬히 바라보고 있으며, 조각상은 화면 우측을 향함.",
        "built_space": "배경에 아치형 입구를 가진 석조 교회 건물이 자리 잡고 있음.",
        "entities": "두 팔을 벌린 돌 재질의 예수 조각상, 천 끈으로 묶인 하회탈, 석조 교회 배경.",
        "hard_violations": [],
        "physics": "하회탈은 굵은 천 끈에 의해 조각상의 머리에 안정적으로 고정되어 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 조각상의 얼굴과 상반신을 올려다보고 있으며, 조각상은 화면 좌측을 향함.",
        "built_space": "화면 우측에 석조 교회가, 좌측 원경에 콘크리트 급수탑이 위치함.",
        "entities": "팔을 벌린 예수 조각상, 밧줄로 묶인 하회탈, 급수탑, 석조 교회.",
        "hard_violations": [],
        "physics": "하회탈이 밧줄로 머리에 묶여 지탱되고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "얼굴을 덮은 탈과 벌린 팔은 구현했지만, 얼굴보다 상체와 주변을 넓게 보여 주며 교회 외관도 우선 적용해야 할 구조 참조와 다릅니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "아래에서 비스듬히 올려다본 하회탈 얼굴의 근접 구도가 더 충실하지만, 교회를 구조 참조의 흰 회벽 건물 대신 석조 고딕 외관으로 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "동상과 탈의 얼굴은 화면 왼쪽 앞을 향하고, 카메라는 얼굴의 정면과 오른쪽 측면을 함께 봅니다. 두 팔은 좌우로 벌어져 있으며 왼쪽 손은 화면 안에, 반대쪽 손은 화면 밖에 있습니다. 특정 대상을 바라보거나 가리키는 장면은 아닙니다. 비스듬한 얼굴 각도는 맞지만 올려다보는 느낌은 약합니다.",
        "built_space": "오른쪽 배경에 회색 석조 교회의 첨두 출입구 하나, 열린 목문 일부, 상부 스테인드글라스, 완전히 보이는 박공 하나와 우측에 잘린 박공 일부가 있습니다. 왼쪽에는 원통형 물탱크와 지지탑 하나가 보입니다. 석조 교회는 장소 사진을 따랐지만, 구조 외형의 우선 권한인 참조 사진의 흰 회벽·계단형 파라펫 교회와는 다릅니다. 반사나 중복 설비 문제는 보이지 않습니다.",
        "entities": "긴 머리와 수염, 긴 옷을 가진 밝은 돌 재질의 예수 동상 하나와 갈색 목제 탈 하나가 보입니다. 탈은 굽은 눈썹과 큰 코, 웃는 입을 갖추었으나 하회탈보다 장식적인 인상이 강하고 동상의 수염과 턱이 많이 드러납니다. 하늘과 교회, 물탱크가 보이며 실제 사람이나 읽을 수 있는 글자는 없습니다. 트럭은 근접 구도 밖이므로 부재를 결함으로 보지 않습니다.",
        "hard_violations": [],
        "physics": "탈은 머리 둘레를 감싼 밧줄과 측면 매듭으로 고정되어 있고, 남은 줄은 아래로 늘어집니다. 팔과 머리는 석조 몸통에 연결되어 있습니다. 받침대는 프레임 밖이지만 동상이 공중에 떠 있다는 징후는 없으며, 지지 없이 떠 있는 물체도 없습니다."
       },
       {
        "label": "B",
        "direction": "탈과 동상의 얼굴은 화면 오른쪽 앞을 향하고, 카메라는 아래에서 정면과 왼쪽 측면을 함께 봅니다. 두 팔은 좌우 화면 밖으로 뻗어 있습니다. 지정된 응시 대상은 없으며, 얼굴을 아래에서 비스듬한 삼사분면 각도로 보여 달라는 요구에 더 가깝습니다.",
        "built_space": "배경에는 회색 석조 교회의 상부 박공 세 개, 첨두 출입구 하나, 그 위 스테인드글라스와 매달린 등 하나, 왼쪽의 좁은 첨두창 하나가 보입니다. 출입구 목문 일부도 보입니다. 장소 사진의 고딕 석조 외관은 재현했지만, 우선 적용해야 하는 구조 참조의 흰 회벽과 계단형 파라펫, 둥근 아치 개구부는 따르지 않았습니다. 중복 설비나 불가능한 반사는 보이지 않습니다.",
        "entities": "긴 머리와 옷 주름을 가진 밝은 석조 예수 동상 하나의 얼굴에 갈색 목제 하회탈 하나가 씌워져 있습니다. 가늘게 굽은 눈, 도드라진 코, 웃는 입과 턱을 갖춘 탈은 하회탈로 더 명확하게 읽힙니다. 교회와 하늘이 배경에 있고 실제 사람이나 읽을 수 있는 글자는 없습니다. 트럭과 물탱크는 이 근접 구도에서 보이지 않아도 무방합니다.",
        "hard_violations": [],
        "physics": "탈 양옆의 밧줄 매듭과 머리를 감싼 띠가 탈을 얼굴에 고정하고 있습니다. 돌로 된 머리와 벌린 팔은 몸통에 이어져 있으며, 하부 지지대는 프레임 밖입니다. 지지 없이 떠 있는 물체나 물리적으로 불가능한 자세는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "얼굴을 덮은 탈과 벌린 팔은 구현했지만, 얼굴보다 상체와 주변을 넓게 보여 주며 교회 외관도 우선 적용해야 할 구조 참조와 다릅니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "아래에서 비스듬히 올려다본 하회탈 얼굴의 근접 구도가 더 충실하지만, 교회를 구조 참조의 흰 회벽 건물 대신 석조 고딕 외관으로 구현했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "동상과 탈의 얼굴은 화면 왼쪽 앞을 향하고, 카메라는 얼굴의 정면과 오른쪽 측면을 함께 봅니다. 두 팔은 좌우로 벌어져 있으며 왼쪽 손은 화면 안에, 반대쪽 손은 화면 밖에 있습니다. 특정 대상을 바라보거나 가리키는 장면은 아닙니다. 비스듬한 얼굴 각도는 맞지만 올려다보는 느낌은 약합니다.",
        "built_space": "오른쪽 배경에 회색 석조 교회의 첨두 출입구 하나, 열린 목문 일부, 상부 스테인드글라스, 완전히 보이는 박공 하나와 우측에 잘린 박공 일부가 있습니다. 왼쪽에는 원통형 물탱크와 지지탑 하나가 보입니다. 석조 교회는 장소 사진을 따랐지만, 구조 외형의 우선 권한인 참조 사진의 흰 회벽·계단형 파라펫 교회와는 다릅니다. 반사나 중복 설비 문제는 보이지 않습니다.",
        "entities": "긴 머리와 수염, 긴 옷을 가진 밝은 돌 재질의 예수 동상 하나와 갈색 목제 탈 하나가 보입니다. 탈은 굽은 눈썹과 큰 코, 웃는 입을 갖추었으나 하회탈보다 장식적인 인상이 강하고 동상의 수염과 턱이 많이 드러납니다. 하늘과 교회, 물탱크가 보이며 실제 사람이나 읽을 수 있는 글자는 없습니다. 트럭은 근접 구도 밖이므로 부재를 결함으로 보지 않습니다.",
        "hard_violations": [],
        "physics": "탈은 머리 둘레를 감싼 밧줄과 측면 매듭으로 고정되어 있고, 남은 줄은 아래로 늘어집니다. 팔과 머리는 석조 몸통에 연결되어 있습니다. 받침대는 프레임 밖이지만 동상이 공중에 떠 있다는 징후는 없으며, 지지 없이 떠 있는 물체도 없습니다."
       },
       {
        "label": "A",
        "direction": "탈과 동상의 얼굴은 화면 오른쪽 앞을 향하고, 카메라는 아래에서 정면과 왼쪽 측면을 함께 봅니다. 두 팔은 좌우 화면 밖으로 뻗어 있습니다. 지정된 응시 대상은 없으며, 얼굴을 아래에서 비스듬한 삼사분면 각도로 보여 달라는 요구에 더 가깝습니다.",
        "built_space": "배경에는 회색 석조 교회의 상부 박공 세 개, 첨두 출입구 하나, 그 위 스테인드글라스와 매달린 등 하나, 왼쪽의 좁은 첨두창 하나가 보입니다. 출입구 목문 일부도 보입니다. 장소 사진의 고딕 석조 외관은 재현했지만, 우선 적용해야 하는 구조 참조의 흰 회벽과 계단형 파라펫, 둥근 아치 개구부는 따르지 않았습니다. 중복 설비나 불가능한 반사는 보이지 않습니다.",
        "entities": "긴 머리와 옷 주름을 가진 밝은 석조 예수 동상 하나의 얼굴에 갈색 목제 하회탈 하나가 씌워져 있습니다. 가늘게 굽은 눈, 도드라진 코, 웃는 입과 턱을 갖춘 탈은 하회탈로 더 명확하게 읽힙니다. 교회와 하늘이 배경에 있고 실제 사람이나 읽을 수 있는 글자는 없습니다. 트럭과 물탱크는 이 근접 구도에서 보이지 않아도 무방합니다.",
        "hard_violations": [],
        "physics": "탈 양옆의 밧줄 매듭과 머리를 감싼 띠가 탈을 얼굴에 고정하고 있습니다. 돌로 된 머리와 벌린 팔은 몸통에 이어져 있으며, 하부 지지대는 프레임 밖입니다. 지지 없이 떠 있는 물체나 물리적으로 불가능한 자세는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.405
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.405
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1405
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 '클로즈업' 프레이밍과 마스크 디테일에 집중하라는 지시를 가장 정확하게 준수했습니다."
   },
   {
    "label": "B",
    "score": 1405,
    "verdict_ko": "급수탑 등 더 많은 배경 요소를 보여주기 위해 프레임을 넓히면서 엄격한 클로즈업 지시를 어겼습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L77B01.png",
    "asset_id": "af710781-9127-4fcc-98c0-94ae3943bb3d",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_ruined_village_church_sel.png",
    "asset_id": "390d0127-8678-4a11-846c-55ca44e9c904",
    "role": "structure_seed_look"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdea-47f2-7e7b-84f9-e1d9b7b95aed",
  "ref_mode": "플레이트+seed만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S57sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:48:04.510390+00:00",
  "fingerprint": "ad36cd6aacf1777113fdde2843b7aa130e0d2a46a7978bc51e9fb93df39875dc",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S57sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S57sh10_sel.png",
  "source_sha256": "44f423be89a32973c172a510affdf16cfc9dfd119064d12118ac0b5c5a9a074e",
  "file": "S57sh10_cine.png",
  "staged_sha256": "c70fb12f3d47a796e588a1a3f7cd47155131feaf235db91bec823e22b4266882",
  "latency_ms": 11496
 },
 "S57sh11::signage": {
  "fp": "9b7c1f8690a095b5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S57sh11": {
  "input_fingerprint": "45b7accf5e022cc1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밧줄에 결박된 이현우와 앰버가 하회탈 마스크의 아시아계 병사들에게 등 떠밀린 힘에 의해 성당 문 안쪽으로 몸이 비틀어지게 쏠린 mid-action 뒷모습.\n\nLOCATION (lock): At the village church's entrance threshold, where the exterior approach opens into a nave lit by daylight through stained glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open doorway receiving the prisoners' movement away from camera in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Building doorway (Open as the bound prisoners are forced inside) — The entrance opening faces the approaching group and camera obliquely; used as Upper-center destination framing the prisoners' forced movement; Binding ropes (Restrain 이현우 and 앰버 during the forced entry) — Visible portions follow their twisting bodies; used as Small narrative details distinguishing loss of balance from voluntary movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Carry exterior daylight naturally to the threshold without specifying an unsupported interior light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The masked, outstretched-armed Jesus statue remains outside the church entrance. The transport truck and its barred cage remain stopped outside; the village water tower is intact. 이현우: He is being taken into the church while still bound, with his treated leg bandaged. His contact card remains concealed in his shoe. 앰버: She is being taken into the church while still bound and without an established new injury.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밧줄에 결박된 이현우와 앰버가 하회탈 마스크의 아시아계 병사들에게 등 떠밀린 힘에 의해 성당 문 안쪽으로 몸이 비틀어지게 쏠린 mid-action 뒷모습.\n\nLOCATION (lock): At the village church's entrance threshold, where the exterior approach opens into a nave lit by daylight through stained glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open doorway receiving the prisoners' movement away from camera in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Building doorway (Open as the bound prisoners are forced inside) — The entrance opening faces the approaching group and camera obliquely; used as Upper-center destination framing the prisoners' forced movement; Binding ropes (Restrain 이현우 and 앰버 during the forced entry) — Visible portions follow their twisting bodies; used as Small narrative details distinguishing loss of balance from voluntary movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Carry exterior daylight naturally to the threshold without specifying an unsupported interior light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The masked, outstretched-armed Jesus statue remains outside the church entrance. The transport truck and its barred cage remain stopped outside; the village water tower is intact. 이현우: He is being taken into the church while still bound, with his treated leg bandaged. His contact card remains concealed in his shoe. 앰버: She is being taken into the church while still bound and without an established new injury.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밧줄에 결박된 이현우와 앰버가 하회탈 마스크의 아시아계 병사들에게 등 떠밀린 힘에 의해 성당 문 안쪽으로 몸이 비틀어지게 쏠린 mid-action 뒷모습.\n\nLOCATION (lock): At the village church's entrance threshold, where the exterior approach opens into a nave lit by daylight through stained glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open doorway receiving the prisoners' movement away from camera in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Building doorway (Open as the bound prisoners are forced inside) — The entrance opening faces the approaching group and camera obliquely; used as Upper-center destination framing the prisoners' forced movement; Binding ropes (Restrain 이현우 and 앰버 during the forced entry) — Visible portions follow their twisting bodies; used as Small narrative details distinguishing loss of balance from voluntary movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Carry exterior daylight naturally to the threshold without specifying an unsupported interior light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The masked, outstretched-armed Jesus statue remains outside the church entrance. The transport truck and its barred cage remain stopped outside; the village water tower is intact. 이현우: He is being taken into the church while still bound, with his treated leg bandaged. His contact card remains concealed in his shoe. 앰버: She is being taken into the church while still bound and without an established new injury.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "병사들이 성당 입구 안쪽에서 죄수들과 카메라를 향해 무기를 겨누고 있으며, 죄수들은 양옆으로 당겨지는 형태입니다.",
    "built_space": "중앙에 열린 성당 문과 스테인드글라스가 보이며, 왼쪽에 예수상과 트럭, 오른쪽에 급수탑이 있습니다.",
    "entities": "이현우와 앰버의 복장과 붕대는 일치하나, 병사들이 죄수의 뒤가 아닌 앞에 서 있습니다. 예수상의 마스크는 나무가 아닌 돌에 새겨진 형태로 변형되었습니다.",
    "hard_violations": [
     "[gemini-pro] 아무도 잡고 있지 않은 밧줄이 허공에 뜬 채로 양옆 화면 밖을 향해 팽팽하게 당겨져 있음 (물리적 지지 기반 부재).",
     "[gemini-pro] 프롬프트의 '등 떠밀린' 묘사와 반대로 병사들이 죄수들 앞에 서서 입구를 막고 있음 (명시된 연출의 정반대 배치).",
     "[gpt-high] 등 뒤에서 포로들을 실내로 밀어야 할 병사들이 모두 포로 앞의 문 안쪽에 배치되어 출입을 가로막는다. 이는 명시된 가해자 위치와 강제 진입 동작을 반대로 만든다."
    ],
    "physics": "죄수들을 양옆으로 강하게 당기는 밧줄이 화면 밖으로 이어지지만, 이를 잡거나 당기는 물리적 주체가 전혀 없습니다."
   },
   {
    "label": "B",
    "direction": "죄수들이 성당 입구를 향해 몸이 쏠려 있으며, 뒤에 위치한 병사가 앰버의 등을 안쪽으로 밀고 있습니다.",
    "built_space": "화면 상단 중앙에 열린 성당 문이 있고 내부에 스테인드글라스가 보입니다. 왼쪽에 마스크를 끈으로 묶은 예수상, 오른쪽에 철창 트럭이 배치되어 있습니다.",
    "entities": "이현우는 오른쪽 다리에 붕대를 감고 밧줄에 묶인 채 비틀거리고 있으며, 앰버 역시 묶인 상태로 밀려납니다. 하회탈을 쓴 아시아계 병사들이 그들의 뒤에 위치합니다.",
    "hard_violations": [],
    "physics": "병사의 손이 앰버의 등에 닿아 미는 힘을 제대로 표현하고 있으며, 인물들의 발과 자세가 지면에 자연스럽게 지지되어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "병사들에게 등 떠밀려 성당 안으로 쏠리는 죄수들의 역동적인 뒷모습과 지정된 배경 요소들을 충실하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "병사들이 뒤에서 미는 대신 앞에서 길을 막고 서 있으며, 허공에서 팽팽하게 당겨지는 밧줄 등 심각한 물리적 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "죄수들이 성당 입구를 향해 몸이 쏠려 있으며, 뒤에 위치한 병사가 앰버의 등을 안쪽으로 밀고 있습니다.",
        "built_space": "화면 상단 중앙에 열린 성당 문이 있고 내부에 스테인드글라스가 보입니다. 왼쪽에 마스크를 끈으로 묶은 예수상, 오른쪽에 철창 트럭이 배치되어 있습니다.",
        "entities": "이현우는 오른쪽 다리에 붕대를 감고 밧줄에 묶인 채 비틀거리고 있으며, 앰버 역시 묶인 상태로 밀려납니다. 하회탈을 쓴 아시아계 병사들이 그들의 뒤에 위치합니다.",
        "hard_violations": [],
        "physics": "병사의 손이 앰버의 등에 닿아 미는 힘을 제대로 표현하고 있으며, 인물들의 발과 자세가 지면에 자연스럽게 지지되어 있습니다."
       },
       {
        "label": "A",
        "direction": "병사들이 성당 입구 안쪽에서 죄수들과 카메라를 향해 무기를 겨누고 있으며, 죄수들은 양옆으로 당겨지는 형태입니다.",
        "built_space": "중앙에 열린 성당 문과 스테인드글라스가 보이며, 왼쪽에 예수상과 트럭, 오른쪽에 급수탑이 있습니다.",
        "entities": "이현우와 앰버의 복장과 붕대는 일치하나, 병사들이 죄수의 뒤가 아닌 앞에 서 있습니다. 예수상의 마스크는 나무가 아닌 돌에 새겨진 형태로 변형되었습니다.",
        "hard_violations": [
         "아무도 잡고 있지 않은 밧줄이 허공에 뜬 채로 양옆 화면 밖을 향해 팽팽하게 당겨져 있음 (물리적 지지 기반 부재).",
         "프롬프트의 '등 떠밀린' 묘사와 반대로 병사들이 죄수들 앞에 서서 입구를 막고 있음 (명시된 연출의 정반대 배치)."
        ],
        "physics": "죄수들을 양옆으로 강하게 당기는 밧줄이 화면 밖으로 이어지지만, 이를 잡거나 당기는 물리적 주체가 전혀 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "병사들에게 등 떠밀려 성당 안으로 쏠리는 죄수들의 역동적인 뒷모습과 지정된 배경 요소들을 충실하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "병사들이 뒤에서 미는 대신 앞에서 길을 막고 서 있으며, 허공에서 팽팽하게 당겨지는 밧줄 등 심각한 물리적 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "죄수들이 성당 입구를 향해 몸이 쏠려 있으며, 뒤에 위치한 병사가 앰버의 등을 안쪽으로 밀고 있습니다.",
        "built_space": "화면 상단 중앙에 열린 성당 문이 있고 내부에 스테인드글라스가 보입니다. 왼쪽에 마스크를 끈으로 묶은 예수상, 오른쪽에 철창 트럭이 배치되어 있습니다.",
        "entities": "이현우는 오른쪽 다리에 붕대를 감고 밧줄에 묶인 채 비틀거리고 있으며, 앰버 역시 묶인 상태로 밀려납니다. 하회탈을 쓴 아시아계 병사들이 그들의 뒤에 위치합니다.",
        "hard_violations": [],
        "physics": "병사의 손이 앰버의 등에 닿아 미는 힘을 제대로 표현하고 있으며, 인물들의 발과 자세가 지면에 자연스럽게 지지되어 있습니다."
       },
       {
        "label": "A",
        "direction": "병사들이 성당 입구 안쪽에서 죄수들과 카메라를 향해 무기를 겨누고 있으며, 죄수들은 양옆으로 당겨지는 형태입니다.",
        "built_space": "중앙에 열린 성당 문과 스테인드글라스가 보이며, 왼쪽에 예수상과 트럭, 오른쪽에 급수탑이 있습니다.",
        "entities": "이현우와 앰버의 복장과 붕대는 일치하나, 병사들이 죄수의 뒤가 아닌 앞에 서 있습니다. 예수상의 마스크는 나무가 아닌 돌에 새겨진 형태로 변형되었습니다.",
        "hard_violations": [
         "아무도 잡고 있지 않은 밧줄이 허공에 뜬 채로 양옆 화면 밖을 향해 팽팽하게 당겨져 있음 (물리적 지지 기반 부재).",
         "프롬프트의 '등 떠밀린' 묘사와 반대로 병사들이 죄수들 앞에 서서 입구를 막고 있음 (명시된 연출의 정반대 배치)."
        ],
        "physics": "죄수들을 양옆으로 강하게 당기는 밧줄이 화면 밖으로 이어지지만, 이를 잡거나 당기는 물리적 주체가 전혀 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "앰버의 등에 닿은 병사의 손과 결박된 두 사람의 비틀린 자세가 강제 진입을 구현하지만, 이현우의 움직임은 문 안쪽보다 옆으로 치우치고 출입구의 사선 구도도 약하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "병사들이 포로들 뒤에서 등을 미는 대신 문 안쪽에서 마주 보고 길을 막아, 장면의 핵심인 강제 진입의 배치와 힘의 방향을 거스른다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 포로의 등은 카메라 쪽으로 보인다. 앰버는 출입구를 향하고 오른쪽 병사의 두 손이 어깨와 윗등을 문 쪽으로 민다. 이현우는 머리와 상체를 왼쪽 문설주 쪽으로 틀고 발도 옆으로 벌려, 안으로 밀려들기보다 옆으로 비틀리는 움직임이 강하다. 전경 병사도 포로 쪽을 향한다. 명확히 식별되는 총구는 없다.",
        "built_space": "회색 석조 입구 하나에 열린 목제 문짝 두 개, 상부 스테인드글라스 하나와 십자가 하나가 있다. 내부에는 세 개의 세로 스테인드글라스 창과 통로 양옆의 목제 좌석 열이 보인다. 포로들은 문턱 바로 바깥 돌계단 상단에 있고 병사들은 그 뒤와 오른쪽에 있어 밀어 넣을 공간이 성립한다. 왼쪽 외부에는 예수상 하나, 오른쪽에는 철창 수송차 일부가 있다. 재료와 고딕 입구는 이전 장면에 가깝지만 카메라는 요구된 사선보다 정면에 가깝다.",
        "entities": "검은 헝클어진 머리의 젊은 남성과 금발 여자아이, 하회탈을 쓴 병사 두 명이 보인다. 이현우의 어두운 후드, 낡은 청바지와 운동화, 앰버의 연녹색 재킷·회색 원피스·갈색 부츠는 참조와 부합한다. 뒷모습이라 정확한 얼굴과 민족적 정체성, 앰버의 마스크 착용 여부는 확인할 수 없다. 병사들의 노출된 옆얼굴과 머리는 아시아계 설정에 어긋나지 않는다. 황갈색 꼬임 밧줄이 몸통과 뒤로 묶인 손목을 감싸며 이현우의 종아리에는 붕대가 있다. 바깥 예수상은 탈을 쓰고 양팔을 벌리고 있다. 급수탑과 신발 속 카드는 이 구도에서 확인되지 않으며 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "이현우의 두 신발과 앰버의 부츠가 돌바닥에 닿아 몸을 지지한다. 앰버의 상체 기울기에는 병사의 손 접촉이라는 외력이 보인다. 이현우에게 직접 닿아 미는 손은 뚜렷하지 않아 그의 비틀림 원인은 덜 명확하지만, 발을 딛고 균형을 잃는 자세 자체는 가능하다. 밧줄은 몸과 손목에 감겨 지지되고 예수상은 석재 받침 위에 서 있다. 지지 없이 공중에 뜬 신체나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "두 포로는 카메라에서 멀어지는 방향으로 문을 향하지만, 병사 세 명은 모두 문 안쪽에서 포로와 카메라 쪽을 마주 본다. 따라서 병사들의 접근·접촉 방향은 포로의 등 뒤에서 실내로 미는 방향이 아니다. 중앙 병사의 총구는 대체로 아래 바닥을 향한다. 두 포로의 손목에서 이어지는 긴 밧줄도 문 안이 아니라 카메라 쪽 양옆으로 뻗는다.",
        "built_space": "중앙 출입구 하나에 열린 목제 문짝 두 개와 상부 스테인드글라스가 있으며, 내부에는 세로 창 세 개와 천장 조명기구 하나가 보인다. 병사 세 명이 문턱과 출입 통로를 점유하고 포로들은 그 바깥 지면에 있다. 출입구는 거의 정면이며, 좌우로 이어지는 성당 벽 대신 좁은 독립 입구 건물처럼 보여 참조의 연속된 석조 성당 외관과 차이가 크다. 왼쪽에 수송차 한 대와 그 뒤 예수상 하나, 오른쪽에 온전한 급수탑 하나가 보인다.",
        "entities": "이현우에 해당하는 검은 머리의 젊은 남성과 앰버에 해당하는 금발 여자아이의 뒷모습, 하회탈과 군복을 착용한 병사 세 명이 보인다. 포로들의 후드·청바지·녹색 재킷·회색 원피스·부츠는 대체로 참조를 따른다. 얼굴은 가려져 정확한 나이와 민족적 정체성은 검증할 수 없다. 이현우의 종아리 붕대와 두 사람의 손목 결박은 보이지만, 몸통을 따르는 결박보다 화면 밖까지 이어지는 긴 밧줄이 강조된다. 탈을 쓴 양팔 벌린 예수상, 철창 수송차, 급수탑은 존재한다. 신발 속 카드는 보이지 않는다.",
        "hard_violations": [
         "등 뒤에서 포로들을 실내로 밀어야 할 병사들이 모두 포로 앞의 문 안쪽에 배치되어 출입을 가로막는다. 이는 명시된 가해자 위치와 강제 진입 동작을 반대로 만든다."
        ],
        "physics": "두 포로는 각각 한쪽 발을 땅에 딛고 반대쪽 발을 들어 전진하는 자세여서 공중에 떠 있지는 않다. 병사들도 문턱과 계단에 발을 딛고 있다. 그러나 포로들의 등에 닿아 앞으로 미는 손은 없으며, 보이는 병사들의 위치로는 요구된 등 떠밀림을 만들 수 없다. 긴 밧줄은 손목 매듭에 연결되어 있지만 화면 밖 끝을 누가 잡는지는 확인되지 않는다. 총은 병사가 들고 있고 수송차는 바퀴로 지지된다. 예수상의 하단은 차에 가려져 받침 접촉을 확인할 수 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "앰버의 등에 닿은 병사의 손과 결박된 두 사람의 비틀린 자세가 강제 진입을 구현하지만, 이현우의 움직임은 문 안쪽보다 옆으로 치우치고 출입구의 사선 구도도 약하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "병사들이 포로들 뒤에서 등을 미는 대신 문 안쪽에서 마주 보고 길을 막아, 장면의 핵심인 강제 진입의 배치와 힘의 방향을 거스른다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 포로의 등은 카메라 쪽으로 보인다. 앰버는 출입구를 향하고 오른쪽 병사의 두 손이 어깨와 윗등을 문 쪽으로 민다. 이현우는 머리와 상체를 왼쪽 문설주 쪽으로 틀고 발도 옆으로 벌려, 안으로 밀려들기보다 옆으로 비틀리는 움직임이 강하다. 전경 병사도 포로 쪽을 향한다. 명확히 식별되는 총구는 없다.",
        "built_space": "회색 석조 입구 하나에 열린 목제 문짝 두 개, 상부 스테인드글라스 하나와 십자가 하나가 있다. 내부에는 세 개의 세로 스테인드글라스 창과 통로 양옆의 목제 좌석 열이 보인다. 포로들은 문턱 바로 바깥 돌계단 상단에 있고 병사들은 그 뒤와 오른쪽에 있어 밀어 넣을 공간이 성립한다. 왼쪽 외부에는 예수상 하나, 오른쪽에는 철창 수송차 일부가 있다. 재료와 고딕 입구는 이전 장면에 가깝지만 카메라는 요구된 사선보다 정면에 가깝다.",
        "entities": "검은 헝클어진 머리의 젊은 남성과 금발 여자아이, 하회탈을 쓴 병사 두 명이 보인다. 이현우의 어두운 후드, 낡은 청바지와 운동화, 앰버의 연녹색 재킷·회색 원피스·갈색 부츠는 참조와 부합한다. 뒷모습이라 정확한 얼굴과 민족적 정체성, 앰버의 마스크 착용 여부는 확인할 수 없다. 병사들의 노출된 옆얼굴과 머리는 아시아계 설정에 어긋나지 않는다. 황갈색 꼬임 밧줄이 몸통과 뒤로 묶인 손목을 감싸며 이현우의 종아리에는 붕대가 있다. 바깥 예수상은 탈을 쓰고 양팔을 벌리고 있다. 급수탑과 신발 속 카드는 이 구도에서 확인되지 않으며 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "이현우의 두 신발과 앰버의 부츠가 돌바닥에 닿아 몸을 지지한다. 앰버의 상체 기울기에는 병사의 손 접촉이라는 외력이 보인다. 이현우에게 직접 닿아 미는 손은 뚜렷하지 않아 그의 비틀림 원인은 덜 명확하지만, 발을 딛고 균형을 잃는 자세 자체는 가능하다. 밧줄은 몸과 손목에 감겨 지지되고 예수상은 석재 받침 위에 서 있다. 지지 없이 공중에 뜬 신체나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "두 포로는 카메라에서 멀어지는 방향으로 문을 향하지만, 병사 세 명은 모두 문 안쪽에서 포로와 카메라 쪽을 마주 본다. 따라서 병사들의 접근·접촉 방향은 포로의 등 뒤에서 실내로 미는 방향이 아니다. 중앙 병사의 총구는 대체로 아래 바닥을 향한다. 두 포로의 손목에서 이어지는 긴 밧줄도 문 안이 아니라 카메라 쪽 양옆으로 뻗는다.",
        "built_space": "중앙 출입구 하나에 열린 목제 문짝 두 개와 상부 스테인드글라스가 있으며, 내부에는 세로 창 세 개와 천장 조명기구 하나가 보인다. 병사 세 명이 문턱과 출입 통로를 점유하고 포로들은 그 바깥 지면에 있다. 출입구는 거의 정면이며, 좌우로 이어지는 성당 벽 대신 좁은 독립 입구 건물처럼 보여 참조의 연속된 석조 성당 외관과 차이가 크다. 왼쪽에 수송차 한 대와 그 뒤 예수상 하나, 오른쪽에 온전한 급수탑 하나가 보인다.",
        "entities": "이현우에 해당하는 검은 머리의 젊은 남성과 앰버에 해당하는 금발 여자아이의 뒷모습, 하회탈과 군복을 착용한 병사 세 명이 보인다. 포로들의 후드·청바지·녹색 재킷·회색 원피스·부츠는 대체로 참조를 따른다. 얼굴은 가려져 정확한 나이와 민족적 정체성은 검증할 수 없다. 이현우의 종아리 붕대와 두 사람의 손목 결박은 보이지만, 몸통을 따르는 결박보다 화면 밖까지 이어지는 긴 밧줄이 강조된다. 탈을 쓴 양팔 벌린 예수상, 철창 수송차, 급수탑은 존재한다. 신발 속 카드는 보이지 않는다.",
        "hard_violations": [
         "등 뒤에서 포로들을 실내로 밀어야 할 병사들이 모두 포로 앞의 문 안쪽에 배치되어 출입을 가로막는다. 이는 명시된 가해자 위치와 강제 진입 동작을 반대로 만든다."
        ],
        "physics": "두 포로는 각각 한쪽 발을 땅에 딛고 반대쪽 발을 들어 전진하는 자세여서 공중에 떠 있지는 않다. 병사들도 문턱과 계단에 발을 딛고 있다. 그러나 포로들의 등에 닿아 앞으로 미는 손은 없으며, 보이는 병사들의 위치로는 요구된 등 떠밀림을 만들 수 없다. 긴 밧줄은 손목 매듭에 연결되어 있지만 화면 밖 끝을 누가 잡는지는 확인되지 않는다. 총은 병사가 들고 있고 수송차는 바퀴로 지지된다. 예수상의 하단은 차에 가려져 받침 접촉을 확인할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.857,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.607,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 아무도 잡고 있지 않은 밧줄이 허공에 뜬 채로 양옆 화면 밖을 향해 팽팽하게 당겨져 있음 (물리적 지지 기반 부재).",
     "[gemini-pro] 프롬프트의 '등 떠밀린' 묘사와 반대로 병사들이 죄수들 앞에 서서 입구를 막고 있음 (명시된 연출의 정반대 배치).",
     "[gpt-high] 등 뒤에서 포로들을 실내로 밀어야 할 병사들이 모두 포로 앞의 문 안쪽에 배치되어 출입을 가로막는다. 이는 명시된 가해자 위치와 강제 진입 동작을 반대로 만든다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 607
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "병사들에게 등 떠밀려 성당 안으로 쏠리는 죄수들의 역동적인 뒷모습과 지정된 배경 요소들을 충실하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 607,
    "verdict_ko": "병사들이 뒤에서 미는 대신 앞에서 길을 막고 서 있으며, 허공에서 팽팽하게 당겨지는 밧줄 등 심각한 물리적 오류가 있습니다.  ★위반: [gemini-pro] 아무도 잡고 있지 않은 밧줄이 허공에 뜬 채로 양옆 화면 밖을 향해 팽팽하게 당겨져 있음 (물리적 지지 기반 부재). / [gemini-pro] 프롬프트의 '등 떠밀린' 묘사와 반대로 병사들이 죄수들 앞에 서서 입구를 막고 있음 (명시된 연출의 정반대 배치). / [gpt-high] 등 뒤에서 포로들을 실내로 밀어야 할 병사들이 모두 포로 앞의 문 안쪽에 배치되어 출입을 가로막는다. 이는 명시된 가해자 위치와 강제 진입 동작을 반대로 만든다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S57sh10_sel.png",
    "asset_id": "50b23d54-d8cb-4a7c-a4b8-8b2798d87d41",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 밧줄: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1074796>",
    "asset_id": "aa0a676d-74b0-43ea-8f55-fc65b425ab8d",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdeb-058e-74b8-b737-384c7b15ea72",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S57sh10"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S57sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:48:16.361038+00:00",
  "fingerprint": "3378c0e6f7a66c6ea249200ca36c66513f50195b83fd989c76144f595bc54555",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S57sh11_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S57sh11_sel.png",
  "source_sha256": "b96a5843b1b5fae5685ba1e8ee5366d6c116f7515f04a781a0219ddeb98bf936",
  "file": "S57sh11_cine.png",
  "staged_sha256": "494b57b13bdec2e178d361519073a7ae92035374b5ca98ec2ef763bbb71e0d38",
  "latency_ms": 11408
 },
 "S58sh6::signage": {
  "fp": "d7a5ba1cac2d5f0f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::3e33e2e068ec62bf": {
  "subjects": [],
  "subject_text": "익산 마을 성당 내부와 단상\n노후한 성당의 높은 실내와 정면 단상. 양옆으로 2층 발코니가 이어지고 스테인드글라스를 통과한 색빛이 들어온다.",
  "identity": "canonical",
  "scope_id": "L78",
  "scope_role": "location_interior",
  "scope_sha": "ef22468ec003ac16"
 },
 "S58sh6::bgfirst_bg": {
  "input_fingerprint": "093d297577ad0980",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 황금빛 하회탈 마스크를 쓴 장백산이 화려한 단상 위 거대한 의자에 육중한 몸을 깊숙이 기대앉은 위압적인 전신.\n\nLOCATION (lock): On the raised dais inside the ruined village church, at a large ceremonial chair beneath daylight filtering through stained glass.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Enormous chair (Occupied by 장백산 as he settles back) — Front and one side are visible from the low oblique camera position; used as Frames his full seated body while maintaining realistic scale; Ornate dais (Supports the chair above the kneeling captives' level) — Front edge and upper surface recede diagonally beneath the chair; used as Establishes the physical elevation underlying his authority; Ruined church interior (Damaged interior surrounding the dais) — Background architecture remains visible around the seated figure; used as Contrasts the ceremonial seating with the damaged setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light entering through the stained glass gives the ruined church a solemn sacred contrast while preserving the golden appearance of 장백산's mask.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 황금빛 하회탈 마스크를 쓴 장백산이 화려한 단상 위 거대한 의자에 육중한 몸을 깊숙이 기대앉은 위압적인 전신.\n\nLOCATION (lock): On the raised dais inside the ruined village church, at a large ceremonial chair beneath daylight filtering through stained glass.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Enormous chair (Occupied by 장백산 as he settles back) — Front and one side are visible from the low oblique camera position; used as Frames his full seated body while maintaining realistic scale; Ornate dais (Supports the chair above the kneeling captives' level) — Front edge and upper surface recede diagonally beneath the chair; used as Establishes the physical elevation underlying his authority; Ruined church interior (Damaged interior surrounding the dais) — Background architecture remains visible around the seated figure; used as Contrasts the ceremonial seating with the damaged setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light entering through the stained glass gives the ruined church a solemn sacred contrast while preserving the golden appearance of 장백산's mask.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S58sh6__bgfirst_bg.png",
  "asset_id": "5509079f-8e1e-4501-9d02-e870e093500e",
  "input_asset_ids": [
   "15eb3a03-dd7f-4b94-a5dd-1d2856db0275",
   "b96507e2-c573-4afc-916c-a1edaf311036"
  ]
 },
 "S58sh6": {
  "input_fingerprint": "b656e2e30278a0db",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 황금빛 하회탈 마스크를 쓴 장백산이 화려한 단상 위 거대한 의자에 육중한 몸을 깊숙이 기대앉은 위압적인 전신.\n\nLOCATION (lock): On the raised dais inside the ruined village church, at a large ceremonial chair beneath daylight filtering through stained glass. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Enormous chair (Occupied by 장백산 as he settles back) — Front and one side are visible from the low oblique camera position; used as Frames his full seated body while maintaining realistic scale; Ornate dais (Supports the chair above the kneeling captives' level) — Front edge and upper surface recede diagonally beneath the chair; used as Establishes the physical elevation underlying his authority; Ruined church interior (Damaged interior surrounding the dais) — Background architecture remains visible around the seated figure; used as Contrasts the ceremonial seating with the damaged setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light entering through the stained glass gives the ruined church a solemn sacred contrast while preserving the golden appearance of 장백산's mask.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Light enters the ruined church through surviving stained glass, and the masked Jesus statue remains outside. Charlie is already confined in the steel cage that will shortly be brought into the church, with no release yet established. 장백산: He sits on the dais wearing an intact golden Hahoe mask and kinglike clothing over his massive body. Severe radiation damage already exists beneath the mask, but his face remains concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 황금빛 하회탈 마스크를 쓴 장백산이 화려한 단상 위 거대한 의자에 육중한 몸을 깊숙이 기대앉은 위압적인 전신.\n\nLOCATION (lock): On the raised dais inside the ruined village church, at a large ceremonial chair beneath daylight filtering through stained glass. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Enormous chair (Occupied by 장백산 as he settles back) — Front and one side are visible from the low oblique camera position; used as Frames his full seated body while maintaining realistic scale; Ornate dais (Supports the chair above the kneeling captives' level) — Front edge and upper surface recede diagonally beneath the chair; used as Establishes the physical elevation underlying his authority; Ruined church interior (Damaged interior surrounding the dais) — Background architecture remains visible around the seated figure; used as Contrasts the ceremonial seating with the damaged setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light entering through the stained glass gives the ruined church a solemn sacred contrast while preserving the golden appearance of 장백산's mask.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Light enters the ruined church through surviving stained glass, and the masked Jesus statue remains outside. Charlie is already confined in the steel cage that will shortly be brought into the church, with no release yet established. 장백산: He sits on the dais wearing an intact golden Hahoe mask and kinglike clothing over his massive body. Severe radiation damage already exists beneath the mask, but his face remains concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 황금빛 하회탈 마스크를 쓴 장백산이 화려한 단상 위 거대한 의자에 육중한 몸을 깊숙이 기대앉은 위압적인 전신.\n\nLOCATION (lock): On the raised dais inside the ruined village church, at a large ceremonial chair beneath daylight filtering through stained glass. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Enormous chair (Occupied by 장백산 as he settles back) — Front and one side are visible from the low oblique camera position; used as Frames his full seated body while maintaining realistic scale; Ornate dais (Supports the chair above the kneeling captives' level) — Front edge and upper surface recede diagonally beneath the chair; used as Establishes the physical elevation underlying his authority; Ruined church interior (Damaged interior surrounding the dais) — Background architecture remains visible around the seated figure; used as Contrasts the ceremonial seating with the damaged setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light entering through the stained glass gives the ruined church a solemn sacred contrast while preserving the golden appearance of 장백산's mask.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Light enters the ruined church through surviving stained glass, and the masked Jesus statue remains outside. Charlie is already confined in the steel cage that will shortly be brought into the church, with no release yet established. 장백산: He sits on the dais wearing an intact golden Hahoe mask and kinglike clothing over his massive body. Severe radiation damage already exists beneath the mask, but his face remains concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S58sh6__bgfirst_bg.png",
     "asset_id": "5509079f-8e1e-4501-9d02-e870e093500e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S58sh6.png",
     "asset_id": "15eb3a03-dd7f-4b94-a5dd-1d2856db0275",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:984613>",
     "asset_id": "aa649eea-0117-4f3d-971e-8c059a9d57fa",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L78B02.png",
     "asset_id": "b96507e2-c573-4afc-916c-a1edaf311036",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:984613>",
     "asset_id": "aa649eea-0117-4f3d-971e-8c059a9d57fa",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선과 정면이 카메라를 향해 올바르게 정렬됨.",
    "built_space": "참조 사진과 동일한 예배당 내부, 계단, 목재 패널, 스테인드글라스가 정확한 위치와 비율로 묘사됨.",
    "entities": "장백산이 황금빛 마스크를 착용 중이나, 참조된 화려한 금박 자수 의상 대신 단순한 붉은 옷으로 변형됨.",
    "hard_violations": [],
    "physics": "의자에 기대어 앉은 자세가 중력에 맞게 자연스럽게 묘사됨."
   },
   {
    "label": "B",
    "direction": "장백산은 정면을 향하고, 무릎 꿇은 인물들은 제단을 향함.",
    "built_space": "예배당 배경은 유사하나 참조에 없는 거대한 대리석 단상이 바닥에 추가되었고, 구조물 일부가 불타고 있음.",
    "entities": "장백산의 붉은 벨벳 옷과 금박 자수 디테일은 참조와 완벽히 일치함. 그러나 샷 텍스트에 언급되지 않은 인물들이 등장함.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 없는 인물들(무릎 꿇은 포로들) 임의 추가",
     "[gemini-pro] 프롬프트에 명시되지 않은 화염 및 파손 효과 임의 추가",
     "[gpt-high] 장백산만 등장해야 하는 화면에 무릎 꿇은 인물 7명을 추가했습니다.",
     "[gpt-high] 참조의 제단 위에 있어야 할 의자와 단상을 기존 제단 계단 앞의 별도 전경 위치에 배치했습니다."
    ],
    "physics": "모든 인물이 바닥과 단상, 의자에 물리적으로 올바르게 지지되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지정된 장소의 구조를 정확히 반영하고 명시된 인물만 배치하여 지침을 준수했으나, 의상의 화려한 자수 디테일이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "의상 구현도는 매우 우수하나, 샷 텍스트에 없는 다수의 인물(포로들)과 화염을 임의로 추가하여 심각한 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선과 정면이 카메라를 향해 올바르게 정렬됨.",
        "built_space": "참조 사진과 동일한 예배당 내부, 계단, 목재 패널, 스테인드글라스가 정확한 위치와 비율로 묘사됨.",
        "entities": "장백산이 황금빛 마스크를 착용 중이나, 참조된 화려한 금박 자수 의상 대신 단순한 붉은 옷으로 변형됨.",
        "hard_violations": [],
        "physics": "의자에 기대어 앉은 자세가 중력에 맞게 자연스럽게 묘사됨."
       },
       {
        "label": "B",
        "direction": "장백산은 정면을 향하고, 무릎 꿇은 인물들은 제단을 향함.",
        "built_space": "예배당 배경은 유사하나 참조에 없는 거대한 대리석 단상이 바닥에 추가되었고, 구조물 일부가 불타고 있음.",
        "entities": "장백산의 붉은 벨벳 옷과 금박 자수 디테일은 참조와 완벽히 일치함. 그러나 샷 텍스트에 언급되지 않은 인물들이 등장함.",
        "hard_violations": [
         "샷 텍스트에 없는 인물들(무릎 꿇은 포로들) 임의 추가",
         "프롬프트에 명시되지 않은 화염 및 파손 효과 임의 추가"
        ],
        "physics": "모든 인물이 바닥과 단상, 의자에 물리적으로 올바르게 지지되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지정된 장소의 구조를 정확히 반영하고 명시된 인물만 배치하여 지침을 준수했으나, 의상의 화려한 자수 디테일이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "의상 구현도는 매우 우수하나, 샷 텍스트에 없는 다수의 인물(포로들)과 화염을 임의로 추가하여 심각한 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선과 정면이 카메라를 향해 올바르게 정렬됨.",
        "built_space": "참조 사진과 동일한 예배당 내부, 계단, 목재 패널, 스테인드글라스가 정확한 위치와 비율로 묘사됨.",
        "entities": "장백산이 황금빛 마스크를 착용 중이나, 참조된 화려한 금박 자수 의상 대신 단순한 붉은 옷으로 변형됨.",
        "hard_violations": [],
        "physics": "의자에 기대어 앉은 자세가 중력에 맞게 자연스럽게 묘사됨."
       },
       {
        "label": "B",
        "direction": "장백산은 정면을 향하고, 무릎 꿇은 인물들은 제단을 향함.",
        "built_space": "예배당 배경은 유사하나 참조에 없는 거대한 대리석 단상이 바닥에 추가되었고, 구조물 일부가 불타고 있음.",
        "entities": "장백산의 붉은 벨벳 옷과 금박 자수 디테일은 참조와 완벽히 일치함. 그러나 샷 텍스트에 언급되지 않은 인물들이 등장함.",
        "hard_violations": [
         "샷 텍스트에 없는 인물들(무릎 꿇은 포로들) 임의 추가",
         "프롬프트에 명시되지 않은 화염 및 파손 효과 임의 추가"
        ],
        "physics": "모든 인물이 바닥과 단상, 의자에 물리적으로 올바르게 지지되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "육중한 전신과 기대앉은 자세는 잘 구현했지만, 금지된 포로 7명을 추가하고 의자와 단상을 원래 제단 앞의 별도 위치로 옮겨 실격입니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "원래 교회 단상에서 장백산 혼자 앉은 전신을 낮은 사선 와이드숏으로 담았으며, 참조보다 단순한 의복과 약한 육중함이 주요 차이입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장백산의 가면 정면은 화면 왼쪽 앞의 예배당 공간을 향하고, 눈의 정확한 시선은 가면 때문에 확인되지 않습니다. 아래 인물들은 단상을 향해 무릎을 꿇고 바닥으로 고개를 숙입니다. 무기나 겨냥하는 소품은 없습니다.",
        "built_space": "큰 의자 1개가 화면 앞쪽의 넓은 조각 장식 단상 위에 있고, 의자의 앞면과 오른쪽 측면이 보입니다. 원래 교회의 제단 계단과 목조 벽은 그 뒤에 따로 남아 있어, 참조의 제단 위가 아니라 그 앞에 새 단상을 놓은 배치입니다. 왼쪽 철제 난간 회랑과 오른쪽 철문 1개는 유지되며, 뒤쪽에는 여러 화염과 파손된 구조물이 추가되었습니다.",
        "entities": "장백산 외에 무릎 꿇은 인물 7명이 보이며, 이는 인물 제한에 어긋납니다. 장백산은 육중한 성인 남성으로 표현되며 얼굴은 온전한 금색 가면으로 가려져 있습니다. 가면의 웃는 표정은 보이지만 하회탈의 특징은 참조보다 약합니다. 자주색 바탕의 촘촘한 금색 자수, 장식 허리띠, 흰 발목 의복과 신발은 참조에 가깝습니다. 다른 인물들의 정확한 국적이나 민족은 외형만으로 확인할 수 없습니다. 예수상과 찰리는 없고, 읽을 수 있는 글자도 없습니다.",
        "hard_violations": [
         "장백산만 등장해야 하는 화면에 무릎 꿇은 인물 7명을 추가했습니다.",
         "참조의 제단 위에 있어야 할 의자와 단상을 기존 제단 계단 앞의 별도 전경 위치에 배치했습니다."
        ],
        "physics": "장백산의 엉덩이와 등은 의자 좌판과 등받이에 받쳐지고, 양팔은 팔걸이에, 양발은 단상 표면에 놓여 있습니다. 의자 다리는 단상에 닿습니다. 아래 인물들도 무릎과 발 또는 손으로 바닥을 지지하며, 보이는 범위에서 공중에 떠 있는 신체나 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "장백산의 가면은 단상 아래 카메라 근처의 예배당 공간을 향합니다. 가면 안의 정확한 눈동자 방향은 판별되지 않습니다. 몸은 의자 정면과 같은 방향으로 놓여 있으며, 겨냥하는 무기나 이동하는 인물은 없습니다.",
        "built_space": "제단 계단 위 중앙에 큰 의자 1개가 있고, 그 뒤로 목조 장식 벽과 주요 첨두형 스테인드글라스 창 3개가 보입니다. 촛대는 높은 것 2개와 바깥쪽 낮은 것 2개로 참조의 배치를 따릅니다. 왼쪽 회랑의 철제 난간과 석조 기둥도 유지됩니다. 낮은 사선 시점에서 의자 앞면과 한쪽 측면, 대각선으로 물러나는 단상 상면과 계단이 보입니다. 바닥의 창빛 반사는 이 시점에서 자연스럽습니다.",
        "entities": "등장인물은 금빛 웃는 하회탈을 쓴 성인 남성 장백산 1명뿐이며 얼굴은 가려져 있습니다. 체격은 넓지만 요청한 육중함은 A보다 약합니다. 자주색 왕 같은 의복과 금색 테두리는 있으나 참조의 전면 자수와 크고 화려한 허리 장식은 간소화되었고, 신발도 참조의 흰 발목 의복과 자수 신발 대신 어두운 장화 형태입니다. 교회, 의자, 단상과 낮의 스테인드글라스 채광이 모두 보입니다. 예수상과 찰리, 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "엉덩이는 좌판에, 등은 등받이 쪽에 놓이고 양팔은 팔걸이에 받쳐져 있습니다. 벌어진 두 다리는 아래로 내려와 양발이 단상 표면에 닿습니다. 의자와 촛대도 각각 바닥에 지지되어 있습니다. 깊이 파묻히는 느낌은 다소 약하지만 실제로 가능한 착석 자세이며, 지지 없이 떠 있는 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "육중한 전신과 기대앉은 자세는 잘 구현했지만, 금지된 포로 7명을 추가하고 의자와 단상을 원래 제단 앞의 별도 위치로 옮겨 실격입니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "원래 교회 단상에서 장백산 혼자 앉은 전신을 낮은 사선 와이드숏으로 담았으며, 참조보다 단순한 의복과 약한 육중함이 주요 차이입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "장백산의 가면 정면은 화면 왼쪽 앞의 예배당 공간을 향하고, 눈의 정확한 시선은 가면 때문에 확인되지 않습니다. 아래 인물들은 단상을 향해 무릎을 꿇고 바닥으로 고개를 숙입니다. 무기나 겨냥하는 소품은 없습니다.",
        "built_space": "큰 의자 1개가 화면 앞쪽의 넓은 조각 장식 단상 위에 있고, 의자의 앞면과 오른쪽 측면이 보입니다. 원래 교회의 제단 계단과 목조 벽은 그 뒤에 따로 남아 있어, 참조의 제단 위가 아니라 그 앞에 새 단상을 놓은 배치입니다. 왼쪽 철제 난간 회랑과 오른쪽 철문 1개는 유지되며, 뒤쪽에는 여러 화염과 파손된 구조물이 추가되었습니다.",
        "entities": "장백산 외에 무릎 꿇은 인물 7명이 보이며, 이는 인물 제한에 어긋납니다. 장백산은 육중한 성인 남성으로 표현되며 얼굴은 온전한 금색 가면으로 가려져 있습니다. 가면의 웃는 표정은 보이지만 하회탈의 특징은 참조보다 약합니다. 자주색 바탕의 촘촘한 금색 자수, 장식 허리띠, 흰 발목 의복과 신발은 참조에 가깝습니다. 다른 인물들의 정확한 국적이나 민족은 외형만으로 확인할 수 없습니다. 예수상과 찰리는 없고, 읽을 수 있는 글자도 없습니다.",
        "hard_violations": [
         "장백산만 등장해야 하는 화면에 무릎 꿇은 인물 7명을 추가했습니다.",
         "참조의 제단 위에 있어야 할 의자와 단상을 기존 제단 계단 앞의 별도 전경 위치에 배치했습니다."
        ],
        "physics": "장백산의 엉덩이와 등은 의자 좌판과 등받이에 받쳐지고, 양팔은 팔걸이에, 양발은 단상 표면에 놓여 있습니다. 의자 다리는 단상에 닿습니다. 아래 인물들도 무릎과 발 또는 손으로 바닥을 지지하며, 보이는 범위에서 공중에 떠 있는 신체나 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "장백산의 가면은 단상 아래 카메라 근처의 예배당 공간을 향합니다. 가면 안의 정확한 눈동자 방향은 판별되지 않습니다. 몸은 의자 정면과 같은 방향으로 놓여 있으며, 겨냥하는 무기나 이동하는 인물은 없습니다.",
        "built_space": "제단 계단 위 중앙에 큰 의자 1개가 있고, 그 뒤로 목조 장식 벽과 주요 첨두형 스테인드글라스 창 3개가 보입니다. 촛대는 높은 것 2개와 바깥쪽 낮은 것 2개로 참조의 배치를 따릅니다. 왼쪽 회랑의 철제 난간과 석조 기둥도 유지됩니다. 낮은 사선 시점에서 의자 앞면과 한쪽 측면, 대각선으로 물러나는 단상 상면과 계단이 보입니다. 바닥의 창빛 반사는 이 시점에서 자연스럽습니다.",
        "entities": "등장인물은 금빛 웃는 하회탈을 쓴 성인 남성 장백산 1명뿐이며 얼굴은 가려져 있습니다. 체격은 넓지만 요청한 육중함은 A보다 약합니다. 자주색 왕 같은 의복과 금색 테두리는 있으나 참조의 전면 자수와 크고 화려한 허리 장식은 간소화되었고, 신발도 참조의 흰 발목 의복과 자수 신발 대신 어두운 장화 형태입니다. 교회, 의자, 단상과 낮의 스테인드글라스 채광이 모두 보입니다. 예수상과 찰리, 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "엉덩이는 좌판에, 등은 등받이 쪽에 놓이고 양팔은 팔걸이에 받쳐져 있습니다. 벌어진 두 다리는 아래로 내려와 양발이 단상 표면에 닿습니다. 의자와 촛대도 각각 바닥에 지지되어 있습니다. 깊이 파묻히는 느낌은 다소 약하지만 실제로 가능한 착석 자세이며, 지지 없이 떠 있는 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.75
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.5
   },
   "violations": {
    "B": [
     "[gemini-pro] 샷 텍스트에 없는 인물들(무릎 꿇은 포로들) 임의 추가",
     "[gemini-pro] 프롬프트에 명시되지 않은 화염 및 파손 효과 임의 추가",
     "[gpt-high] 장백산만 등장해야 하는 화면에 무릎 꿇은 인물 7명을 추가했습니다.",
     "[gpt-high] 참조의 제단 위에 있어야 할 의자와 단상을 기존 제단 계단 앞의 별도 전경 위치에 배치했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 장소의 구조를 정확히 반영하고 명시된 인물만 배치하여 지침을 준수했으나, 의상의 화려한 자수 디테일이 누락되었습니다."
   },
   {
    "label": "B",
    "score": 500,
    "verdict_ko": "의상 구현도는 매우 우수하나, 샷 텍스트에 없는 다수의 인물(포로들)과 화염을 임의로 추가하여 심각한 위반이 발생했습니다.  ★위반: [gemini-pro] 샷 텍스트에 없는 인물들(무릎 꿇은 포로들) 임의 추가 / [gemini-pro] 프롬프트에 명시되지 않은 화염 및 파손 효과 임의 추가 / [gpt-high] 장백산만 등장해야 하는 화면에 무릎 꿇은 인물 7명을 추가했습니다. / [gpt-high] 참조의 제단 위에 있어야 할 의자와 단상을 기존 제단 계단 앞의 별도 전경 위치에 배치했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L78B02.png",
    "asset_id": "b96507e2-c573-4afc-916c-a1edaf311036",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:984613>",
    "asset_id": "aa649eea-0117-4f3d-971e-8c059a9d57fa",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdeb-c18a-7505-9fa2-3cd826395dd5",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S58sh6__bgfirst_bg.png",
   "bg_asset_id": "5509079f-8e1e-4501-9d02-e870e093500e",
   "bg_record_key": "S58sh6::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S58sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:48:28.184512+00:00",
  "fingerprint": "cb6296f2e54737fc04dc0bb0be3e7c3ee1f3a344fe730ee306d683736657d2cd",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S58sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S58sh6_sel.png",
  "source_sha256": "457ded57833b3bb449746333d01aee0967a47bf04009a803094028a2b2cfadde",
  "file": "S58sh6_cine.png",
  "staged_sha256": "fdbb7fffcb3ac18806158e5a82b9e22436f1c6ae4b4684ccd92b554d527343e5",
  "latency_ms": 12757
 },
 "S58sh15::signage": {
  "fp": "4c9797b0eabf9668",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S58sh15": {
  "input_fingerprint": "30b392c1f210dfd5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하회탈 마스크의 아시아계 병사들이 거대한 찰리가 갇힌 강철 케이지에 두 손을 짚고 성당 안쪽을 향해 힘껏 밀고 있는 mid-action 구도.\n\nLOCATION (lock): In the ruined church's interior approach to the dais, where a steel cage is pushed through under stained-glass daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 강철 케이지 (Being pushed into the church with 찰리 confined inside) — The near side and trailing end are visible obliquely, with gaps between the bars exposing 찰리; used as Connects the soldiers' pushing hands to the captive without obscuring either; 성당 내부 (Ruined; the cage route leads farther inside) — The interior recedes toward screen right behind the advancing group; used as Provides enough surrounding space to preserve the cage's scale and direction of travel; 하회탈 마스크 (Worn by the pushing soldiers) — Different oblique mask faces turn along the shared route rather than toward the lens; used as Makes the escort identifiable without turning the soldiers into identical frontal figures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight entering through the stained glass gives the ruined church solemn contrast while preserving detail in the soldiers' hands and 찰리 inside the cage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Stained-glass light enters the ruined church, where an opened barred doorway admits a steel cage containing Charlie. Charlie retains his large straw hat, colorful raincoat and oversized boots; his worn body bears a peeling Yubik chest logo.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 하회탈 마스크의 아시아계 병사들 right now, so 하회탈 마스크의 아시아계 병사들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 하회탈 마스크의 아시아계 병사들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하회탈 마스크의 아시아계 병사들이 거대한 찰리가 갇힌 강철 케이지에 두 손을 짚고 성당 안쪽을 향해 힘껏 밀고 있는 mid-action 구도.\n\nLOCATION (lock): In the ruined church's interior approach to the dais, where a steel cage is pushed through under stained-glass daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 강철 케이지 (Being pushed into the church with 찰리 confined inside) — The near side and trailing end are visible obliquely, with gaps between the bars exposing 찰리; used as Connects the soldiers' pushing hands to the captive without obscuring either; 성당 내부 (Ruined; the cage route leads farther inside) — The interior recedes toward screen right behind the advancing group; used as Provides enough surrounding space to preserve the cage's scale and direction of travel; 하회탈 마스크 (Worn by the pushing soldiers) — Different oblique mask faces turn along the shared route rather than toward the lens; used as Makes the escort identifiable without turning the soldiers into identical frontal figures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight entering through the stained glass gives the ruined church solemn contrast while preserving detail in the soldiers' hands and 찰리 inside the cage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Stained-glass light enters the ruined church, where an opened barred doorway admits a steel cage containing Charlie. Charlie retains his large straw hat, colorful raincoat and oversized boots; his worn body bears a peeling Yubik chest logo.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 하회탈 마스크의 아시아계 병사들 right now, so 하회탈 마스크의 아시아계 병사들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 하회탈 마스크의 아시아계 병사들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하회탈 마스크의 아시아계 병사들이 거대한 찰리가 갇힌 강철 케이지에 두 손을 짚고 성당 안쪽을 향해 힘껏 밀고 있는 mid-action 구도.\n\nLOCATION (lock): In the ruined church's interior approach to the dais, where a steel cage is pushed through under stained-glass daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 강철 케이지 (Being pushed into the church with 찰리 confined inside) — The near side and trailing end are visible obliquely, with gaps between the bars exposing 찰리; used as Connects the soldiers' pushing hands to the captive without obscuring either; 성당 내부 (Ruined; the cage route leads farther inside) — The interior recedes toward screen right behind the advancing group; used as Provides enough surrounding space to preserve the cage's scale and direction of travel; 하회탈 마스크 (Worn by the pushing soldiers) — Different oblique mask faces turn along the shared route rather than toward the lens; used as Makes the escort identifiable without turning the soldiers into identical frontal figures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight entering through the stained glass gives the ruined church solemn contrast while preserving detail in the soldiers' hands and 찰리 inside the cage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Stained-glass light enters the ruined church, where an opened barred doorway admits a steel cage containing Charlie. Charlie retains his large straw hat, colorful raincoat and oversized boots; his worn body bears a peeling Yubik chest logo.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 하회탈 마스크의 아시아계 병사들 right now, so 하회탈 마스크의 아시아계 병사들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 하회탈 마스크의 아시아계 병사들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "좌우 병사들이 서로 마주보고 반대 방향으로 케이지를 밀고 있어 성당 안쪽으로 이동하는 방향성이 상실됨.",
    "built_space": "성당 내부는 유사하나 좌측 바닥에 레퍼런스에 없는 돌 난간이 임의로 추가됨. 케이지는 지시대로 비스듬히 배치됨.",
    "entities": "찰리는 지정된 복장을 입고 케이지 안에 있음. 병사들은 군복을 입었으나 마스크 형태가 전통 하회탈에서 다소 변형됨.",
    "hard_violations": [
     "[gemini-pro] 서로 반대 방향으로 미는 병사들 배치 (물리적/상황적 모순)",
     "[gemini-pro] 레퍼런스에 없는 좌측 돌 난간 임의 생성",
     "[gpt-high] 참조에서 개방되어 있던 제단 접근 공간에 고정 석조 난간 두 구간을 새로 만들어, 잠금된 장소의 구조와 통로를 변경했다."
    ],
    "physics": "찰리는 바닥에 발을 딛고 쪼그려 있으나, 케이지 양쪽에서 반대로 미는 힘의 작용이 물리적으로 불가능한 상황을 만듦."
   },
   {
    "label": "B",
    "direction": "A와 마찬가지로 병사들이 양쪽에서 서로 반대 방향으로 밀고 있어 전진하는 방향성이 없음.",
    "built_space": "성당 배경은 일치하나, 케이지가 비스듬하지 않고 카메라를 향해 정면으로 배치되어 프레이밍 지시를 위반함.",
    "entities": "찰리는 묘사된 의상을 입고 있으나, 병사들의 마스크가 하회탈보다는 타 국가의 목재 가면 형상에 가까움.",
    "hard_violations": [
     "[gemini-pro] 지지대 없이 허공에 떠서 앉아 있는 찰리 (물리적 오류)",
     "[gemini-pro] 비스듬한 각도 지시를 무시한 케이지의 정면 배치",
     "[gemini-pro] 서로 반대 방향으로 미는 병사들 배치 (물리적/상황적 모순)"
    ],
    "physics": "찰리가 의자 등 지지대가 전혀 없이 엉덩이가 공중에 뜬 채 앉은 자세를 취하고 있음. 우측 뒤편 병사는 허공을 밀고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "양쪽에서 반대로 밀어 액션이 성립하지 않고 없는 난간이 추가되었으나, B보다 물리적 결함이 적습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리가 지지대 없이 허공에 앉아 있고, 케이지 정면 배치로 구도를 어겨 치명적인 결함이 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "좌우 병사들이 서로 마주보고 반대 방향으로 케이지를 밀고 있어 성당 안쪽으로 이동하는 방향성이 상실됨.",
        "built_space": "성당 내부는 유사하나 좌측 바닥에 레퍼런스에 없는 돌 난간이 임의로 추가됨. 케이지는 지시대로 비스듬히 배치됨.",
        "entities": "찰리는 지정된 복장을 입고 케이지 안에 있음. 병사들은 군복을 입었으나 마스크 형태가 전통 하회탈에서 다소 변형됨.",
        "hard_violations": [
         "서로 반대 방향으로 미는 병사들 배치 (물리적/상황적 모순)",
         "레퍼런스에 없는 좌측 돌 난간 임의 생성"
        ],
        "physics": "찰리는 바닥에 발을 딛고 쪼그려 있으나, 케이지 양쪽에서 반대로 미는 힘의 작용이 물리적으로 불가능한 상황을 만듦."
       },
       {
        "label": "B",
        "direction": "A와 마찬가지로 병사들이 양쪽에서 서로 반대 방향으로 밀고 있어 전진하는 방향성이 없음.",
        "built_space": "성당 배경은 일치하나, 케이지가 비스듬하지 않고 카메라를 향해 정면으로 배치되어 프레이밍 지시를 위반함.",
        "entities": "찰리는 묘사된 의상을 입고 있으나, 병사들의 마스크가 하회탈보다는 타 국가의 목재 가면 형상에 가까움.",
        "hard_violations": [
         "지지대 없이 허공에 떠서 앉아 있는 찰리 (물리적 오류)",
         "비스듬한 각도 지시를 무시한 케이지의 정면 배치",
         "서로 반대 방향으로 미는 병사들 배치 (물리적/상황적 모순)"
        ],
        "physics": "찰리가 의자 등 지지대가 전혀 없이 엉덩이가 공중에 뜬 채 앉은 자세를 취하고 있음. 우측 뒤편 병사는 허공을 밀고 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "양쪽에서 반대로 밀어 액션이 성립하지 않고 없는 난간이 추가되었으나, B보다 물리적 결함이 적습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리가 지지대 없이 허공에 앉아 있고, 케이지 정면 배치로 구도를 어겨 치명적인 결함이 있습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "좌우 병사들이 서로 마주보고 반대 방향으로 케이지를 밀고 있어 성당 안쪽으로 이동하는 방향성이 상실됨.",
        "built_space": "성당 내부는 유사하나 좌측 바닥에 레퍼런스에 없는 돌 난간이 임의로 추가됨. 케이지는 지시대로 비스듬히 배치됨.",
        "entities": "찰리는 지정된 복장을 입고 케이지 안에 있음. 병사들은 군복을 입었으나 마스크 형태가 전통 하회탈에서 다소 변형됨.",
        "hard_violations": [
         "서로 반대 방향으로 미는 병사들 배치 (물리적/상황적 모순)",
         "레퍼런스에 없는 좌측 돌 난간 임의 생성"
        ],
        "physics": "찰리는 바닥에 발을 딛고 쪼그려 있으나, 케이지 양쪽에서 반대로 미는 힘의 작용이 물리적으로 불가능한 상황을 만듦."
       },
       {
        "label": "B",
        "direction": "A와 마찬가지로 병사들이 양쪽에서 서로 반대 방향으로 밀고 있어 전진하는 방향성이 없음.",
        "built_space": "성당 배경은 일치하나, 케이지가 비스듬하지 않고 카메라를 향해 정면으로 배치되어 프레이밍 지시를 위반함.",
        "entities": "찰리는 묘사된 의상을 입고 있으나, 병사들의 마스크가 하회탈보다는 타 국가의 목재 가면 형상에 가까움.",
        "hard_violations": [
         "지지대 없이 허공에 떠서 앉아 있는 찰리 (물리적 오류)",
         "비스듬한 각도 지시를 무시한 케이지의 정면 배치",
         "서로 반대 방향으로 미는 병사들 배치 (물리적/상황적 모순)"
        ],
        "physics": "찰리가 의자 등 지지대가 전혀 없이 엉덩이가 공중에 뜬 채 앉은 자세를 취하고 있음. 우측 뒤편 병사는 허공을 밀고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "기존 성당의 구조와 전신 와이드 구도는 더 충실하지만, 병사들이 케이지 양쪽에서 서로 반대 방향으로 밀어 화면 오른쪽으로 전진하는 핵심 동작을 실현하지 못했다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "참조에 없던 석조 난간을 이동 통로에 추가했고, 병사들이 제단이 있는 오른쪽 안쪽이 아니라 화면 왼쪽으로 미는 구도여서 장소와 이동 방향 모두 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 병사들은 케이지를 향해 오른쪽으로 몸을 기울이고, 오른쪽 병사들은 왼쪽으로 몸을 기울여 민다. 따라서 힘의 방향이 서로 맞서며, 일행이 화면 오른쪽 성당 안쪽으로 함께 전진하는 동작으로 읽히지 않는다. 탈은 대체로 케이지 쪽을 향하지만 찰리의 얼굴은 카메라 쪽을 향한다. 무기나 조준 대상은 없다.",
        "built_space": "중앙의 케이지 한 개에는 닫힌 문 한 개가 있고, 넓은 문 쪽 면과 좁은 오른쪽 면이 보인다. 왼쪽 상층 철제 난간, 석재 기둥, 뒤쪽 제단 계단 한 벌, 목조 제단벽과 높은 의자 한 개가 참조와 대체로 이어진다. 촛대는 오른쪽에 두 개가 드러나며 나머지 제단 부분은 케이지에 가려진다. 병사들은 케이지 양옆과 뒤쪽에 분산되어 있고, 오른쪽으로 이어지는 진입 통로보다 제단을 정면으로 마주한 배치가 강하다. 바닥의 빛 반사는 창문과 광택 석재에 부합한다.",
        "entities": "찰리 한 명과 하회탈을 쓴 병사 다섯 명이 보인다. 병사들은 성인 남성형 체격이며 일부 노출된 얼굴과 피부는 아시아계 설정에 어긋나지 않지만, 탈 때문에 정확한 정체성은 확인하기 어렵다. 찰리는 각진 기계 머리, 흰 얼굴판, 작은 눈과 선형 입, 밀짚모자, 다색 우비, 줄무늬 상의, 녹색 바지와 노란 장화를 갖췄다. 다만 앉아 있는 찰리의 체구에서 병사들보다 거대하다는 인상은 약하다. 케이지는 참조처럼 낡은 강철 격자와 경첩 달린 문을 갖춘다. 가슴의 벗겨진 표식과 열린 철창 출입문은 식별되지 않으며, 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "병사들은 벌린 발로 바닥을 딛고 손을 철창과 기둥에 대고 있어 개별적으로 미는 자세는 가능하다. 찰리의 엉덩이 아래에는 어두운 좌판으로 보이는 지지물이 있고 장화도 케이지 바닥에 놓여 있다. 케이지 하부는 성당 바닥에 닿아 있어 떠 있지 않다. 바퀴 없이 무거운 케이지를 미끄러뜨리는 동작 자체는 가능하지만, 양쪽 병사들의 상반된 힘은 요청된 공동 전진을 설명하지 못한다."
       },
       {
        "label": "B",
        "direction": "오른쪽 병사들은 손을 철창에 대고 몸을 왼쪽으로 기울여 케이지를 화면 왼쪽으로 미는 모습이다. 제단과 성당 안쪽은 그들의 뒤쪽인 화면 오른쪽에 있어 요청된 진입 방향과 반대다. 앞쪽 금색 탈 병사는 얼굴을 카메라 쪽으로 돌리고, 뒤 병사들의 탈도 상당히 정면을 향한다. 찰리는 왼쪽을 바라본다. 무기나 조준 대상은 없다.",
        "built_space": "케이지 한 개의 닫힌 문과 긴 오른쪽 측면이 비스듬히 보인다. 기존의 왼쪽 상층 철제 난간, 기둥, 스테인드글라스, 제단 계단과 목조 의자는 이어지지만, 전경 통로 양옆에 석조 난간 두 구간이 새로 생겼다. 참조에서 열린 바닥과 계단으로 보이던 접근 공간을 이 난간들이 좁힌다. 오른쪽 제단에는 촛대 두 개가 보인다. 병사들은 주로 케이지 오른쪽과 뒤쪽에 있고 일부는 왼쪽 가장자리에 있다. 바닥 반사는 광원과 재질상 가능하다.",
        "entities": "찰리 한 명과 탈을 쓴 병사 다섯 명이 보이며, 병사들은 성인 체격이다. 얼굴이 대부분 가려져 민족적 정체성을 확정하기 어렵다. 탈의 종류는 다양하지만 여러 얼굴이 이동 경로보다 렌즈 쪽을 향한다. 찰리의 기계 얼굴, 밀짚모자, 다색 우비, 줄무늬 상의, 녹색 바지와 큰 노란 장화는 참조에 가깝다. 웅크린 체구는 거대한 포로라는 크기 관계를 충분히 드러내지 않는다. 강철 케이지의 재료와 격자 구조는 유사하나 참조보다 깊게 늘어난 형태다. 열린 철창 출입문과 벗겨진 가슴 표식은 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참조에서 개방되어 있던 제단 접근 공간에 고정 석조 난간 두 구간을 새로 만들어, 잠금된 장소의 구조와 통로를 변경했다."
        ],
        "physics": "찰리는 무릎을 굽히고 두 장화를 케이지 바닥에 댄 낮은 자세이며 손으로 철창을 잡고 있어 지지 관계가 보인다. 병사들도 발을 바닥에 딛고 손을 철창에 대고 있다. 케이지는 하부 틀이 바닥에 닿아 있어 공중에 뜬 물체는 없다. 병사들의 기울기와 팔의 접촉은 왼쪽으로 미는 힘을 설명하지만, 그것은 요청된 오른쪽 안쪽으로의 이동과 반대다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "기존 성당의 구조와 전신 와이드 구도는 더 충실하지만, 병사들이 케이지 양쪽에서 서로 반대 방향으로 밀어 화면 오른쪽으로 전진하는 핵심 동작을 실현하지 못했다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "참조에 없던 석조 난간을 이동 통로에 추가했고, 병사들이 제단이 있는 오른쪽 안쪽이 아니라 화면 왼쪽으로 미는 구도여서 장소와 이동 방향 모두 어긋난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 병사들은 케이지를 향해 오른쪽으로 몸을 기울이고, 오른쪽 병사들은 왼쪽으로 몸을 기울여 민다. 따라서 힘의 방향이 서로 맞서며, 일행이 화면 오른쪽 성당 안쪽으로 함께 전진하는 동작으로 읽히지 않는다. 탈은 대체로 케이지 쪽을 향하지만 찰리의 얼굴은 카메라 쪽을 향한다. 무기나 조준 대상은 없다.",
        "built_space": "중앙의 케이지 한 개에는 닫힌 문 한 개가 있고, 넓은 문 쪽 면과 좁은 오른쪽 면이 보인다. 왼쪽 상층 철제 난간, 석재 기둥, 뒤쪽 제단 계단 한 벌, 목조 제단벽과 높은 의자 한 개가 참조와 대체로 이어진다. 촛대는 오른쪽에 두 개가 드러나며 나머지 제단 부분은 케이지에 가려진다. 병사들은 케이지 양옆과 뒤쪽에 분산되어 있고, 오른쪽으로 이어지는 진입 통로보다 제단을 정면으로 마주한 배치가 강하다. 바닥의 빛 반사는 창문과 광택 석재에 부합한다.",
        "entities": "찰리 한 명과 하회탈을 쓴 병사 다섯 명이 보인다. 병사들은 성인 남성형 체격이며 일부 노출된 얼굴과 피부는 아시아계 설정에 어긋나지 않지만, 탈 때문에 정확한 정체성은 확인하기 어렵다. 찰리는 각진 기계 머리, 흰 얼굴판, 작은 눈과 선형 입, 밀짚모자, 다색 우비, 줄무늬 상의, 녹색 바지와 노란 장화를 갖췄다. 다만 앉아 있는 찰리의 체구에서 병사들보다 거대하다는 인상은 약하다. 케이지는 참조처럼 낡은 강철 격자와 경첩 달린 문을 갖춘다. 가슴의 벗겨진 표식과 열린 철창 출입문은 식별되지 않으며, 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "병사들은 벌린 발로 바닥을 딛고 손을 철창과 기둥에 대고 있어 개별적으로 미는 자세는 가능하다. 찰리의 엉덩이 아래에는 어두운 좌판으로 보이는 지지물이 있고 장화도 케이지 바닥에 놓여 있다. 케이지 하부는 성당 바닥에 닿아 있어 떠 있지 않다. 바퀴 없이 무거운 케이지를 미끄러뜨리는 동작 자체는 가능하지만, 양쪽 병사들의 상반된 힘은 요청된 공동 전진을 설명하지 못한다."
       },
       {
        "label": "A",
        "direction": "오른쪽 병사들은 손을 철창에 대고 몸을 왼쪽으로 기울여 케이지를 화면 왼쪽으로 미는 모습이다. 제단과 성당 안쪽은 그들의 뒤쪽인 화면 오른쪽에 있어 요청된 진입 방향과 반대다. 앞쪽 금색 탈 병사는 얼굴을 카메라 쪽으로 돌리고, 뒤 병사들의 탈도 상당히 정면을 향한다. 찰리는 왼쪽을 바라본다. 무기나 조준 대상은 없다.",
        "built_space": "케이지 한 개의 닫힌 문과 긴 오른쪽 측면이 비스듬히 보인다. 기존의 왼쪽 상층 철제 난간, 기둥, 스테인드글라스, 제단 계단과 목조 의자는 이어지지만, 전경 통로 양옆에 석조 난간 두 구간이 새로 생겼다. 참조에서 열린 바닥과 계단으로 보이던 접근 공간을 이 난간들이 좁힌다. 오른쪽 제단에는 촛대 두 개가 보인다. 병사들은 주로 케이지 오른쪽과 뒤쪽에 있고 일부는 왼쪽 가장자리에 있다. 바닥 반사는 광원과 재질상 가능하다.",
        "entities": "찰리 한 명과 탈을 쓴 병사 다섯 명이 보이며, 병사들은 성인 체격이다. 얼굴이 대부분 가려져 민족적 정체성을 확정하기 어렵다. 탈의 종류는 다양하지만 여러 얼굴이 이동 경로보다 렌즈 쪽을 향한다. 찰리의 기계 얼굴, 밀짚모자, 다색 우비, 줄무늬 상의, 녹색 바지와 큰 노란 장화는 참조에 가깝다. 웅크린 체구는 거대한 포로라는 크기 관계를 충분히 드러내지 않는다. 강철 케이지의 재료와 격자 구조는 유사하나 참조보다 깊게 늘어난 형태다. 열린 철창 출입문과 벗겨진 가슴 표식은 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참조에서 개방되어 있던 제단 접근 공간에 고정 석조 난간 두 구간을 새로 만들어, 잠금된 장소의 구조와 통로를 변경했다."
        ],
        "physics": "찰리는 무릎을 굽히고 두 장화를 케이지 바닥에 댄 낮은 자세이며 손으로 철창을 잡고 있어 지지 관계가 보인다. 병사들도 발을 바닥에 딛고 손을 철창에 대고 있다. 케이지는 하부 틀이 바닥에 닿아 있어 공중에 뜬 물체는 없다. 병사들의 기울기와 팔의 접촉은 왼쪽으로 미는 힘을 설명하지만, 그것은 요청된 오른쪽 안쪽으로의 이동과 반대다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 서로 반대 방향으로 미는 병사들 배치 (물리적/상황적 모순)",
     "[gemini-pro] 레퍼런스에 없는 좌측 돌 난간 임의 생성",
     "[gpt-high] 참조에서 개방되어 있던 제단 접근 공간에 고정 석조 난간 두 구간을 새로 만들어, 잠금된 장소의 구조와 통로를 변경했다."
    ],
    "B": [
     "[gemini-pro] 지지대 없이 허공에 떠서 앉아 있는 찰리 (물리적 오류)",
     "[gemini-pro] 비스듬한 각도 지시를 무시한 케이지의 정면 배치",
     "[gemini-pro] 서로 반대 방향으로 미는 병사들 배치 (물리적/상황적 모순)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1500
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "양쪽에서 반대로 밀어 액션이 성립하지 않고 없는 난간이 추가되었으나, B보다 물리적 결함이 적습니다.  ★위반: [gemini-pro] 서로 반대 방향으로 미는 병사들 배치 (물리적/상황적 모순) / [gemini-pro] 레퍼런스에 없는 좌측 돌 난간 임의 생성 / [gpt-high] 참조에서 개방되어 있던 제단 접근 공간에 고정 석조 난간 두 구간을 새로 만들어, 잠금된 장소의 구조와 통로를 변경했다."
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "찰리가 지지대 없이 허공에 앉아 있고, 케이지 정면 배치로 구도를 어겨 치명적인 결함이 있습니다.  ★위반: [gemini-pro] 지지대 없이 허공에 떠서 앉아 있는 찰리 (물리적 오류) / [gemini-pro] 비스듬한 각도 지시를 무시한 케이지의 정면 배치 / [gemini-pro] 서로 반대 방향으로 미는 병사들 배치 (물리적/상황적 모순)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S58sh6_sel.png",
    "asset_id": "d5396759-9ead-4130-8986-3e4302fdcd1a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:885644>",
    "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 철창 케이지: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1148866>",
    "asset_id": "3e096b98-d7fb-48f0-b57f-d01d3f21144a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdec-9454-7eb1-8bb5-6cf8cf777eec",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S58sh6"
  }
 },
 "S58sh15::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:48:41.283468+00:00",
  "fingerprint": "dcfef881e3af72f720992b97470c63b957b637f5ace417908c5438f87cc875fc",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S58sh15_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S58sh15_sel.png",
  "source_sha256": "ed2d38fa5f908d6759e1f2d91ae3024a86a0465e532f3c5d0d04978e27e1b72f",
  "file": "S58sh15_cine.png",
  "staged_sha256": "2a7d6f37e4896d7920025b97f4f0198d072c1ca915edb2104f71104fa88fac67",
  "latency_ms": 9146
 },
 "S58sh27::signage": {
  "fp": "cb60a39c980fd6cd",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S58sh27": {
  "input_fingerprint": "1e1a1362614451a6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공을 날아온 날카로운 화살촉이 이현우 품에 안긴 아시아계 병사의 가슴팍에 정통으로 꽂힌 타격의 순간.\n\nLOCATION (lock): On the nave floor before the ruined church's dais, beneath the overlooking upper balcony and stained-glass daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 화살 (Embedded in the soldier's chest at impact) — The embedded end meets the chest while the exposed shaft projects away from it; its origin remains outside the image; used as A small, sharply readable impact detail surrounded by the soldier's torso and 이현우's arms; 파괴된 성당 내부 (Present behind the struggle); used as A narrow, defocused background margin maintains spatial continuity without suggesting a shooter.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the church's stained-glass daylight and restrained contrast, making the impact legible without introducing a new flash or light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The hostage soldier is held against Hyunwoo with an arrow embedded in his chest at the instant of death. His exact head angle, torso inclination, limb positions, and remaining support against the floor are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steel cage containing Charlie remains inside the ruined, stained-glass-lit church, with its door not yet closed by the final capture order. Charlie retains his straw hat, colorful raincoat, oversized boots and worn chest logo. 이현우: He has risen from the floor and is maintaining a chokehold, his previously treated leg still injured.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공을 날아온 날카로운 화살촉이 이현우 품에 안긴 아시아계 병사의 가슴팍에 정통으로 꽂힌 타격의 순간.\n\nLOCATION (lock): On the nave floor before the ruined church's dais, beneath the overlooking upper balcony and stained-glass daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 화살 (Embedded in the soldier's chest at impact) — The embedded end meets the chest while the exposed shaft projects away from it; its origin remains outside the image; used as A small, sharply readable impact detail surrounded by the soldier's torso and 이현우's arms; 파괴된 성당 내부 (Present behind the struggle); used as A narrow, defocused background margin maintains spatial continuity without suggesting a shooter.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the church's stained-glass daylight and restrained contrast, making the impact legible without introducing a new flash or light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The hostage soldier is held against Hyunwoo with an arrow embedded in his chest at the instant of death. His exact head angle, torso inclination, limb positions, and remaining support against the floor are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steel cage containing Charlie remains inside the ruined, stained-glass-lit church, with its door not yet closed by the final capture order. Charlie retains his straw hat, colorful raincoat, oversized boots and worn chest logo. 이현우: He has risen from the floor and is maintaining a chokehold, his previously treated leg still injured.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공을 날아온 날카로운 화살촉이 이현우 품에 안긴 아시아계 병사의 가슴팍에 정통으로 꽂힌 타격의 순간.\n\nLOCATION (lock): On the nave floor before the ruined church's dais, beneath the overlooking upper balcony and stained-glass daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 화살 (Embedded in the soldier's chest at impact) — The embedded end meets the chest while the exposed shaft projects away from it; its origin remains outside the image; used as A small, sharply readable impact detail surrounded by the soldier's torso and 이현우's arms; 파괴된 성당 내부 (Present behind the struggle); used as A narrow, defocused background margin maintains spatial continuity without suggesting a shooter.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the church's stained-glass daylight and restrained contrast, making the impact legible without introducing a new flash or light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The hostage soldier is held against Hyunwoo with an arrow embedded in his chest at the instant of death. His exact head angle, torso inclination, limb positions, and remaining support against the floor are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steel cage containing Charlie remains inside the ruined, stained-glass-lit church, with its door not yet closed by the final capture order. Charlie retains his straw hat, colorful raincoat, oversized boots and worn chest logo. 이현우: He has risen from the floor and is maintaining a chokehold, his previously treated leg still injured.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "화살촉이 병사의 가슴이 아닌 등에 꽂혀 있음. 이현우의 시선은 정면을 향함.",
    "built_space": "부서진 성당 내부로 제단과 스테인드글라스 창이 배경에 위치하며 공간 연속성이 유지됨.",
    "entities": "이현우는 레퍼런스와 일치하나, 병사는 뒷모습만 보이고 아시아계 외모로 식별하기 어려움.",
    "hard_violations": [
     "[gemini-pro] 화살이 가슴이 아닌 등에 꽂힘",
     "[gemini-pro] 초크홀드가 아닌 마주 안고 있는 잘못된 자세"
    ],
    "physics": "두 인물은 서로 기대어 서 있고, 화살은 조끼에 박힌 채 물리적으로 유지됨."
   },
   {
    "label": "B",
    "direction": "화살촉이 아시아계 병사의 가슴 정면에 정확히 꽂혀 있음. 두 인물의 시선은 앞쪽 허공을 향함.",
    "built_space": "성당 내부의 제단과 스테인드글라스 창문이 올바른 위치와 비율로 배경에 흐리게 묘사됨.",
    "entities": "이현우의 얼굴과 복장이 레퍼런스와 일치하며, 병사 역시 아시아계 외모를 명확히 보여줌.",
    "hard_violations": [],
    "physics": "이현우가 뒤에서 초크홀드 자세로 체중을 실어 병사를 지탱하고 있으며, 화살은 가슴에 박혀 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "아시아계 병사의 가슴에 화살이 꽂힌 타격 순간과 이현우의 초크홀드 자세를 정확히 구현함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화살이 가슴이 아닌 등에 꽂혔고, 지시된 초크홀드 자세가 아닌 마주 안은 포즈로 제시문을 심각하게 위반함."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "화살촉이 아시아계 병사의 가슴 정면에 정확히 꽂혀 있음. 두 인물의 시선은 앞쪽 허공을 향함.",
        "built_space": "성당 내부의 제단과 스테인드글라스 창문이 올바른 위치와 비율로 배경에 흐리게 묘사됨.",
        "entities": "이현우의 얼굴과 복장이 레퍼런스와 일치하며, 병사 역시 아시아계 외모를 명확히 보여줌.",
        "hard_violations": [],
        "physics": "이현우가 뒤에서 초크홀드 자세로 체중을 실어 병사를 지탱하고 있으며, 화살은 가슴에 박혀 있음."
       },
       {
        "label": "A",
        "direction": "화살촉이 병사의 가슴이 아닌 등에 꽂혀 있음. 이현우의 시선은 정면을 향함.",
        "built_space": "부서진 성당 내부로 제단과 스테인드글라스 창이 배경에 위치하며 공간 연속성이 유지됨.",
        "entities": "이현우는 레퍼런스와 일치하나, 병사는 뒷모습만 보이고 아시아계 외모로 식별하기 어려움.",
        "hard_violations": [
         "화살이 가슴이 아닌 등에 꽂힘",
         "초크홀드가 아닌 마주 안고 있는 잘못된 자세"
        ],
        "physics": "두 인물은 서로 기대어 서 있고, 화살은 조끼에 박힌 채 물리적으로 유지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "아시아계 병사의 가슴에 화살이 꽂힌 타격 순간과 이현우의 초크홀드 자세를 정확히 구현함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화살이 가슴이 아닌 등에 꽂혔고, 지시된 초크홀드 자세가 아닌 마주 안은 포즈로 제시문을 심각하게 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "화살촉이 아시아계 병사의 가슴 정면에 정확히 꽂혀 있음. 두 인물의 시선은 앞쪽 허공을 향함.",
        "built_space": "성당 내부의 제단과 스테인드글라스 창문이 올바른 위치와 비율로 배경에 흐리게 묘사됨.",
        "entities": "이현우의 얼굴과 복장이 레퍼런스와 일치하며, 병사 역시 아시아계 외모를 명확히 보여줌.",
        "hard_violations": [],
        "physics": "이현우가 뒤에서 초크홀드 자세로 체중을 실어 병사를 지탱하고 있으며, 화살은 가슴에 박혀 있음."
       },
       {
        "label": "A",
        "direction": "화살촉이 병사의 가슴이 아닌 등에 꽂혀 있음. 이현우의 시선은 정면을 향함.",
        "built_space": "부서진 성당 내부로 제단과 스테인드글라스 창이 배경에 위치하며 공간 연속성이 유지됨.",
        "entities": "이현우는 레퍼런스와 일치하나, 병사는 뒷모습만 보이고 아시아계 외모로 식별하기 어려움.",
        "hard_violations": [
         "화살이 가슴이 아닌 등에 꽂힘",
         "초크홀드가 아닌 마주 안고 있는 잘못된 자세"
        ],
        "physics": "두 인물은 서로 기대어 서 있고, 화살은 조끼에 박힌 채 물리적으로 유지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "가슴에 박힌 화살과 뒤에서 목·몸통을 붙잡는 관계는 정확하지만, 요구한 타격부 클로즈업보다 넓고 성당 배경과 화살이 지나치게 크게 드러난다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "화살이 가슴이 아니라 등에 박혀 있으며, 목 조르기도 마주 안는 자세로 바뀌어 핵심 타격 위치와 연속 동작을 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화살의 삽입부는 병사의 앞가슴 상처에 닿고, 노출된 자루와 깃은 화면 오른쪽 위로 뻗는다. 가슴을 향해 날아온 방향으로 읽히며 발사자는 보이지 않는다. 현우는 오른쪽 화면 밖을 보고, 병사는 눈을 반쯤 뜬 채 오른쪽 아래로 시선이 풀려 있다.",
        "built_space": "왼쪽 위 난간 발코니 한 곳, 오른쪽 뒤 제단 단상 한 곳과 여러 계단, 긴 촛대 두 개가 보인다. 뒤쪽의 큰 스테인드글라스 창 두 개와 가장자리의 부분 창들, 낡은 석재 기둥이 참고 성당과 이어진다. 두 사람은 단상 앞 본당 쪽에 있으며 구조적 모순은 보이지 않는다. 다만 바닥과 건축물이 화면의 넓은 부분을 차지하고 비교적 선명하여, 좁고 흐릿한 배경 여백이라는 요구에 맞지 않는다.",
        "entities": "두 사람만 보인다. 현우는 헝클어진 검은 머리의 젊은 동아시아계 남성이고 해진 짙은 후드 상의를 입어 참고와 대체로 맞지만 얼굴은 참고보다 성숙해 보인다. 병사는 짧은 검은 머리의 성인 아시아계 남성으로 밝은 군복을 입었다. 화살은 밝은 목재 자루, 갈색 줄무늬 깃, 붉은 감김을 갖춰 소품 참고와 대체로 일치한다. 촉 끝은 가슴 안에 가려져 있다. 읽을 수 있는 글자나 추가 인물은 없다. 철창과 찰리, 현우의 다리는 이 구도로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "현우의 한쪽 팔은 병사의 목 아래와 어깨를 감싸고 다른 팔은 복부를 붙들어 몸통을 지탱한다. 병사는 현우 쪽으로 기대고 팔은 아래로 처져 있으며, 기울어진 머리의 목 부근에도 현우의 팔이 받쳐져 있다. 하체와 바닥 접촉은 잘려 있어 판단할 수 있으나 공중에 떠 있다는 증거는 없다. 화살은 가슴에 박힌 부분으로 지지되고, 상처 주변의 작은 혈액 방울은 충격 순간의 비산으로 설명된다."
       },
       {
        "label": "B",
        "direction": "화살의 삽입부가 병사의 등 중앙에 닿고 자루는 오른쪽 위로 뻗는다. 보이는 명중 대상은 가슴이 아니라 등이다. 현우는 오른쪽 화면 밖을 바라보고 병사의 얼굴은 왼쪽 아래로 숙여져 있다. 발사자는 보이지 않는다.",
        "built_space": "왼쪽 위 발코니 한 곳, 뒤쪽 단상 한 곳과 여러 계단, 긴 촛대 두 개가 보이며, 큰 스테인드글라스 창 두 개와 부분적으로 잘린 창들이 참고의 석재 성당을 이어 준다. 두 사람의 위치는 단상 앞 본당으로 읽히고 불가능한 반사나 중복 설비는 없다. 그러나 배경과 바닥이 넓게 노출되어 요구된 좁고 흐릿한 배경 여백을 충족하지 못한다.",
        "entities": "현우와 병사 두 명만 보인다. 현우의 젊은 동아시아계 남성 얼굴, 검은 머리와 해진 짙은 후드는 참고에 대체로 부합한다. 병사는 짧은 머리의 성인 남성이고 얼굴이 일부 가려져 아시아계라는 설정을 명확히 확인하기는 어렵다. 올리브색 군복과 전술 조끼를 착용한다. 화살의 목재 자루와 줄무늬 깃, 붉은 감김은 참고와 유사하지만 박힌 위치가 다르다. 읽을 수 있는 글자는 없으며 철창과 찰리, 다리 부상은 화면 밖이다.",
        "hard_violations": [],
        "physics": "병사의 몸통은 현우의 가슴에 기대고 현우의 양손과 팔이 등을 감싸 지지한다. 병사의 머리는 아래로 숙여지고 보이는 팔도 처져 있어 중력에 따른 자세로 읽힌다. 화살은 등에 박힌 접점으로 지지되며 떠 있는 물체는 없다. 신체 지지는 가능하지만 현우가 뒤에서 유지해야 하는 목 조르기가 아니라 서로 마주 안는 배치다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "가슴에 박힌 화살과 뒤에서 목·몸통을 붙잡는 관계는 정확하지만, 요구한 타격부 클로즈업보다 넓고 성당 배경과 화살이 지나치게 크게 드러난다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화살이 가슴이 아니라 등에 박혀 있으며, 목 조르기도 마주 안는 자세로 바뀌어 핵심 타격 위치와 연속 동작을 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "화살의 삽입부는 병사의 앞가슴 상처에 닿고, 노출된 자루와 깃은 화면 오른쪽 위로 뻗는다. 가슴을 향해 날아온 방향으로 읽히며 발사자는 보이지 않는다. 현우는 오른쪽 화면 밖을 보고, 병사는 눈을 반쯤 뜬 채 오른쪽 아래로 시선이 풀려 있다.",
        "built_space": "왼쪽 위 난간 발코니 한 곳, 오른쪽 뒤 제단 단상 한 곳과 여러 계단, 긴 촛대 두 개가 보인다. 뒤쪽의 큰 스테인드글라스 창 두 개와 가장자리의 부분 창들, 낡은 석재 기둥이 참고 성당과 이어진다. 두 사람은 단상 앞 본당 쪽에 있으며 구조적 모순은 보이지 않는다. 다만 바닥과 건축물이 화면의 넓은 부분을 차지하고 비교적 선명하여, 좁고 흐릿한 배경 여백이라는 요구에 맞지 않는다.",
        "entities": "두 사람만 보인다. 현우는 헝클어진 검은 머리의 젊은 동아시아계 남성이고 해진 짙은 후드 상의를 입어 참고와 대체로 맞지만 얼굴은 참고보다 성숙해 보인다. 병사는 짧은 검은 머리의 성인 아시아계 남성으로 밝은 군복을 입었다. 화살은 밝은 목재 자루, 갈색 줄무늬 깃, 붉은 감김을 갖춰 소품 참고와 대체로 일치한다. 촉 끝은 가슴 안에 가려져 있다. 읽을 수 있는 글자나 추가 인물은 없다. 철창과 찰리, 현우의 다리는 이 구도로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "현우의 한쪽 팔은 병사의 목 아래와 어깨를 감싸고 다른 팔은 복부를 붙들어 몸통을 지탱한다. 병사는 현우 쪽으로 기대고 팔은 아래로 처져 있으며, 기울어진 머리의 목 부근에도 현우의 팔이 받쳐져 있다. 하체와 바닥 접촉은 잘려 있어 판단할 수 있으나 공중에 떠 있다는 증거는 없다. 화살은 가슴에 박힌 부분으로 지지되고, 상처 주변의 작은 혈액 방울은 충격 순간의 비산으로 설명된다."
       },
       {
        "label": "A",
        "direction": "화살의 삽입부가 병사의 등 중앙에 닿고 자루는 오른쪽 위로 뻗는다. 보이는 명중 대상은 가슴이 아니라 등이다. 현우는 오른쪽 화면 밖을 바라보고 병사의 얼굴은 왼쪽 아래로 숙여져 있다. 발사자는 보이지 않는다.",
        "built_space": "왼쪽 위 발코니 한 곳, 뒤쪽 단상 한 곳과 여러 계단, 긴 촛대 두 개가 보이며, 큰 스테인드글라스 창 두 개와 부분적으로 잘린 창들이 참고의 석재 성당을 이어 준다. 두 사람의 위치는 단상 앞 본당으로 읽히고 불가능한 반사나 중복 설비는 없다. 그러나 배경과 바닥이 넓게 노출되어 요구된 좁고 흐릿한 배경 여백을 충족하지 못한다.",
        "entities": "현우와 병사 두 명만 보인다. 현우의 젊은 동아시아계 남성 얼굴, 검은 머리와 해진 짙은 후드는 참고에 대체로 부합한다. 병사는 짧은 머리의 성인 남성이고 얼굴이 일부 가려져 아시아계라는 설정을 명확히 확인하기는 어렵다. 올리브색 군복과 전술 조끼를 착용한다. 화살의 목재 자루와 줄무늬 깃, 붉은 감김은 참고와 유사하지만 박힌 위치가 다르다. 읽을 수 있는 글자는 없으며 철창과 찰리, 다리 부상은 화면 밖이다.",
        "hard_violations": [],
        "physics": "병사의 몸통은 현우의 가슴에 기대고 현우의 양손과 팔이 등을 감싸 지지한다. 병사의 머리는 아래로 숙여지고 보이는 팔도 처져 있어 중력에 따른 자세로 읽힌다. 화살은 등에 박힌 접점으로 지지되며 떠 있는 물체는 없다. 신체 지지는 가능하지만 현우가 뒤에서 유지해야 하는 목 조르기가 아니라 서로 마주 안는 배치다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.857,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.607,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 화살이 가슴이 아닌 등에 꽂힘",
     "[gemini-pro] 초크홀드가 아닌 마주 안고 있는 잘못된 자세"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 607
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "아시아계 병사의 가슴에 화살이 꽂힌 타격 순간과 이현우의 초크홀드 자세를 정확히 구현함."
   },
   {
    "label": "A",
    "score": 607,
    "verdict_ko": "화살이 가슴이 아닌 등에 꽂혔고, 지시된 초크홀드 자세가 아닌 마주 안은 포즈로 제시문을 심각하게 위반함.  ★위반: [gemini-pro] 화살이 가슴이 아닌 등에 꽂힘 / [gemini-pro] 초크홀드가 아닌 마주 안고 있는 잘못된 자세"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S58sh15_sel.png",
    "asset_id": "49584c4d-7466-4300-843e-17a401ed7e9d",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 화살: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:883226>",
    "asset_id": "2134b4ba-9e55-444f-adb9-a28ecdc7f924",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacded-2c16-7e47-954c-fef9c4d42fc4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S58sh15"
  }
 },
 "S58sh27::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:48:50.768115+00:00",
  "fingerprint": "e79377ac5673ff7dec9a17016352cb6a4699c50041c2b5d010c623b6d4c743bd",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S58sh27_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S58sh27_sel.png",
  "source_sha256": "eefd7ca9d6466c96b24d7560130e68a9c16bef17247ca7e604f36b33a1e85af2",
  "file": "S58sh27_cine.png",
  "staged_sha256": "5b18f8c4f31b4ca85bf13b8ad7495fc8eefd517db9b185824feeba29a0ff977d",
  "latency_ms": 10439
 },
 "S59sh8::signage": {
  "fp": "43ab25d07b5c0502",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::ac016a0231fef2e0": {
  "subjects": [],
  "subject_text": "익산 마을 지하감옥 감방\n굵은 쇠창살로 막힌 어두운 지하 감방. 바닥 구석에 건초더미가 놓여 있고 낮은 곳에는 물이 고여 있다.",
  "identity": "canonical",
  "scope_id": "L79",
  "scope_role": "location_interior",
  "scope_sha": "d806381fc38370c4"
 },
 "S59sh8::bgfirst_bg": {
  "input_fingerprint": "9d3610fa4321b740",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수빈이 젖은 건초 더미를 앰버의 코앞에 바짝 가져다 댄 손끝.\n\nLOCATION (lock): In the damp corner of a village underground prison cell, beside wet straw and pooled water in faint light from the prison's small window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wet hay held just beneath 앰버's nose, not covering her face in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 물에 적신 건초 (Held in 수빈's fingertips immediately beneath 앰버's nose); used as A compact foreground detail linking the helping hand to 앰버's recovering breath; 감옥 내부 (The characters remain confined inside); used as A soft peripheral margin keeps the hand detail within the present-day cell.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the cell gently separates the fingertips and 앰버's face, without prescribing an unsupported source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수빈이 젖은 건초 더미를 앰버의 코앞에 바짝 가져다 댄 손끝.\n\nLOCATION (lock): In the damp corner of a village underground prison cell, beside wet straw and pooled water in faint light from the prison's small window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wet hay held just beneath 앰버's nose, not covering her face in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 물에 적신 건초 (Held in 수빈's fingertips immediately beneath 앰버's nose); used as A compact foreground detail linking the helping hand to 앰버's recovering breath; 감옥 내부 (The characters remain confined inside); used as A soft peripheral margin keeps the hand detail within the present-day cell.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the cell gently separates the fingertips and 앰버's face, without prescribing an unsupported source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S59sh8__bgfirst_bg.png",
  "asset_id": "8e06d7b9-d2c5-44d1-946a-9cc423ddb943",
  "input_asset_ids": [
   "5c645fc6-5658-4006-bf81-501a214760ad",
   "c45c80b3-4a14-44ff-869e-6a809de13b16"
  ]
 },
 "S59sh8": {
  "input_fingerprint": "46123ecf21c12a50",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈이 젖은 건초 더미를 앰버의 코앞에 바짝 가져다 댄 손끝.\n\nLOCATION (lock): In the damp corner of a village underground prison cell, beside wet straw and pooled water in faint light from the prison's small window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wet hay held just beneath 앰버's nose, not covering her face in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 물에 적신 건초 (Held in 수빈's fingertips immediately beneath 앰버's nose); used as A compact foreground detail linking the helping hand to 앰버's recovering breath; 감옥 내부 (The characters remain confined inside); used as A soft peripheral margin keeps the hand detail within the present-day cell.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the cell gently separates the fingertips and 앰버's face, without prescribing an unsupported source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground cell remains shut behind rigid iron bars, with hay and pooled water on its floor. 수빈: She has facial wounds and accumulated grime, and holds a handful of water-soaked hay. Radiation-damaged skin on her abdomen is already present beneath her T-shirt. 앰버: She is confined in the cell, with her severe coughing beginning to settle into steadier breathing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈이 젖은 건초 더미를 앰버의 코앞에 바짝 가져다 댄 손끝.\n\nLOCATION (lock): In the damp corner of a village underground prison cell, beside wet straw and pooled water in faint light from the prison's small window. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wet hay held just beneath 앰버's nose, not covering her face in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 물에 적신 건초 (Held in 수빈's fingertips immediately beneath 앰버's nose); used as A compact foreground detail linking the helping hand to 앰버's recovering breath; 감옥 내부 (The characters remain confined inside); used as A soft peripheral margin keeps the hand detail within the present-day cell.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the cell gently separates the fingertips and 앰버's face, without prescribing an unsupported source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground cell remains shut behind rigid iron bars, with hay and pooled water on its floor. 수빈: She has facial wounds and accumulated grime, and holds a handful of water-soaked hay. Radiation-damaged skin on her abdomen is already present beneath her T-shirt. 앰버: She is confined in the cell, with her severe coughing beginning to settle into steadier breathing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈이 젖은 건초 더미를 앰버의 코앞에 바짝 가져다 댄 손끝.\n\nLOCATION (lock): In the damp corner of a village underground prison cell, beside wet straw and pooled water in faint light from the prison's small window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wet hay held just beneath 앰버's nose, not covering her face in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 물에 적신 건초 (Held in 수빈's fingertips immediately beneath 앰버's nose); used as A compact foreground detail linking the helping hand to 앰버's recovering breath; 감옥 내부 (The characters remain confined inside); used as A soft peripheral margin keeps the hand detail within the present-day cell.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the cell gently separates the fingertips and 앰버's face, without prescribing an unsupported source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground cell remains shut behind rigid iron bars, with hay and pooled water on its floor. 수빈: She has facial wounds and accumulated grime, and holds a handful of water-soaked hay. Radiation-damaged skin on her abdomen is already present beneath her T-shirt. 앰버: She is confined in the cell, with her severe coughing beginning to settle into steadier breathing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S59sh8__bgfirst_bg.png",
     "asset_id": "8e06d7b9-d2c5-44d1-946a-9cc423ddb943",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S59sh8.png",
     "asset_id": "5c645fc6-5658-4006-bf81-501a214760ad",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904512>",
     "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L79B01.png",
     "asset_id": "c45c80b3-4a14-44ff-869e-6a809de13b16",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904512>",
     "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "수빈의 손이 건초를 앰버의 코밑으로 향함. 앰버는 건초를 내려다봄.",
    "built_space": "왼쪽의 쇠창살, 오른쪽 돌벽, 바닥의 물과 짚 등 감옥의 입체적인 공간이 확인됨.",
    "entities": "앰버의 외형(금발, 이목구비)이 일치함. 젖은 건초 더미가 묘사됨. 수빈의 팔은 레퍼런스의 재킷 소매가 없는 맨팔임.",
    "hard_violations": [],
    "physics": "손가락이 젖은 건초를 자연스럽게 쥐고 있으며 물방울이 떨어짐. 앰버는 바닥에 안정적으로 앉아 있음."
   },
   {
    "label": "B",
    "direction": "수빈의 손이 건초를 앰버의 코밑에 위치시킴. 앰버는 정면을 무심하게 응시함.",
    "built_space": "앰버 바로 뒤에 쇠창살과 자물쇠가 위치하여 감옥 구조를 보이나 공간감이 다소 평면적임.",
    "entities": "앰버의 외형이 일치함. 수빈의 팔에 올리브색 소매가 확인됨. 건초가 바닥의 더미가 아닌 끈으로 정갈하게 묶인 다발 형태임.",
    "hard_violations": [
     "[gemini-pro] invented objects: 프롬프트가 지시한 젖은 건초 더미가 아닌, 인위적으로 묶은 건초 다발이 등장함.",
     "[gpt-high] 앰버를 닫힌 감방 철창의 바깥쪽에 배치하여, 인물이 감방 내부에 갇혀 있어야 한다는 공간 지시를 위반합니다."
    ],
    "physics": "수빈의 손가락이 건초 다발을 쥐고 지탱하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "젖은 건초 더미와 앰버의 연출이 지침에 부합하나, 수빈의 손목 의상(소매)이 누락되어 감점됨."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "수빈의 소매는 표현되었으나, 인위적으로 묶인 건초 다발이라는 발명된 사물이 등장하여 심각한 위반임."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수빈의 손이 건초를 앰버의 코밑으로 향함. 앰버는 건초를 내려다봄.",
        "built_space": "왼쪽의 쇠창살, 오른쪽 돌벽, 바닥의 물과 짚 등 감옥의 입체적인 공간이 확인됨.",
        "entities": "앰버의 외형(금발, 이목구비)이 일치함. 젖은 건초 더미가 묘사됨. 수빈의 팔은 레퍼런스의 재킷 소매가 없는 맨팔임.",
        "hard_violations": [],
        "physics": "손가락이 젖은 건초를 자연스럽게 쥐고 있으며 물방울이 떨어짐. 앰버는 바닥에 안정적으로 앉아 있음."
       },
       {
        "label": "B",
        "direction": "수빈의 손이 건초를 앰버의 코밑에 위치시킴. 앰버는 정면을 무심하게 응시함.",
        "built_space": "앰버 바로 뒤에 쇠창살과 자물쇠가 위치하여 감옥 구조를 보이나 공간감이 다소 평면적임.",
        "entities": "앰버의 외형이 일치함. 수빈의 팔에 올리브색 소매가 확인됨. 건초가 바닥의 더미가 아닌 끈으로 정갈하게 묶인 다발 형태임.",
        "hard_violations": [
         "invented objects: 프롬프트가 지시한 젖은 건초 더미가 아닌, 인위적으로 묶은 건초 다발이 등장함."
        ],
        "physics": "수빈의 손가락이 건초 다발을 쥐고 지탱하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "젖은 건초 더미와 앰버의 연출이 지침에 부합하나, 수빈의 손목 의상(소매)이 누락되어 감점됨."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "수빈의 소매는 표현되었으나, 인위적으로 묶인 건초 다발이라는 발명된 사물이 등장하여 심각한 위반임."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "수빈의 손이 건초를 앰버의 코밑으로 향함. 앰버는 건초를 내려다봄.",
        "built_space": "왼쪽의 쇠창살, 오른쪽 돌벽, 바닥의 물과 짚 등 감옥의 입체적인 공간이 확인됨.",
        "entities": "앰버의 외형(금발, 이목구비)이 일치함. 젖은 건초 더미가 묘사됨. 수빈의 팔은 레퍼런스의 재킷 소매가 없는 맨팔임.",
        "hard_violations": [],
        "physics": "손가락이 젖은 건초를 자연스럽게 쥐고 있으며 물방울이 떨어짐. 앰버는 바닥에 안정적으로 앉아 있음."
       },
       {
        "label": "B",
        "direction": "수빈의 손이 건초를 앰버의 코밑에 위치시킴. 앰버는 정면을 무심하게 응시함.",
        "built_space": "앰버 바로 뒤에 쇠창살과 자물쇠가 위치하여 감옥 구조를 보이나 공간감이 다소 평면적임.",
        "entities": "앰버의 외형이 일치함. 수빈의 팔에 올리브색 소매가 확인됨. 건초가 바닥의 더미가 아닌 끈으로 정갈하게 묶인 다발 형태임.",
        "hard_violations": [
         "invented objects: 프롬프트가 지시한 젖은 건초 더미가 아닌, 인위적으로 묶은 건초 다발이 등장함."
        ],
        "physics": "수빈의 손가락이 건초 다발을 쥐고 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "건초를 코 밑에 대는 행동은 보이지만, 앰버가 감방 철창 바깥에 놓였고 큰 건초 묶음이 입과 얼굴 하단을 가려 핵심 공간·구도 지시를 어깁니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "수빈의 손끝이 젖은 건초를 앰버의 코 바로 아래로 가져가는 근접 구도와 드러난 얼굴이 더 충실하지만, 건초가 다소 길고 앰버의 참고 의상인 녹색 재킷이 보이지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 정면보다 아래쪽의 건초와 손을 내려다봅니다. 왼쪽 위에서 내려온 손이 건초를 코 아래로 내밀어 대상은 맞지만, 가로로 넓은 묶음이 입과 얼굴 하단 앞을 막습니다.",
        "built_space": "뒤쪽에 주요 세로 철창 약 7개와 가로대 1개, 오른쪽 문틀과 잠금장치 1개, 높은 작은 창 1개가 보입니다. 얼룩진 벽과 물 고인 바닥은 참고 장소와 유사합니다. 그러나 참고 사진의 감방 내부가 철창 너머로 펼쳐지고 앰버는 그 철창보다 카메라 쪽에 있어, 감방 안에 갇힌 인물이 아니라 문 바깥에 놓인 배치입니다.",
        "entities": "금발의 어린 여자아이, 큰 눈, 녹색 재킷과 회색 속옷은 앰버의 참고 외형에 대체로 맞습니다. 혼혈 정체성은 외형만으로 확정할 수 없습니다. 수빈은 손과 팔목만 보이며 상처와 때, 짙은 시보리 소매가 있습니다. 손은 참고 인물보다 다소 크고 굵게 읽힙니다. 젖은 건초와 물방울은 보이지만, 한 줌의 느슨한 건초 대신 가운데를 묶어 정돈한 다발입니다. 추가 인물이나 읽을 수 있는 문자는 없습니다.",
        "hard_violations": [
         "앰버를 닫힌 감방 철창의 바깥쪽에 배치하여, 인물이 감방 내부에 갇혀 있어야 한다는 공간 지시를 위반합니다."
        ],
        "physics": "건초는 손바닥과 아래쪽으로 굽힌 손가락에 받쳐져 있고 팔목과 팔도 연결되어 있습니다. 물방울은 건초와 손끝에서 아래로 떨어집니다. 앰버의 하체 지지면은 근접 구도 밖이지만, 보이는 상체에 부유나 불가능한 자세는 없습니다."
       },
       {
        "label": "B",
        "direction": "앰버는 고개를 숙여 왼쪽 아래의 젖은 건초를 바라봅니다. 왼쪽에서 들어온 손이 건초 끝을 코 바로 아래와 앞쪽으로 향하게 잡고 있어 행동의 목표가 분명하며, 코와 입을 가리지 않습니다.",
        "built_space": "왼쪽과 뒤쪽에 주요 세로 철창 약 6개, 가로대 1개, 왼쪽 잠금장치 1개와 높은 작은 창 1개가 보입니다. 오른쪽에는 앰버가 기댄 거친 벽이 있고 아래에는 물 고인 바닥과 건초가 있습니다. 문 옆 벽 모서리에 몸을 웅크린 배치로 읽히며, 참고 장소의 축축한 벽·철창·바닥 재질을 유지합니다. 다만 철창 너머 공간도 크게 보여 정확한 감방 경계는 선명하지 않습니다.",
        "entities": "금발과 큰 눈을 가진 어린 여자아이로 앰버의 나이·성별·주요 외형에 대체로 부합합니다. 혼혈 정체성은 외형만으로 확정할 수 없습니다. 보이는 옷은 낡은 회색 옷이며 참고의 녹색 재킷은 없습니다. 수빈의 손과 팔은 때가 묻은 가는 성인 손으로 표현되고, 손끝이 실제 젖은 건초를 잡습니다. 얼굴과 머리는 프레임 밖이므로 수빈의 얼굴 상처나 단발은 평가할 수 없습니다. 건초의 젖은 광택과 물방울, 바닥의 물이 보이며 추가 인물과 읽을 수 있는 문자는 없습니다.",
        "hard_violations": [],
        "physics": "엄지와 굽힌 손가락이 건초를 집어 지지하고, 긴 줄기는 중력 방향으로 늘어집니다. 물방울도 아래로 떨어집니다. 앰버는 무릎을 끌어올리고 벽에 기대어 웅크린 자세로, 보이는 몸과 주변 벽·바닥 관계에 물리적 모순이 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "건초를 코 밑에 대는 행동은 보이지만, 앰버가 감방 철창 바깥에 놓였고 큰 건초 묶음이 입과 얼굴 하단을 가려 핵심 공간·구도 지시를 어깁니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "수빈의 손끝이 젖은 건초를 앰버의 코 바로 아래로 가져가는 근접 구도와 드러난 얼굴이 더 충실하지만, 건초가 다소 길고 앰버의 참고 의상인 녹색 재킷이 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 정면보다 아래쪽의 건초와 손을 내려다봅니다. 왼쪽 위에서 내려온 손이 건초를 코 아래로 내밀어 대상은 맞지만, 가로로 넓은 묶음이 입과 얼굴 하단 앞을 막습니다.",
        "built_space": "뒤쪽에 주요 세로 철창 약 7개와 가로대 1개, 오른쪽 문틀과 잠금장치 1개, 높은 작은 창 1개가 보입니다. 얼룩진 벽과 물 고인 바닥은 참고 장소와 유사합니다. 그러나 참고 사진의 감방 내부가 철창 너머로 펼쳐지고 앰버는 그 철창보다 카메라 쪽에 있어, 감방 안에 갇힌 인물이 아니라 문 바깥에 놓인 배치입니다.",
        "entities": "금발의 어린 여자아이, 큰 눈, 녹색 재킷과 회색 속옷은 앰버의 참고 외형에 대체로 맞습니다. 혼혈 정체성은 외형만으로 확정할 수 없습니다. 수빈은 손과 팔목만 보이며 상처와 때, 짙은 시보리 소매가 있습니다. 손은 참고 인물보다 다소 크고 굵게 읽힙니다. 젖은 건초와 물방울은 보이지만, 한 줌의 느슨한 건초 대신 가운데를 묶어 정돈한 다발입니다. 추가 인물이나 읽을 수 있는 문자는 없습니다.",
        "hard_violations": [
         "앰버를 닫힌 감방 철창의 바깥쪽에 배치하여, 인물이 감방 내부에 갇혀 있어야 한다는 공간 지시를 위반합니다."
        ],
        "physics": "건초는 손바닥과 아래쪽으로 굽힌 손가락에 받쳐져 있고 팔목과 팔도 연결되어 있습니다. 물방울은 건초와 손끝에서 아래로 떨어집니다. 앰버의 하체 지지면은 근접 구도 밖이지만, 보이는 상체에 부유나 불가능한 자세는 없습니다."
       },
       {
        "label": "A",
        "direction": "앰버는 고개를 숙여 왼쪽 아래의 젖은 건초를 바라봅니다. 왼쪽에서 들어온 손이 건초 끝을 코 바로 아래와 앞쪽으로 향하게 잡고 있어 행동의 목표가 분명하며, 코와 입을 가리지 않습니다.",
        "built_space": "왼쪽과 뒤쪽에 주요 세로 철창 약 6개, 가로대 1개, 왼쪽 잠금장치 1개와 높은 작은 창 1개가 보입니다. 오른쪽에는 앰버가 기댄 거친 벽이 있고 아래에는 물 고인 바닥과 건초가 있습니다. 문 옆 벽 모서리에 몸을 웅크린 배치로 읽히며, 참고 장소의 축축한 벽·철창·바닥 재질을 유지합니다. 다만 철창 너머 공간도 크게 보여 정확한 감방 경계는 선명하지 않습니다.",
        "entities": "금발과 큰 눈을 가진 어린 여자아이로 앰버의 나이·성별·주요 외형에 대체로 부합합니다. 혼혈 정체성은 외형만으로 확정할 수 없습니다. 보이는 옷은 낡은 회색 옷이며 참고의 녹색 재킷은 없습니다. 수빈의 손과 팔은 때가 묻은 가는 성인 손으로 표현되고, 손끝이 실제 젖은 건초를 잡습니다. 얼굴과 머리는 프레임 밖이므로 수빈의 얼굴 상처나 단발은 평가할 수 없습니다. 건초의 젖은 광택과 물방울, 바닥의 물이 보이며 추가 인물과 읽을 수 있는 문자는 없습니다.",
        "hard_violations": [],
        "physics": "엄지와 굽힌 손가락이 건초를 집어 지지하고, 긴 줄기는 중력 방향으로 늘어집니다. 물방울도 아래로 떨어집니다. 앰버는 무릎을 끌어올리고 벽에 기대어 웅크린 자세로, 보이는 몸과 주변 벽·바닥 관계에 물리적 모순이 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.071
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.821
   },
   "violations": {
    "B": [
     "[gemini-pro] invented objects: 프롬프트가 지시한 젖은 건초 더미가 아닌, 인위적으로 묶은 건초 다발이 등장함.",
     "[gpt-high] 앰버를 닫힌 감방 철창의 바깥쪽에 배치하여, 인물이 감방 내부에 갇혀 있어야 한다는 공간 지시를 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 821
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "젖은 건초 더미와 앰버의 연출이 지침에 부합하나, 수빈의 손목 의상(소매)이 누락되어 감점됨."
   },
   {
    "label": "B",
    "score": 821,
    "verdict_ko": "수빈의 소매는 표현되었으나, 인위적으로 묶인 건초 다발이라는 발명된 사물이 등장하여 심각한 위반임.  ★위반: [gemini-pro] invented objects: 프롬프트가 지시한 젖은 건초 더미가 아닌, 인위적으로 묶은 건초 다발이 등장함. / [gpt-high] 앰버를 닫힌 감방 철창의 바깥쪽에 배치하여, 인물이 감방 내부에 갇혀 있어야 한다는 공간 지시를 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L79B01.png",
    "asset_id": "c45c80b3-4a14-44ff-869e-6a809de13b16",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:904512>",
    "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacded-d896-754d-8ff5-c8425d536a6e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S59sh8__bgfirst_bg.png",
   "bg_asset_id": "8e06d7b9-d2c5-44d1-946a-9cc423ddb943",
   "bg_record_key": "S59sh8::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S59sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:49:01.628657+00:00",
  "fingerprint": "ed74f143c39c1b6522fc2c5894af3c89d5b5ae222fbfe13671d0d31b732ab707",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S59sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S59sh8_sel.png",
  "source_sha256": "ce896b28430456055939b8924f6585d2f813b286a39d3d374f83431d69f6d39c",
  "file": "S59sh8_cine.png",
  "staged_sha256": "b5df29b3d94c8d0da52c39877c5862d182a6a44369ea7e1517543373358a3197",
  "latency_ms": 12867
 },
 "S59sh24::signage": {
  "fp": "b4531f5adea01281",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S59sh24": {
  "input_fingerprint": "041537f3fad8d418",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 벌겋게 짓무르고 피부가 흉측하게 벗겨진 수빈의 피폭된 복부 상처 클로즈업.\n\nLOCATION (lock): Inside the damp underground prison cell, near the prisoners' resting place in weak ambient daylight from the prison window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 수빈의 티셔츠 (Lifted to expose her abdominal injury) — The raised front hem appears along the upper edge, held clear of the injured area; used as Provides the action's context and an upper framing boundary without concealing the injury; 감옥 내부 (Present-day cell, before the recollection begins); used as Thin peripheral background prevents the close detail from reading as a detached medical image or flashback.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the cell's neutral ambient illumination consistent, rendering the injury's stated redness clearly without stylized glow or an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground cell remains locked, with hay and standing water on the floor; the iron bars remain undamaged despite the escape attempts. 수빈: She has lifted her T-shirt to expose radiation-damaged skin across her abdomen. Her facial wounds and grime remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 벌겋게 짓무르고 피부가 흉측하게 벗겨진 수빈의 피폭된 복부 상처 클로즈업.\n\nLOCATION (lock): Inside the damp underground prison cell, near the prisoners' resting place in weak ambient daylight from the prison window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 수빈의 티셔츠 (Lifted to expose her abdominal injury) — The raised front hem appears along the upper edge, held clear of the injured area; used as Provides the action's context and an upper framing boundary without concealing the injury; 감옥 내부 (Present-day cell, before the recollection begins); used as Thin peripheral background prevents the close detail from reading as a detached medical image or flashback.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the cell's neutral ambient illumination consistent, rendering the injury's stated redness clearly without stylized glow or an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground cell remains locked, with hay and standing water on the floor; the iron bars remain undamaged despite the escape attempts. 수빈: She has lifted her T-shirt to expose radiation-damaged skin across her abdomen. Her facial wounds and grime remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 벌겋게 짓무르고 피부가 흉측하게 벗겨진 수빈의 피폭된 복부 상처 클로즈업.\n\nLOCATION (lock): Inside the damp underground prison cell, near the prisoners' resting place in weak ambient daylight from the prison window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 수빈의 티셔츠 (Lifted to expose her abdominal injury) — The raised front hem appears along the upper edge, held clear of the injured area; used as Provides the action's context and an upper framing boundary without concealing the injury; 감옥 내부 (Present-day cell, before the recollection begins); used as Thin peripheral background prevents the close detail from reading as a detached medical image or flashback.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the cell's neutral ambient illumination consistent, rendering the injury's stated redness clearly without stylized glow or an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground cell remains locked, with hay and standing water on the floor; the iron bars remain undamaged despite the escape attempts. 수빈: She has lifted her T-shirt to expose radiation-damaged skin across her abdomen. Her facial wounds and grime remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 수빈의 복부를 정면에서 바라보고 있음.",
    "built_space": "배경의 감옥 철창, 볏짚, 젖은 바닥이 레퍼런스 이미지의 환경과 일치하게 배치되어 있음.",
    "entities": "올리브색 의상을 입은 몸통. '벌겋게 짓무른 상처'는 전혀 보이지 않고 복부 전체가 붕대로 덮여 있음. 프롬프트에 없는 다른 사람의 손이 등장함.",
    "hard_violations": [
     "[gemini-pro] 지정된 상처를 완전히 가리는 붕대(invented object) 생성",
     "[gemini-pro] 프롬프트에 명시되지 않은 제3자의 손(invented people) 추가",
     "[gpt-high] 지시되지 않은 넓은 복부 붕대를 추가하여 필수 피폭 상처를 완전히 가렸습니다.",
     "[gpt-high] 양옆으로 내려온 두 팔과 별도로 복부를 덮는 세 번째 팔이 보여, 추가 인물의 신체 일부 또는 불가능한 팔 중복이 발생했습니다."
    ],
    "physics": "셔츠가 위로 말려 올라가 있으며, 화면 밖에서 들어온 손이 복부의 붕대 위에 안착해 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 수빈의 복부를 정면에서 바라보고 있음.",
    "built_space": "배경에 철창, 볏짚, 물웅덩이가 있는 감옥 내부가 레퍼런스와 일치하게 나타남.",
    "entities": "올리브색 셔츠를 입은 수빈의 몸통. 프롬프트가 명시한 '벌겋게 짓무르고 벗겨진 상처'가 없고 크고 하얀 사각형 반창고가 복부를 덮고 있음.",
    "hard_violations": [
     "[gemini-pro] 지정된 상처를 가리는 대형 반창고(invented object) 생성",
     "[gpt-high] 지시되지 않은 대형 접착식 드레싱을 복부에 추가하여, 반드시 보여야 하는 상처를 완전히 가렸습니다."
    ],
    "physics": "수빈의 오른손이 셔츠 밑단을 단단히 쥐고 위로 들어 올리고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "수빈이 셔츠를 들어 올리는 동작은 맞으나, 붉고 벗겨진 피폭 상처 대신 하얀 반창고가 붙어 있어 핵심 피사체 묘사에 완전히 실패했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "요구된 붉은 상처 대신 복부 전체에 붕대가 감겨 있으며, 프롬프트에 없는 제3자의 손이 등장하여 치명적인 규정 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 수빈의 복부를 정면에서 바라보고 있음.",
        "built_space": "배경의 감옥 철창, 볏짚, 젖은 바닥이 레퍼런스 이미지의 환경과 일치하게 배치되어 있음.",
        "entities": "올리브색 의상을 입은 몸통. '벌겋게 짓무른 상처'는 전혀 보이지 않고 복부 전체가 붕대로 덮여 있음. 프롬프트에 없는 다른 사람의 손이 등장함.",
        "hard_violations": [
         "지정된 상처를 완전히 가리는 붕대(invented object) 생성",
         "프롬프트에 명시되지 않은 제3자의 손(invented people) 추가"
        ],
        "physics": "셔츠가 위로 말려 올라가 있으며, 화면 밖에서 들어온 손이 복부의 붕대 위에 안착해 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 수빈의 복부를 정면에서 바라보고 있음.",
        "built_space": "배경에 철창, 볏짚, 물웅덩이가 있는 감옥 내부가 레퍼런스와 일치하게 나타남.",
        "entities": "올리브색 셔츠를 입은 수빈의 몸통. 프롬프트가 명시한 '벌겋게 짓무르고 벗겨진 상처'가 없고 크고 하얀 사각형 반창고가 복부를 덮고 있음.",
        "hard_violations": [
         "지정된 상처를 가리는 대형 반창고(invented object) 생성"
        ],
        "physics": "수빈의 오른손이 셔츠 밑단을 단단히 쥐고 위로 들어 올리고 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "수빈이 셔츠를 들어 올리는 동작은 맞으나, 붉고 벗겨진 피폭 상처 대신 하얀 반창고가 붙어 있어 핵심 피사체 묘사에 완전히 실패했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "요구된 붉은 상처 대신 복부 전체에 붕대가 감겨 있으며, 프롬프트에 없는 제3자의 손이 등장하여 치명적인 규정 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 수빈의 복부를 정면에서 바라보고 있음.",
        "built_space": "배경의 감옥 철창, 볏짚, 젖은 바닥이 레퍼런스 이미지의 환경과 일치하게 배치되어 있음.",
        "entities": "올리브색 의상을 입은 몸통. '벌겋게 짓무른 상처'는 전혀 보이지 않고 복부 전체가 붕대로 덮여 있음. 프롬프트에 없는 다른 사람의 손이 등장함.",
        "hard_violations": [
         "지정된 상처를 완전히 가리는 붕대(invented object) 생성",
         "프롬프트에 명시되지 않은 제3자의 손(invented people) 추가"
        ],
        "physics": "셔츠가 위로 말려 올라가 있으며, 화면 밖에서 들어온 손이 복부의 붕대 위에 안착해 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 수빈의 복부를 정면에서 바라보고 있음.",
        "built_space": "배경에 철창, 볏짚, 물웅덩이가 있는 감옥 내부가 레퍼런스와 일치하게 나타남.",
        "entities": "올리브색 셔츠를 입은 수빈의 몸통. 프롬프트가 명시한 '벌겋게 짓무르고 벗겨진 상처'가 없고 크고 하얀 사각형 반창고가 복부를 덮고 있음.",
        "hard_violations": [
         "지정된 상처를 가리는 대형 반창고(invented object) 생성"
        ],
        "physics": "수빈의 오른손이 셔츠 밑단을 단단히 쥐고 위로 들어 올리고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "티셔츠를 손으로 들어 올리는 동작은 맞지만, 임의의 드레싱으로 핵심인 붉게 짓무르고 벗겨진 복부 상처를 완전히 가렸으며 구도도 지나치게 넓습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "복부 전체를 임의의 붕대로 가려 상처가 없고, 양옆으로 내려온 두 팔 외에 복부를 덮는 팔까지 추가되어 인물 제한과 신체 구성을 위반합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 화면 밖입니다. 화면 왼쪽 손이 티셔츠 앞자락을 잡아 위로 당기며 복부를 노출하는 방향은 맞지만, 드러난 대상은 상처가 아니라 흰 드레싱입니다.",
        "built_space": "왼쪽 배경에 녹슨 수직 철창 여러 개, 문 잠금장치 하나, 뒤쪽 채광창 하나가 보입니다. 바닥의 짚과 물기도 이전 장면의 재질 및 낮 조명과 대체로 이어집니다. 다만 감옥 배경과 상체가 상당한 면적을 차지하여, 배경을 얇게 남기는 복부 상처 클로즈업보다 넓습니다.",
        "entities": "한 사람의 몸통과 두 팔 일부가 보이고, 얼굴과 머리는 제외되어 한국인 여성의 얼굴·나이·단발 일치 여부는 확인할 수 없습니다. 때 묻은 회녹색 티셔츠와 바지, 복부 피부가 보이지만 요구된 붉은 짓무름과 피부 박리는 보이지 않습니다. 대신 지시되지 않은 커다란 접착식 드레싱 하나가 붙어 있습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "지시되지 않은 대형 접착식 드레싱을 복부에 추가하여, 반드시 보여야 하는 상처를 완전히 가렸습니다."
        ],
        "physics": "손가락이 티셔츠를 실제로 움켜쥐고 있어 올라간 옷자락의 지지가 보입니다. 드레싱은 접착 테두리로 피부에 부착되어 있습니다. 팔과 몸통 연결에 명백한 물리적 모순은 없으며, 하체 지지점은 클로즈업 밖이라 판단할 수 없습니다."
       },
       {
        "label": "B",
        "direction": "얼굴과 시선은 보이지 않습니다. 전경의 손은 복부를 가로질러 화면 오른쪽으로 향하며 붕대를 누르고 있고, 티셔츠 앞자락을 들어 올리는 손동작은 보이지 않습니다. 손과 붕대 모두 상처를 드러내기보다 가리는 배치입니다.",
        "built_space": "왼쪽에 잠금장치 하나가 달린 철창문과 가로대, 여러 수직 철봉이 있고, 뒤쪽에 채광창 하나가 보입니다. 짚과 고인 물, 거친 벽의 재질은 이전 장면과 대체로 부합합니다. 그러나 감옥문과 바닥이 넓게 드러나며, 요구된 상처 중심의 밀착 구도와 얇은 주변 배경을 충족하지 못합니다.",
        "entities": "때 묻은 티셔츠와 바지를 입은 몸통이 보이지만 얼굴과 머리는 화면 밖이라 인물 정체성의 세부 일치는 확인할 수 없습니다. 복부에는 붉은 상처 대신 몸통을 넓게 감싼 흰 붕대가 있습니다. 몸통 양옆의 내려온 두 팔 외에, 화면 아래 왼쪽에서 들어와 복부를 덮는 별도의 팔과 손이 보입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "지시되지 않은 넓은 복부 붕대를 추가하여 필수 피폭 상처를 완전히 가렸습니다.",
         "양옆으로 내려온 두 팔과 별도로 복부를 덮는 세 번째 팔이 보여, 추가 인물의 신체 일부 또는 불가능한 팔 중복이 발생했습니다."
        ],
        "physics": "붕대는 몸통을 둘러 감겨 지지되고, 티셔츠는 말아 올린 주름이 몸통에 걸쳐 있습니다. 전경 손은 붕대에 접촉하지만, 양쪽 팔이 이미 몸통 옆으로 내려와 있어 이 손의 팔을 수빈의 신체에 정상적으로 연결할 수 없습니다. 발과 좌면은 화면 밖입니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "티셔츠를 손으로 들어 올리는 동작은 맞지만, 임의의 드레싱으로 핵심인 붉게 짓무르고 벗겨진 복부 상처를 완전히 가렸으며 구도도 지나치게 넓습니다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "복부 전체를 임의의 붕대로 가려 상처가 없고, 양옆으로 내려온 두 팔 외에 복부를 덮는 팔까지 추가되어 인물 제한과 신체 구성을 위반합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 화면 밖입니다. 화면 왼쪽 손이 티셔츠 앞자락을 잡아 위로 당기며 복부를 노출하는 방향은 맞지만, 드러난 대상은 상처가 아니라 흰 드레싱입니다.",
        "built_space": "왼쪽 배경에 녹슨 수직 철창 여러 개, 문 잠금장치 하나, 뒤쪽 채광창 하나가 보입니다. 바닥의 짚과 물기도 이전 장면의 재질 및 낮 조명과 대체로 이어집니다. 다만 감옥 배경과 상체가 상당한 면적을 차지하여, 배경을 얇게 남기는 복부 상처 클로즈업보다 넓습니다.",
        "entities": "한 사람의 몸통과 두 팔 일부가 보이고, 얼굴과 머리는 제외되어 한국인 여성의 얼굴·나이·단발 일치 여부는 확인할 수 없습니다. 때 묻은 회녹색 티셔츠와 바지, 복부 피부가 보이지만 요구된 붉은 짓무름과 피부 박리는 보이지 않습니다. 대신 지시되지 않은 커다란 접착식 드레싱 하나가 붙어 있습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "지시되지 않은 대형 접착식 드레싱을 복부에 추가하여, 반드시 보여야 하는 상처를 완전히 가렸습니다."
        ],
        "physics": "손가락이 티셔츠를 실제로 움켜쥐고 있어 올라간 옷자락의 지지가 보입니다. 드레싱은 접착 테두리로 피부에 부착되어 있습니다. 팔과 몸통 연결에 명백한 물리적 모순은 없으며, 하체 지지점은 클로즈업 밖이라 판단할 수 없습니다."
       },
       {
        "label": "A",
        "direction": "얼굴과 시선은 보이지 않습니다. 전경의 손은 복부를 가로질러 화면 오른쪽으로 향하며 붕대를 누르고 있고, 티셔츠 앞자락을 들어 올리는 손동작은 보이지 않습니다. 손과 붕대 모두 상처를 드러내기보다 가리는 배치입니다.",
        "built_space": "왼쪽에 잠금장치 하나가 달린 철창문과 가로대, 여러 수직 철봉이 있고, 뒤쪽에 채광창 하나가 보입니다. 짚과 고인 물, 거친 벽의 재질은 이전 장면과 대체로 부합합니다. 그러나 감옥문과 바닥이 넓게 드러나며, 요구된 상처 중심의 밀착 구도와 얇은 주변 배경을 충족하지 못합니다.",
        "entities": "때 묻은 티셔츠와 바지를 입은 몸통이 보이지만 얼굴과 머리는 화면 밖이라 인물 정체성의 세부 일치는 확인할 수 없습니다. 복부에는 붉은 상처 대신 몸통을 넓게 감싼 흰 붕대가 있습니다. 몸통 양옆의 내려온 두 팔 외에, 화면 아래 왼쪽에서 들어와 복부를 덮는 별도의 팔과 손이 보입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "지시되지 않은 넓은 복부 붕대를 추가하여 필수 피폭 상처를 완전히 가렸습니다.",
         "양옆으로 내려온 두 팔과 별도로 복부를 덮는 세 번째 팔이 보여, 추가 인물의 신체 일부 또는 불가능한 팔 중복이 발생했습니다."
        ],
        "physics": "붕대는 몸통을 둘러 감겨 지지되고, 티셔츠는 말아 올린 주름이 몸통에 걸쳐 있습니다. 전경 손은 붕대에 접촉하지만, 양쪽 팔이 이미 몸통 옆으로 내려와 있어 이 손의 팔을 수빈의 신체에 정상적으로 연결할 수 없습니다. 발과 좌면은 화면 밖입니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.167,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.917,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 지정된 상처를 완전히 가리는 붕대(invented object) 생성",
     "[gemini-pro] 프롬프트에 명시되지 않은 제3자의 손(invented people) 추가",
     "[gpt-high] 지시되지 않은 넓은 복부 붕대를 추가하여 필수 피폭 상처를 완전히 가렸습니다.",
     "[gpt-high] 양옆으로 내려온 두 팔과 별도로 복부를 덮는 세 번째 팔이 보여, 추가 인물의 신체 일부 또는 불가능한 팔 중복이 발생했습니다."
    ],
    "B": [
     "[gemini-pro] 지정된 상처를 가리는 대형 반창고(invented object) 생성",
     "[gpt-high] 지시되지 않은 대형 접착식 드레싱을 복부에 추가하여, 반드시 보여야 하는 상처를 완전히 가렸습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 917
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "수빈이 셔츠를 들어 올리는 동작은 맞으나, 붉고 벗겨진 피폭 상처 대신 하얀 반창고가 붙어 있어 핵심 피사체 묘사에 완전히 실패했습니다.  ★위반: [gemini-pro] 지정된 상처를 가리는 대형 반창고(invented object) 생성 / [gpt-high] 지시되지 않은 대형 접착식 드레싱을 복부에 추가하여, 반드시 보여야 하는 상처를 완전히 가렸습니다."
   },
   {
    "label": "A",
    "score": 917,
    "verdict_ko": "요구된 붉은 상처 대신 복부 전체에 붕대가 감겨 있으며, 프롬프트에 없는 제3자의 손이 등장하여 치명적인 규정 위반이 발생했습니다.  ★위반: [gemini-pro] 지정된 상처를 완전히 가리는 붕대(invented object) 생성 / [gemini-pro] 프롬프트에 명시되지 않은 제3자의 손(invented people) 추가 / [gpt-high] 지시되지 않은 넓은 복부 붕대를 추가하여 필수 피폭 상처를 완전히 가렸습니다. / [gpt-high] 양옆으로 내려온 두 팔과 별도로 복부를 덮는 세 번째 팔이 보여, 추가 인물의 신체 일부 또는 불가능한 팔 중복이 발생했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S59sh8_sel.png",
    "asset_id": "925ff5b0-d1c3-4826-84b1-2842a55a3380",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1268814>",
    "asset_id": "1fd7271e-b207-4382-a59d-3e1cfc4b27df",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdee-ad5f-76ba-830a-e348a75c51dc",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S59sh8"
  }
 },
 "S59sh24::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:49:14.846704+00:00",
  "fingerprint": "5f1d11fb8e0e5f5a1d1fc55d815d45b47af736abd11893d5ffea1838b6d6c70f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S59sh24_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S59sh24_sel.png",
  "source_sha256": "5362586c5f9d713e6bb0116b261cfa55d07f9b0fef6bb64c0152e551340cf79e",
  "file": "S59sh24_cine.png",
  "staged_sha256": "9fef7a06a89dab6303e2967bccd5e91a48a15b6ce108a919b8c119623a55c1de",
  "latency_ms": 12123
 },
 "S59sh48::signage": {
  "fp": "95e4db2c78c0f681",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S59sh48": {
  "input_fingerprint": "a365356dae1263d7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 둥둥거리는 북소리가 울리는 감옥 밖을 향해, 이현우가 고개를 휙 돌린 채 두 눈을 크게 치켜뜬 긴장된 얼굴.\n\nLOCATION (lock): Inside the barred underground prison cell, facing the corridor and the faint daylight near the prison's small window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Cell bars at the right edge in 이현우's outward looking direction in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 쇠창살 (Still enclosing the cell) — A narrow oblique section appears at the right edge, beyond 이현우's face; used as Anchors the outward eyeline while leaving open looking space between the face and frame edge.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral cell ambience and readable eye contrast, with no exterior lighting change to illustrate the off-screen drums.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell remains locked and its bars intact; hay and pooled water remain on the floor. 이현우: He remains inside the cell, holding and eating a partially consumed potato. His previously treated leg remains injured.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 둥둥거리는 북소리가 울리는 감옥 밖을 향해, 이현우가 고개를 휙 돌린 채 두 눈을 크게 치켜뜬 긴장된 얼굴.\n\nLOCATION (lock): Inside the barred underground prison cell, facing the corridor and the faint daylight near the prison's small window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Cell bars at the right edge in 이현우's outward looking direction in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 쇠창살 (Still enclosing the cell) — A narrow oblique section appears at the right edge, beyond 이현우's face; used as Anchors the outward eyeline while leaving open looking space between the face and frame edge.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral cell ambience and readable eye contrast, with no exterior lighting change to illustrate the off-screen drums.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell remains locked and its bars intact; hay and pooled water remain on the floor. 이현우: He remains inside the cell, holding and eating a partially consumed potato. His previously treated leg remains injured.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 둥둥거리는 북소리가 울리는 감옥 밖을 향해, 이현우가 고개를 휙 돌린 채 두 눈을 크게 치켜뜬 긴장된 얼굴.\n\nLOCATION (lock): Inside the barred underground prison cell, facing the corridor and the faint daylight near the prison's small window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Cell bars at the right edge in 이현우's outward looking direction in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 쇠창살 (Still enclosing the cell) — A narrow oblique section appears at the right edge, beyond 이현우's face; used as Anchors the outward eyeline while leaving open looking space between the face and frame edge.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral cell ambience and readable eye contrast, with no exterior lighting change to illustrate the off-screen drums.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell remains locked and its bars intact; hay and pooled water remain on the floor. 이현우: He remains inside the cell, holding and eating a partially consumed potato. His previously treated leg remains injured.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 프롬프트의 지시대로 화면 우측을 향해 시선을 강하게 던지고 있으며, 우측 가장자리에 배치된 쇠창살 너머로 시선이 이어집니다.",
    "built_space": "이전 샷에서 확립된 지하 감옥의 환경(작은 창문, 뒤편과 우측의 쇠창살, 바닥에 깔린 건초와 고인 물)이 일관되게 구현되어 있으며 인물이 공간 내에 자연스럽게 위치해 있습니다.",
    "entities": "캐릭터 레퍼런스의 이현우(18세 한국계 남성, 헝클어진 검은 머리, 낡은 후드티)와 완벽히 일치하며, 먹다 남은 감자를 들고 있는 모습과 긴장하여 눈을 치켜뜬 표정이 정확히 묘사되었습니다.",
    "hard_violations": [],
    "physics": "오른손으로 감자를 쥐고 있는 자세와 손가락의 형태가 자연스러우며, 바닥에 앉아 몸을 지탱하는 물리적 제약에 어긋남이 없습니다."
   },
   {
    "label": "B",
    "direction": "이현우의 시선이 화면 좌측을 향하고 있어, 우측 가장자리의 창살 너머를 보아야 한다는 레이아웃 및 시선 방향 지시를 완전히 위반했습니다.",
    "built_space": "배경의 쇠창살, 창문, 바닥의 건초와 물 등 감옥 환경은 적절히 묘사되어 있습니다.",
    "entities": "이현우의 외모와 의상은 레퍼런스와 일치하나, 고개를 휙 돌려 긴장한 표정이라기보다는 평온하게 감자를 베어 무는 모습에 가깝습니다.",
    "hard_violations": [
     "[gemini-pro] 중복된 신체 부위: 인물이 오른손으로 감자를 들고 있음에도, 프레임 우측 하단의 무릎 위에 또 다른 오른손(또는 존재하지 않아야 할 타인의 손)이 놓여 있습니다."
    ],
    "physics": "감자를 쥐고 입에 가져가는 주된 팔의 자세는 무리가 없으나, 추가로 생성된 손이 무릎에 얹혀 있어 물리적, 해부학적으로 불가능한 상태입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시된 프레임 레이아웃에 맞춰 우측 가장자리의 창살을 향해 고개를 돌린 긴장된 표정을 완벽히 연출했으며, 캐릭터의 외모와 감옥 환경도 지시사항에 맞게 훌륭하게 재현되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트가 요구한 시선 방향(우측)과 반대인 좌측을 향하고 있으며, 화면 우측 하단에 여분의 손이 생성되는 치명적인 신체 구조 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 프롬프트의 지시대로 화면 우측을 향해 시선을 강하게 던지고 있으며, 우측 가장자리에 배치된 쇠창살 너머로 시선이 이어집니다.",
        "built_space": "이전 샷에서 확립된 지하 감옥의 환경(작은 창문, 뒤편과 우측의 쇠창살, 바닥에 깔린 건초와 고인 물)이 일관되게 구현되어 있으며 인물이 공간 내에 자연스럽게 위치해 있습니다.",
        "entities": "캐릭터 레퍼런스의 이현우(18세 한국계 남성, 헝클어진 검은 머리, 낡은 후드티)와 완벽히 일치하며, 먹다 남은 감자를 들고 있는 모습과 긴장하여 눈을 치켜뜬 표정이 정확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "오른손으로 감자를 쥐고 있는 자세와 손가락의 형태가 자연스러우며, 바닥에 앉아 몸을 지탱하는 물리적 제약에 어긋남이 없습니다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 화면 좌측을 향하고 있어, 우측 가장자리의 창살 너머를 보아야 한다는 레이아웃 및 시선 방향 지시를 완전히 위반했습니다.",
        "built_space": "배경의 쇠창살, 창문, 바닥의 건초와 물 등 감옥 환경은 적절히 묘사되어 있습니다.",
        "entities": "이현우의 외모와 의상은 레퍼런스와 일치하나, 고개를 휙 돌려 긴장한 표정이라기보다는 평온하게 감자를 베어 무는 모습에 가깝습니다.",
        "hard_violations": [
         "중복된 신체 부위: 인물이 오른손으로 감자를 들고 있음에도, 프레임 우측 하단의 무릎 위에 또 다른 오른손(또는 존재하지 않아야 할 타인의 손)이 놓여 있습니다."
        ],
        "physics": "감자를 쥐고 입에 가져가는 주된 팔의 자세는 무리가 없으나, 추가로 생성된 손이 무릎에 얹혀 있어 물리적, 해부학적으로 불가능한 상태입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시된 프레임 레이아웃에 맞춰 우측 가장자리의 창살을 향해 고개를 돌린 긴장된 표정을 완벽히 연출했으며, 캐릭터의 외모와 감옥 환경도 지시사항에 맞게 훌륭하게 재현되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트가 요구한 시선 방향(우측)과 반대인 좌측을 향하고 있으며, 화면 우측 하단에 여분의 손이 생성되는 치명적인 신체 구조 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 프롬프트의 지시대로 화면 우측을 향해 시선을 강하게 던지고 있으며, 우측 가장자리에 배치된 쇠창살 너머로 시선이 이어집니다.",
        "built_space": "이전 샷에서 확립된 지하 감옥의 환경(작은 창문, 뒤편과 우측의 쇠창살, 바닥에 깔린 건초와 고인 물)이 일관되게 구현되어 있으며 인물이 공간 내에 자연스럽게 위치해 있습니다.",
        "entities": "캐릭터 레퍼런스의 이현우(18세 한국계 남성, 헝클어진 검은 머리, 낡은 후드티)와 완벽히 일치하며, 먹다 남은 감자를 들고 있는 모습과 긴장하여 눈을 치켜뜬 표정이 정확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "오른손으로 감자를 쥐고 있는 자세와 손가락의 형태가 자연스러우며, 바닥에 앉아 몸을 지탱하는 물리적 제약에 어긋남이 없습니다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 화면 좌측을 향하고 있어, 우측 가장자리의 창살 너머를 보아야 한다는 레이아웃 및 시선 방향 지시를 완전히 위반했습니다.",
        "built_space": "배경의 쇠창살, 창문, 바닥의 건초와 물 등 감옥 환경은 적절히 묘사되어 있습니다.",
        "entities": "이현우의 외모와 의상은 레퍼런스와 일치하나, 고개를 휙 돌려 긴장한 표정이라기보다는 평온하게 감자를 베어 무는 모습에 가깝습니다.",
        "hard_violations": [
         "중복된 신체 부위: 인물이 오른손으로 감자를 들고 있음에도, 프레임 우측 하단의 무릎 위에 또 다른 오른손(또는 존재하지 않아야 할 타인의 손)이 놓여 있습니다."
        ],
        "physics": "감자를 쥐고 입에 가져가는 주된 팔의 자세는 무리가 없으나, 추가로 생성된 손이 무릎에 얹혀 있어 물리적, 해부학적으로 불가능한 상태입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "감자를 먹는 상태와 감옥 재질은 잘 유지했지만, 시선이 지정된 오른쪽 밖이 아닌 왼쪽으로 향하고 무릎까지 보이는 넓은 구도로 얼굴 클로즈업을 놓쳤다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "오른쪽 밖을 향해 눈을 치켜뜬 긴장된 반응은 더 정확하지만, 상반신까지 넓게 담았고 굵은 전경 창살이 얼굴 옆을 가려 요구된 여백과 배경 창살 배치를 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개와 두 눈이 화면 왼쪽의 문틀 쪽을 향한다. 요구된 오른쪽 가장자리 창살 너머가 시선의 목표가 아니다. 감자는 손에 잡혀 입으로 향하며 먹는 동작과 맞는다. 눈은 옆을 경계하지만 크게 치켜뜬 반응은 비교적 약하다.",
        "built_space": "뒤쪽 창살 면 하나와 오른쪽에서 비스듬히 이어지는 창살 면 하나, 왼쪽 문틀의 잠금장치 하나, 뒤쪽 위의 작은 창 하나가 보인다. 녹슨 금속, 벗겨진 벽, 바닥의 짚과 고인 물은 장소 참고와 부합한다. 인물은 창살 안쪽 바닥에 앉아 있으나, 창살이 오른쪽의 좁은 배경 조각에 그치지 않고 화면 뒤와 오른쪽을 넓게 차지한다. 얼굴뿐 아니라 몸통과 세운 무릎까지 담겨 클로즈업보다 넓다.",
        "entities": "검고 헝클어진 머리의 한국계 청년으로 보이는 남성 한 명만 있다. 얼굴과 머리, 해진 짙은 회색 후드와 찢어진 청바지는 인물 참고에 대체로 맞으며, 정확히 18세인지는 외관만으로 확정할 수 없다. 손에는 속살이 드러난 일부 먹은 감자 하나가 있다. 드러난 무릎에서 치료 흔적은 뚜렷하지 않지만 이것만으로 부상 지속 여부를 판정할 수 없다. 이전 장면의 여성이나 추가 인물, 읽을 수 있는 글자는 없다. 작은 창의 낮빛과 중성적인 실내 조명도 유지된다.",
        "hard_violations": [],
        "physics": "인물의 몸은 바닥에 앉아 무릎을 세운 자세로 읽힌다. 엉덩이의 직접 접촉점은 하단 밖이지만 공중에 뜬 몸으로 보이지 않는다. 감자는 입 앞의 손가락과 손바닥이 받치고 있으며, 다른 손은 세운 다리 부근에 놓여 있다. 목을 왼쪽으로 돌린 자세와 감자를 입에 대는 동작은 해부학적으로 가능하다."
       },
       {
        "label": "B",
        "direction": "두 눈은 화면 오른쪽 위, 얼굴 옆 창살 너머의 화면 밖을 향한다. 몸통에서 고개를 돌려 밖의 소리에 반응하는 방향이 요구와 맞으며, 벌어진 눈과 입에 긴장이 보인다. 감자는 입에서 내려와 있어 먹다가 멈춘 순간으로 읽힌다.",
        "built_space": "뒤쪽 창살 면 하나와 오른쪽 가까운 창살 면 하나, 뒤쪽 위의 작은 창 하나가 보이며 잠금장치는 프레임에서 확인되지 않는다. 낡은 벽, 녹슨 창살, 짚과 물이 남은 바닥은 참고 장소와 대체로 일치한다. 인물은 가까운 창살 뒤에 있지만, 굵은 세로 부재가 얼굴 오른쪽 윤곽 일부를 가리는 전경으로 들어온다. 따라서 얼굴 너머 오른쪽 가장자리에 좁은 배경 창살을 두고 시선 여백을 남기라는 배치와 다르다. 얼굴과 상반신, 감자를 든 손 및 바닥이 함께 보여 요구된 얼굴 클로즈업보다 넓다.",
        "entities": "검은 헝클어진 머리와 참고 인물에 가까운 얼굴을 가진 한국계 청년으로 보이는 남성 한 명이다. 해진 짙은 회색 후드가 참고와 맞는다. 손에 든 감자 하나는 윗부분을 먹어 속살이 드러나 있다. 다리와 치료 부위는 프레임 밖이므로 평가하지 않는다. 추가 인물이나 읽을 수 있는 글자는 없고, 작은 창에서 들어오는 낮빛과 중성적인 감옥 분위기가 유지된다.",
        "hard_violations": [],
        "physics": "상체는 프레임 아래로 이어지며 앉거나 낮게 몸을 둔 자세로 읽힌다. 하체와 바닥 접촉점은 잘렸지만 몸이 떠 있다는 징후는 없다. 감자는 손가락과 손바닥으로 확실히 지지된다. 어깨에 대해 목을 돌리고 눈을 오른쪽 위로 움직인 자세는 자연스럽게 가능한 반응이며, 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "감자를 먹는 상태와 감옥 재질은 잘 유지했지만, 시선이 지정된 오른쪽 밖이 아닌 왼쪽으로 향하고 무릎까지 보이는 넓은 구도로 얼굴 클로즈업을 놓쳤다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "오른쪽 밖을 향해 눈을 치켜뜬 긴장된 반응은 더 정확하지만, 상반신까지 넓게 담았고 굵은 전경 창살이 얼굴 옆을 가려 요구된 여백과 배경 창살 배치를 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "고개와 두 눈이 화면 왼쪽의 문틀 쪽을 향한다. 요구된 오른쪽 가장자리 창살 너머가 시선의 목표가 아니다. 감자는 손에 잡혀 입으로 향하며 먹는 동작과 맞는다. 눈은 옆을 경계하지만 크게 치켜뜬 반응은 비교적 약하다.",
        "built_space": "뒤쪽 창살 면 하나와 오른쪽에서 비스듬히 이어지는 창살 면 하나, 왼쪽 문틀의 잠금장치 하나, 뒤쪽 위의 작은 창 하나가 보인다. 녹슨 금속, 벗겨진 벽, 바닥의 짚과 고인 물은 장소 참고와 부합한다. 인물은 창살 안쪽 바닥에 앉아 있으나, 창살이 오른쪽의 좁은 배경 조각에 그치지 않고 화면 뒤와 오른쪽을 넓게 차지한다. 얼굴뿐 아니라 몸통과 세운 무릎까지 담겨 클로즈업보다 넓다.",
        "entities": "검고 헝클어진 머리의 한국계 청년으로 보이는 남성 한 명만 있다. 얼굴과 머리, 해진 짙은 회색 후드와 찢어진 청바지는 인물 참고에 대체로 맞으며, 정확히 18세인지는 외관만으로 확정할 수 없다. 손에는 속살이 드러난 일부 먹은 감자 하나가 있다. 드러난 무릎에서 치료 흔적은 뚜렷하지 않지만 이것만으로 부상 지속 여부를 판정할 수 없다. 이전 장면의 여성이나 추가 인물, 읽을 수 있는 글자는 없다. 작은 창의 낮빛과 중성적인 실내 조명도 유지된다.",
        "hard_violations": [],
        "physics": "인물의 몸은 바닥에 앉아 무릎을 세운 자세로 읽힌다. 엉덩이의 직접 접촉점은 하단 밖이지만 공중에 뜬 몸으로 보이지 않는다. 감자는 입 앞의 손가락과 손바닥이 받치고 있으며, 다른 손은 세운 다리 부근에 놓여 있다. 목을 왼쪽으로 돌린 자세와 감자를 입에 대는 동작은 해부학적으로 가능하다."
       },
       {
        "label": "A",
        "direction": "두 눈은 화면 오른쪽 위, 얼굴 옆 창살 너머의 화면 밖을 향한다. 몸통에서 고개를 돌려 밖의 소리에 반응하는 방향이 요구와 맞으며, 벌어진 눈과 입에 긴장이 보인다. 감자는 입에서 내려와 있어 먹다가 멈춘 순간으로 읽힌다.",
        "built_space": "뒤쪽 창살 면 하나와 오른쪽 가까운 창살 면 하나, 뒤쪽 위의 작은 창 하나가 보이며 잠금장치는 프레임에서 확인되지 않는다. 낡은 벽, 녹슨 창살, 짚과 물이 남은 바닥은 참고 장소와 대체로 일치한다. 인물은 가까운 창살 뒤에 있지만, 굵은 세로 부재가 얼굴 오른쪽 윤곽 일부를 가리는 전경으로 들어온다. 따라서 얼굴 너머 오른쪽 가장자리에 좁은 배경 창살을 두고 시선 여백을 남기라는 배치와 다르다. 얼굴과 상반신, 감자를 든 손 및 바닥이 함께 보여 요구된 얼굴 클로즈업보다 넓다.",
        "entities": "검은 헝클어진 머리와 참고 인물에 가까운 얼굴을 가진 한국계 청년으로 보이는 남성 한 명이다. 해진 짙은 회색 후드가 참고와 맞는다. 손에 든 감자 하나는 윗부분을 먹어 속살이 드러나 있다. 다리와 치료 부위는 프레임 밖이므로 평가하지 않는다. 추가 인물이나 읽을 수 있는 글자는 없고, 작은 창에서 들어오는 낮빛과 중성적인 감옥 분위기가 유지된다.",
        "hard_violations": [],
        "physics": "상체는 프레임 아래로 이어지며 앉거나 낮게 몸을 둔 자세로 읽힌다. 하체와 바닥 접촉점은 잘렸지만 몸이 떠 있다는 징후는 없다. 감자는 손가락과 손바닥으로 확실히 지지된다. 어깨에 대해 목을 돌리고 눈을 오른쪽 위로 움직인 자세는 자연스럽게 가능한 반응이며, 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.889
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.639
   },
   "violations": {
    "B": [
     "[gemini-pro] 중복된 신체 부위: 인물이 오른손으로 감자를 들고 있음에도, 프레임 우측 하단의 무릎 위에 또 다른 오른손(또는 존재하지 않아야 할 타인의 손)이 놓여 있습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 639
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 프레임 레이아웃에 맞춰 우측 가장자리의 창살을 향해 고개를 돌린 긴장된 표정을 완벽히 연출했으며, 캐릭터의 외모와 감옥 환경도 지시사항에 맞게 훌륭하게 재현되었습니다."
   },
   {
    "label": "B",
    "score": 639,
    "verdict_ko": "프롬프트가 요구한 시선 방향(우측)과 반대인 좌측을 향하고 있으며, 화면 우측 하단에 여분의 손이 생성되는 치명적인 신체 구조 오류가 발생했습니다.  ★위반: [gemini-pro] 중복된 신체 부위: 인물이 오른손으로 감자를 들고 있음에도, 프레임 우측 하단의 무릎 위에 또 다른 오른손(또는 존재하지 않아야 할 타인의 손)이 놓여 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S59sh8_sel.png",
    "asset_id": "925ff5b0-d1c3-4826-84b1-2842a55a3380",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdef-74cf-71a1-b0d8-d3b5b087075e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S59sh8"
  }
 },
 "S59sh48::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:49:27.311552+00:00",
  "fingerprint": "b97d16a97493f5e63196616a5677d823510642750ca381e3e08b803854566ff8",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S59sh48_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S59sh48_sel.png",
  "source_sha256": "2346c26cb6dea5dd80c256114b7bc030aed97707cc639680c719555626934596",
  "file": "S59sh48_cine.png",
  "staged_sha256": "4dac3e4442d6e46255723c8e38fadd75fbe8e143c0ac24011808335732562ccb",
  "latency_ms": 17249
 },
 "S60sh75::signage": {
  "fp": "9e5a96e88fcba654",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S60sh75": {
  "input_fingerprint": "cc655cc2e021bdd0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): sunset, red sky.\n\nSHOT TEXT (authoritative, Korean): 폭발 연기가 자욱한 격투장 한가운데, 찰리와 B-200이 등을 맞댄 채 주변을 경계하며 바라보는 고립된 구도.\n\nLOCATION (lock): In the open center of the village's outdoor fighting arena before the church, surrounded by explosion smoke and spectator stands. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 격투장 바닥 (The two robots are isolated near the middle during the surrounding battle) — Seen diagonally from above, extending visibly around both figures; used as Negative space establishes isolation and preserves the difference in the robots' scale; 폭발 연기 (Lingering thickly within the arena); used as Softens the distant arena boundaries while leaving both robots and their back-to-back contact legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk light and the arena's established torchlight are softened by explosion smoke, preserving distinct robot contours within the surrounding confusion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Explosion smoke and mortar damage mark the torch-lined arena; Charlie is out of his cage and battered from the fight, with his earlier disguise not explicitly removed. B-200 retains his modified gun arms, and the water tower is still standing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): sunset, red sky.\n\nSHOT TEXT (authoritative, Korean): 폭발 연기가 자욱한 격투장 한가운데, 찰리와 B-200이 등을 맞댄 채 주변을 경계하며 바라보는 고립된 구도.\n\nLOCATION (lock): In the open center of the village's outdoor fighting arena before the church, surrounded by explosion smoke and spectator stands. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 격투장 바닥 (The two robots are isolated near the middle during the surrounding battle) — Seen diagonally from above, extending visibly around both figures; used as Negative space establishes isolation and preserves the difference in the robots' scale; 폭발 연기 (Lingering thickly within the arena); used as Softens the distant arena boundaries while leaving both robots and their back-to-back contact legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk light and the arena's established torchlight are softened by explosion smoke, preserving distinct robot contours within the surrounding confusion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Explosion smoke and mortar damage mark the torch-lined arena; Charlie is out of his cage and battered from the fight, with his earlier disguise not explicitly removed. B-200 retains his modified gun arms, and the water tower is still standing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): sunset, red sky.\n\nSHOT TEXT (authoritative, Korean): 폭발 연기가 자욱한 격투장 한가운데, 찰리와 B-200이 등을 맞댄 채 주변을 경계하며 바라보는 고립된 구도.\n\nLOCATION (lock): In the open center of the village's outdoor fighting arena before the church, surrounded by explosion smoke and spectator stands. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 격투장 바닥 (The two robots are isolated near the middle during the surrounding battle) — Seen diagonally from above, extending visibly around both figures; used as Negative space establishes isolation and preserves the difference in the robots' scale; 폭발 연기 (Lingering thickly within the arena); used as Softens the distant arena boundaries while leaving both robots and their back-to-back contact legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk light and the arena's established torchlight are softened by explosion smoke, preserving distinct robot contours within the surrounding confusion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Explosion smoke and mortar damage mark the torch-lined arena; Charlie is out of his cage and battered from the fight, with his earlier disguise not explicitly removed. B-200 retains his modified gun arms, and the water tower is still standing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리는 왼쪽을 바라보고 있으며, B-200은 오른쪽을 향해 총이 장착된 팔을 겨누고 있음.",
    "built_space": "흙바닥 격투장, 횃불, 관중석이 보이고 우측 스카이라인에 급수탑이 서 있으나, 레퍼런스에 있는 좌측 전경의 철창 구조물이 생략되었고 지시된 부감(대각선 위) 구도가 아님.",
    "entities": "찰리는 흰 얼굴판을 가졌으나 머리 외장이 각지지 않고 둥글며, B-200은 총기 팔을 가졌으나 묘사된 것보다 인간형에 가까운 머리 형태를 띰.",
    "hard_violations": [],
    "physics": "두 캐릭터 모두 격투장 흙바닥에 두 발을 딛고 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "찰리는 왼쪽을 경계하고, B-200은 오른쪽을 향해 총기 팔을 뻗어 겨냥하고 있음.",
    "built_space": "레퍼런스와 동일한 좌측 철창, 정면의 계단과 관중석 등 장소의 물리적 구조를 완벽하게 재현했으나, 지시된 급수탑이 없고 카메라가 부감 구도 대신 레퍼런스와 유사한 눈높이에 머묾.",
    "entities": "찰리는 지시된 대로 각진 머리 외장과 흰 얼굴판을 지닌 고릴라형 기계로 묘사되었고, B-200은 짙은 회색 철제 외장, 기계식 머리, 총기 팔을 정확히 갖춤.",
    "hard_violations": [],
    "physics": "두 개체 모두 격투장 바닥에 발을 딛고 있으며 자세를 지탱하는 데 무리가 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트에 명시된 캐릭터의 세부 외형(찰리의 각진 머리, B-200의 기계식 머리)과 장소의 디테일을 잘 살렸으나, 대각선 위에서 내려다보는 구도와 급수탑이 누락된 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "배경에 급수탑을 포함한 점은 좋으나 찰리의 머리가 각지지 않고 둥글며, B-200의 디자인이 명세와 다르고 구도 지시(부감)를 따르지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 왼쪽을 바라보고 있으며, B-200은 오른쪽을 향해 총이 장착된 팔을 겨누고 있음.",
        "built_space": "흙바닥 격투장, 횃불, 관중석이 보이고 우측 스카이라인에 급수탑이 서 있으나, 레퍼런스에 있는 좌측 전경의 철창 구조물이 생략되었고 지시된 부감(대각선 위) 구도가 아님.",
        "entities": "찰리는 흰 얼굴판을 가졌으나 머리 외장이 각지지 않고 둥글며, B-200은 총기 팔을 가졌으나 묘사된 것보다 인간형에 가까운 머리 형태를 띰.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 격투장 흙바닥에 두 발을 딛고 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "찰리는 왼쪽을 경계하고, B-200은 오른쪽을 향해 총기 팔을 뻗어 겨냥하고 있음.",
        "built_space": "레퍼런스와 동일한 좌측 철창, 정면의 계단과 관중석 등 장소의 물리적 구조를 완벽하게 재현했으나, 지시된 급수탑이 없고 카메라가 부감 구도 대신 레퍼런스와 유사한 눈높이에 머묾.",
        "entities": "찰리는 지시된 대로 각진 머리 외장과 흰 얼굴판을 지닌 고릴라형 기계로 묘사되었고, B-200은 짙은 회색 철제 외장, 기계식 머리, 총기 팔을 정확히 갖춤.",
        "hard_violations": [],
        "physics": "두 개체 모두 격투장 바닥에 발을 딛고 있으며 자세를 지탱하는 데 무리가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트에 명시된 캐릭터의 세부 외형(찰리의 각진 머리, B-200의 기계식 머리)과 장소의 디테일을 잘 살렸으나, 대각선 위에서 내려다보는 구도와 급수탑이 누락된 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "배경에 급수탑을 포함한 점은 좋으나 찰리의 머리가 각지지 않고 둥글며, B-200의 디자인이 명세와 다르고 구도 지시(부감)를 따르지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 왼쪽을 바라보고 있으며, B-200은 오른쪽을 향해 총이 장착된 팔을 겨누고 있음.",
        "built_space": "흙바닥 격투장, 횃불, 관중석이 보이고 우측 스카이라인에 급수탑이 서 있으나, 레퍼런스에 있는 좌측 전경의 철창 구조물이 생략되었고 지시된 부감(대각선 위) 구도가 아님.",
        "entities": "찰리는 흰 얼굴판을 가졌으나 머리 외장이 각지지 않고 둥글며, B-200은 총기 팔을 가졌으나 묘사된 것보다 인간형에 가까운 머리 형태를 띰.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 격투장 흙바닥에 두 발을 딛고 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "찰리는 왼쪽을 경계하고, B-200은 오른쪽을 향해 총기 팔을 뻗어 겨냥하고 있음.",
        "built_space": "레퍼런스와 동일한 좌측 철창, 정면의 계단과 관중석 등 장소의 물리적 구조를 완벽하게 재현했으나, 지시된 급수탑이 없고 카메라가 부감 구도 대신 레퍼런스와 유사한 눈높이에 머묾.",
        "entities": "찰리는 지시된 대로 각진 머리 외장과 흰 얼굴판을 지닌 고릴라형 기계로 묘사되었고, B-200은 짙은 회색 철제 외장, 기계식 머리, 총기 팔을 정확히 갖춤.",
        "hard_violations": [],
        "physics": "두 개체 모두 격투장 바닥에 발을 딛고 있으며 자세를 지탱하는 데 무리가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "등을 맞댄 경계와 장소 구조는 충실하지만, 내려다보는 각도가 약하고 두 로봇이 더 크게 보여 바닥의 여백과 고립감이 B보다 부족합니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "비스듬히 내려다보는 와이드 구도, 사방의 바닥 여백, 등을 맞댄 반대 방향 경계와 짙은 연기가 핵심 연출을 가장 충실하게 구현합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 화면 왼쪽 앞의 빈 경기장 쪽을 보고, B-200의 머리와 총구는 화면 오른쪽 경기장 바깥 방향을 향합니다. 서로 반대쪽을 경계하며, 조준 대상인 적은 보이지 않습니다. 특정 적을 맞추라는 요구는 없습니다.",
        "built_space": "좌우의 계단식 관람석, 후면 중앙 계단 한 줄과 상부 출입구, 천막과 기둥이 보입니다. 불빛은 좌우 상단 두 곳, 후면 중앙 양옆 두 곳, 경기장 측면 두 곳으로 총 여섯 곳이 식별됩니다. 왼쪽 전경에는 철창과 낮은 벽으로 둘러싸인 함몰 공간 일부가 있어 참조 장소와 연결됩니다. 두 로봇은 시설물과 떨어진 중앙 흙바닥에 서 있습니다. 다만 카메라의 하향 각도가 완만하여 요구된 대각선 상부 시점이 약합니다.",
        "entities": "로봇은 정확히 두 대이고 다른 사람은 없습니다. 찰리는 고릴라형 머리와 흰 얼굴판, 작은 눈을 갖췄지만 얼굴의 코와 입 주변 조형이 복잡해 단순한 점 눈·선 입 인상은 약합니다. B-200은 각진 기계 머리와 짙은 회색 외장, 총열이 달린 팔을 갖췄습니다. 눈의 크기 변화는 정지 화면에서 확인할 수 없습니다. 두 외장에는 마모와 오염이 있으며, 붉은 석양과 폭발 연기도 보입니다. 급수탑과 이전 위장의 구체적인 상태는 식별되지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 로봇 모두 발을 흙바닥에 대고 무릎을 굽혀 서 있으며, 몸통 뒤쪽이 서로 맞닿아 있습니다. B-200의 수평 총팔은 어깨와 팔꿈치 관절에 연결되어 지지됩니다. 찰리의 내려진 팔도 몸통에 정상적으로 연결됩니다. 공중에 떠 있거나 지지체가 없는 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "찰리는 화면 왼쪽 앞을, B-200은 오른쪽을 바라봅니다. B-200의 뻗은 총팔과 총구도 오른쪽의 연기 낀 경기장 방향을 향해 시선과 일치합니다. 두 로봇은 등을 맞대고 반대 구역을 경계하며, 별도의 적이나 표적은 보이지 않습니다.",
        "built_space": "좌우 계단식 관람석과 낮은 경기장 경계벽, 후면 중앙 계단 한 줄과 상부 출입구, 천막 구조가 보입니다. 횃불은 좌우 상단 두 곳, 후면 중앙 양옆 두 곳, 경기장 측면 두 곳으로 총 여섯 곳이 식별됩니다. 후면 오른쪽에는 다리로 지지된 급수탑 한 기가 서 있습니다. 두 로봇은 중앙의 열린 흙바닥에 있고, 카메라는 비스듬한 상부에서 전후좌우의 바닥 여백을 보여줍니다. 연기가 먼 관람석 경계를 흐리지만 두 로봇의 접촉은 남겨 둡니다. 참조의 전경 철창 공간은 이 프레임에서 보이지 않습니다.",
        "entities": "찰리와 B-200 두 로봇만 있습니다. 찰리는 더 육중한 고릴라형 체격, 흰 얼굴판과 작은 점 형태의 눈, 짧은 입선을 갖췄으나 머리 외장은 요구보다 둥글게 보입니다. B-200은 더 가는 체격의 회색 금속 로봇이며 기계 머리와 총팔이 식별됩니다. 눈의 크기 변화는 확인할 수 없습니다. 외장의 흠집과 오염, 짙은 연기, 붉은 석양, 서 있는 급수탑이 보입니다. 바닥은 패이고 거칠지만 명확한 박격포 탄착구는 불분명하며, 이전 위장은 판단할 정보가 없습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "찰리는 양발로 바닥을 딛고 있고, B-200도 벌린 다리의 발로 지면을 지지합니다. 두 몸통의 등 쪽이 맞닿아 있으며 무게중심은 지지 범위 안에 있습니다. 총팔은 어깨에서 팔꿈치와 전완으로 이어져 지지됩니다. 급수탑은 가시적인 기둥들 위에 놓여 있습니다. 지지 없이 떠 있는 몸이나 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "등을 맞댄 경계와 장소 구조는 충실하지만, 내려다보는 각도가 약하고 두 로봇이 더 크게 보여 바닥의 여백과 고립감이 B보다 부족합니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "비스듬히 내려다보는 와이드 구도, 사방의 바닥 여백, 등을 맞댄 반대 방향 경계와 짙은 연기가 핵심 연출을 가장 충실하게 구현합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 화면 왼쪽 앞의 빈 경기장 쪽을 보고, B-200의 머리와 총구는 화면 오른쪽 경기장 바깥 방향을 향합니다. 서로 반대쪽을 경계하며, 조준 대상인 적은 보이지 않습니다. 특정 적을 맞추라는 요구는 없습니다.",
        "built_space": "좌우의 계단식 관람석, 후면 중앙 계단 한 줄과 상부 출입구, 천막과 기둥이 보입니다. 불빛은 좌우 상단 두 곳, 후면 중앙 양옆 두 곳, 경기장 측면 두 곳으로 총 여섯 곳이 식별됩니다. 왼쪽 전경에는 철창과 낮은 벽으로 둘러싸인 함몰 공간 일부가 있어 참조 장소와 연결됩니다. 두 로봇은 시설물과 떨어진 중앙 흙바닥에 서 있습니다. 다만 카메라의 하향 각도가 완만하여 요구된 대각선 상부 시점이 약합니다.",
        "entities": "로봇은 정확히 두 대이고 다른 사람은 없습니다. 찰리는 고릴라형 머리와 흰 얼굴판, 작은 눈을 갖췄지만 얼굴의 코와 입 주변 조형이 복잡해 단순한 점 눈·선 입 인상은 약합니다. B-200은 각진 기계 머리와 짙은 회색 외장, 총열이 달린 팔을 갖췄습니다. 눈의 크기 변화는 정지 화면에서 확인할 수 없습니다. 두 외장에는 마모와 오염이 있으며, 붉은 석양과 폭발 연기도 보입니다. 급수탑과 이전 위장의 구체적인 상태는 식별되지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 로봇 모두 발을 흙바닥에 대고 무릎을 굽혀 서 있으며, 몸통 뒤쪽이 서로 맞닿아 있습니다. B-200의 수평 총팔은 어깨와 팔꿈치 관절에 연결되어 지지됩니다. 찰리의 내려진 팔도 몸통에 정상적으로 연결됩니다. 공중에 떠 있거나 지지체가 없는 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "찰리는 화면 왼쪽 앞을, B-200은 오른쪽을 바라봅니다. B-200의 뻗은 총팔과 총구도 오른쪽의 연기 낀 경기장 방향을 향해 시선과 일치합니다. 두 로봇은 등을 맞대고 반대 구역을 경계하며, 별도의 적이나 표적은 보이지 않습니다.",
        "built_space": "좌우 계단식 관람석과 낮은 경기장 경계벽, 후면 중앙 계단 한 줄과 상부 출입구, 천막 구조가 보입니다. 횃불은 좌우 상단 두 곳, 후면 중앙 양옆 두 곳, 경기장 측면 두 곳으로 총 여섯 곳이 식별됩니다. 후면 오른쪽에는 다리로 지지된 급수탑 한 기가 서 있습니다. 두 로봇은 중앙의 열린 흙바닥에 있고, 카메라는 비스듬한 상부에서 전후좌우의 바닥 여백을 보여줍니다. 연기가 먼 관람석 경계를 흐리지만 두 로봇의 접촉은 남겨 둡니다. 참조의 전경 철창 공간은 이 프레임에서 보이지 않습니다.",
        "entities": "찰리와 B-200 두 로봇만 있습니다. 찰리는 더 육중한 고릴라형 체격, 흰 얼굴판과 작은 점 형태의 눈, 짧은 입선을 갖췄으나 머리 외장은 요구보다 둥글게 보입니다. B-200은 더 가는 체격의 회색 금속 로봇이며 기계 머리와 총팔이 식별됩니다. 눈의 크기 변화는 확인할 수 없습니다. 외장의 흠집과 오염, 짙은 연기, 붉은 석양, 서 있는 급수탑이 보입니다. 바닥은 패이고 거칠지만 명확한 박격포 탄착구는 불분명하며, 이전 위장은 판단할 정보가 없습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "찰리는 양발로 바닥을 딛고 있고, B-200도 벌린 다리의 발로 지면을 지지합니다. 두 몸통의 등 쪽이 맞닿아 있으며 무게중심은 지지 범위 안에 있습니다. 총팔은 어깨에서 팔꿈치와 전완으로 이어져 지지됩니다. 급수탑은 가시적인 기둥들 위에 놓여 있습니다. 지지 없이 떠 있는 몸이나 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.778
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.778
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1778,
   "A": 1714
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1778,
    "verdict_ko": "프롬프트에 명시된 캐릭터의 세부 외형(찰리의 각진 머리, B-200의 기계식 머리)과 장소의 디테일을 잘 살렸으나, 대각선 위에서 내려다보는 구도와 급수탑이 누락된 점이 아쉽습니다."
   },
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "배경에 급수탑을 포함한 점은 좋으나 찰리의 머리가 각지지 않고 둥글며, B-200의 디자인이 명세와 다르고 구도 지시(부감)를 따르지 않았습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L87B01.png",
    "asset_id": "6a2920c7-71d8-4f64-81c8-4fabca1a804c",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdf0-8e65-7226-828a-47dc06121540",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S60sh75::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:49:44.912056+00:00",
  "fingerprint": "566ec38170439afbcf55eccbdff55c6e1e0f77b8fbbffea2b2849932f7a0b58b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S60sh75_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S60sh75_sel.png",
  "source_sha256": "07d6f620418b5c09cf8a52759e79a3ffd6c242d5d8db04bb26a04eec4054bf54",
  "file": "S60sh75_cine.png",
  "staged_sha256": "286d37e8347961fd3adfa99a4f3e8b661949d3bdeba8efe0343f8391b535fc87",
  "latency_ms": 12042
 },
 "S60sh104::signage": {
  "fp": "b4015888da255753",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::3e690a510c774585": {
  "subjects": [],
  "subject_text": "익산 마을 격투장과 관중석·단상\n성당 앞 마을 중앙에 조성된 야외 격투장. 중앙 경기 공간 주위로 관중석과 높은 단상이 있고 곳곳에 횃불이 꽂혀 있다.",
  "identity": "canonical",
  "scope_id": "L87",
  "scope_role": "location_exterior",
  "scope_sha": "b66d299178674b36"
 },
 "S60sh104::bgfirst_bg": {
  "input_fingerprint": "225075fce6829c78",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 부서진 가면 너머로 피부가 붉게 짓무르고 흉측하게 녹아내린 장백산의 끔찍한 얼굴이 드러난 클로즈업.\n\nLOCATION (lock): On the soaked ground beside the collapsed village water tower in the evening light.\n\nTIME OF DAY (lock): sunset, red sky.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 황금 하회탈 (Half broken, leaving the radiation-damaged face exposed) — The remaining outer mask face is seen obliquely beside the uncovered skin; used as Keeps the evidence of concealed authority adjacent to the face; the fragment occupies only a small part of the image; 장백산의 옷 (Soaked after the water tank collapse) — Only the shoulder and upper chest portions appear at the lower edge; used as Retains the immediate physical aftermath without widening away from the revelation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established evening sunset reveals the wet face and broken golden mask with restrained warmth and clear, unsensational detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 부서진 가면 너머로 피부가 붉게 짓무르고 흉측하게 녹아내린 장백산의 끔찍한 얼굴이 드러난 클로즈업.\n\nLOCATION (lock): On the soaked ground beside the collapsed village water tower in the evening light.\n\nTIME OF DAY (lock): sunset, red sky.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 황금 하회탈 (Half broken, leaving the radiation-damaged face exposed) — The remaining outer mask face is seen obliquely beside the uncovered skin; used as Keeps the evidence of concealed authority adjacent to the face; the fragment occupies only a small part of the image; 장백산의 옷 (Soaked after the water tank collapse) — Only the shoulder and upper chest portions appear at the lower edge; used as Retains the immediate physical aftermath without widening away from the revelation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established evening sunset reveals the wet face and broken golden mask with restrained warmth and clear, unsensational detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S60sh104__bgfirst_bg.png",
  "asset_id": "550f9eb9-cb11-4dc8-8df7-067efb40ca0c",
  "input_asset_ids": [
   "06f56e22-4eae-4819-a3d5-24a510ac0b12",
   "f6a54720-c3b4-41fd-a06f-736ac241da58"
  ]
 },
 "S60sh104": {
  "input_fingerprint": "1b666e91edea0ce6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): sunset, red sky.\n\nSHOT TEXT (authoritative, Korean): 부서진 가면 너머로 피부가 붉게 짓무르고 흉측하게 녹아내린 장백산의 끔찍한 얼굴이 드러난 클로즈업.\n\nLOCATION (lock): On the soaked ground beside the collapsed village water tower in the evening light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 황금 하회탈 (Half broken, leaving the radiation-damaged face exposed) — The remaining outer mask face is seen obliquely beside the uncovered skin; used as Keeps the evidence of concealed authority adjacent to the face; the fragment occupies only a small part of the image; 장백산의 옷 (Soaked after the water tank collapse) — Only the shoulder and upper chest portions appear at the lower edge; used as Retains the immediate physical aftermath without widening away from the revelation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established evening sunset reveals the wet face and broken golden mask with restrained warmth and clear, unsensational detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The water tower has been severed and collapsed, releasing a large volume of water across the ground. Charlie remains battered and B-200 retains his gun arms; the ruined arena and its combat damage remain nearby in sunset light. 장백산: He is soaked, with his golden mask half broken and his radiation-disfigured face exposed. His regal clothing has not yet been stripped away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): sunset, red sky.\n\nSHOT TEXT (authoritative, Korean): 부서진 가면 너머로 피부가 붉게 짓무르고 흉측하게 녹아내린 장백산의 끔찍한 얼굴이 드러난 클로즈업.\n\nLOCATION (lock): On the soaked ground beside the collapsed village water tower in the evening light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 황금 하회탈 (Half broken, leaving the radiation-damaged face exposed) — The remaining outer mask face is seen obliquely beside the uncovered skin; used as Keeps the evidence of concealed authority adjacent to the face; the fragment occupies only a small part of the image; 장백산의 옷 (Soaked after the water tank collapse) — Only the shoulder and upper chest portions appear at the lower edge; used as Retains the immediate physical aftermath without widening away from the revelation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established evening sunset reveals the wet face and broken golden mask with restrained warmth and clear, unsensational detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The water tower has been severed and collapsed, releasing a large volume of water across the ground. Charlie remains battered and B-200 retains his gun arms; the ruined arena and its combat damage remain nearby in sunset light. 장백산: He is soaked, with his golden mask half broken and his radiation-disfigured face exposed. His regal clothing has not yet been stripped away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): sunset, red sky.\n\nSHOT TEXT (authoritative, Korean): 부서진 가면 너머로 피부가 붉게 짓무르고 흉측하게 녹아내린 장백산의 끔찍한 얼굴이 드러난 클로즈업.\n\nLOCATION (lock): On the soaked ground beside the collapsed village water tower in the evening light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 황금 하회탈 (Half broken, leaving the radiation-damaged face exposed) — The remaining outer mask face is seen obliquely beside the uncovered skin; used as Keeps the evidence of concealed authority adjacent to the face; the fragment occupies only a small part of the image; 장백산의 옷 (Soaked after the water tank collapse) — Only the shoulder and upper chest portions appear at the lower edge; used as Retains the immediate physical aftermath without widening away from the revelation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established evening sunset reveals the wet face and broken golden mask with restrained warmth and clear, unsensational detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The water tower has been severed and collapsed, releasing a large volume of water across the ground. Charlie remains battered and B-200 retains his gun arms; the ruined arena and its combat damage remain nearby in sunset light. 장백산: He is soaked, with his golden mask half broken and his radiation-disfigured face exposed. His regal clothing has not yet been stripped away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S60sh104__bgfirst_bg.png",
     "asset_id": "550f9eb9-cb11-4dc8-8df7-067efb40ca0c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S60sh104.png",
     "asset_id": "06f56e22-4eae-4819-a3d5-24a510ac0b12",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1531415>",
     "asset_id": "184110cd-a0d4-49f1-a438-62fe0b127791",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L87B03.png",
     "asset_id": "f6a54720-c3b4-41fd-a06f-736ac241da58",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1531415>",
     "asset_id": "184110cd-a0d4-49f1-a438-62fe0b127791",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 왼쪽 밖을 향함.",
    "built_space": "물웅덩이가 있는 폐허가 된 경기장 구조. 위치 레퍼런스와 일치함.",
    "entities": "장백산의 얼굴과 의상은 레퍼런스와 일치하나(화상 추가됨), 가면은 하회탈이 아닌 일반 판타지 가면임. 배경 왼쪽에 프롬프트에 없는 무장한 인물 2명이 존재함.",
    "hard_violations": [
     "[gemini-pro] invented people (프롬프트에 지시되지 않은 배경의 군인들)",
     "[gemini-pro] unsupported object (어떤 지지대나 끈 없이 얼굴 표면에 임의로 붙어 있는 가면 조각)",
     "[gpt-high] 장백산만 보여야 하는 장면에 배경 인물 두 명과 왼쪽 가장자리의 별도 인물 하반신을 추가했다.",
     "[gpt-high] 샷에 등장하도록 지정되지 않은 배경 인물들의 총기를 추가했다."
    ],
    "physics": "가면 조각이 물리적 지지 없이 얼굴에 붙어 있음. 양손은 허공에 들려 있음."
   },
   {
    "label": "B",
    "direction": "시선은 화면 오른쪽 밖을 향함.",
    "built_space": "금속 잔해와 얕은 물웅덩이가 있는 폐허. 위치 레퍼런스의 환경과 일치함.",
    "entities": "장백산의 외모와 의상이 레퍼런스와 일치하며 피부가 짓무른 묘사가 있음. 금색 하회탈 조각이 얼굴 옆에 정확히 묘사됨. 추가된 인물 없음.",
    "hard_violations": [],
    "physics": "가면 조각이 캐릭터의 옷깃과 쇄골 부근에 얹혀 얼굴에 기대어져 지지받고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트에 지시된 인물 스케일, 하회탈의 형태, 그리고 정확한 배경을 잘 구현했으며 불필요한 요소를 추가하지 않아 우수합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트에 명시되지 않은 인물(군인)을 배경에 추가하는 하드 위반을 범했으며, 가면의 형태도 지시된 하회탈이 아닙니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 왼쪽 밖을 향함.",
        "built_space": "물웅덩이가 있는 폐허가 된 경기장 구조. 위치 레퍼런스와 일치함.",
        "entities": "장백산의 얼굴과 의상은 레퍼런스와 일치하나(화상 추가됨), 가면은 하회탈이 아닌 일반 판타지 가면임. 배경 왼쪽에 프롬프트에 없는 무장한 인물 2명이 존재함.",
        "hard_violations": [
         "invented people (프롬프트에 지시되지 않은 배경의 군인들)",
         "unsupported object (어떤 지지대나 끈 없이 얼굴 표면에 임의로 붙어 있는 가면 조각)"
        ],
        "physics": "가면 조각이 물리적 지지 없이 얼굴에 붙어 있음. 양손은 허공에 들려 있음."
       },
       {
        "label": "B",
        "direction": "시선은 화면 오른쪽 밖을 향함.",
        "built_space": "금속 잔해와 얕은 물웅덩이가 있는 폐허. 위치 레퍼런스의 환경과 일치함.",
        "entities": "장백산의 외모와 의상이 레퍼런스와 일치하며 피부가 짓무른 묘사가 있음. 금색 하회탈 조각이 얼굴 옆에 정확히 묘사됨. 추가된 인물 없음.",
        "hard_violations": [],
        "physics": "가면 조각이 캐릭터의 옷깃과 쇄골 부근에 얹혀 얼굴에 기대어져 지지받고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트에 지시된 인물 스케일, 하회탈의 형태, 그리고 정확한 배경을 잘 구현했으며 불필요한 요소를 추가하지 않아 우수합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트에 명시되지 않은 인물(군인)을 배경에 추가하는 하드 위반을 범했으며, 가면의 형태도 지시된 하회탈이 아닙니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 왼쪽 밖을 향함.",
        "built_space": "물웅덩이가 있는 폐허가 된 경기장 구조. 위치 레퍼런스와 일치함.",
        "entities": "장백산의 얼굴과 의상은 레퍼런스와 일치하나(화상 추가됨), 가면은 하회탈이 아닌 일반 판타지 가면임. 배경 왼쪽에 프롬프트에 없는 무장한 인물 2명이 존재함.",
        "hard_violations": [
         "invented people (프롬프트에 지시되지 않은 배경의 군인들)",
         "unsupported object (어떤 지지대나 끈 없이 얼굴 표면에 임의로 붙어 있는 가면 조각)"
        ],
        "physics": "가면 조각이 물리적 지지 없이 얼굴에 붙어 있음. 양손은 허공에 들려 있음."
       },
       {
        "label": "B",
        "direction": "시선은 화면 오른쪽 밖을 향함.",
        "built_space": "금속 잔해와 얕은 물웅덩이가 있는 폐허. 위치 레퍼런스의 환경과 일치함.",
        "entities": "장백산의 외모와 의상이 레퍼런스와 일치하며 피부가 짓무른 묘사가 있음. 금색 하회탈 조각이 얼굴 옆에 정확히 묘사됨. 추가된 인물 없음.",
        "hard_violations": [],
        "physics": "가면 조각이 캐릭터의 옷깃과 쇄골 부근에 얹혀 얼굴에 기대어져 지지받고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "장백산만 등장하며 젖은 폐허에서 얼굴과 깨진 금빛 가면을 드러내는 클로즈업을 충실히 구현했지만, 피부 손상은 녹아내린 짓무름보다 찢어진 상처에 가깝다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "붉게 짓무른 얼굴과 작은 가면 잔편은 적절하지만, 금지된 추가 인물들과 총기를 배경에 넣어 단독 인물 장면의 필수 조건을 위반했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장백산은 화면 오른쪽의 프레임 밖을 바라본다. 지정된 응시 대상은 없으므로 충돌하지 않는다. 금빛 가면의 바깥 얼굴은 카메라 쪽으로 비스듬히 보이고, 깨진 가장자리가 노출된 뺨에 인접한다. 무기나 이동하는 인물은 없다.",
        "built_space": "왼쪽 뒤에는 계단식 관람석과 기둥들이 있고, 오른쪽에는 큰 곡면 물탱크 잔해 한 부분과 무너진 철골이 있다. 아래와 왼쪽 지면에는 물웅덩이와 파편이 보인다. 장소 참조의 재료와 배치에 부합하며, 중복된 고정 설비나 불가능한 반사는 보이지 않는다. 얼굴 중심 클로즈업이지만 배경과 가슴 부분이 비교적 넓게 남아 있다.",
        "entities": "짧은 회색 머리의 중년 남성 한 명으로, 참조의 장백산과 얼굴 윤곽 및 머리가 대체로 일치한다. 자주색 바탕의 금색 자수 예복이 어깨와 윗가슴에 남아 있다. 얼굴과 머리에는 젖은 광택이 있고 피부가 붉게 벗겨져 있으나, 방사선으로 흉측하게 녹아내린 인상은 다소 약하다. 금빛 하회탈 형태의 깨진 잔편이 얼굴 옆에 있으나 눈·코·입이 상당 부분 남아 있어 작은 반쪽 잔편이라는 요구에는 덜 정확하다. 노을은 붉은색보다 연한 주황색이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "상체를 뒤로 기댄 자세로 목이 머리를 지탱한다. 하체와 등 접점은 프레임 밖이므로 정확한 지지면은 확인할 수 없지만 공중에 떠 있는 모습은 아니다. 가면 잔편은 뺨에 밀착되고 아래끝이 옷깃과 어깨 부위에 걸쳐 있어 지지 없이 떠 있다고 볼 근거가 없다. 철골과 탱크 잔해는 지면 및 다른 잔해에 기대어 있다."
       },
       {
        "label": "B",
        "direction": "장백산은 화면 왼쪽 위를 올려다보며, 그 방향에 무장 인물들이 있다. 배경의 선명한 소총은 오른쪽 아래를 향하지만 장백산을 직접 겨냥한다고 확정하기는 어렵다. 가면 잔편의 금빛 바깥 면은 카메라에 비스듬히 노출된다.",
        "built_space": "뒤쪽에 계단식 관람석과 기둥들, 중앙 계단 한 구간이 보이며 오른쪽에는 붕괴한 철골과 곡면 탱크 잔해가 있다. 물이 고인 바닥과 붉은 노을은 장소 참조에 부합한다. 다만 장백산 뒤의 젖은 지면에 추가 인물들을 배치했다. 얼굴은 크게 잡혔지만 하단에 양손까지 들어와 지정된 어깨·윗가슴 중심의 가장자리 구성에서 벗어난다.",
        "entities": "중앙에는 참조와 비슷한 짧은 회색 머리의 중년 남성이 있고, 자주색과 금색 자수 예복을 입고 있다. 노출된 얼굴 한쪽의 붉은 궤양과 벗겨진 피부는 요구된 짓무름을 A보다 강하게 드러낸다. 정상적인 홍채와 동공이 보이며, 작은 금빛 가면 잔편이 얼굴 옆에 남아 있다. 그러나 왼쪽 배경에 두 사람과 가장자리의 별도 하반신까지 보여 추가 인물이 최소 세 명이며, 총기도 등장한다. 이들은 이 샷에 허용되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장백산만 보여야 하는 장면에 배경 인물 두 명과 왼쪽 가장자리의 별도 인물 하반신을 추가했다.",
         "샷에 등장하도록 지정되지 않은 배경 인물들의 총기를 추가했다."
        ],
        "physics": "장백산의 머리는 기울어진 목과 상체로 지지되며 양손은 가슴과 옷깃 앞에 올라와 있다. 가면은 얼굴 측면에 밀착된 착용 잔편으로 보인다. 배경 인물들의 발은 젖은 지면에 닿아 있고 선명한 소총은 손으로 받쳐 들고 있다. 잔해도 지면에 놓여 있어 명백한 부유나 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "장백산만 등장하며 젖은 폐허에서 얼굴과 깨진 금빛 가면을 드러내는 클로즈업을 충실히 구현했지만, 피부 손상은 녹아내린 짓무름보다 찢어진 상처에 가깝다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "붉게 짓무른 얼굴과 작은 가면 잔편은 적절하지만, 금지된 추가 인물들과 총기를 배경에 넣어 단독 인물 장면의 필수 조건을 위반했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "장백산은 화면 오른쪽의 프레임 밖을 바라본다. 지정된 응시 대상은 없으므로 충돌하지 않는다. 금빛 가면의 바깥 얼굴은 카메라 쪽으로 비스듬히 보이고, 깨진 가장자리가 노출된 뺨에 인접한다. 무기나 이동하는 인물은 없다.",
        "built_space": "왼쪽 뒤에는 계단식 관람석과 기둥들이 있고, 오른쪽에는 큰 곡면 물탱크 잔해 한 부분과 무너진 철골이 있다. 아래와 왼쪽 지면에는 물웅덩이와 파편이 보인다. 장소 참조의 재료와 배치에 부합하며, 중복된 고정 설비나 불가능한 반사는 보이지 않는다. 얼굴 중심 클로즈업이지만 배경과 가슴 부분이 비교적 넓게 남아 있다.",
        "entities": "짧은 회색 머리의 중년 남성 한 명으로, 참조의 장백산과 얼굴 윤곽 및 머리가 대체로 일치한다. 자주색 바탕의 금색 자수 예복이 어깨와 윗가슴에 남아 있다. 얼굴과 머리에는 젖은 광택이 있고 피부가 붉게 벗겨져 있으나, 방사선으로 흉측하게 녹아내린 인상은 다소 약하다. 금빛 하회탈 형태의 깨진 잔편이 얼굴 옆에 있으나 눈·코·입이 상당 부분 남아 있어 작은 반쪽 잔편이라는 요구에는 덜 정확하다. 노을은 붉은색보다 연한 주황색이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "상체를 뒤로 기댄 자세로 목이 머리를 지탱한다. 하체와 등 접점은 프레임 밖이므로 정확한 지지면은 확인할 수 없지만 공중에 떠 있는 모습은 아니다. 가면 잔편은 뺨에 밀착되고 아래끝이 옷깃과 어깨 부위에 걸쳐 있어 지지 없이 떠 있다고 볼 근거가 없다. 철골과 탱크 잔해는 지면 및 다른 잔해에 기대어 있다."
       },
       {
        "label": "A",
        "direction": "장백산은 화면 왼쪽 위를 올려다보며, 그 방향에 무장 인물들이 있다. 배경의 선명한 소총은 오른쪽 아래를 향하지만 장백산을 직접 겨냥한다고 확정하기는 어렵다. 가면 잔편의 금빛 바깥 면은 카메라에 비스듬히 노출된다.",
        "built_space": "뒤쪽에 계단식 관람석과 기둥들, 중앙 계단 한 구간이 보이며 오른쪽에는 붕괴한 철골과 곡면 탱크 잔해가 있다. 물이 고인 바닥과 붉은 노을은 장소 참조에 부합한다. 다만 장백산 뒤의 젖은 지면에 추가 인물들을 배치했다. 얼굴은 크게 잡혔지만 하단에 양손까지 들어와 지정된 어깨·윗가슴 중심의 가장자리 구성에서 벗어난다.",
        "entities": "중앙에는 참조와 비슷한 짧은 회색 머리의 중년 남성이 있고, 자주색과 금색 자수 예복을 입고 있다. 노출된 얼굴 한쪽의 붉은 궤양과 벗겨진 피부는 요구된 짓무름을 A보다 강하게 드러낸다. 정상적인 홍채와 동공이 보이며, 작은 금빛 가면 잔편이 얼굴 옆에 남아 있다. 그러나 왼쪽 배경에 두 사람과 가장자리의 별도 하반신까지 보여 추가 인물이 최소 세 명이며, 총기도 등장한다. 이들은 이 샷에 허용되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장백산만 보여야 하는 장면에 배경 인물 두 명과 왼쪽 가장자리의 별도 인물 하반신을 추가했다.",
         "샷에 등장하도록 지정되지 않은 배경 인물들의 총기를 추가했다."
        ],
        "physics": "장백산의 머리는 기울어진 목과 상체로 지지되며 양손은 가슴과 옷깃 앞에 올라와 있다. 가면은 얼굴 측면에 밀착된 착용 잔편으로 보인다. 배경 인물들의 발은 젖은 지면에 닿아 있고 선명한 소총은 손으로 받쳐 들고 있다. 잔해도 지면에 놓여 있어 명백한 부유나 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.804,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.554,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people (프롬프트에 지시되지 않은 배경의 군인들)",
     "[gemini-pro] unsupported object (어떤 지지대나 끈 없이 얼굴 표면에 임의로 붙어 있는 가면 조각)",
     "[gpt-high] 장백산만 보여야 하는 장면에 배경 인물 두 명과 왼쪽 가장자리의 별도 인물 하반신을 추가했다.",
     "[gpt-high] 샷에 등장하도록 지정되지 않은 배경 인물들의 총기를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 554
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트에 지시된 인물 스케일, 하회탈의 형태, 그리고 정확한 배경을 잘 구현했으며 불필요한 요소를 추가하지 않아 우수합니다."
   },
   {
    "label": "A",
    "score": 554,
    "verdict_ko": "프롬프트에 명시되지 않은 인물(군인)을 배경에 추가하는 하드 위반을 범했으며, 가면의 형태도 지시된 하회탈이 아닙니다.  ★위반: [gemini-pro] invented people (프롬프트에 지시되지 않은 배경의 군인들) / [gemini-pro] unsupported object (어떤 지지대나 끈 없이 얼굴 표면에 임의로 붙어 있는 가면 조각) / [gpt-high] 장백산만 보여야 하는 장면에 배경 인물 두 명과 왼쪽 가장자리의 별도 인물 하반신을 추가했다. / [gpt-high] 샷에 등장하도록 지정되지 않은 배경 인물들의 총기를 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L87B03.png",
    "asset_id": "f6a54720-c3b4-41fd-a06f-736ac241da58",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1531415>",
    "asset_id": "184110cd-a0d4-49f1-a438-62fe0b127791",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdf1-548a-7127-b732-c5b05ec90e01",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S60sh104__bgfirst_bg.png",
   "bg_asset_id": "550f9eb9-cb11-4dc8-8df7-067efb40ca0c",
   "bg_record_key": "S60sh104::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S60sh104::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:49:57.372976+00:00",
  "fingerprint": "78e7ad4b65a63112e9dbfa459c7d01b69ecec57db99f521fb70bbdc8d0973d4d",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S60sh104_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S60sh104_sel.png",
  "source_sha256": "bf8294cb43881bd3b9a555f657f655d2c16a3ec809f63a03aaaac0ec82218f19",
  "file": "S60sh104_cine.png",
  "staged_sha256": "1255f043ec1485d115be09943e192a9173c5c75365e73fe585bb80f96757f72a",
  "latency_ms": 19683
 },
 "S60sh129::signage": {
  "fp": "39a02943cfed677e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S60sh129::bgfirst_bg": {
  "input_fingerprint": "181eb34fe5b7a05c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 지는 해를 바라보며 찰리와 이현우가 손을 맞잡은 채 어깨를 나란히 한 감동적인 후면 실루엣.\n\nLOCATION (lock): On a ridge overlooking the damaged village, with the setting sun ahead of the seated companions.\n\nTIME OF DAY (lock): sunset, red sky.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (찰리 and 이현우 sit together after escaping the village) — The ridge extends beneath their seated bodies across the lower frame; used as A shared supporting plane anchors the pair while the withdrawal opens space around them; 지는 해와 하늘 (The sun is setting beyond the ridge); used as Provides the distant destination of both eyelines and the open background behind their silhouettes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun creates tender rear silhouettes, retaining enough edge separation for the joined hands and their battered outlines to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 지는 해를 바라보며 찰리와 이현우가 손을 맞잡은 채 어깨를 나란히 한 감동적인 후면 실루엣.\n\nLOCATION (lock): On a ridge overlooking the damaged village, with the setting sun ahead of the seated companions.\n\nTIME OF DAY (lock): sunset, red sky.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (찰리 and 이현우 sit together after escaping the village) — The ridge extends beneath their seated bodies across the lower frame; used as A shared supporting plane anchors the pair while the withdrawal opens space around them; 지는 해와 하늘 (The sun is setting beyond the ridge); used as Provides the distant destination of both eyelines and the open background behind their silhouettes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun creates tender rear silhouettes, retaining enough edge separation for the joined hands and their battered outlines to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S60sh129__bgfirst_bg.png",
  "asset_id": "c0aba654-9c0e-49ab-8a23-bcb1f707fb5a",
  "input_asset_ids": [
   "05004506-85e2-4264-9110-26dfd49074e4",
   "419b257c-78bd-4139-b801-d4259c549b9a"
  ]
 },
 "S60sh129": {
  "input_fingerprint": "1bf99cc54770fbba",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): sunset, red sky.\n\nSHOT TEXT (authoritative, Korean): 지는 해를 바라보며 찰리와 이현우가 손을 맞잡은 채 어깨를 나란히 한 감동적인 후면 실루엣.\n\nLOCATION (lock): On a ridge overlooking the damaged village, with the setting sun ahead of the seated companions. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (찰리 and 이현우 sit together after escaping the village) — The ridge extends beneath their seated bodies across the lower frame; used as A shared supporting plane anchors the pair while the withdrawal opens space around them; 지는 해와 하늘 (The sun is setting beyond the ridge); used as Provides the distant destination of both eyelines and the open background behind their silhouettes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun creates tender rear silhouettes, retaining enough edge separation for the joined hands and their battered outlines to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie sits on the ridge with holes and dents in his body and malfunctioning sensors; B-200's gun arms remain intact. The devastated village and collapsed water tower remain below in the fading sunset. 이현우: He sits on the ridge, visibly battered and wounded, with one hand extended in a clasp.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): sunset, red sky.\n\nSHOT TEXT (authoritative, Korean): 지는 해를 바라보며 찰리와 이현우가 손을 맞잡은 채 어깨를 나란히 한 감동적인 후면 실루엣.\n\nLOCATION (lock): On a ridge overlooking the damaged village, with the setting sun ahead of the seated companions. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (찰리 and 이현우 sit together after escaping the village) — The ridge extends beneath their seated bodies across the lower frame; used as A shared supporting plane anchors the pair while the withdrawal opens space around them; 지는 해와 하늘 (The sun is setting beyond the ridge); used as Provides the distant destination of both eyelines and the open background behind their silhouettes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun creates tender rear silhouettes, retaining enough edge separation for the joined hands and their battered outlines to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie sits on the ridge with holes and dents in his body and malfunctioning sensors; B-200's gun arms remain intact. The devastated village and collapsed water tower remain below in the fading sunset. 이현우: He sits on the ridge, visibly battered and wounded, with one hand extended in a clasp.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): sunset, red sky.\n\nSHOT TEXT (authoritative, Korean): 지는 해를 바라보며 찰리와 이현우가 손을 맞잡은 채 어깨를 나란히 한 감동적인 후면 실루엣.\n\nLOCATION (lock): On a ridge overlooking the damaged village, with the setting sun ahead of the seated companions. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (찰리 and 이현우 sit together after escaping the village) — The ridge extends beneath their seated bodies across the lower frame; used as A shared supporting plane anchors the pair while the withdrawal opens space around them; 지는 해와 하늘 (The sun is setting beyond the ridge); used as Provides the distant destination of both eyelines and the open background behind their silhouettes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun creates tender rear silhouettes, retaining enough edge separation for the joined hands and their battered outlines to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie sits on the ridge with holes and dents in his body and malfunctioning sensors; B-200's gun arms remain intact. The devastated village and collapsed water tower remain below in the fading sunset. 이현우: He sits on the ridge, visibly battered and wounded, with one hand extended in a clasp.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S60sh129__bgfirst_bg.png",
     "asset_id": "c0aba654-9c0e-49ab-8a23-bcb1f707fb5a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S60sh129.png",
     "asset_id": "05004506-85e2-4264-9110-26dfd49074e4",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1486510>",
     "asset_id": "786398f5-97e1-4b9a-9afe-a9b796604793",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L87B02.png",
     "asset_id": "419b257c-78bd-4139-b801-d4259c549b9a",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1486510>",
     "asset_id": "786398f5-97e1-4b9a-9afe-a9b796604793",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 화면 중앙에 위치한 지는 해를 향해 시선을 두고 있습니다.",
    "built_space": "레퍼런스 사진과 동일한 바위와 마른 풀이 있는 능선에 앉아 있으며, 아래 계곡 마을에는 온전하게 서 있는 급수탑이 보입니다.",
    "entities": "이현우(왼쪽, 후드티 착용)와 찰리(오른쪽, 금속 로봇 외장 및 등에 밝게 빛나는 포트)가 등장하며 두 사람이 손을 잡고 있습니다.",
    "hard_violations": [],
    "physics": "두 사람 모두 지면에 안정적으로 앉아 체중을 지탱하고 있으며, 손을 맞잡은 자세가 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 붉은 노을이 지는 정면의 태양을 향해 시선을 고정하고 있습니다.",
    "built_space": "흙으로 덮인 능선 위에 앉아 있으며(레퍼런스의 세부 바위 지형과는 차이가 있음), 아래쪽 폐허가 된 마을에는 프롬프트대로 기울어져 무너진 급수탑이 존재합니다.",
    "entities": "찰리(왼쪽, 파손된 자국과 전선이 드러난 로봇)와 이현우(오른쪽, 후드티 착용 및 상처 입은 얼굴 측면)가 어두운 실루엣으로 올바르게 묘사됩니다.",
    "hard_violations": [],
    "physics": "인물들은 능선 바닥에 앉아 무게중심을 잡고 있으며, 서로 손을 맞잡은 상태로 안정적으로 지지되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 후면 실루엣, 붉은 하늘, 무너진 급수탑의 디테일을 정확히 구현했으나, 로케이션 지형이 레퍼런스와 다소 다릅니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "로케이션 지형은 레퍼런스와 완벽히 일치하지만, 핵심 지시사항인 후면 실루엣을 무시하고 인물의 등이 밝게 묘사되었으며 급수탑도 멀쩡합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 화면 중앙에 위치한 지는 해를 향해 시선을 두고 있습니다.",
        "built_space": "레퍼런스 사진과 동일한 바위와 마른 풀이 있는 능선에 앉아 있으며, 아래 계곡 마을에는 온전하게 서 있는 급수탑이 보입니다.",
        "entities": "이현우(왼쪽, 후드티 착용)와 찰리(오른쪽, 금속 로봇 외장 및 등에 밝게 빛나는 포트)가 등장하며 두 사람이 손을 잡고 있습니다.",
        "hard_violations": [],
        "physics": "두 사람 모두 지면에 안정적으로 앉아 체중을 지탱하고 있으며, 손을 맞잡은 자세가 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 붉은 노을이 지는 정면의 태양을 향해 시선을 고정하고 있습니다.",
        "built_space": "흙으로 덮인 능선 위에 앉아 있으며(레퍼런스의 세부 바위 지형과는 차이가 있음), 아래쪽 폐허가 된 마을에는 프롬프트대로 기울어져 무너진 급수탑이 존재합니다.",
        "entities": "찰리(왼쪽, 파손된 자국과 전선이 드러난 로봇)와 이현우(오른쪽, 후드티 착용 및 상처 입은 얼굴 측면)가 어두운 실루엣으로 올바르게 묘사됩니다.",
        "hard_violations": [],
        "physics": "인물들은 능선 바닥에 앉아 무게중심을 잡고 있으며, 서로 손을 맞잡은 상태로 안정적으로 지지되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 후면 실루엣, 붉은 하늘, 무너진 급수탑의 디테일을 정확히 구현했으나, 로케이션 지형이 레퍼런스와 다소 다릅니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "로케이션 지형은 레퍼런스와 완벽히 일치하지만, 핵심 지시사항인 후면 실루엣을 무시하고 인물의 등이 밝게 묘사되었으며 급수탑도 멀쩡합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 화면 중앙에 위치한 지는 해를 향해 시선을 두고 있습니다.",
        "built_space": "레퍼런스 사진과 동일한 바위와 마른 풀이 있는 능선에 앉아 있으며, 아래 계곡 마을에는 온전하게 서 있는 급수탑이 보입니다.",
        "entities": "이현우(왼쪽, 후드티 착용)와 찰리(오른쪽, 금속 로봇 외장 및 등에 밝게 빛나는 포트)가 등장하며 두 사람이 손을 잡고 있습니다.",
        "hard_violations": [],
        "physics": "두 사람 모두 지면에 안정적으로 앉아 체중을 지탱하고 있으며, 손을 맞잡은 자세가 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 붉은 노을이 지는 정면의 태양을 향해 시선을 고정하고 있습니다.",
        "built_space": "흙으로 덮인 능선 위에 앉아 있으며(레퍼런스의 세부 바위 지형과는 차이가 있음), 아래쪽 폐허가 된 마을에는 프롬프트대로 기울어져 무너진 급수탑이 존재합니다.",
        "entities": "찰리(왼쪽, 파손된 자국과 전선이 드러난 로봇)와 이현우(오른쪽, 후드티 착용 및 상처 입은 얼굴 측면)가 어두운 실루엣으로 올바르게 묘사됩니다.",
        "hard_violations": [],
        "physics": "인물들은 능선 바닥에 앉아 무게중심을 잡고 있으며, 서로 손을 맞잡은 상태로 안정적으로 지지되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "맞잡은 손과 붉은 역광은 명확하지만, 두 인물이 해보다 서로를 향하고 인물과 마을이 크게 잡혀 요구된 넓은 후면 구도와 능선 장소성이 약하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "능선에 나란히 앉아 손잡고 일몰을 향하는 후면 와이드 구도와 참조 장소를 더 충실히 구현했으나, 급수탑이 무너지지 않았고 현우의 부상 표현은 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 머리를 오른쪽 현우 쪽으로 돌렸고, 현우는 왼쪽 찰리의 얼굴 쪽을 바라본다. 해는 두 사람 사이의 먼 지평선에 있지만 두 사람 모두 해를 바라보는 순간으로 읽히지는 않는다. 안쪽 팔은 서로에게 뻗어 손을 맞잡는다. 별도의 총구는 식별되지 않는다.",
        "built_space": "두 사람은 화면 아래를 가로지르는 흙 능선 위에 찰리 왼쪽, 현우 오른쪽으로 앉아 있다. 바로 아래에는 여러 파손 주택과 왼쪽으로 기울어진 급수탑 한 기가 보인다. 급수탑은 탱크와 지지대가 여전히 서 있어 완전히 붕괴한 상태는 아니다. 참조의 돌·마른 풀·흙 재질은 있으나, 가까운 마을이 배경을 크게 차지해 참조의 넓고 겹겹이 펼쳐진 산지 전망과 차이가 크다.",
        "entities": "찰리와 현우 두 명만 보인다. 찰리는 둥근 어깨 장갑과 분절된 금속 몸체를 갖추었고 등에 여러 구멍과 손상이 있다. 뒤돌린 머리 때문에 흰 얼굴판과 점 눈, 선 입은 확인할 수 없다. 현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보이며 어두운 후드와 청바지가 참조에 부합하지만 정확한 나이와 얼굴 동일성은 확정하기 어렵다. 후드의 마모는 보이나 뚜렷한 신체 상처는 드러나지 않는다. 센서 고장과 명시된 총기 팔의 상태도 확인되지 않는다. 붉은 하늘, 지는 해, 파손 마을은 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 엉덩이와 접힌 다리가 능선 지면에 닿아 체중을 지탱한다. 맞잡은 손은 각각 몸통에 연결된 팔에서 자연스럽게 내려와 서로 접촉한다. 찰리의 돌출 배선은 등 장갑에 연결되어 있다. 공중에 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "두 사람은 등을 카메라에 두고 앞쪽 산지와 일몰이 있는 지평선을 향한다. 현우의 머리는 약간 오른쪽으로 돌아가지만 A처럼 상대 얼굴을 직접 바라보는 옆얼굴은 아니다. 해는 전방 오른쪽에 있으며 눈이 보이지 않아 정확한 주시점은 확인할 수 없다. 현우의 오른손과 찰리의 왼손이 두 사람 사이에서 맞잡혀 있다. 식별 가능한 총구는 없다.",
        "built_space": "마른 풀과 관목, 흙바닥, 양옆 바위가 있는 능선이 하단 전체에 이어지고, 현우는 왼쪽, 찰리는 오른쪽에서 같은 지면에 앉아 있다. 능선 아래 멀리 작은 건물들이 모인 마을과 급수탑 한 기가 보인다. 급수탑은 곧게 서 있어 붕괴 조건에는 맞지 않는다. 전경의 암석과 식생, 열린 골짜기와 겹치는 산줄기는 참조 장소에 가깝고, 마을도 먼 거리의 작은 크기로 유지된다.",
        "entities": "추가 인물 없이 현우와 찰리만 보인다. 현우의 헝클어진 검은 머리, 젊고 마른 체격, 어두운 후드와 청바지는 참조와 부합한다. 후면이므로 한국계 얼굴과 정확한 나이는 확인할 수 없다. 찰리는 분절된 금속 머리와 장갑 몸체를 갖추고 등에 파손 구멍이 있으나 참조보다 상체가 다소 가늘고 등 장갑 형태도 다르다. 숨겨진 흰 얼굴판과 눈·입은 평가할 수 없다. 센서 고장과 총기 팔은 명확히 식별되지 않으며 현우의 부상도 두드러지지 않는다. 먼 마을에는 연기가 보이고 하늘은 붉은 주황빛 일몰이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 인물의 엉덩이와 접힌 다리가 흙 능선에 직접 닿아 앉은 몸을 지지한다. 안쪽 팔은 아래로 이어져 지면 가까이에서 손을 맞잡으며 관절 연결과 접촉이 자연스럽다. 장갑과 배선은 몸체에 부착되어 있고 바위와 건물도 지면에 놓여 있다. 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "맞잡은 손과 붉은 역광은 명확하지만, 두 인물이 해보다 서로를 향하고 인물과 마을이 크게 잡혀 요구된 넓은 후면 구도와 능선 장소성이 약하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "능선에 나란히 앉아 손잡고 일몰을 향하는 후면 와이드 구도와 참조 장소를 더 충실히 구현했으나, 급수탑이 무너지지 않았고 현우의 부상 표현은 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 머리를 오른쪽 현우 쪽으로 돌렸고, 현우는 왼쪽 찰리의 얼굴 쪽을 바라본다. 해는 두 사람 사이의 먼 지평선에 있지만 두 사람 모두 해를 바라보는 순간으로 읽히지는 않는다. 안쪽 팔은 서로에게 뻗어 손을 맞잡는다. 별도의 총구는 식별되지 않는다.",
        "built_space": "두 사람은 화면 아래를 가로지르는 흙 능선 위에 찰리 왼쪽, 현우 오른쪽으로 앉아 있다. 바로 아래에는 여러 파손 주택과 왼쪽으로 기울어진 급수탑 한 기가 보인다. 급수탑은 탱크와 지지대가 여전히 서 있어 완전히 붕괴한 상태는 아니다. 참조의 돌·마른 풀·흙 재질은 있으나, 가까운 마을이 배경을 크게 차지해 참조의 넓고 겹겹이 펼쳐진 산지 전망과 차이가 크다.",
        "entities": "찰리와 현우 두 명만 보인다. 찰리는 둥근 어깨 장갑과 분절된 금속 몸체를 갖추었고 등에 여러 구멍과 손상이 있다. 뒤돌린 머리 때문에 흰 얼굴판과 점 눈, 선 입은 확인할 수 없다. 현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보이며 어두운 후드와 청바지가 참조에 부합하지만 정확한 나이와 얼굴 동일성은 확정하기 어렵다. 후드의 마모는 보이나 뚜렷한 신체 상처는 드러나지 않는다. 센서 고장과 명시된 총기 팔의 상태도 확인되지 않는다. 붉은 하늘, 지는 해, 파손 마을은 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 엉덩이와 접힌 다리가 능선 지면에 닿아 체중을 지탱한다. 맞잡은 손은 각각 몸통에 연결된 팔에서 자연스럽게 내려와 서로 접촉한다. 찰리의 돌출 배선은 등 장갑에 연결되어 있다. 공중에 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "두 사람은 등을 카메라에 두고 앞쪽 산지와 일몰이 있는 지평선을 향한다. 현우의 머리는 약간 오른쪽으로 돌아가지만 A처럼 상대 얼굴을 직접 바라보는 옆얼굴은 아니다. 해는 전방 오른쪽에 있으며 눈이 보이지 않아 정확한 주시점은 확인할 수 없다. 현우의 오른손과 찰리의 왼손이 두 사람 사이에서 맞잡혀 있다. 식별 가능한 총구는 없다.",
        "built_space": "마른 풀과 관목, 흙바닥, 양옆 바위가 있는 능선이 하단 전체에 이어지고, 현우는 왼쪽, 찰리는 오른쪽에서 같은 지면에 앉아 있다. 능선 아래 멀리 작은 건물들이 모인 마을과 급수탑 한 기가 보인다. 급수탑은 곧게 서 있어 붕괴 조건에는 맞지 않는다. 전경의 암석과 식생, 열린 골짜기와 겹치는 산줄기는 참조 장소에 가깝고, 마을도 먼 거리의 작은 크기로 유지된다.",
        "entities": "추가 인물 없이 현우와 찰리만 보인다. 현우의 헝클어진 검은 머리, 젊고 마른 체격, 어두운 후드와 청바지는 참조와 부합한다. 후면이므로 한국계 얼굴과 정확한 나이는 확인할 수 없다. 찰리는 분절된 금속 머리와 장갑 몸체를 갖추고 등에 파손 구멍이 있으나 참조보다 상체가 다소 가늘고 등 장갑 형태도 다르다. 숨겨진 흰 얼굴판과 눈·입은 평가할 수 없다. 센서 고장과 총기 팔은 명확히 식별되지 않으며 현우의 부상도 두드러지지 않는다. 먼 마을에는 연기가 보이고 하늘은 붉은 주황빛 일몰이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 인물의 엉덩이와 접힌 다리가 흙 능선에 직접 닿아 앉은 몸을 지지한다. 안쪽 팔은 아래로 이어져 지면 가까이에서 손을 맞잡으며 관절 연결과 접촉이 자연스럽다. 장갑과 배선은 몸체에 부착되어 있고 바위와 건물도 지면에 놓여 있다. 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.625
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.625
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1625,
   "A": 1571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "프롬프트가 요구한 후면 실루엣, 붉은 하늘, 무너진 급수탑의 디테일을 정확히 구현했으나, 로케이션 지형이 레퍼런스와 다소 다릅니다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "로케이션 지형은 레퍼런스와 완벽히 일치하지만, 핵심 지시사항인 후면 실루엣을 무시하고 인물의 등이 밝게 묘사되었으며 급수탑도 멀쩡합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L87B02.png",
    "asset_id": "419b257c-78bd-4139-b801-d4259c549b9a",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1486510>",
    "asset_id": "786398f5-97e1-4b9a-9afe-a9b796604793",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdf2-9634-7f72-8e4b-b657f88496e7",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S60sh129__bgfirst_bg.png",
   "bg_asset_id": "c0aba654-9c0e-49ab-8a23-bcb1f707fb5a",
   "bg_record_key": "S60sh129::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S60sh129::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:50:17.481995+00:00",
  "fingerprint": "4ea2e67c3602b6bbd730711625921052430cf884d2864e6e1968ce1022b6dd56",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S60sh129_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S60sh129_sel.png",
  "source_sha256": "5beb4b4b144086d6b45603ef9878f6032ae32e0efc4312164febf43590924fe9",
  "file": "S60sh129_cine.png",
  "staged_sha256": "7b5488e429c49d33ff4f79367cf57d773577a0e65b62ca1cb47e5c5e50947d8f",
  "latency_ms": 11698
 },
 "S61sh1::signage": {
  "fp": "a225a709bd822b15",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S61sh1::bgfirst_bg": {
  "input_fingerprint": "09a14ed5b3b2124c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 늪지대, 박철진이 방치된 낡은 캠핑카의 녹슨 차체를 한 손으로 짚고 선 전신.\n\nLOCATION (lock): In the nighttime marsh clearing beside the abandoned camper stuck in mud.\n\nTIME OF DAY (lock): night, moon visible.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 낡은 캠핑카 (Left in the swamp with rusty bodywork; being touched and inspected) — Its near side recedes obliquely from 박철진 toward the inspecting militia; used as Connects the searchers through a shared physical object while remaining within roughly a third of the frame; 늪지대 (The abandoned vehicle and search party are located here); used as Surrounding negative space establishes the isolated search location without adding weather or surface effects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the swamp at the scene's stated nighttime darkness, using restrained ambient visibility without specifying an unsupported lamp, beam, or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 늪지대, 박철진이 방치된 낡은 캠핑카의 녹슨 차체를 한 손으로 짚고 선 전신.\n\nLOCATION (lock): In the nighttime marsh clearing beside the abandoned camper stuck in mud.\n\nTIME OF DAY (lock): night, moon visible.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 낡은 캠핑카 (Left in the swamp with rusty bodywork; being touched and inspected) — Its near side recedes obliquely from 박철진 toward the inspecting militia; used as Connects the searchers through a shared physical object while remaining within roughly a third of the frame; 늪지대 (The abandoned vehicle and search party are located here); used as Surrounding negative space establishes the isolated search location without adding weather or surface effects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the swamp at the scene's stated nighttime darkness, using restrained ambient visibility without specifying an unsupported lamp, beam, or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S61sh1__bgfirst_bg.png",
  "asset_id": "03734688-7bcc-44a6-9ca9-c40f40385a32",
  "input_asset_ids": [
   "8128b673-2497-4d1c-b053-db42f874a9cf",
   "21b3c525-8f13-484a-9366-828766ce2823"
  ]
 },
 "S61sh1": {
  "input_fingerprint": "99fcbdadbd5c81a6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 어두운 늪지대, 박철진이 방치된 낡은 캠핑카의 녹슨 차체를 한 손으로 짚고 선 전신.\n\nLOCATION (lock): In the nighttime marsh clearing beside the abandoned camper stuck in mud. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 낡은 캠핑카 (Left in the swamp with rusty bodywork; being touched and inspected) — Its near side recedes obliquely from 박철진 toward the inspecting militia; used as Connects the searchers through a shared physical object while remaining within roughly a third of the frame; 늪지대 (The abandoned vehicle and search party are located here); used as Surrounding negative space establishes the isolated search location without adding weather or surface effects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the swamp at the scene's stated nighttime darkness, using restrained ambient visibility without specifying an unsupported lamp, beam, or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned camper remains stuck in the swamp where it was left. A no-entry sign and a corroded, barely legible radiation warning remain nearby under the night sky. 박철진: He is beside the abandoned camper, inspecting it.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 어두운 늪지대, 박철진이 방치된 낡은 캠핑카의 녹슨 차체를 한 손으로 짚고 선 전신.\n\nLOCATION (lock): In the nighttime marsh clearing beside the abandoned camper stuck in mud. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 낡은 캠핑카 (Left in the swamp with rusty bodywork; being touched and inspected) — Its near side recedes obliquely from 박철진 toward the inspecting militia; used as Connects the searchers through a shared physical object while remaining within roughly a third of the frame; 늪지대 (The abandoned vehicle and search party are located here); used as Surrounding negative space establishes the isolated search location without adding weather or surface effects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the swamp at the scene's stated nighttime darkness, using restrained ambient visibility without specifying an unsupported lamp, beam, or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned camper remains stuck in the swamp where it was left. A no-entry sign and a corroded, barely legible radiation warning remain nearby under the night sky. 박철진: He is beside the abandoned camper, inspecting it.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 어두운 늪지대, 박철진이 방치된 낡은 캠핑카의 녹슨 차체를 한 손으로 짚고 선 전신.\n\nLOCATION (lock): In the nighttime marsh clearing beside the abandoned camper stuck in mud. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 낡은 캠핑카 (Left in the swamp with rusty bodywork; being touched and inspected) — Its near side recedes obliquely from 박철진 toward the inspecting militia; used as Connects the searchers through a shared physical object while remaining within roughly a third of the frame; 늪지대 (The abandoned vehicle and search party are located here); used as Surrounding negative space establishes the isolated search location without adding weather or surface effects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the swamp at the scene's stated nighttime darkness, using restrained ambient visibility without specifying an unsupported lamp, beam, or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned camper remains stuck in the swamp where it was left. A no-entry sign and a corroded, barely legible radiation warning remain nearby under the night sky. 박철진: He is beside the abandoned camper, inspecting it.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S61sh1__bgfirst_bg.png",
     "asset_id": "03734688-7bcc-44a6-9ca9-c40f40385a32",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S61sh1.png",
     "asset_id": "8128b673-2497-4d1c-b053-db42f874a9cf",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822047>",
     "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_swamp_mud_clearing_sel.png",
     "asset_id": "21b3c525-8f13-484a-9366-828766ce2823",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822047>",
     "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "박철진의 시선은 손이 닿은 캠핑카의 표면을 향하며, 배경의 한 인물은 왼쪽을 가리킴.",
    "built_space": "진흙과 갈대가 있는 늪지대이며 왼쪽에 녹슨 방사선 표지판이 세워져 있음. 오른쪽에 위치한 낡은 캠핑카가 뒤쪽으로 비스듬히 이어짐.",
    "entities": "박철진(얼굴과 군복은 일치하나 모자는 없음), 낡은 캠핑카, 샷 텍스트에 명시되지 않은 4명의 추가 인물.",
    "hard_violations": [
     "[gemini-pro] invented people",
     "[gpt-high] 박철진만 등장해야 하는 장면에 추가 인물 네 명을 배치했다.",
     "[gpt-high] 허용되지 않은 추가 인물들에게 총기와 배낭을 부여했다."
    ],
    "physics": "박철진은 진흙을 딛고 서서 오른손으로 캠핑카를 짚고 있으며, 배경 인물들도 정상적으로 땅에 서 있음."
   },
   {
    "label": "B",
    "direction": "박철진은 화면 왼쪽 밖을 응시하고, 배경 인물들은 앞쪽을 향함.",
    "built_space": "늪지대 환경이며 왼쪽에 낡은 캠핑카가 배치됨. 배경 중앙 우측에 방사선 표지판이 있음.",
    "entities": "박철진(얼굴, 모자, 군복 일치), 낡은 캠핑카, 샷 텍스트에 없는 플래시라이트를 든 4명의 추가 인물.",
    "hard_violations": [
     "[gemini-pro] invented people",
     "[gemini-pro] leaked text",
     "[gpt-high] 박철진만 등장해야 하는 장면에 추가 인물 네 명을 배치했다.",
     "[gpt-high] 허용되지 않은 추가 인물들에게 총기와 켜진 손전등을 부여했다.",
     "[gpt-high] 완장에 읽을 수 있는 '헌병'과 'MP' 글자가 노출되어 있다."
    ],
    "physics": "박철진은 풀과 진흙 위에 서서 오른손으로 캠핑카를 짚고 지탱함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지침의 모순으로 인해 샷 텍스트에 없는 인물들이 배경에 등장하는 치명적 오류가 있으나, 캠핑카의 비스듬한 구도와 텍스트 비노출 규칙을 잘 지켰습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "텍스트에 없는 추가 인물이 등장하고 완장에 읽을 수 있는 텍스트('헌병 MP')가 노출되는 등 다수의 치명적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 시선은 손이 닿은 캠핑카의 표면을 향하며, 배경의 한 인물은 왼쪽을 가리킴.",
        "built_space": "진흙과 갈대가 있는 늪지대이며 왼쪽에 녹슨 방사선 표지판이 세워져 있음. 오른쪽에 위치한 낡은 캠핑카가 뒤쪽으로 비스듬히 이어짐.",
        "entities": "박철진(얼굴과 군복은 일치하나 모자는 없음), 낡은 캠핑카, 샷 텍스트에 명시되지 않은 4명의 추가 인물.",
        "hard_violations": [
         "invented people"
        ],
        "physics": "박철진은 진흙을 딛고 서서 오른손으로 캠핑카를 짚고 있으며, 배경 인물들도 정상적으로 땅에 서 있음."
       },
       {
        "label": "B",
        "direction": "박철진은 화면 왼쪽 밖을 응시하고, 배경 인물들은 앞쪽을 향함.",
        "built_space": "늪지대 환경이며 왼쪽에 낡은 캠핑카가 배치됨. 배경 중앙 우측에 방사선 표지판이 있음.",
        "entities": "박철진(얼굴, 모자, 군복 일치), 낡은 캠핑카, 샷 텍스트에 없는 플래시라이트를 든 4명의 추가 인물.",
        "hard_violations": [
         "invented people",
         "leaked text"
        ],
        "physics": "박철진은 풀과 진흙 위에 서서 오른손으로 캠핑카를 짚고 지탱함."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지침의 모순으로 인해 샷 텍스트에 없는 인물들이 배경에 등장하는 치명적 오류가 있으나, 캠핑카의 비스듬한 구도와 텍스트 비노출 규칙을 잘 지켰습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "텍스트에 없는 추가 인물이 등장하고 완장에 읽을 수 있는 텍스트('헌병 MP')가 노출되는 등 다수의 치명적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 시선은 손이 닿은 캠핑카의 표면을 향하며, 배경의 한 인물은 왼쪽을 가리킴.",
        "built_space": "진흙과 갈대가 있는 늪지대이며 왼쪽에 녹슨 방사선 표지판이 세워져 있음. 오른쪽에 위치한 낡은 캠핑카가 뒤쪽으로 비스듬히 이어짐.",
        "entities": "박철진(얼굴과 군복은 일치하나 모자는 없음), 낡은 캠핑카, 샷 텍스트에 명시되지 않은 4명의 추가 인물.",
        "hard_violations": [
         "invented people"
        ],
        "physics": "박철진은 진흙을 딛고 서서 오른손으로 캠핑카를 짚고 있으며, 배경 인물들도 정상적으로 땅에 서 있음."
       },
       {
        "label": "B",
        "direction": "박철진은 화면 왼쪽 밖을 응시하고, 배경 인물들은 앞쪽을 향함.",
        "built_space": "늪지대 환경이며 왼쪽에 낡은 캠핑카가 배치됨. 배경 중앙 우측에 방사선 표지판이 있음.",
        "entities": "박철진(얼굴, 모자, 군복 일치), 낡은 캠핑카, 샷 텍스트에 없는 플래시라이트를 든 4명의 추가 인물.",
        "hard_violations": [
         "invented people",
         "leaked text"
        ],
        "physics": "박철진은 풀과 진흙 위에 서서 오른손으로 캠핑카를 짚고 지탱함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "전신과 한 손의 차체 접촉, 약 3분의 1을 차지하는 캠핑카와 늪의 여백은 더 충실하지만, 금지된 추가 인물 네 명 때문에 실격입니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "모자와 복장은 참조에 더 가깝지만, 추가 인물 네 명과 읽히는 글자·손전등을 넣었고 캠핑카도 요구보다 크게 배치했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진은 왼쪽의 캠핑카 창과 차체를 보며 한 손바닥을 창 가장자리의 금속 부분에 댄다. 뒤의 네 사람은 대체로 차량과 박철진 쪽을 향한다. 앞쪽 추가 남성의 소총 총구는 오른쪽 아래 지면을 향하고, 뒤쪽 손전등 세 개는 카메라 쪽으로 밝은 점을 드러낸다. 차량 측면은 왼쪽 전경에서 화면 중앙 뒤쪽으로 물러난다.",
        "built_space": "캠핑카 한 대가 왼쪽 화면의 거의 절반을 차지하며 위쪽과 왼쪽이 잘렸다. 큰 후면 창 하나, 측면 창 두 개와 측면 문 하나가 보인다. 박철진은 차량 바로 옆 진흙 지면에 서 있고 추가 인물 네 명은 오른쪽 뒤에 있다. 두 기둥으로 받친 방사능 표지판 하나가 중앙 뒤에 있다. 갈대와 산 능선은 참조와 유사하지만 참조의 포장 진입로와 물이 넓게 드러난 배치는 확인되지 않는다.",
        "entities": "박철진에 해당하는 중년 남성 한 명은 짧은 회색 섞인 수염, 녹색 모자와 군복, 검은 완장 및 전투화를 착용해 인물 참조에 비교적 가깝다. 그러나 허용되지 않은 사람 네 명과 그들의 총기·손전등도 보인다. 녹슨 폐캠핑카, 늪 식생, 달, 방사능 표지판은 있다. 별도의 출입금지 표지는 식별되지 않는다. 완장의 '헌병'과 'MP'가 읽혀 글자 금지 조건에 어긋난다.",
        "hard_violations": [
         "박철진만 등장해야 하는 장면에 추가 인물 네 명을 배치했다.",
         "허용되지 않은 추가 인물들에게 총기와 켜진 손전등을 부여했다.",
         "완장에 읽을 수 있는 '헌병'과 'MP' 글자가 노출되어 있다."
        ],
        "physics": "박철진의 양발은 진흙 지면에 닿고 손바닥은 차량에 접촉한다. 추가 인물들도 지면 위에 서 있으며 총기와 손전등은 손이나 멜빵으로 지지된다. 차량 하부는 진흙과 식생에 잠겨 있고 표지판은 기둥으로 지지된다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "박철진은 오른쪽의 녹슨 차체와 손을 댄 지점을 내려다본다. 한 손은 창 아래쪽 차체에 닿고 다른 손은 몸 옆에 내려와 있다. 뒤에서 웅크린 남성은 차량 하부를 살피고, 중앙의 모자 쓴 남성은 오른쪽 차량 방향을 가리킨다. 왼쪽 인물의 등에 멘 총은 총구가 왼쪽 위로 향하며 중앙 인물의 총은 아래쪽으로 향한다. 차량 측면은 오른쪽 전경에서 중앙 뒤쪽으로 물러난다.",
        "built_space": "캠핑카 한 대가 오른쪽 약 3분의 1을 차지하고, 창 여러 개와 문 하나, 측면 거울 하나, 진흙에 잠긴 바퀴 두 개가 보인다. 박철진은 차량 옆에 전신으로 서 있다. 추가 인물 한 명은 차량 하부 옆에 웅크리고 세 명은 뒤쪽 늪에 서 있다. 왼쪽에는 두 기둥으로 받친 방사능 표지판 하나가 있다. 물웅덩이와 갈대, 산 능선이 참조의 늪 환경에 가깝지만 포장 진입로는 보이지 않아 정확한 장소 배치는 확인하기 어렵다.",
        "entities": "주인공은 회색 섞인 짧은 머리와 수염을 가진 중년 남성이며 녹색 군복, 검은 완장, 전투화를 착용한다. 참조와 대체로 비슷한 인상이지만 참조의 모자가 없다. 허용되지 않은 추가 남성 네 명과 총기·배낭이 보인다. 녹슨 캠핑카와 진흙 늪, 달, 부식된 방사능 표지판은 확인된다. 출입금지 표시는 식별되지 않으며 명확히 읽히는 문구도 보이지 않는다.",
        "hard_violations": [
         "박철진만 등장해야 하는 장면에 추가 인물 네 명을 배치했다.",
         "허용되지 않은 추가 인물들에게 총기와 배낭을 부여했다."
        ],
        "physics": "박철진의 두 부츠는 진흙에 닿고 뻗은 손은 차체를 짚어 자연스러운 지지 관계를 이룬다. 웅크린 남성은 굽힌 다리와 지면에 닿은 부츠로 몸을 지탱한다. 나머지 인물도 발로 지면을 딛고, 총기와 배낭은 손 또는 끈으로 지지된다. 캠핑카 바퀴는 진흙에 박혀 있으며 표지판은 기둥 위에 고정되어 있다. 떠 있는 인체나 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "전신과 한 손의 차체 접촉, 약 3분의 1을 차지하는 캠핑카와 늪의 여백은 더 충실하지만, 금지된 추가 인물 네 명 때문에 실격입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "모자와 복장은 참조에 더 가깝지만, 추가 인물 네 명과 읽히는 글자·손전등을 넣었고 캠핑카도 요구보다 크게 배치했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박철진은 왼쪽의 캠핑카 창과 차체를 보며 한 손바닥을 창 가장자리의 금속 부분에 댄다. 뒤의 네 사람은 대체로 차량과 박철진 쪽을 향한다. 앞쪽 추가 남성의 소총 총구는 오른쪽 아래 지면을 향하고, 뒤쪽 손전등 세 개는 카메라 쪽으로 밝은 점을 드러낸다. 차량 측면은 왼쪽 전경에서 화면 중앙 뒤쪽으로 물러난다.",
        "built_space": "캠핑카 한 대가 왼쪽 화면의 거의 절반을 차지하며 위쪽과 왼쪽이 잘렸다. 큰 후면 창 하나, 측면 창 두 개와 측면 문 하나가 보인다. 박철진은 차량 바로 옆 진흙 지면에 서 있고 추가 인물 네 명은 오른쪽 뒤에 있다. 두 기둥으로 받친 방사능 표지판 하나가 중앙 뒤에 있다. 갈대와 산 능선은 참조와 유사하지만 참조의 포장 진입로와 물이 넓게 드러난 배치는 확인되지 않는다.",
        "entities": "박철진에 해당하는 중년 남성 한 명은 짧은 회색 섞인 수염, 녹색 모자와 군복, 검은 완장 및 전투화를 착용해 인물 참조에 비교적 가깝다. 그러나 허용되지 않은 사람 네 명과 그들의 총기·손전등도 보인다. 녹슨 폐캠핑카, 늪 식생, 달, 방사능 표지판은 있다. 별도의 출입금지 표지는 식별되지 않는다. 완장의 '헌병'과 'MP'가 읽혀 글자 금지 조건에 어긋난다.",
        "hard_violations": [
         "박철진만 등장해야 하는 장면에 추가 인물 네 명을 배치했다.",
         "허용되지 않은 추가 인물들에게 총기와 켜진 손전등을 부여했다.",
         "완장에 읽을 수 있는 '헌병'과 'MP' 글자가 노출되어 있다."
        ],
        "physics": "박철진의 양발은 진흙 지면에 닿고 손바닥은 차량에 접촉한다. 추가 인물들도 지면 위에 서 있으며 총기와 손전등은 손이나 멜빵으로 지지된다. 차량 하부는 진흙과 식생에 잠겨 있고 표지판은 기둥으로 지지된다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "박철진은 오른쪽의 녹슨 차체와 손을 댄 지점을 내려다본다. 한 손은 창 아래쪽 차체에 닿고 다른 손은 몸 옆에 내려와 있다. 뒤에서 웅크린 남성은 차량 하부를 살피고, 중앙의 모자 쓴 남성은 오른쪽 차량 방향을 가리킨다. 왼쪽 인물의 등에 멘 총은 총구가 왼쪽 위로 향하며 중앙 인물의 총은 아래쪽으로 향한다. 차량 측면은 오른쪽 전경에서 중앙 뒤쪽으로 물러난다.",
        "built_space": "캠핑카 한 대가 오른쪽 약 3분의 1을 차지하고, 창 여러 개와 문 하나, 측면 거울 하나, 진흙에 잠긴 바퀴 두 개가 보인다. 박철진은 차량 옆에 전신으로 서 있다. 추가 인물 한 명은 차량 하부 옆에 웅크리고 세 명은 뒤쪽 늪에 서 있다. 왼쪽에는 두 기둥으로 받친 방사능 표지판 하나가 있다. 물웅덩이와 갈대, 산 능선이 참조의 늪 환경에 가깝지만 포장 진입로는 보이지 않아 정확한 장소 배치는 확인하기 어렵다.",
        "entities": "주인공은 회색 섞인 짧은 머리와 수염을 가진 중년 남성이며 녹색 군복, 검은 완장, 전투화를 착용한다. 참조와 대체로 비슷한 인상이지만 참조의 모자가 없다. 허용되지 않은 추가 남성 네 명과 총기·배낭이 보인다. 녹슨 캠핑카와 진흙 늪, 달, 부식된 방사능 표지판은 확인된다. 출입금지 표시는 식별되지 않으며 명확히 읽히는 문구도 보이지 않는다.",
        "hard_violations": [
         "박철진만 등장해야 하는 장면에 추가 인물 네 명을 배치했다.",
         "허용되지 않은 추가 인물들에게 총기와 배낭을 부여했다."
        ],
        "physics": "박철진의 두 부츠는 진흙에 닿고 뻗은 손은 차체를 짚어 자연스러운 지지 관계를 이룬다. 웅크린 남성은 굽힌 다리와 지면에 닿은 부츠로 몸을 지탱한다. 나머지 인물도 발로 지면을 딛고, 총기와 배낭은 손 또는 끈으로 지지된다. 캠핑카 바퀴는 진흙에 박혀 있으며 표지판은 기둥 위에 고정되어 있다. 떠 있는 인체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.417
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.167
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people",
     "[gpt-high] 박철진만 등장해야 하는 장면에 추가 인물 네 명을 배치했다.",
     "[gpt-high] 허용되지 않은 추가 인물들에게 총기와 배낭을 부여했다."
    ],
    "B": [
     "[gemini-pro] invented people",
     "[gemini-pro] leaked text",
     "[gpt-high] 박철진만 등장해야 하는 장면에 추가 인물 네 명을 배치했다.",
     "[gpt-high] 허용되지 않은 추가 인물들에게 총기와 켜진 손전등을 부여했다.",
     "[gpt-high] 완장에 읽을 수 있는 '헌병'과 'MP' 글자가 노출되어 있다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1167
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지침의 모순으로 인해 샷 텍스트에 없는 인물들이 배경에 등장하는 치명적 오류가 있으나, 캠핑카의 비스듬한 구도와 텍스트 비노출 규칙을 잘 지켰습니다.  ★위반: [gemini-pro] invented people / [gpt-high] 박철진만 등장해야 하는 장면에 추가 인물 네 명을 배치했다. / [gpt-high] 허용되지 않은 추가 인물들에게 총기와 배낭을 부여했다."
   },
   {
    "label": "B",
    "score": 1167,
    "verdict_ko": "텍스트에 없는 추가 인물이 등장하고 완장에 읽을 수 있는 텍스트('헌병 MP')가 노출되는 등 다수의 치명적 오류가 발생했습니다.  ★위반: [gemini-pro] invented people / [gemini-pro] leaked text / [gpt-high] 박철진만 등장해야 하는 장면에 추가 인물 네 명을 배치했다. / [gpt-high] 허용되지 않은 추가 인물들에게 총기와 켜진 손전등을 부여했다. / [gpt-high] 완장에 읽을 수 있는 '헌병'과 'MP' 글자가 노출되어 있다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_swamp_mud_clearing_sel.png",
    "asset_id": "21b3c525-8f13-484a-9366-828766ce2823",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdf3-583b-756b-b810-4d784556af95",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S61sh1__bgfirst_bg.png",
   "bg_asset_id": "03734688-7bcc-44a6-9ca9-c40f40385a32",
   "bg_record_key": "S61sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S61sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:50:29.601508+00:00",
  "fingerprint": "8633537fab84daf78b5ceb62343d3face6398145832cb38aa69caf6ad12a42ca",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S61sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S61sh1_sel.png",
  "source_sha256": "44c940fb3115a3f1a33c91a1995602f0bc89189c597741bd04a66584525efe98",
  "file": "S61sh1_cine.png",
  "staged_sha256": "8d96ef0085a53b36139cc53b6dac3ecd36b6e4cb1ddeaf2534326001569a91cc",
  "latency_ms": 22596
 },
 "S61sh3::signage": {
  "fp": "2e7a79a61dfccff3",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "방사능 구역 팻말"
   }
  ],
  "dropped": []
 },
 "S61sh3": {
  "input_fingerprint": "0ea05dc7fe42997a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 박철진이 바닥의 녹슨 방사능 구역 팻말을 내려다보며 한쪽 입꼬리를 비스듬히 올린 비웃는 얼굴.\n\nLOCATION (lock): At the edge of the marsh clearing near the stranded camper and a corroded radiation warning sign. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 자동차 차체 (The inspected vehicle remains beside 박철진) — A small oblique portion of its side remains behind his shoulder; used as A soft background fragment links the facial close-up to the established vehicle position without competing with the smile.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding shot's subdued nighttime ambience, allowing the asymmetrical smile and lowered eyes to read without introducing a new source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the abandoned camper, muddy ground, surrounding marsh, and nighttime lighting. EXCLUDE fresh tire spray or airborne mud; the vehicle is stationary.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains abandoned and stuck in swamp mud beside the no-entry and corroded radiation-warning signs. Crows gather around the warning sign before taking flight into the moonlit sky. 박철진: He remains at the vehicle inspection site.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 박철진이 바닥의 녹슨 방사능 구역 팻말을 내려다보며 한쪽 입꼬리를 비스듬히 올린 비웃는 얼굴.\n\nLOCATION (lock): At the edge of the marsh clearing near the stranded camper and a corroded radiation warning sign. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 자동차 차체 (The inspected vehicle remains beside 박철진) — A small oblique portion of its side remains behind his shoulder; used as A soft background fragment links the facial close-up to the established vehicle position without competing with the smile.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding shot's subdued nighttime ambience, allowing the asymmetrical smile and lowered eyes to read without introducing a new source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the abandoned camper, muddy ground, surrounding marsh, and nighttime lighting. EXCLUDE fresh tire spray or airborne mud; the vehicle is stationary.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains abandoned and stuck in swamp mud beside the no-entry and corroded radiation-warning signs. Crows gather around the warning sign before taking flight into the moonlit sky. 박철진: He remains at the vehicle inspection site.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 박철진이 바닥의 녹슨 방사능 구역 팻말을 내려다보며 한쪽 입꼬리를 비스듬히 올린 비웃는 얼굴.\n\nLOCATION (lock): At the edge of the marsh clearing near the stranded camper and a corroded radiation warning sign. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 자동차 차체 (The inspected vehicle remains beside 박철진) — A small oblique portion of its side remains behind his shoulder; used as A soft background fragment links the facial close-up to the established vehicle position without competing with the smile.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding shot's subdued nighttime ambience, allowing the asymmetrical smile and lowered eyes to read without introducing a new source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the abandoned camper, muddy ground, surrounding marsh, and nighttime lighting. EXCLUDE fresh tire spray or airborne mud; the vehicle is stationary.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains abandoned and stuck in swamp mud beside the no-entry and corroded radiation-warning signs. Crows gather around the warning sign before taking flight into the moonlit sky. 박철진: He remains at the vehicle inspection site.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 좌측 전경의 방사능 팻말을 향해 아래로 향함.",
    "built_space": "카메라는 팻말 근처에서 인물과 배경의 캠핑카를 바라보며 공간 배치를 유지함(단, 팻말 앞면이 카메라를 향해 방향이 뒤집힘).",
    "entities": "인물의 얼굴과 복장은 참조와 일치하나, 완장에 참조에 없는 방패 마크가 생성됨.",
    "hard_violations": [
     "[gpt-high] 오른쪽 아래 완장에 판독 가능한 한글이 남아 있어, 읽을 수 있는 글자를 전부 배제하라는 금지 조건을 위반한다."
    ],
    "physics": "인물은 보이지 않는 지면에 서 있고, 팻말은 화면 밖 하단에 고정되어 수직으로 세워져 있음."
   },
   {
    "label": "B",
    "direction": "시선은 바닥에 눕혀진 팻말을 향해 아래로 향함.",
    "built_space": "팻말이 인물 바로 앞 바닥에 눕혀져 있어, 멀리 세워져 있던 참조 이미지의 고정 요소 배치 및 거리감과 어긋남.",
    "entities": "인물의 얼굴, 복장, 그리고 완장(헌병 MP) 텍스트까지 참조와 정확히 일치함.",
    "hard_violations": [
     "[gpt-high] 완장의 ‘헌병’과 영문 약자가 선명하게 읽혀, 이미지 어디에도 판독 가능한 글자를 두지 말라는 금지 조건을 위반한다."
    ],
    "physics": "인물은 바닥에 서 있고, 팻말은 진흙 바닥 위에 평평하게 놓여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "텍스트 판독 불가 지침을 준수하고 세트의 공간 배치를 잘 유지했으나, 완장 마크가 참조와 다르게 변형됨."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "완장 디자인은 정확하나, 명확히 읽을 수 있는 텍스트 금지 지침을 위반했으며 고정된 팻말을 바닥에 눕혀 공간 연속성을 훼손함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 좌측 전경의 방사능 팻말을 향해 아래로 향함.",
        "built_space": "카메라는 팻말 근처에서 인물과 배경의 캠핑카를 바라보며 공간 배치를 유지함(단, 팻말 앞면이 카메라를 향해 방향이 뒤집힘).",
        "entities": "인물의 얼굴과 복장은 참조와 일치하나, 완장에 참조에 없는 방패 마크가 생성됨.",
        "hard_violations": [],
        "physics": "인물은 보이지 않는 지면에 서 있고, 팻말은 화면 밖 하단에 고정되어 수직으로 세워져 있음."
       },
       {
        "label": "B",
        "direction": "시선은 바닥에 눕혀진 팻말을 향해 아래로 향함.",
        "built_space": "팻말이 인물 바로 앞 바닥에 눕혀져 있어, 멀리 세워져 있던 참조 이미지의 고정 요소 배치 및 거리감과 어긋남.",
        "entities": "인물의 얼굴, 복장, 그리고 완장(헌병 MP) 텍스트까지 참조와 정확히 일치함.",
        "hard_violations": [],
        "physics": "인물은 바닥에 서 있고, 팻말은 진흙 바닥 위에 평평하게 놓여 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "텍스트 판독 불가 지침을 준수하고 세트의 공간 배치를 잘 유지했으나, 완장 마크가 참조와 다르게 변형됨."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "완장 디자인은 정확하나, 명확히 읽을 수 있는 텍스트 금지 지침을 위반했으며 고정된 팻말을 바닥에 눕혀 공간 연속성을 훼손함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 좌측 전경의 방사능 팻말을 향해 아래로 향함.",
        "built_space": "카메라는 팻말 근처에서 인물과 배경의 캠핑카를 바라보며 공간 배치를 유지함(단, 팻말 앞면이 카메라를 향해 방향이 뒤집힘).",
        "entities": "인물의 얼굴과 복장은 참조와 일치하나, 완장에 참조에 없는 방패 마크가 생성됨.",
        "hard_violations": [],
        "physics": "인물은 보이지 않는 지면에 서 있고, 팻말은 화면 밖 하단에 고정되어 수직으로 세워져 있음."
       },
       {
        "label": "B",
        "direction": "시선은 바닥에 눕혀진 팻말을 향해 아래로 향함.",
        "built_space": "팻말이 인물 바로 앞 바닥에 눕혀져 있어, 멀리 세워져 있던 참조 이미지의 고정 요소 배치 및 거리감과 어긋남.",
        "entities": "인물의 얼굴, 복장, 그리고 완장(헌병 MP) 텍스트까지 참조와 정확히 일치함.",
        "hard_violations": [],
        "physics": "인물은 바닥에 서 있고, 팻말은 진흙 바닥 위에 평평하게 놓여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "바닥의 녹슨 팻말을 내려다보는 시선과 비대칭 비웃음, 어두운 밤의 연속성은 더 정확하지만, 완장에 읽을 수 있는 글자가 있어 실격이다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "팻말을 바닥에 눕히지 않고 얼굴 옆에 크게 세워 핵심 행동과 얼굴 중심 구도를 약화했으며, 완장에 판독 가능한 글자도 남아 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개를 숙이고 눈을 화면 왼쪽 아래로 내려 바닥에 누운 방사능 팻말 쪽을 본다. 카메라를 바라보지 않으며 한쪽 입꼬리를 올린 비웃음이 보인다. 무기나 이동하는 물체는 없다.",
        "built_space": "오른쪽 어깨 뒤에 녹슨 캠핑카 한 대의 측면 일부와 창틀, 돌출된 거울이 보인다. 왼쪽 아래에는 경고판 한 장이 진흙과 갈대 위에 비스듬히 누워 있다. 뒤쪽 습지와 산줄기는 이전 장소와 연결되며, 중복 차량이나 불가능한 반사는 없다. 다만 차체가 요구된 부드러운 배경 조각보다 선명하고, 팻말과 풍경의 비중도 크다.",
        "entities": "중년 남성 한 명만 보이며, 짧은 회색 섞인 머리와 수염, 올리브색 제복과 검은 완장은 참고의 전반적 외형을 따른다. 얼굴의 눈매와 윤곽은 참고 인물과 완전히 같지는 않다. 녹슨 금속판에는 방사능 표식이 있고 달도 보인다. 완장의 ‘헌병’과 영문 약자는 읽을 수 있다. 출입금지 표지와 까마귀는 이 얼굴 중심 화면에 보이지 않는다.",
        "hard_violations": [
         "완장의 ‘헌병’과 영문 약자가 선명하게 읽혀, 이미지 어디에도 판독 가능한 글자를 두지 말라는 금지 조건을 위반한다."
        ],
        "physics": "남성의 머리와 상체는 자연스럽게 연결되어 앞으로 기울어 있으며, 하체와 발은 화면 밖이라 접지 상태는 확인할 수 없다. 팻말은 진흙과 식생에 받쳐져 있어 공중에 뜨지 않는다. 손은 보이지 않으므로 물체와의 손 접촉은 확인되지 않는다. 비산하는 진흙이나 비현실적인 신체 부유는 없다."
       },
       {
        "label": "B",
        "direction": "눈은 화면 왼쪽 아래의 세워진 경고판 쪽을 향하지만, 바닥에 놓인 팻말을 내려다보는 구체적인 행동은 아니다. 고개 숙임과 한쪽 입꼬리의 미소는 보이나 A보다 약하다. 무기나 이동하는 물체는 없다.",
        "built_space": "오른쪽 어깨 뒤로 캠핑카 한 대의 녹슨 측면과 창문, 상부 난간이 흐리게 보인다. 경고판 한 장은 화면 왼쪽 전경에 수직으로 서서 얼굴과 경쟁할 만큼 크게 차지한다. 습지와 산 배경은 이어지지만, 바닥의 팻말이라는 배치와 작은 차체 배경 조각이라는 요구에서 벗어난다. 중복 구조물이나 불가능한 반사는 보이지 않는다.",
        "entities": "중년 남성 한 명이며 회색 섞인 짧은 머리, 수염, 올리브색 제복과 검은 완장을 갖췄다. 참고의 전반적 외형은 따르지만 얼굴 윤곽과 눈매의 정확한 일치는 부족하다. 녹슨 방사능 경고판과 달이 보인다. 오른쪽 아래 완장에는 잘렸지만 판독 가능한 한글 ‘헌’이 남아 있다. 이전 장면보다 하늘과 얼굴의 조도가 높다. 출입금지 표지와 까마귀는 보이지 않는다.",
        "hard_violations": [
         "오른쪽 아래 완장에 판독 가능한 한글이 남아 있어, 읽을 수 있는 글자를 전부 배제하라는 금지 조건을 위반한다."
        ],
        "physics": "남성의 머리와 상체 자세는 해부학적으로 가능하며 발은 화면 밖이다. 수직 경고판은 하단이 화면 밖으로 이어져 받침이나 지면 접촉을 확인할 수 없지만, 공중에 떠 있다고 단정할 근거도 없다. 손과 팻말의 접촉은 보이지 않는다. 차량은 정지해 있고 공중의 진흙이나 지지 없는 신체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "바닥의 녹슨 팻말을 내려다보는 시선과 비대칭 비웃음, 어두운 밤의 연속성은 더 정확하지만, 완장에 읽을 수 있는 글자가 있어 실격이다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "팻말을 바닥에 눕히지 않고 얼굴 옆에 크게 세워 핵심 행동과 얼굴 중심 구도를 약화했으며, 완장에 판독 가능한 글자도 남아 있다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "고개를 숙이고 눈을 화면 왼쪽 아래로 내려 바닥에 누운 방사능 팻말 쪽을 본다. 카메라를 바라보지 않으며 한쪽 입꼬리를 올린 비웃음이 보인다. 무기나 이동하는 물체는 없다.",
        "built_space": "오른쪽 어깨 뒤에 녹슨 캠핑카 한 대의 측면 일부와 창틀, 돌출된 거울이 보인다. 왼쪽 아래에는 경고판 한 장이 진흙과 갈대 위에 비스듬히 누워 있다. 뒤쪽 습지와 산줄기는 이전 장소와 연결되며, 중복 차량이나 불가능한 반사는 없다. 다만 차체가 요구된 부드러운 배경 조각보다 선명하고, 팻말과 풍경의 비중도 크다.",
        "entities": "중년 남성 한 명만 보이며, 짧은 회색 섞인 머리와 수염, 올리브색 제복과 검은 완장은 참고의 전반적 외형을 따른다. 얼굴의 눈매와 윤곽은 참고 인물과 완전히 같지는 않다. 녹슨 금속판에는 방사능 표식이 있고 달도 보인다. 완장의 ‘헌병’과 영문 약자는 읽을 수 있다. 출입금지 표지와 까마귀는 이 얼굴 중심 화면에 보이지 않는다.",
        "hard_violations": [
         "완장의 ‘헌병’과 영문 약자가 선명하게 읽혀, 이미지 어디에도 판독 가능한 글자를 두지 말라는 금지 조건을 위반한다."
        ],
        "physics": "남성의 머리와 상체는 자연스럽게 연결되어 앞으로 기울어 있으며, 하체와 발은 화면 밖이라 접지 상태는 확인할 수 없다. 팻말은 진흙과 식생에 받쳐져 있어 공중에 뜨지 않는다. 손은 보이지 않으므로 물체와의 손 접촉은 확인되지 않는다. 비산하는 진흙이나 비현실적인 신체 부유는 없다."
       },
       {
        "label": "A",
        "direction": "눈은 화면 왼쪽 아래의 세워진 경고판 쪽을 향하지만, 바닥에 놓인 팻말을 내려다보는 구체적인 행동은 아니다. 고개 숙임과 한쪽 입꼬리의 미소는 보이나 A보다 약하다. 무기나 이동하는 물체는 없다.",
        "built_space": "오른쪽 어깨 뒤로 캠핑카 한 대의 녹슨 측면과 창문, 상부 난간이 흐리게 보인다. 경고판 한 장은 화면 왼쪽 전경에 수직으로 서서 얼굴과 경쟁할 만큼 크게 차지한다. 습지와 산 배경은 이어지지만, 바닥의 팻말이라는 배치와 작은 차체 배경 조각이라는 요구에서 벗어난다. 중복 구조물이나 불가능한 반사는 보이지 않는다.",
        "entities": "중년 남성 한 명이며 회색 섞인 짧은 머리, 수염, 올리브색 제복과 검은 완장을 갖췄다. 참고의 전반적 외형은 따르지만 얼굴 윤곽과 눈매의 정확한 일치는 부족하다. 녹슨 방사능 경고판과 달이 보인다. 오른쪽 아래 완장에는 잘렸지만 판독 가능한 한글 ‘헌’이 남아 있다. 이전 장면보다 하늘과 얼굴의 조도가 높다. 출입금지 표지와 까마귀는 보이지 않는다.",
        "hard_violations": [
         "오른쪽 아래 완장에 판독 가능한 한글이 남아 있어, 읽을 수 있는 글자를 전부 배제하라는 금지 조건을 위반한다."
        ],
        "physics": "남성의 머리와 상체 자세는 해부학적으로 가능하며 발은 화면 밖이다. 수직 경고판은 하단이 화면 밖으로 이어져 받침이나 지면 접촉을 확인할 수 없지만, 공중에 떠 있다고 단정할 근거도 없다. 손과 팻말의 접촉은 보이지 않는다. 차량은 정지해 있고 공중의 진흙이나 지지 없는 신체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.464
   },
   "violations": {
    "B": [
     "[gpt-high] 완장의 ‘헌병’과 영문 약자가 선명하게 읽혀, 이미지 어디에도 판독 가능한 글자를 두지 말라는 금지 조건을 위반한다."
    ],
    "A": [
     "[gpt-high] 오른쪽 아래 완장에 판독 가능한 한글이 남아 있어, 읽을 수 있는 글자를 전부 배제하라는 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1417,
   "B": 1464
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "텍스트 판독 불가 지침을 준수하고 세트의 공간 배치를 잘 유지했으나, 완장 마크가 참조와 다르게 변형됨.  ★위반: [gpt-high] 오른쪽 아래 완장에 판독 가능한 한글이 남아 있어, 읽을 수 있는 글자를 전부 배제하라는 금지 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1464,
    "verdict_ko": "완장 디자인은 정확하나, 명확히 읽을 수 있는 텍스트 금지 지침을 위반했으며 고정된 팻말을 바닥에 눕혀 공간 연속성을 훼손함.  ★위반: [gpt-high] 완장의 ‘헌병’과 영문 약자가 선명하게 읽혀, 이미지 어디에도 판독 가능한 글자를 두지 말라는 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S61sh1_sel.png",
    "asset_id": "62a4aa48-54c7-4e4b-9d71-57fca2dcbff8",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdf4-c8d8-74e3-8bde-c7fd5ee2b440",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S61sh1"
  }
 },
 "S61sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:50:52.563044+00:00",
  "fingerprint": "d2e8f90d8b230788f9538a88f3337a875e0d2b53eb5d14c870ea010ee2a69654",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S61sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S61sh3_sel.png",
  "source_sha256": "21a507c09fda7681be041f47b4823ce137ac07ebb6bdcecd7b0e75949552149c",
  "file": "S61sh3_cine.png",
  "staged_sha256": "c853f14e4c360a4fba304a8ce68ea4056f6cbe6c15541e6f4aa738f7b7ef25ed",
  "latency_ms": 12269
 },
 "S62sh2::signage": {
  "fp": "84d590b304195d61",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S62sh2": {
  "input_fingerprint": "7c9e1920edaecd35",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 철문 너머로 쌀 포대와 감자 포대 몇 개만이 덩그러니 놓인 텅 빈 식량창고 바닥 구도.\n\nLOCATION (lock): Inside the village food store just beyond its opened iron door, with daylight falling on a nearly empty floor and a few sacks. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Storehouse iron door (Old and open) — Seen obliquely along one foreground edge, leaving the interior unobstructed; used as Frames the threshold without obscuring the remaining food; Remaining food sacks (A few rice sacks and one potato sack remain on the floor) — Grouped beyond the doorway in an oblique view; used as Small focal cluster establishing the shortage; no individual sack exceeds a small portion of the frame; Storehouse floor (Otherwise empty); used as Broad negative space makes the remaining food appear insufficient.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime light and restrained contrast give the empty interior a quiet, depleted feeling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The food store's old iron door stands open, revealing only a few sacks of rice and one sack of potatoes. Elsewhere in the village courtyard, an old truck awaits repair; Charlie's dented, perforated body remains unrepaired.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 철문 너머로 쌀 포대와 감자 포대 몇 개만이 덩그러니 놓인 텅 빈 식량창고 바닥 구도.\n\nLOCATION (lock): Inside the village food store just beyond its opened iron door, with daylight falling on a nearly empty floor and a few sacks. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Storehouse iron door (Old and open) — Seen obliquely along one foreground edge, leaving the interior unobstructed; used as Frames the threshold without obscuring the remaining food; Remaining food sacks (A few rice sacks and one potato sack remain on the floor) — Grouped beyond the doorway in an oblique view; used as Small focal cluster establishing the shortage; no individual sack exceeds a small portion of the frame; Storehouse floor (Otherwise empty); used as Broad negative space makes the remaining food appear insufficient.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime light and restrained contrast give the empty interior a quiet, depleted feeling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The food store's old iron door stands open, revealing only a few sacks of rice and one sack of potatoes. Elsewhere in the village courtyard, an old truck awaits repair; Charlie's dented, perforated body remains unrepaired.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 철문 너머로 쌀 포대와 감자 포대 몇 개만이 덩그러니 놓인 텅 빈 식량창고 바닥 구도.\n\nLOCATION (lock): Inside the village food store just beyond its opened iron door, with daylight falling on a nearly empty floor and a few sacks. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Storehouse iron door (Old and open) — Seen obliquely along one foreground edge, leaving the interior unobstructed; used as Frames the threshold without obscuring the remaining food; Remaining food sacks (A few rice sacks and one potato sack remain on the floor) — Grouped beyond the doorway in an oblique view; used as Small focal cluster establishing the shortage; no individual sack exceeds a small portion of the frame; Storehouse floor (Otherwise empty); used as Broad negative space makes the remaining food appear insufficient.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime light and restrained contrast give the empty interior a quiet, depleted feeling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The food store's old iron door stands open, revealing only a few sacks of rice and one sack of potatoes. Elsewhere in the village courtyard, an old truck awaits repair; Charlie's dented, perforated body remains unrepaired.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 좌측의 열린 철문을 지나 창고 중앙 바닥에 놓인 포대들을 향해 시선을 둠.",
    "built_space": "콘크리트 바닥과 블록 벽면으로 이루어진 창고 내부. 좌측 전경에 비스듬히 열린 낡은 철문 1개가 공간의 입구를 형성함.",
    "entities": "바닥 중앙에 쌀 포대들과 감자 그림이 그려진 포대 1개가 지시된 대로 모여 있음.",
    "hard_violations": [
     "[gemini-pro] 포대 전면과 우측 벽면에 명확히 읽을 수 있는 텍스트(한글)가 존재함 (읽을 수 있는 글자 금지 지시 위반)",
     "[gpt-high] 포대의 '20kg'와 한글 인쇄가 읽을 수 있게 노출되어, 무문자 조건을 위반한다."
    ],
    "physics": "열린 철문과 바닥에 쌓인 포대들 모두 중력에 맞게 안정적으로 지지되어 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 철창 너머 어두운 창고 안쪽 벽면을 향하고 있으며, 그곳에 놓인 포대들을 비춤.",
    "built_space": "흙바닥으로 된 창고 내부. 좌측에는 수직 철창이 위치하고, 우측 상단에는 허공을 사선으로 가로지르는 불규칙한 철창 구조물이 있음.",
    "entities": "안쪽 벽에 4개의 포대가 모여 있으나, 감자와 쌀 포대의 구분이 명확하지 않음.",
    "hard_violations": [
     "[gemini-pro] 우측 상단을 가로지르는 지지대 없는 사선형 철창이 물리적으로 불가능한 건축 구조를 형성함",
     "[gpt-high] 포대에 '쌀' 등 읽을 수 있는 인쇄 문자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
    ],
    "physics": "포대는 바닥에 놓여 있으나, 우측의 사선형 철창은 정상적인 문이나 천장 구조로 볼 수 없어 물리적으로 어긋남."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도와 조명 등 전반적인 연출은 프롬프트를 훌륭히 따랐으나, 화면 내에서 읽을 수 있는 텍스트를 엄격히 금지한 지시를 어기고 포대에 뚜렷한 한글이 렌더링되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "지정된 낡은 철문 대신 창살 형태의 문이 나타났으며, 우측 상단에 물리적으로 불가능한 사선형 철창 구조물이 생성되어 사용할 수 없는 결과물입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 좌측의 열린 철문을 지나 창고 중앙 바닥에 놓인 포대들을 향해 시선을 둠.",
        "built_space": "콘크리트 바닥과 블록 벽면으로 이루어진 창고 내부. 좌측 전경에 비스듬히 열린 낡은 철문 1개가 공간의 입구를 형성함.",
        "entities": "바닥 중앙에 쌀 포대들과 감자 그림이 그려진 포대 1개가 지시된 대로 모여 있음.",
        "hard_violations": [
         "포대 전면과 우측 벽면에 명확히 읽을 수 있는 텍스트(한글)가 존재함 (읽을 수 있는 글자 금지 지시 위반)"
        ],
        "physics": "열린 철문과 바닥에 쌓인 포대들 모두 중력에 맞게 안정적으로 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 철창 너머 어두운 창고 안쪽 벽면을 향하고 있으며, 그곳에 놓인 포대들을 비춤.",
        "built_space": "흙바닥으로 된 창고 내부. 좌측에는 수직 철창이 위치하고, 우측 상단에는 허공을 사선으로 가로지르는 불규칙한 철창 구조물이 있음.",
        "entities": "안쪽 벽에 4개의 포대가 모여 있으나, 감자와 쌀 포대의 구분이 명확하지 않음.",
        "hard_violations": [
         "우측 상단을 가로지르는 지지대 없는 사선형 철창이 물리적으로 불가능한 건축 구조를 형성함"
        ],
        "physics": "포대는 바닥에 놓여 있으나, 우측의 사선형 철창은 정상적인 문이나 천장 구조로 볼 수 없어 물리적으로 어긋남."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도와 조명 등 전반적인 연출은 프롬프트를 훌륭히 따랐으나, 화면 내에서 읽을 수 있는 텍스트를 엄격히 금지한 지시를 어기고 포대에 뚜렷한 한글이 렌더링되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "지정된 낡은 철문 대신 창살 형태의 문이 나타났으며, 우측 상단에 물리적으로 불가능한 사선형 철창 구조물이 생성되어 사용할 수 없는 결과물입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 좌측의 열린 철문을 지나 창고 중앙 바닥에 놓인 포대들을 향해 시선을 둠.",
        "built_space": "콘크리트 바닥과 블록 벽면으로 이루어진 창고 내부. 좌측 전경에 비스듬히 열린 낡은 철문 1개가 공간의 입구를 형성함.",
        "entities": "바닥 중앙에 쌀 포대들과 감자 그림이 그려진 포대 1개가 지시된 대로 모여 있음.",
        "hard_violations": [
         "포대 전면과 우측 벽면에 명확히 읽을 수 있는 텍스트(한글)가 존재함 (읽을 수 있는 글자 금지 지시 위반)"
        ],
        "physics": "열린 철문과 바닥에 쌓인 포대들 모두 중력에 맞게 안정적으로 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 철창 너머 어두운 창고 안쪽 벽면을 향하고 있으며, 그곳에 놓인 포대들을 비춤.",
        "built_space": "흙바닥으로 된 창고 내부. 좌측에는 수직 철창이 위치하고, 우측 상단에는 허공을 사선으로 가로지르는 불규칙한 철창 구조물이 있음.",
        "entities": "안쪽 벽에 4개의 포대가 모여 있으나, 감자와 쌀 포대의 구분이 명확하지 않음.",
        "hard_violations": [
         "우측 상단을 가로지르는 지지대 없는 사선형 철창이 물리적으로 불가능한 건축 구조를 형성함"
        ],
        "physics": "포대는 바닥에 놓여 있으나, 우측의 사선형 철창은 정상적인 문이나 천장 구조로 볼 수 없어 물리적으로 어긋남."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "작은 포대 무리와 넓은 빈 바닥은 맞지만, 철문이 상부를 대각선으로 크게 가로지르고 포대의 읽을 수 있는 글자가 무문자 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "열린 철문을 왼쪽 가장자리에 두고 빈 창고와 포대를 드러내는 구도가 더 정확하지만, 포대의 선명한 문자와 숫자 때문에 최종 사용 조건에는 미달한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "문턱 밖에서 안쪽 바닥과 뒤편 포대 무리를 비스듬히 내려다본다. 왼쪽 문짝은 가장자리를 따라 열려 있으나 오른쪽 격자 문짝은 화면 위쪽을 대각선으로 크게 가로지른다. 사람의 시선이나 이동 방향은 없다.",
        "built_space": "전경에 문턱 하나, 왼쪽의 좁게 보이는 금속 문짝 하나, 오른쪽의 큰 격자 문짝 하나가 보인다. 안쪽에는 거친 벽과 넓은 빈 바닥이 있다. 참고 사진의 낡은 미장과 금속 격자 재질에는 일부 대응하지만, 정확히 같은 출입구인지는 확인하기 어렵다. 문을 한쪽 가장자리에서만 프레임으로 쓰라는 지시와 달리 큰 문짝이 상부 공간을 차지한다.",
        "entities": "뒤편에 큰 포대 세 개와 작은 묶인 포대 한 개가 보이며 사람은 없다. 큰 포대에는 쌀을 나타내는 글자가 보이지만 내용물은 직접 보이지 않는다. 작은 포대가 감자 포대인지는 외형만으로 확정하기 어렵다. 포대의 '쌀' 등 식별 가능한 인쇄 문자가 무문자 지시와 충돌한다. 다른 장소의 트럭이나 인물을 넣지 않은 것은 맞다.",
        "hard_violations": [
         "포대에 '쌀' 등 읽을 수 있는 인쇄 문자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
        ],
        "physics": "네 포대 모두 밑면이 바닥에 닿아 서 있고 서로 기대어 있어 지지가 자연스럽다. 문짝은 출입구 양쪽으로 이어지며, 오른쪽 문짝의 정확한 경첩은 화면에서 확인되지 않는다. 대각선 문짝이 어색하게 보이지만 그것만으로 공중 부유나 물리적 불가능이라고 단정할 근거는 없다."
       },
       {
        "label": "B",
        "direction": "열린 문턱 너머로 창고 안을 비스듬히 바라본다. 철문은 왼쪽 전경으로 열려 있고 중앙의 포대 무리를 가리지 않는다. 사람이나 움직이는 물체는 없다.",
        "built_space": "왼쪽에 열린 철문 한 짝과 손잡이 하나, 아래에 문턱 하나가 보인다. 내부에는 뒤벽, 오른쪽 블록 벽과 수직 구조부, 오른쪽 위의 작은 개구부가 있으며 바닥 대부분은 비어 있다. 참고 사진의 낡은 미장·블록 재질은 이어지지만, 보이는 철문은 참고 사진의 격자 출입문과 다른 판형이다. 문 가장자리로 내부를 unobstructed하게 보여주는 배치는 A보다 지시에 가깝다.",
        "entities": "왼쪽에 눕혀 쌓은 포대 세 개, 뒤쪽의 작은 묶인 포대 하나, 오른쪽에 세운 큰 포대 하나, 앞쪽의 작은 포대 하나로 총 여섯 개가 보인다. 앞 포대에는 감자 모양 인쇄가 있어 감자 포대 한 개라는 해석이 가능하다. 다른 포대의 실제 내용물은 보이지 않는다. 사람이나 트럭은 없다. 오른쪽 큰 포대의 '20kg'와 앞 포대의 한글 등 읽을 수 있는 인쇄가 있다.",
        "hard_violations": [
         "포대의 '20kg'와 한글 인쇄가 읽을 수 있게 노출되어, 무문자 조건을 위반한다."
        ],
        "physics": "눕힌 포대는 바닥과 아래 포대가 받치고, 세워진 포대들은 바닥에 직접 닿아 있다. 천의 눌림과 접힘도 적재 상태에 맞는다. 열린 철문은 왼쪽 문설주의 경첩으로 지지되며, 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "작은 포대 무리와 넓은 빈 바닥은 맞지만, 철문이 상부를 대각선으로 크게 가로지르고 포대의 읽을 수 있는 글자가 무문자 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "열린 철문을 왼쪽 가장자리에 두고 빈 창고와 포대를 드러내는 구도가 더 정확하지만, 포대의 선명한 문자와 숫자 때문에 최종 사용 조건에는 미달한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "문턱 밖에서 안쪽 바닥과 뒤편 포대 무리를 비스듬히 내려다본다. 왼쪽 문짝은 가장자리를 따라 열려 있으나 오른쪽 격자 문짝은 화면 위쪽을 대각선으로 크게 가로지른다. 사람의 시선이나 이동 방향은 없다.",
        "built_space": "전경에 문턱 하나, 왼쪽의 좁게 보이는 금속 문짝 하나, 오른쪽의 큰 격자 문짝 하나가 보인다. 안쪽에는 거친 벽과 넓은 빈 바닥이 있다. 참고 사진의 낡은 미장과 금속 격자 재질에는 일부 대응하지만, 정확히 같은 출입구인지는 확인하기 어렵다. 문을 한쪽 가장자리에서만 프레임으로 쓰라는 지시와 달리 큰 문짝이 상부 공간을 차지한다.",
        "entities": "뒤편에 큰 포대 세 개와 작은 묶인 포대 한 개가 보이며 사람은 없다. 큰 포대에는 쌀을 나타내는 글자가 보이지만 내용물은 직접 보이지 않는다. 작은 포대가 감자 포대인지는 외형만으로 확정하기 어렵다. 포대의 '쌀' 등 식별 가능한 인쇄 문자가 무문자 지시와 충돌한다. 다른 장소의 트럭이나 인물을 넣지 않은 것은 맞다.",
        "hard_violations": [
         "포대에 '쌀' 등 읽을 수 있는 인쇄 문자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
        ],
        "physics": "네 포대 모두 밑면이 바닥에 닿아 서 있고 서로 기대어 있어 지지가 자연스럽다. 문짝은 출입구 양쪽으로 이어지며, 오른쪽 문짝의 정확한 경첩은 화면에서 확인되지 않는다. 대각선 문짝이 어색하게 보이지만 그것만으로 공중 부유나 물리적 불가능이라고 단정할 근거는 없다."
       },
       {
        "label": "A",
        "direction": "열린 문턱 너머로 창고 안을 비스듬히 바라본다. 철문은 왼쪽 전경으로 열려 있고 중앙의 포대 무리를 가리지 않는다. 사람이나 움직이는 물체는 없다.",
        "built_space": "왼쪽에 열린 철문 한 짝과 손잡이 하나, 아래에 문턱 하나가 보인다. 내부에는 뒤벽, 오른쪽 블록 벽과 수직 구조부, 오른쪽 위의 작은 개구부가 있으며 바닥 대부분은 비어 있다. 참고 사진의 낡은 미장·블록 재질은 이어지지만, 보이는 철문은 참고 사진의 격자 출입문과 다른 판형이다. 문 가장자리로 내부를 unobstructed하게 보여주는 배치는 A보다 지시에 가깝다.",
        "entities": "왼쪽에 눕혀 쌓은 포대 세 개, 뒤쪽의 작은 묶인 포대 하나, 오른쪽에 세운 큰 포대 하나, 앞쪽의 작은 포대 하나로 총 여섯 개가 보인다. 앞 포대에는 감자 모양 인쇄가 있어 감자 포대 한 개라는 해석이 가능하다. 다른 포대의 실제 내용물은 보이지 않는다. 사람이나 트럭은 없다. 오른쪽 큰 포대의 '20kg'와 앞 포대의 한글 등 읽을 수 있는 인쇄가 있다.",
        "hard_violations": [
         "포대의 '20kg'와 한글 인쇄가 읽을 수 있게 노출되어, 무문자 조건을 위반한다."
        ],
        "physics": "눕힌 포대는 바닥과 아래 포대가 받치고, 세워진 포대들은 바닥에 직접 닿아 있다. 천의 눌림과 접힘도 적재 상태에 맞는다. 열린 철문은 왼쪽 문설주의 경첩으로 지지되며, 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.167
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.917
   },
   "violations": {
    "A": [
     "[gemini-pro] 포대 전면과 우측 벽면에 명확히 읽을 수 있는 텍스트(한글)가 존재함 (읽을 수 있는 글자 금지 지시 위반)",
     "[gpt-high] 포대의 '20kg'와 한글 인쇄가 읽을 수 있게 노출되어, 무문자 조건을 위반한다."
    ],
    "B": [
     "[gemini-pro] 우측 상단을 가로지르는 지지대 없는 사선형 철창이 물리적으로 불가능한 건축 구조를 형성함",
     "[gpt-high] 포대에 '쌀' 등 읽을 수 있는 인쇄 문자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 917
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "구도와 조명 등 전반적인 연출은 프롬프트를 훌륭히 따랐으나, 화면 내에서 읽을 수 있는 텍스트를 엄격히 금지한 지시를 어기고 포대에 뚜렷한 한글이 렌더링되었습니다.  ★위반: [gemini-pro] 포대 전면과 우측 벽면에 명확히 읽을 수 있는 텍스트(한글)가 존재함 (읽을 수 있는 글자 금지 지시 위반) / [gpt-high] 포대의 '20kg'와 한글 인쇄가 읽을 수 있게 노출되어, 무문자 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 917,
    "verdict_ko": "지정된 낡은 철문 대신 창살 형태의 문이 나타났으며, 우측 상단에 물리적으로 불가능한 사선형 철창 구조물이 생성되어 사용할 수 없는 결과물입니다.  ★위반: [gemini-pro] 우측 상단을 가로지르는 지지대 없는 사선형 철창이 물리적으로 불가능한 건축 구조를 형성함 / [gpt-high] 포대에 '쌀' 등 읽을 수 있는 인쇄 문자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_village_bonfire_clearing_sel.png",
    "asset_id": "476b4788-828b-41ae-bff5-dcf6666051ef",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdf5-92b3-79e2-8668-8b74d18c7024",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S62sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:51:05.178513+00:00",
  "fingerprint": "edf59eba1705c067b465a751b7e960e1e20cd801b3e4f56aa8c2b2d9f1ad1c16",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S62sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S62sh2_sel.png",
  "source_sha256": "10b0bec483da2d05844fa137ee7183a692592767eac4ee6171c9a6f4415b27ec",
  "file": "S62sh2_cine.png",
  "staged_sha256": "74f61cbf43dcc1648af282738806490e85bd87e98e27c95188eb0657696e8f69",
  "latency_ms": 30583
 },
 "S62sh12::signage": {
  "fp": "3bc22e63ca3f05a1",
  "inscriptions": [
   {
    "text_native": "서남쪽 25.6km 백화점/마트",
    "source": "scene_text_quoted",
    "reason_ko": "찰리의 시야 인터페이스 상에 지도 정보로 직접 표시되는 텍스트이다.",
    "source_quote": "서남쪽 25.6km 백화점/마트"
   }
  ],
  "cues": [],
  "dropped": []
 },
 "S62sh12": {
  "input_fingerprint": "bbacc2d1e73268fb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 시야 인터페이스 속, 낡은 지도 이미지 위에 '서남쪽 25.6km 백화점/마트'라는 붉은 글씨가 선명하게 떠오른 주관 시점 찰나.\n\nLOCATION (lock): In the village's open communal lot beside the old truck, with a navigation map overlaid in the robot's view. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: Navigation map image (Displayed as an aged map within the visual interface) — Its content-bearing face is presented directly to the viewer, showing the navigation map; used as Provides geographic context around the destination information without introducing a physical display object; Destination label (Clearly displaying '서남쪽 25.6km 백화점/마트' in red) — The readable text faces the viewer in the image plane; used as Primary information focus for the brief subjective insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled interface contrast separates the sharp red destination text from the subdued map image without simulated exterior illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck remains in the courtyard under repair, with an old football in use nearby; Charlie retains holes, dents and malfunctioning sensors. His navigation display locates the department store and mart 25.6 km to the southwest.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"서남쪽 25.6km 백화점/마트\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 시야 인터페이스 속, 낡은 지도 이미지 위에 '서남쪽 25.6km 백화점/마트'라는 붉은 글씨가 선명하게 떠오른 주관 시점 찰나.\n\nLOCATION (lock): In the village's open communal lot beside the old truck, with a navigation map overlaid in the robot's view. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: Navigation map image (Displayed as an aged map within the visual interface) — Its content-bearing face is presented directly to the viewer, showing the navigation map; used as Provides geographic context around the destination information without introducing a physical display object; Destination label (Clearly displaying '서남쪽 25.6km 백화점/마트' in red) — The readable text faces the viewer in the image plane; used as Primary information focus for the brief subjective insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled interface contrast separates the sharp red destination text from the subdued map image without simulated exterior illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck remains in the courtyard under repair, with an old football in use nearby; Charlie retains holes, dents and malfunctioning sensors. His navigation display locates the department store and mart 25.6 km to the southwest.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"서남쪽 25.6km 백화점/마트\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 시야 인터페이스 속, 낡은 지도 이미지 위에 '서남쪽 25.6km 백화점/마트'라는 붉은 글씨가 선명하게 떠오른 주관 시점 찰나.\n\nLOCATION (lock): In the village's open communal lot beside the old truck, with a navigation map overlaid in the robot's view. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: Navigation map image (Displayed as an aged map within the visual interface) — Its content-bearing face is presented directly to the viewer, showing the navigation map; used as Provides geographic context around the destination information without introducing a physical display object; Destination label (Clearly displaying '서남쪽 25.6km 백화점/마트' in red) — The readable text faces the viewer in the image plane; used as Primary information focus for the brief subjective insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled interface contrast separates the sharp red destination text from the subdued map image without simulated exterior illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck remains in the courtyard under repair, with an old football in use nearby; Charlie retains holes, dents and malfunctioning sensors. His navigation display locates the department store and mart 25.6 km to the southwest.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"서남쪽 25.6km 백화점/마트\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 지도 인터페이스를 정면으로 향하고 있음.",
    "built_space": "해당 없음 (평면적인 시야 인터페이스 화면).",
    "entities": "낡은 지도 배경, '서남쪽 25.6km 백화점/마트'라는 붉은색 텍스트. 지도 위에 프롬프트에 없는 다수의 한글 지명들이 렌더링됨.",
    "hard_violations": [
     "[gemini-pro] 지시되지 않은 읽을 수 있는 텍스트(지도 위의 수많은 한글 지명들)가 포함됨.",
     "[gpt-high] 허용된 목적지 문구 외에 판독 가능한 지도 지명과 한글 표기가 다수 노출되어 추가 가독 문자 금지 조건을 위반한다."
    ],
    "physics": "텍스트가 평면 지도 이미지 위에 오버레이되어 떠 있음."
   },
   {
    "label": "B",
    "direction": "시선은 지도 인터페이스를 정면으로 향하고 있음.",
    "built_space": "해당 없음 (평면적인 시야 인터페이스 화면).",
    "entities": "텍스트가 없는 낡은 지도 배경, '서남쪽 25.6km 백화점/마트'라는 붉은색 윤곽선 텍스트.",
    "hard_violations": [],
    "physics": "텍스트가 평면 지도 이미지 위에 오버레이되어 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 붉은색 텍스트를 정확히 렌더링했으며, 프롬프트의 금지 사항에 따라 지도 배경에 다른 읽을 수 있는 글자를 전혀 포함하지 않아 지시를 완벽히 따랐습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 텍스트는 정확하게 출력되었으나, 금지 사항을 어기고 지도 배경에 수많은 다른 한글 지명(읽을 수 있는 텍스트)을 무단으로 추가하여 중대한 감점 요인이 되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 지도 인터페이스를 정면으로 향하고 있음.",
        "built_space": "해당 없음 (평면적인 시야 인터페이스 화면).",
        "entities": "낡은 지도 배경, '서남쪽 25.6km 백화점/마트'라는 붉은색 텍스트. 지도 위에 프롬프트에 없는 다수의 한글 지명들이 렌더링됨.",
        "hard_violations": [
         "지시되지 않은 읽을 수 있는 텍스트(지도 위의 수많은 한글 지명들)가 포함됨."
        ],
        "physics": "텍스트가 평면 지도 이미지 위에 오버레이되어 떠 있음."
       },
       {
        "label": "B",
        "direction": "시선은 지도 인터페이스를 정면으로 향하고 있음.",
        "built_space": "해당 없음 (평면적인 시야 인터페이스 화면).",
        "entities": "텍스트가 없는 낡은 지도 배경, '서남쪽 25.6km 백화점/마트'라는 붉은색 윤곽선 텍스트.",
        "hard_violations": [],
        "physics": "텍스트가 평면 지도 이미지 위에 오버레이되어 떠 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 붉은색 텍스트를 정확히 렌더링했으며, 프롬프트의 금지 사항에 따라 지도 배경에 다른 읽을 수 있는 글자를 전혀 포함하지 않아 지시를 완벽히 따랐습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 텍스트는 정확하게 출력되었으나, 금지 사항을 어기고 지도 배경에 수많은 다른 한글 지명(읽을 수 있는 텍스트)을 무단으로 추가하여 중대한 감점 요인이 되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 지도 인터페이스를 정면으로 향하고 있음.",
        "built_space": "해당 없음 (평면적인 시야 인터페이스 화면).",
        "entities": "낡은 지도 배경, '서남쪽 25.6km 백화점/마트'라는 붉은색 텍스트. 지도 위에 프롬프트에 없는 다수의 한글 지명들이 렌더링됨.",
        "hard_violations": [
         "지시되지 않은 읽을 수 있는 텍스트(지도 위의 수많은 한글 지명들)가 포함됨."
        ],
        "physics": "텍스트가 평면 지도 이미지 위에 오버레이되어 떠 있음."
       },
       {
        "label": "B",
        "direction": "시선은 지도 인터페이스를 정면으로 향하고 있음.",
        "built_space": "해당 없음 (평면적인 시야 인터페이스 화면).",
        "entities": "텍스트가 없는 낡은 지도 배경, '서남쪽 25.6km 백화점/마트'라는 붉은색 윤곽선 텍스트.",
        "hard_violations": [],
        "physics": "텍스트가 평면 지도 이미지 위에 오버레이되어 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "낡은 지도를 정면으로 채운 주관 시점 삽입 화면과 정확한 붉은 목적지 문구를 구현했으며, 추가로 읽히는 글자가 없어 지시에 가장 충실하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "목적지 문구는 정확하지만 지도에 읽을 수 있는 지명을 다수 추가했고, 인터페이스보다 실물 종이지도를 비스듬히 촬영한 화면에 가까워 핵심 지시를 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지도 내용과 중앙의 '서남쪽 25.6km 백화점/마트' 문구가 관객을 정면으로 향한다. 서남쪽이라는 방향은 문구로 명시되며, 별도의 방향 화살표나 시선·총구·이동 인물은 없다.",
        "built_space": "낡은 지도 이미지가 화면 전체를 차지한다. 도로와 건물 윤곽은 지도 내부의 기호이며 실제 공간의 시설물이 아니다. 물리적 화면 장치, 마당, 트럭, 사람, 반사는 보이지 않아 지정된 인터페이스 세부 삽입 구도에 맞는다.",
        "entities": "접힘과 마모 흔적이 있는 누런 지도 위에 지정된 문구가 정확히 한 줄로 표시된다. 글자는 검은 내부와 붉게 빛나는 윤곽으로 표현되어 붉은 목적지 정보로 읽힌다. 다른 판독 가능한 문구는 없다. 찰리의 외형과 이현우, 트럭, 축구공은 이 구도의 화면 밖이므로 불일치가 아니다.",
        "hard_violations": [],
        "physics": "지도와 문구는 명시된 시야 인터페이스의 표시 내용으로 읽히며, 지지 없이 떠 있는 실물 물체가 아니다. 공중에 뜬 신체나 움직이는 물체가 없고, 별도의 지지·도약·착지 검증이 필요한 대상도 없다."
       },
       {
        "label": "B",
        "direction": "중앙의 목적지 문구는 관객을 정면으로 향하지만, 배경 지도는 비스듬히 누운 면으로 보인다. 서남쪽 방향은 문구로만 제시된다. 시선이나 무기, 이동하는 신체는 없다.",
        "built_space": "지도 한 장의 왼쪽 가장자리와 그 바깥의 어두운 바탕이 보이며, 원근과 주변부 흐림 때문에 실제 종이지도를 가까이서 촬영한 인상이 강하다. 화면 장치나 시설물, 사람, 반사는 없지만 지도 이미지 자체를 보여주는 주관 인터페이스 구도와는 거리가 있다.",
        "entities": "낡고 접힌 지도와 정확한 붉은 '서남쪽 25.6km 백화점/마트' 문구가 있다. 그러나 지도 곳곳에 '회산', '노대강둑' 등 읽히는 한글 표기가 추가되어 지정 문구만 허용한다는 조건을 어긴다. 사람이나 로봇의 신체, 트럭, 축구공은 보이지 않으며 해당 삽입 구도에서 필수 노출 대상은 아니다.",
        "hard_violations": [
         "허용된 목적지 문구 외에 판독 가능한 지도 지명과 한글 표기가 다수 노출되어 추가 가독 문자 금지 조건을 위반한다."
        ],
        "physics": "종이지도로 해석하면 왼쪽 가장자리 밖의 어두운 바탕에 놓인 모습이며, 공중 부양을 나타내는 근거는 없다. 붉은 문구는 인터페이스 표시로 해석할 수 있으므로 물리적 지지물이 필요하지 않다. 신체나 운동 물체의 불가능한 자세는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "낡은 지도를 정면으로 채운 주관 시점 삽입 화면과 정확한 붉은 목적지 문구를 구현했으며, 추가로 읽히는 글자가 없어 지시에 가장 충실하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "목적지 문구는 정확하지만 지도에 읽을 수 있는 지명을 다수 추가했고, 인터페이스보다 실물 종이지도를 비스듬히 촬영한 화면에 가까워 핵심 지시를 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "지도 내용과 중앙의 '서남쪽 25.6km 백화점/마트' 문구가 관객을 정면으로 향한다. 서남쪽이라는 방향은 문구로 명시되며, 별도의 방향 화살표나 시선·총구·이동 인물은 없다.",
        "built_space": "낡은 지도 이미지가 화면 전체를 차지한다. 도로와 건물 윤곽은 지도 내부의 기호이며 실제 공간의 시설물이 아니다. 물리적 화면 장치, 마당, 트럭, 사람, 반사는 보이지 않아 지정된 인터페이스 세부 삽입 구도에 맞는다.",
        "entities": "접힘과 마모 흔적이 있는 누런 지도 위에 지정된 문구가 정확히 한 줄로 표시된다. 글자는 검은 내부와 붉게 빛나는 윤곽으로 표현되어 붉은 목적지 정보로 읽힌다. 다른 판독 가능한 문구는 없다. 찰리의 외형과 이현우, 트럭, 축구공은 이 구도의 화면 밖이므로 불일치가 아니다.",
        "hard_violations": [],
        "physics": "지도와 문구는 명시된 시야 인터페이스의 표시 내용으로 읽히며, 지지 없이 떠 있는 실물 물체가 아니다. 공중에 뜬 신체나 움직이는 물체가 없고, 별도의 지지·도약·착지 검증이 필요한 대상도 없다."
       },
       {
        "label": "A",
        "direction": "중앙의 목적지 문구는 관객을 정면으로 향하지만, 배경 지도는 비스듬히 누운 면으로 보인다. 서남쪽 방향은 문구로만 제시된다. 시선이나 무기, 이동하는 신체는 없다.",
        "built_space": "지도 한 장의 왼쪽 가장자리와 그 바깥의 어두운 바탕이 보이며, 원근과 주변부 흐림 때문에 실제 종이지도를 가까이서 촬영한 인상이 강하다. 화면 장치나 시설물, 사람, 반사는 없지만 지도 이미지 자체를 보여주는 주관 인터페이스 구도와는 거리가 있다.",
        "entities": "낡고 접힌 지도와 정확한 붉은 '서남쪽 25.6km 백화점/마트' 문구가 있다. 그러나 지도 곳곳에 '회산', '노대강둑' 등 읽히는 한글 표기가 추가되어 지정 문구만 허용한다는 조건을 어긴다. 사람이나 로봇의 신체, 트럭, 축구공은 보이지 않으며 해당 삽입 구도에서 필수 노출 대상은 아니다.",
        "hard_violations": [
         "허용된 목적지 문구 외에 판독 가능한 지도 지명과 한글 표기가 다수 노출되어 추가 가독 문자 금지 조건을 위반한다."
        ],
        "physics": "종이지도로 해석하면 왼쪽 가장자리 밖의 어두운 바탕에 놓인 모습이며, 공중 부양을 나타내는 근거는 없다. 붉은 문구는 인터페이스 표시로 해석할 수 있으므로 물리적 지지물이 필요하지 않다. 신체나 운동 물체의 불가능한 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.762,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.512,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지시되지 않은 읽을 수 있는 텍스트(지도 위의 수많은 한글 지명들)가 포함됨.",
     "[gpt-high] 허용된 목적지 문구 외에 판독 가능한 지도 지명과 한글 표기가 다수 노출되어 추가 가독 문자 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 512
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 붉은색 텍스트를 정확히 렌더링했으며, 프롬프트의 금지 사항에 따라 지도 배경에 다른 읽을 수 있는 글자를 전혀 포함하지 않아 지시를 완벽히 따랐습니다."
   },
   {
    "label": "A",
    "score": 512,
    "verdict_ko": "지정된 텍스트는 정확하게 출력되었으나, 금지 사항을 어기고 지도 배경에 수많은 다른 한글 지명(읽을 수 있는 텍스트)을 무단으로 추가하여 중대한 감점 요인이 되었습니다.  ★위반: [gemini-pro] 지시되지 않은 읽을 수 있는 텍스트(지도 위의 수많은 한글 지명들)가 포함됨. / [gpt-high] 허용된 목적지 문구 외에 판독 가능한 지도 지명과 한글 표기가 다수 노출되어 추가 가독 문자 금지 조건을 위반한다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdf7-817d-7d00-ab3b-0ac0c4300493",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S62sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:51:36.101611+00:00",
  "fingerprint": "27f3b273ba5e8c1f4277b4b48c6cd300a9404bbdba787f38c88d57a88ac209af",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S62sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S62sh12_sel.png",
  "source_sha256": "0704ba532bfc74c71fc6ead68dc4a1aba595ec738e2b990592ff36f0d8741809",
  "file": "S62sh12_cine.png",
  "staged_sha256": "0a239451a32028fb80660e40b23d9516504eef690d8e7b67f094c80422fca30f",
  "latency_ms": 12072
 },
 "S62sh15::signage": {
  "fp": "e9000c67638c4a2c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::7830353bb19693cb": {
  "subjects": [],
  "subject_text": "익산 마을 공터와 모닥불 주변\n한옥 건물들 사이의 넓은 공터. 한쪽에 모닥불 자리가 있고 인접한 창고 출입구와 주변 지붕들이 보인다.",
  "identity": "canonical",
  "scope_id": "L93",
  "scope_role": "location_exterior",
  "scope_sha": "7bc7c5074345d667"
 },
 "groupbg::village_common": {
  "input_fingerprint": "56f8514626c201a0",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "village_common",
    "tags": [
     "S62sh12",
     "S62sh15",
     "S67sh53"
    ]
   },
   "context_sig": "1151d5b60119277c"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: Beside the old truck in the village's outdoor communal lot in daylight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 마을 공터와 모닥불 주변: 아이들이 낡은 축구공을 차고 놀거나, 밤에 주민들이 모여 앉아 고기를 나눠 먹고 춤을 추는 마을 중심 마당. (특징: 흙바닥을 구르는 낡은 가죽 축구공; 타오르는 모닥불과 식사가 담긴 투박한 접시들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- / 마을 공터\n- 한편, 공터에서 낡은 트럭의 시동을 부르릉 부르릉 걸고 있는 앰버..\n- / 마을 공터 - 해질무렵\n- 마을 사람들이 다 모여 맘껏 음식을 나눠먹는다.\n- 이때 어디선가 끌려 나오는 현우와 앰버. 다가가는 박철진.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: Beside the old truck in the village's outdoor communal lot in daylight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 마을 공터와 모닥불 주변: 아이들이 낡은 축구공을 차고 놀거나, 밤에 주민들이 모여 앉아 고기를 나눠 먹고 춤을 추는 마을 중심 마당. (특징: 흙바닥을 구르는 낡은 가죽 축구공; 타오르는 모닥불과 식사가 담긴 투박한 접시들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- / 마을 공터\n- 한편, 공터에서 낡은 트럭의 시동을 부르릉 부르릉 걸고 있는 앰버..\n- / 마을 공터 - 해질무렵\n- 마을 사람들이 다 모여 맘껏 음식을 나눠먹는다.\n- 이때 어디선가 끌려 나오는 현우와 앰버. 다가가는 박철진.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_village_common_7675fe.png",
  "asset_id": "0a5bd9fe-7778-4b7d-83a4-bd1f31ad9726",
  "input_asset_ids": [
   "aae6607a-a0d6-487c-b5ef-e6f254ddfe25"
  ],
  "origin_tag": "S62sh15",
  "place_text": "Beside the old truck in the village's outdoor communal lot in daylight.",
  "origin_inputs": {
   "place_text": "Beside the old truck in the village's outdoor communal lot in daylight.",
   "time_of_day_en": "day",
   "conti_asset_id": "aae6607a-a0d6-487c-b5ef-e6f254ddfe25"
  }
 },
 "S62sh15::bgfirst_bg": {
  "input_fingerprint": "31edb3ced27545b5",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 수빈을 향해 한쪽 눈을 질끈 감고 윙크를 한 채 미소 짓는 여유로운 얼굴 클로즈업.\n\nLOCATION (lock): Beside the old truck in the village's outdoor communal lot in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the clearing's ambient daytime illumination, with gentle facial contrast supporting the reassuring smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 수빈을 향해 한쪽 눈을 질끈 감고 윙크를 한 채 미소 짓는 여유로운 얼굴 클로즈업.\n\nLOCATION (lock): Beside the old truck in the village's outdoor communal lot in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the clearing's ambient daytime illumination, with gentle facial contrast supporting the reassuring smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S62sh15__bgfirst_bg.png",
  "asset_id": "d0ef3269-e358-4e21-8ba6-b8092e2bd584",
  "input_asset_ids": [
   "aae6607a-a0d6-487c-b5ef-e6f254ddfe25",
   "0a5bd9fe-7778-4b7d-83a4-bd1f31ad9726"
  ]
 },
 "S62sh15": {
  "input_fingerprint": "6814a76a4544ff51",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 수빈을 향해 한쪽 눈을 질끈 감고 윙크를 한 채 미소 짓는 여유로운 얼굴 클로즈업.\n\nLOCATION (lock): Beside the old truck in the village's outdoor communal lot in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the clearing's ambient daytime illumination, with gentle facial contrast supporting the reassuring smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck remains in the courtyard under repair, and the old football remains in use nearby. Charlie's holes, dents and faulty sensors persist without a completed repair. 이현우: He stands in the courtyard, still carrying the wounds from the fighting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 수빈을 향해 한쪽 눈을 질끈 감고 윙크를 한 채 미소 짓는 여유로운 얼굴 클로즈업.\n\nLOCATION (lock): Beside the old truck in the village's outdoor communal lot in daylight. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the clearing's ambient daytime illumination, with gentle facial contrast supporting the reassuring smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck remains in the courtyard under repair, and the old football remains in use nearby. Charlie's holes, dents and faulty sensors persist without a completed repair. 이현우: He stands in the courtyard, still carrying the wounds from the fighting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 수빈을 향해 한쪽 눈을 질끈 감고 윙크를 한 채 미소 짓는 여유로운 얼굴 클로즈업.\n\nLOCATION (lock): Beside the old truck in the village's outdoor communal lot in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the clearing's ambient daytime illumination, with gentle facial contrast supporting the reassuring smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck remains in the courtyard under repair, and the old football remains in use nearby. Charlie's holes, dents and faulty sensors persist without a completed repair. 이현우: He stands in the courtyard, still carrying the wounds from the fighting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S62sh15__bgfirst_bg.png",
     "asset_id": "d0ef3269-e358-4e21-8ba6-b8092e2bd584",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S62sh15.png",
     "asset_id": "aae6607a-a0d6-487c-b5ef-e6f254ddfe25",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_village_common_7675fe.png",
     "asset_id": "0a5bd9fe-7778-4b7d-83a4-bd1f31ad9726",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선이 전경의 인물(수빈)을 향함.",
    "built_space": "야외 공터. 트럭, 빨간 공구함, 타이어, 벤치, 축구공이 레퍼런스와 완벽히 일치함.",
    "entities": "이현우의 외모와 의상이 일치하며 상처와 윙크 표정을 잘 묘사함. 전경에 수빈의 뒷모습 배치.",
    "hard_violations": [
     "[gpt-high] 전경 왼쪽에 이현우가 아닌 두 번째 사람의 머리와 어깨를 추가하여, 등장 가능한 인물을 이현우로 제한한 명시적 조건을 위반했다."
    ],
    "physics": "자연스럽게 서 있으며 공중에 떠 있는 사물 없음."
   },
   {
    "label": "B",
    "direction": "이현우가 화면 밖 좌측을 바라봄.",
    "built_space": "야외 공터. 트럭 헤드라이트가 변형되었고 레퍼런스에 없는 높은 스툴이 등장함.",
    "entities": "이현우의 외모, 의상, 상처, 윙크 표정은 일치함.",
    "hard_violations": [
     "[gemini-pro] 발명된 사물(레퍼런스에 없는 높은 스툴)",
     "[gemini-pro] 잘못된 트럭 디자인"
    ],
    "physics": "자연스럽게 서 있으며 지지되지 않은 사물 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "샷 텍스트에 맞게 수빈을 향한 시선과 윙크를 잘 표현했으며, 배경의 트럭과 소품들을 레퍼런스와 완벽하게 일치시켰습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "표정 연기는 좋으나, 트럭 디자인이 레퍼런스와 다르고 현장에 없는 높은 스툴이 등장하여 배경 일관성에서 탈락입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 전경의 인물(수빈)을 향함.",
        "built_space": "야외 공터. 트럭, 빨간 공구함, 타이어, 벤치, 축구공이 레퍼런스와 완벽히 일치함.",
        "entities": "이현우의 외모와 의상이 일치하며 상처와 윙크 표정을 잘 묘사함. 전경에 수빈의 뒷모습 배치.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있으며 공중에 떠 있는 사물 없음."
       },
       {
        "label": "B",
        "direction": "이현우가 화면 밖 좌측을 바라봄.",
        "built_space": "야외 공터. 트럭 헤드라이트가 변형되었고 레퍼런스에 없는 높은 스툴이 등장함.",
        "entities": "이현우의 외모, 의상, 상처, 윙크 표정은 일치함.",
        "hard_violations": [
         "발명된 사물(레퍼런스에 없는 높은 스툴)",
         "잘못된 트럭 디자인"
        ],
        "physics": "자연스럽게 서 있으며 지지되지 않은 사물 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "샷 텍스트에 맞게 수빈을 향한 시선과 윙크를 잘 표현했으며, 배경의 트럭과 소품들을 레퍼런스와 완벽하게 일치시켰습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "표정 연기는 좋으나, 트럭 디자인이 레퍼런스와 다르고 현장에 없는 높은 스툴이 등장하여 배경 일관성에서 탈락입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 전경의 인물(수빈)을 향함.",
        "built_space": "야외 공터. 트럭, 빨간 공구함, 타이어, 벤치, 축구공이 레퍼런스와 완벽히 일치함.",
        "entities": "이현우의 외모와 의상이 일치하며 상처와 윙크 표정을 잘 묘사함. 전경에 수빈의 뒷모습 배치.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있으며 공중에 떠 있는 사물 없음."
       },
       {
        "label": "B",
        "direction": "이현우가 화면 밖 좌측을 바라봄.",
        "built_space": "야외 공터. 트럭 헤드라이트가 변형되었고 레퍼런스에 없는 높은 스툴이 등장함.",
        "entities": "이현우의 외모, 의상, 상처, 윙크 표정은 일치함.",
        "hard_violations": [
         "발명된 사물(레퍼런스에 없는 높은 스툴)",
         "잘못된 트럭 디자인"
        ],
        "physics": "자연스럽게 서 있으며 지지되지 않은 사물 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "트럭 바로 옆에서 상처 입은 이현우의 윙크와 미소를 얼굴 중심 클로즈업으로 담았으며, 수빈을 향한 시선은 불명확하지만 금지된 추가 인물이 없다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "상대를 향한 윙크는 명확하지만 허용되지 않은 두 번째 인물을 전경에 추가했고, 얼굴 클로즈업을 어깨너머 상반신 구도로 바꾸었다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 오른쪽 눈을 감고 다른 눈은 카메라 부근을 보며 이를 드러내고 미소 짓는다. 시선의 대상은 화면 밖이므로 수빈이라고 확인할 수 없으며, 렌즈를 향한 윙크처럼도 읽힌다. 무기나 방향을 확인할 휴대 물체는 없다.",
        "built_space": "왼쪽 가까이에 낡은 트럭 한 대의 앞부분과 열린 보닛 하나가 보인다. 그 옆에는 포개진 타이어 두 개와 높은 의자 하나가 있고, 뒤로 콘크리트 담장·전봇대·나무·산과 파란 지붕이 보인다. 오른쪽 흙바닥에는 축구공 하나가 있다. 이현우가 트럭 옆에 있다는 배치와 낮의 공동 마당 재질은 참조에 부합한다. 클로즈업 밖의 건물과 벤치는 확인 대상이 아니며, 불가능한 반사나 중복된 고정 시설은 보이지 않는다.",
        "entities": "인물은 한 명뿐이며 검고 헝클어진 머리의 젊은 동아시아계 남성으로, 한국계 이현우 설정과 대체로 부합한다. 참조보다 얼굴이 다소 길고 성숙해 보이지만 낡고 구멍 난 짙은 회색 후드티는 일치한다. 코와 양 볼, 목에 싸움의 상처가 남아 있다. 녹슨 트럭과 낡은 축구공이 보이며, 찰리는 프레임에 없다. 읽을 수 있는 글자나 덧씌운 그래픽은 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되어 있고 윙크와 미소도 정상적인 얼굴 근육의 움직임으로 보인다. 하체와 발은 클로즈업 밖이므로 접지는 확인할 수 없지만 공중에 떠 있다는 단서는 없다. 공과 타이어, 의자는 지면에 놓여 있고 열린 보닛은 트럭에 연결되어 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "화면 오른쪽 눈을 감고 뜬 눈과 얼굴을 화면 왼쪽 전경의 상대에게 향한다. 따라서 윙크의 대상은 눈앞의 두 번째 인물로 명확하다. 다만 그 사람이 수빈인지 식별할 근거는 없으며, 인물 등장 제한에는 어긋난다.",
        "built_space": "중경 왼쪽에 열린 보닛 하나를 가진 트럭 한 대, 포개진 타이어 두 개, 높은 의자 하나와 공구함이 보인다. 뒤에는 담장과 기와집, 전봇대가 있고, 오른쪽에는 벤치 두 개·드럼통 하나·축구공 하나가 보인다. 참조 장소의 구성을 많이 유지하지만 이현우는 트럭에서 떨어진 전경에 있으며, 두 번째 인물의 뒤통수와 어깨까지 넣어 어깨너머 상반신 구도가 되었다. 불가능한 반사는 보이지 않는다.",
        "entities": "중앙의 이현우는 검고 헝클어진 머리, 젊은 동아시아계 남성 얼굴, 해진 짙은 회색 후드티로 참조와 대체로 부합한다. 볼의 긁힌 상처와 작은 반창고가 보인다. 왼쪽에는 머리를 묶은 다른 사람의 뒤통수·귀·목·어깨가 추가되어 있으며, 허용된 이현우 한 명만 등장시키라는 조건을 위반한다. 낡은 트럭과 축구공은 식별되고 찰리는 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "전경 왼쪽에 이현우가 아닌 두 번째 사람의 머리와 어깨를 추가하여, 등장 가능한 인물을 이현우로 제한한 명시적 조건을 위반했다."
        ],
        "physics": "이현우의 머리와 목, 어깨는 자연스럽게 이어지고 윙크도 해부학적으로 가능한 표정이다. 전경 인물 역시 보이는 범위에서 머리와 목이 몸통에 연결되어 있다. 두 사람의 발은 프레임 밖이라 접지는 확인되지 않지만 부유의 증거는 없다. 트럭은 바퀴로, 벤치와 의자는 다리로 지면에 지지되며 공과 타이어도 바닥에 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "트럭 바로 옆에서 상처 입은 이현우의 윙크와 미소를 얼굴 중심 클로즈업으로 담았으며, 수빈을 향한 시선은 불명확하지만 금지된 추가 인물이 없다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "상대를 향한 윙크는 명확하지만 허용되지 않은 두 번째 인물을 전경에 추가했고, 얼굴 클로즈업을 어깨너머 상반신 구도로 바꾸었다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "화면 오른쪽 눈을 감고 다른 눈은 카메라 부근을 보며 이를 드러내고 미소 짓는다. 시선의 대상은 화면 밖이므로 수빈이라고 확인할 수 없으며, 렌즈를 향한 윙크처럼도 읽힌다. 무기나 방향을 확인할 휴대 물체는 없다.",
        "built_space": "왼쪽 가까이에 낡은 트럭 한 대의 앞부분과 열린 보닛 하나가 보인다. 그 옆에는 포개진 타이어 두 개와 높은 의자 하나가 있고, 뒤로 콘크리트 담장·전봇대·나무·산과 파란 지붕이 보인다. 오른쪽 흙바닥에는 축구공 하나가 있다. 이현우가 트럭 옆에 있다는 배치와 낮의 공동 마당 재질은 참조에 부합한다. 클로즈업 밖의 건물과 벤치는 확인 대상이 아니며, 불가능한 반사나 중복된 고정 시설은 보이지 않는다.",
        "entities": "인물은 한 명뿐이며 검고 헝클어진 머리의 젊은 동아시아계 남성으로, 한국계 이현우 설정과 대체로 부합한다. 참조보다 얼굴이 다소 길고 성숙해 보이지만 낡고 구멍 난 짙은 회색 후드티는 일치한다. 코와 양 볼, 목에 싸움의 상처가 남아 있다. 녹슨 트럭과 낡은 축구공이 보이며, 찰리는 프레임에 없다. 읽을 수 있는 글자나 덧씌운 그래픽은 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되어 있고 윙크와 미소도 정상적인 얼굴 근육의 움직임으로 보인다. 하체와 발은 클로즈업 밖이므로 접지는 확인할 수 없지만 공중에 떠 있다는 단서는 없다. 공과 타이어, 의자는 지면에 놓여 있고 열린 보닛은 트럭에 연결되어 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "화면 오른쪽 눈을 감고 뜬 눈과 얼굴을 화면 왼쪽 전경의 상대에게 향한다. 따라서 윙크의 대상은 눈앞의 두 번째 인물로 명확하다. 다만 그 사람이 수빈인지 식별할 근거는 없으며, 인물 등장 제한에는 어긋난다.",
        "built_space": "중경 왼쪽에 열린 보닛 하나를 가진 트럭 한 대, 포개진 타이어 두 개, 높은 의자 하나와 공구함이 보인다. 뒤에는 담장과 기와집, 전봇대가 있고, 오른쪽에는 벤치 두 개·드럼통 하나·축구공 하나가 보인다. 참조 장소의 구성을 많이 유지하지만 이현우는 트럭에서 떨어진 전경에 있으며, 두 번째 인물의 뒤통수와 어깨까지 넣어 어깨너머 상반신 구도가 되었다. 불가능한 반사는 보이지 않는다.",
        "entities": "중앙의 이현우는 검고 헝클어진 머리, 젊은 동아시아계 남성 얼굴, 해진 짙은 회색 후드티로 참조와 대체로 부합한다. 볼의 긁힌 상처와 작은 반창고가 보인다. 왼쪽에는 머리를 묶은 다른 사람의 뒤통수·귀·목·어깨가 추가되어 있으며, 허용된 이현우 한 명만 등장시키라는 조건을 위반한다. 낡은 트럭과 축구공은 식별되고 찰리는 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "전경 왼쪽에 이현우가 아닌 두 번째 사람의 머리와 어깨를 추가하여, 등장 가능한 인물을 이현우로 제한한 명시적 조건을 위반했다."
        ],
        "physics": "이현우의 머리와 목, 어깨는 자연스럽게 이어지고 윙크도 해부학적으로 가능한 표정이다. 전경 인물 역시 보이는 범위에서 머리와 목이 몸통에 연결되어 있다. 두 사람의 발은 프레임 밖이라 접지는 확인되지 않지만 부유의 증거는 없다. 트럭은 바퀴로, 벤치와 의자는 다리로 지면에 지지되며 공과 타이어도 바닥에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.375,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.125,
    "B": 1.321
   },
   "violations": {
    "B": [
     "[gemini-pro] 발명된 사물(레퍼런스에 없는 높은 스툴)",
     "[gemini-pro] 잘못된 트럭 디자인"
    ],
    "A": [
     "[gpt-high] 전경 왼쪽에 이현우가 아닌 두 번째 사람의 머리와 어깨를 추가하여, 등장 가능한 인물을 이현우로 제한한 명시적 조건을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1125,
   "B": 1321
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "샷 텍스트에 맞게 수빈을 향한 시선과 윙크를 잘 표현했으며, 배경의 트럭과 소품들을 레퍼런스와 완벽하게 일치시켰습니다.  ★위반: [gpt-high] 전경 왼쪽에 이현우가 아닌 두 번째 사람의 머리와 어깨를 추가하여, 등장 가능한 인물을 이현우로 제한한 명시적 조건을 위반했다."
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "표정 연기는 좋으나, 트럭 디자인이 레퍼런스와 다르고 현장에 없는 높은 스툴이 등장하여 배경 일관성에서 탈락입니다.  ★위반: [gemini-pro] 발명된 사물(레퍼런스에 없는 높은 스툴) / [gemini-pro] 잘못된 트럭 디자인"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_village_common_7675fe.png",
    "asset_id": "0a5bd9fe-7778-4b7d-83a4-bd1f31ad9726",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdf8-4821-79da-8642-c991bed14d9a",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S62sh15__bgfirst_bg.png",
   "bg_asset_id": "d0ef3269-e358-4e21-8ba6-b8092e2bd584",
   "bg_record_key": "S62sh15::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "village_common",
   "groupbg_asset_id": "0a5bd9fe-7778-4b7d-83a4-bd1f31ad9726"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S62sh15::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:51:48.663283+00:00",
  "fingerprint": "5e7f43a7443aaf18fdc4a50aee6a769f46ed27115edc87608cf6ddac8ba14024",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S62sh15_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S62sh15_sel.png",
  "source_sha256": "8b641c99c63d686d7be884f0aa941e1a81345f42e0964ac2aa718c8f6598a3c1",
  "file": "S62sh15_cine.png",
  "staged_sha256": "65dfaaed7f00cc750dd6b39b7845a33ec99629c110b3e7d6919d7d955bc2b661",
  "latency_ms": 12096
 },
 "S63sh7::confined_fp_apt": {
  "applies": true,
  "reason_ko": "트럭 조수석에 앉은 수빈이 운전석에 있는 이현우를 향해 고함을 치는 장면으로, 차량 내부의 좌석 배치와 인물 간의 위치 및 시선 방향을 정확하게 표현해야 하므로 평면도 배치가 필요합니다.",
  "input_fingerprint": "a3f29c765c5e099d"
 },
 "S63sh7::signage": {
  "fp": "6dfa4d76bb1ec05f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::2ae3272f74c6": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/confinedfp_base_2ae3272f74c6.png",
  "place_text": "Inside the truck's compact front cab, at the passenger seat beside the driver with daylight through the windshield.",
  "input_fingerprint": "8c34ca827305eb54"
 },
 "S63sh7::confined_fp": {
  "reads": {
   "controls": "Steering wheel is located at the front left seat (driver's side).",
   "mirrors": "No mirrors are present in the diagram.",
   "camera": "Positioned behind the center gap between the seats, angled forward and to the right, pointing directly at the passenger.",
   "occupants": "The left seat is occupied by 'Driver'. The right seat is occupied by '수빈'."
  },
  "mismatches": [],
  "scene_description_en": "The camera views the truck cab obliquely from just behind the center gap between the seats. In the near left foreground, the back of the 'Driver' occupies the driver's seat facing forward. Beyond the driver on the far left, the steering wheel is positioned. In the right foreground and center, '수빈' occupies the passenger seat. She is turned sharply to face left toward the driver, her profile visible to the camera. The windshield stretches across the far background from left to right. No mirrors are present in the frame.",
  "fixed": true,
  "input_fingerprint": "ec707973c258b47c"
 },
 "S63sh7": {
  "input_fingerprint": "f1e9990185e4bc0e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈이 이현우를 향해 두 눈을 번쩍 치켜뜬 채 목에 핏대를 세우고 입을 크게 벌려 고함치는 분노의 표정.\n\nLOCATION (lock): Inside the truck's compact front cab, at the passenger seat beside the driver with daylight through the windshield. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Truck seats (Occupied by the driver and passenger) — The inner seat edges flank the camera's oblique view across the cab; used as Retain enough cabin context to establish the compressed distance between the two occupants.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime illumination inside the cab retains facial detail without introducing a dramatic lighting change for the outburst.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck is now operational and traveling along the road in daylight. 수빈: She occupies the passenger seat. Her facial wounds and radiation-damaged abdominal skin remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera views the truck cab obliquely from just behind the center gap between the seats. In the near left foreground, the back of the 'Driver' occupies the driver's seat facing forward. Beyond the driver on the far left, the steering wheel is positioned. In the right foreground and center, '수빈' occupies the passenger seat. She is turned sharply to face left toward the driver, her profile visible to the camera. The windshield stretches across the far background from left to right. No mirrors are present in the frame.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈이 이현우를 향해 두 눈을 번쩍 치켜뜬 채 목에 핏대를 세우고 입을 크게 벌려 고함치는 분노의 표정.\n\nLOCATION (lock): Inside the truck's compact front cab, at the passenger seat beside the driver with daylight through the windshield. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime illumination inside the cab retains facial detail without introducing a dramatic lighting change for the outburst.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck is now operational and traveling along the road in daylight. 수빈: She occupies the passenger seat. Her facial wounds and radiation-damaged abdominal skin remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera views the truck cab obliquely from just behind the center gap between the seats. In the near left foreground, the back of the 'Driver' occupies the driver's seat facing forward. Beyond the driver on the far left, the steering wheel is positioned. In the right foreground and center, '수빈' occupies the passenger seat. She is turned sharply to face left toward the driver, her profile visible to the camera. The windshield stretches across the far background from left to right. No mirrors are present in the frame.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈이 이현우를 향해 두 눈을 번쩍 치켜뜬 채 목에 핏대를 세우고 입을 크게 벌려 고함치는 분노의 표정.\n\nLOCATION (lock): Inside the truck's compact front cab, at the passenger seat beside the driver with daylight through the windshield. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime illumination inside the cab retains facial detail without introducing a dramatic lighting change for the outburst.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck is now operational and traveling along the road in daylight. 수빈: She occupies the passenger seat. Her facial wounds and radiation-damaged abdominal skin remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S63sh7_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "수빈",
     "path": "<bytes:904512>",
     "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S63sh7_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "수빈",
     "path": "<bytes:904512>",
     "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "수빈은 왼쪽의 운전자를 향해 고개를 돌려 고함치고 있으며, 운전자는 앞을 향함.",
    "built_space": "트럭 운전석 내부. 좌우 좌석 사이에 넓은 콘솔이 있어 비좁은 거리감이 떨어짐.",
    "entities": "수빈은 레퍼런스의 의상과 스카프를 착용했으나, 눈을 크게 뜨지 않음.",
    "hard_violations": [],
    "physics": "두 인물 모두 좌석에 안정적으로 앉아 있음."
   },
   {
    "label": "B",
    "direction": "수빈의 시선과 고함치는 방향이 정확히 운전자를 향함.",
    "built_space": "트럭 운전석 내부. 카메라 양옆으로 좌석 안쪽이 배치되어 압축된 거리감을 줌.",
    "entities": "수빈은 크게 뜬 눈과 목의 핏대를 정확히 보여주며 가슴 부위에 피부 손상이 확인됨.",
    "hard_violations": [],
    "physics": "수빈은 상체를 기울인 채 하체로 지지하고 있으며 운전자는 두 손으로 운전대를 쥠."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "크게 뜬 눈과 목의 핏대 등 요구된 표정 연기를 정확히 구현했으며, 좌석 사이의 비좁은 공간감도 지시에 맞게 포착함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "눈을 찡그리고 있어 '두 눈을 번쩍 치켜뜬 채'라는 핵심 표정 지시를 위반했으며 목의 핏대 묘사도 부족함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수빈은 왼쪽의 운전자를 향해 고개를 돌려 고함치고 있으며, 운전자는 앞을 향함.",
        "built_space": "트럭 운전석 내부. 좌우 좌석 사이에 넓은 콘솔이 있어 비좁은 거리감이 떨어짐.",
        "entities": "수빈은 레퍼런스의 의상과 스카프를 착용했으나, 눈을 크게 뜨지 않음.",
        "hard_violations": [],
        "physics": "두 인물 모두 좌석에 안정적으로 앉아 있음."
       },
       {
        "label": "B",
        "direction": "수빈의 시선과 고함치는 방향이 정확히 운전자를 향함.",
        "built_space": "트럭 운전석 내부. 카메라 양옆으로 좌석 안쪽이 배치되어 압축된 거리감을 줌.",
        "entities": "수빈은 크게 뜬 눈과 목의 핏대를 정확히 보여주며 가슴 부위에 피부 손상이 확인됨.",
        "hard_violations": [],
        "physics": "수빈은 상체를 기울인 채 하체로 지지하고 있으며 운전자는 두 손으로 운전대를 쥠."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "크게 뜬 눈과 목의 핏대 등 요구된 표정 연기를 정확히 구현했으며, 좌석 사이의 비좁은 공간감도 지시에 맞게 포착함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "눈을 찡그리고 있어 '두 눈을 번쩍 치켜뜬 채'라는 핵심 표정 지시를 위반했으며 목의 핏대 묘사도 부족함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "수빈은 왼쪽의 운전자를 향해 고개를 돌려 고함치고 있으며, 운전자는 앞을 향함.",
        "built_space": "트럭 운전석 내부. 좌우 좌석 사이에 넓은 콘솔이 있어 비좁은 거리감이 떨어짐.",
        "entities": "수빈은 레퍼런스의 의상과 스카프를 착용했으나, 눈을 크게 뜨지 않음.",
        "hard_violations": [],
        "physics": "두 인물 모두 좌석에 안정적으로 앉아 있음."
       },
       {
        "label": "B",
        "direction": "수빈의 시선과 고함치는 방향이 정확히 운전자를 향함.",
        "built_space": "트럭 운전석 내부. 카메라 양옆으로 좌석 안쪽이 배치되어 압축된 거리감을 줌.",
        "entities": "수빈은 크게 뜬 눈과 목의 핏대를 정확히 보여주며 가슴 부위에 피부 손상이 확인됨.",
        "hard_violations": [],
        "physics": "수빈은 상체를 기울인 채 하체로 지지하고 있으며 운전자는 두 손으로 운전대를 쥠."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "클로즈업보다 넓지만, 이현우를 향해 부릅뜬 눈·크게 벌린 입·목의 핏대가 명확하여 핵심 분노 연기를 더 충실히 구현한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "나란한 좌석 배치와 운전자를 향한 고함은 맞지만, 실내를 지나치게 넓게 보여 주고 눈을 치켜뜨는 연기와 목의 핏대가 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수빈은 화면 왼쪽의 운전자 쪽으로 얼굴과 상체를 돌리고, 크게 뜬 두 눈으로 그의 얼굴 방향을 바라보며 입을 벌리고 있다. 운전자는 앞유리 너머 도로를 향하고 양손으로 운전대를 잡는다. 수빈의 고함 대상은 화면 안의 이현우로 읽힌다.",
        "built_space": "앞좌석 등받이 두 개가 각각 왼쪽 전경과 오른쪽 아래에 보이고, 운전자는 왼쪽 운전대 앞, 수빈은 오른쪽 조수석에 있다. 운전대 하나, 중앙 대시보드 하나, 앞유리 하나, 중앙 실내거울 하나와 좌우 차창이 보인다. 좌석 뒤쪽에서 비스듬히 보는 배치는 평면도의 관계와 부합한다. 다만 수빈의 허리 부근과 운전자의 상체, 넓은 앞유리까지 포함해 요구된 얼굴 클로즈업이 아니다. 거울의 흐릿한 부분상에서 불가능한 반사를 확정할 근거는 없다.",
        "entities": "수빈은 젊은 동아시아 여성으로 보이며 한국인 20세 설정과 모순되지 않는다. 검은 단발, 얼굴 윤곽, 낡은 갈색 가죽 재킷과 올리브색 후드는 참고 인물과 대체로 맞는다. 볼의 상처가 보이지만 참고의 목도리는 없고 가슴 위쪽에 손상 피부가 드러난다. 복부는 화면 밖이므로 복부 손상의 유지 여부는 판단할 수 없다. 이현우로 읽히는 짧은 검은 머리의 남성 운전자 한 명 외에 추가 인물은 없다. 눈은 정상적인 홍채와 동공을 유지하며, 크게 벌린 입과 도드라진 목의 힘줄이 보인다. 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "수빈의 골반 부근은 조수석 위치에 남아 있고 몸통만 운전자 쪽으로 기울어져 있다. 화면 오른쪽 아래의 손은 좌석 가장자리에 닿아 몸을 지탱하는 것으로 보인다. 운전자의 몸은 운전석에 받쳐지고 양손은 운전대에 접촉한다. 고함치며 목에 힘을 주고 상체를 돌리는 동작은 가능하며, 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "수빈은 왼쪽 운전자의 옆얼굴 방향으로 고개를 돌려 고함친다. 시선 역시 운전자 쪽이지만 두 눈은 번쩍 뜨기보다 좁혀져 있다. 운전자는 대체로 앞쪽 도로를 향하고, 보이는 손 하나로 운전대를 잡는다.",
        "built_space": "왼쪽 운전석 등받이와 오른쪽 조수석 등받이가 각각 한 개씩 보이며 두 사람은 각자의 좌석을 차지한다. 왼쪽 운전대 하나, 중앙 대시보드 하나, 기어 레버 하나, 두 개의 컵 구멍이 있는 중앙 콘솔, 앞유리와 좌우 차창이 보인다. 좌석 사이 뒤쪽에서 바라본 공간 관계는 타당하지만, 수빈의 무릎과 중앙 콘솔까지 드러나는 넓은 구도여서 클로즈업 요구에서 크게 벗어난다. 중복 좌석이나 명백히 불가능한 반사는 보이지 않는다.",
        "entities": "수빈은 젊은 동아시아 여성으로 보이며 검은 단발과 볼의 상처가 있다. 갈색 가죽 재킷, 올리브색 후드와 목도리는 참고 의상에 가깝다. 복부는 옷에 가려져 손상 피부를 확인할 수 없다. 이현우로 읽히는 남성 운전자 외에 추가 인물은 없다. 입은 크게 벌어져 있으나 눈을 치켜뜬 표정은 약하고 목도리와 옷 때문에 목의 핏대도 확인하기 어렵다. 실내와 도로는 낮으로 보이며 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "수빈은 조수석에 골반을 두고 무릎을 굽힌 채 상체와 목을 운전자 쪽으로 돌린다. 아래쪽 손은 허벅지 부근에 놓여 있고 등받이는 몸 뒤에 있다. 운전자는 좌석에 앉아 보이는 손으로 운전대를 잡는다. 두 사람의 무게를 좌석이 받치는 자연스러운 착석이며, 고함 동작을 불가능하게 만드는 관절 변형이나 부유 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "클로즈업보다 넓지만, 이현우를 향해 부릅뜬 눈·크게 벌린 입·목의 핏대가 명확하여 핵심 분노 연기를 더 충실히 구현한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "나란한 좌석 배치와 운전자를 향한 고함은 맞지만, 실내를 지나치게 넓게 보여 주고 눈을 치켜뜨는 연기와 목의 핏대가 부족하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "수빈은 화면 왼쪽의 운전자 쪽으로 얼굴과 상체를 돌리고, 크게 뜬 두 눈으로 그의 얼굴 방향을 바라보며 입을 벌리고 있다. 운전자는 앞유리 너머 도로를 향하고 양손으로 운전대를 잡는다. 수빈의 고함 대상은 화면 안의 이현우로 읽힌다.",
        "built_space": "앞좌석 등받이 두 개가 각각 왼쪽 전경과 오른쪽 아래에 보이고, 운전자는 왼쪽 운전대 앞, 수빈은 오른쪽 조수석에 있다. 운전대 하나, 중앙 대시보드 하나, 앞유리 하나, 중앙 실내거울 하나와 좌우 차창이 보인다. 좌석 뒤쪽에서 비스듬히 보는 배치는 평면도의 관계와 부합한다. 다만 수빈의 허리 부근과 운전자의 상체, 넓은 앞유리까지 포함해 요구된 얼굴 클로즈업이 아니다. 거울의 흐릿한 부분상에서 불가능한 반사를 확정할 근거는 없다.",
        "entities": "수빈은 젊은 동아시아 여성으로 보이며 한국인 20세 설정과 모순되지 않는다. 검은 단발, 얼굴 윤곽, 낡은 갈색 가죽 재킷과 올리브색 후드는 참고 인물과 대체로 맞는다. 볼의 상처가 보이지만 참고의 목도리는 없고 가슴 위쪽에 손상 피부가 드러난다. 복부는 화면 밖이므로 복부 손상의 유지 여부는 판단할 수 없다. 이현우로 읽히는 짧은 검은 머리의 남성 운전자 한 명 외에 추가 인물은 없다. 눈은 정상적인 홍채와 동공을 유지하며, 크게 벌린 입과 도드라진 목의 힘줄이 보인다. 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "수빈의 골반 부근은 조수석 위치에 남아 있고 몸통만 운전자 쪽으로 기울어져 있다. 화면 오른쪽 아래의 손은 좌석 가장자리에 닿아 몸을 지탱하는 것으로 보인다. 운전자의 몸은 운전석에 받쳐지고 양손은 운전대에 접촉한다. 고함치며 목에 힘을 주고 상체를 돌리는 동작은 가능하며, 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "수빈은 왼쪽 운전자의 옆얼굴 방향으로 고개를 돌려 고함친다. 시선 역시 운전자 쪽이지만 두 눈은 번쩍 뜨기보다 좁혀져 있다. 운전자는 대체로 앞쪽 도로를 향하고, 보이는 손 하나로 운전대를 잡는다.",
        "built_space": "왼쪽 운전석 등받이와 오른쪽 조수석 등받이가 각각 한 개씩 보이며 두 사람은 각자의 좌석을 차지한다. 왼쪽 운전대 하나, 중앙 대시보드 하나, 기어 레버 하나, 두 개의 컵 구멍이 있는 중앙 콘솔, 앞유리와 좌우 차창이 보인다. 좌석 사이 뒤쪽에서 바라본 공간 관계는 타당하지만, 수빈의 무릎과 중앙 콘솔까지 드러나는 넓은 구도여서 클로즈업 요구에서 크게 벗어난다. 중복 좌석이나 명백히 불가능한 반사는 보이지 않는다.",
        "entities": "수빈은 젊은 동아시아 여성으로 보이며 검은 단발과 볼의 상처가 있다. 갈색 가죽 재킷, 올리브색 후드와 목도리는 참고 의상에 가깝다. 복부는 옷에 가려져 손상 피부를 확인할 수 없다. 이현우로 읽히는 남성 운전자 외에 추가 인물은 없다. 입은 크게 벌어져 있으나 눈을 치켜뜬 표정은 약하고 목도리와 옷 때문에 목의 핏대도 확인하기 어렵다. 실내와 도로는 낮으로 보이며 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "수빈은 조수석에 골반을 두고 무릎을 굽힌 채 상체와 목을 운전자 쪽으로 돌린다. 아래쪽 손은 허벅지 부근에 놓여 있고 등받이는 몸 뒤에 있다. 운전자는 좌석에 앉아 보이는 손으로 운전대를 잡는다. 두 사람의 무게를 좌석이 받치는 자연스러운 착석이며, 고함 동작을 불가능하게 만드는 관절 변형이나 부유 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.238,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.238,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1238
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "크게 뜬 눈과 목의 핏대 등 요구된 표정 연기를 정확히 구현했으며, 좌석 사이의 비좁은 공간감도 지시에 맞게 포착함."
   },
   {
    "label": "A",
    "score": 1238,
    "verdict_ko": "눈을 찡그리고 있어 '두 눈을 번쩍 치켜뜬 채'라는 핵심 표정 지시를 위반했으며 목의 핏대 묘사도 부족함."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S63sh7_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "수빈",
    "path": "<bytes:904512>",
    "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdf9-1154-75d8-a603-4566773adeaa",
  "confined_fp": {
   "base_key": "confinedfp::2ae3272f74c6",
   "apt_reason": "트럭 조수석에 앉은 수빈이 운전석에 있는 이현우를 향해 고함을 치는 장면으로, 차량 내부의 좌석 배치와 인물 간의 위치 및 시선 방향을 정확하게 표현해야 하므로 평면도 배치가 필요합니다.",
   "fixed": true,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S63sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:52:01.164396+00:00",
  "fingerprint": "e9036d880851f6d477efdd51e1fe92152d19eb3198e25146148c5f05d0c1696c",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S63sh7_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S63sh7_sel.png",
  "source_sha256": "1b1c8c9b9edb8345a4d245bdd194d79f63ecebc908aec49dce11aad9d38fd36c",
  "file": "S63sh7_cine.png",
  "staged_sha256": "6f04535677d0ce26fb7a74f5c556c11bd7226d0924ee0df41d3f552f43f9291d",
  "latency_ms": 11551
 },
 "S63sh12::signage": {
  "fp": "93687d479ff7dfd0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S63sh12": {
  "input_fingerprint": "ebdb052901de73e2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 회상 속, 피투성이가 된 미연이 눈꺼풀을 힘겹게 들어 올린 채 애절하게 입을 벌린 찰나.\n\nLOCATION (lock): On a container panel floating in open floodwater at dawn, revisited as a memory of the mother's final moments. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained illumination keeps the blood and failing facial effort legible, with no added dream glow or invented light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon is reclining helplessly on a floating container panel, supported by the panel and bleeding from her abdomen. Her precise head angle, arm and leg positions, and face-up or sideways orientation are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the dawn flashback, a container panel floats on the sea amid the flood aftermath. 미연: She is reclining helplessly on a floating container panel, supported by the panel, with a bleeding abdominal puncture wound and a previously swollen, beaten face. She is still alive but barely able to open her eyes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 회상 속, 피투성이가 된 미연이 눈꺼풀을 힘겹게 들어 올린 채 애절하게 입을 벌린 찰나.\n\nLOCATION (lock): On a container panel floating in open floodwater at dawn, revisited as a memory of the mother's final moments. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained illumination keeps the blood and failing facial effort legible, with no added dream glow or invented light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon is reclining helplessly on a floating container panel, supported by the panel and bleeding from her abdomen. Her precise head angle, arm and leg positions, and face-up or sideways orientation are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the dawn flashback, a container panel floats on the sea amid the flood aftermath. 미연: She is reclining helplessly on a floating container panel, supported by the panel, with a bleeding abdominal puncture wound and a previously swollen, beaten face. She is still alive but barely able to open her eyes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 회상 속, 피투성이가 된 미연이 눈꺼풀을 힘겹게 들어 올린 채 애절하게 입을 벌린 찰나.\n\nLOCATION (lock): On a container panel floating in open floodwater at dawn, revisited as a memory of the mother's final moments. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained illumination keeps the blood and failing facial effort legible, with no added dream glow or invented light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon is reclining helplessly on a floating container panel, supported by the panel and bleeding from her abdomen. Her precise head angle, arm and leg positions, and face-up or sideways orientation are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the dawn flashback, a container panel floats on the sea amid the flood aftermath. 미연: She is reclining helplessly on a floating container panel, supported by the panel, with a bleeding abdominal puncture wound and a previously swollen, beaten face. She is still alive but barely able to open her eyes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성 얼굴, 40대 후반, 중년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 허공을 향해 다소 멍하게 풀려 있으며, 입은 아주 살짝 벌려진 상태입니다.",
    "built_space": "녹슨 평평한 컨테이너 패널 위에 위치하며 주변으로 수면이 보입니다.",
    "entities": "미연의 이목구비와 의상은 레퍼런스와 일치하지만, 지문에 명시된 출혈 흔적이 전혀 없습니다.",
    "hard_violations": [],
    "physics": "머리와 상체가 평평한 패널 표면에 자연스럽게 밀착되어 중력에 맞게 지지되고 있습니다."
   },
   {
    "label": "B",
    "direction": "시선은 고통스럽게 위를 향하고 있으며, 지문대로 입을 애절하게 벌리고 있습니다.",
    "built_space": "단차가 있는 녹슨 컨테이너 패널에 기대어 있으며, 뒤로는 새벽 빛의 바다가 보입니다.",
    "entities": "미연의 이목구비와 의상이 레퍼런스와 일치하며, 얼굴과 복부 의상에 피가 묻어 있어 지문과 부합합니다.",
    "hard_violations": [],
    "physics": "머리와 어깨가 컨테이너의 턱 부분에 단단히 기대어 지지되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "피투성이가 된 얼굴, 힘겹게 뜬 눈꺼풀, 애절하게 벌린 입 등 샷 텍스트의 핵심 묘사와 새벽의 조명 분위기를 매우 충실히 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴에 피가 전혀 없고 표정이 평온하여 '피투성이', '애절하게 입을 벌린 찰나'라는 핵심 지문을 완전히 누락했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 허공을 향해 다소 멍하게 풀려 있으며, 입은 아주 살짝 벌려진 상태입니다.",
        "built_space": "녹슨 평평한 컨테이너 패널 위에 위치하며 주변으로 수면이 보입니다.",
        "entities": "미연의 이목구비와 의상은 레퍼런스와 일치하지만, 지문에 명시된 출혈 흔적이 전혀 없습니다.",
        "hard_violations": [],
        "physics": "머리와 상체가 평평한 패널 표면에 자연스럽게 밀착되어 중력에 맞게 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "시선은 고통스럽게 위를 향하고 있으며, 지문대로 입을 애절하게 벌리고 있습니다.",
        "built_space": "단차가 있는 녹슨 컨테이너 패널에 기대어 있으며, 뒤로는 새벽 빛의 바다가 보입니다.",
        "entities": "미연의 이목구비와 의상이 레퍼런스와 일치하며, 얼굴과 복부 의상에 피가 묻어 있어 지문과 부합합니다.",
        "hard_violations": [],
        "physics": "머리와 어깨가 컨테이너의 턱 부분에 단단히 기대어 지지되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "피투성이가 된 얼굴, 힘겹게 뜬 눈꺼풀, 애절하게 벌린 입 등 샷 텍스트의 핵심 묘사와 새벽의 조명 분위기를 매우 충실히 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴에 피가 전혀 없고 표정이 평온하여 '피투성이', '애절하게 입을 벌린 찰나'라는 핵심 지문을 완전히 누락했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 허공을 향해 다소 멍하게 풀려 있으며, 입은 아주 살짝 벌려진 상태입니다.",
        "built_space": "녹슨 평평한 컨테이너 패널 위에 위치하며 주변으로 수면이 보입니다.",
        "entities": "미연의 이목구비와 의상은 레퍼런스와 일치하지만, 지문에 명시된 출혈 흔적이 전혀 없습니다.",
        "hard_violations": [],
        "physics": "머리와 상체가 평평한 패널 표면에 자연스럽게 밀착되어 중력에 맞게 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "시선은 고통스럽게 위를 향하고 있으며, 지문대로 입을 애절하게 벌리고 있습니다.",
        "built_space": "단차가 있는 녹슨 컨테이너 패널에 기대어 있으며, 뒤로는 새벽 빛의 바다가 보입니다.",
        "entities": "미연의 이목구비와 의상이 레퍼런스와 일치하며, 얼굴과 복부 의상에 피가 묻어 있어 지문과 부합합니다.",
        "hard_violations": [],
        "physics": "머리와 어깨가 컨테이너의 턱 부분에 단단히 기대어 지지되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "얼굴 중심 클로즈업에서 피와 타박상, 힘겹게 뜬 눈과 애절하게 벌린 입을 구현해 결정적인 순간에 충실하지만, 눈은 지시보다 다소 크게 떠 있다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "패널에 기대어 누운 클로즈업과 인물 외형은 맞지만, 입이 거의 닫혀 있고 얼굴의 피와 부종이 없어 요청한 마지막 순간의 행동과 상태를 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연의 얼굴은 위로 돌아가 있고 두 눈은 화면 위쪽의 프레임 밖을 향한다. 시선이 닿는 사람이나 물체는 보이지 않으며, 지문도 특정 시선 대상을 정하지 않았다. 입은 분명히 벌어져 있고 눈썹과 눈가에 고통스러운 긴장이 보인다. 무기나 방향성 있는 소품은 없다.",
        "built_space": "녹슨 청회색 컨테이너 패널 한 장의 상판, 가로 보강선들과 가까운 두꺼운 테두리가 보인다. 오른쪽 뒤 모서리에는 돌출된 모서리 부품 하나가 보인다. 미연의 머리는 가까운 테두리에 놓였고 몸통은 오른쪽 아래로 이어진다. 배경은 탁 트인 바다와 따뜻한 수평선으로, 이전 장면의 금속 재질과 새벽 자연광을 대체로 유지한다. 추가 구조물이나 중복 설비, 문제 되는 거울 반사는 없다.",
        "entities": "중년 한국인 여성으로 제시된 미연 한 명만 보인다. 검은 머리, 얼굴 윤곽, 낡고 얼룩진 갈색 작업 재킷과 회색 속옷이 참조와 대체로 일치하며, 이전 장면처럼 모자는 없다. 얼굴에는 피 묻은 긁힌 상처와 눈가 타박상이 있고 화면 아래쪽 옷에도 혈흔이 보인다. 복부 관통상 자체는 잘려 있어 확인할 수 없다. 눈은 정상적인 사람의 눈이며, 읽을 수 있는 글자나 다른 인물은 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리카락이 패널 테두리에 닿아 머리를 받치고 있다. 목과 몸통은 오른쪽 아래로 비스듬히 이어지는 기대어 누운 자세이며, 몸통 하부의 정확한 받침과 팔다리는 화면 밖이다. 보이는 부분에서 허공에 떠 있는 신체나 물체는 없다. 패널은 주변 수면에 떠 있으며, 얼굴의 움직임은 아직 살아 있는 인물이 힘겹게 눈과 입을 여는 동작으로 가능하다."
       },
       {
        "label": "B",
        "direction": "미연은 고개를 옆으로 기울인 채 카메라 가까운 프레임 밖을 흐릿하게 바라본다. 특정 시선 대상은 보이지 않는다. 눈꺼풀은 낮게 내려와 있지만 입술은 거의 붙어 있어, 애절하게 입을 벌린 순간으로는 읽히지 않는다. 방향을 확인할 무기나 사용 중인 소품은 없다.",
        "built_space": "녹슨 청회색 컨테이너 패널 한 장의 상판과 보강선, 가까운 테두리가 보이고 왼쪽 뒤에는 구멍 난 돌출 부품 하나가 있다. 미연의 머리는 가까운 패널 가장자리에 놓이고 어깨와 몸통은 오른쪽 아래로 잘린다. 바다와 수평선이 뒤에 있으며 재질과 장소는 참조에 부합한다. 다만 빛은 이전 장면보다 밝고 중성적이다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "중년 한국인 여성으로 제시된 미연 한 명만 등장한다. 검은 머리와 중년의 얼굴, 얼룩진 갈색 재킷은 참조와 잘 맞고 모자는 없다. 그러나 보이는 얼굴은 대체로 깨끗해 피투성이 상태와 구타로 부은 얼굴이 구현되지 않았다. 복부와 상처는 클로즈업 밖이므로 평가할 수 없다. 눈은 정상적인 해부학적 형태이며 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리카락이 금속 상판 및 가장자리에 눌려 있어 머리의 지지가 확인된다. 목은 어깨로 자연스럽게 이어지고 몸통의 나머지와 팔다리는 프레임 밖이다. 보이는 신체가 지지 없이 들려 있거나 떠 있는 징후는 없다. 패널은 바닷물에 떠 있는 것으로 읽힌다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "얼굴 중심 클로즈업에서 피와 타박상, 힘겹게 뜬 눈과 애절하게 벌린 입을 구현해 결정적인 순간에 충실하지만, 눈은 지시보다 다소 크게 떠 있다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "패널에 기대어 누운 클로즈업과 인물 외형은 맞지만, 입이 거의 닫혀 있고 얼굴의 피와 부종이 없어 요청한 마지막 순간의 행동과 상태를 놓쳤다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "미연의 얼굴은 위로 돌아가 있고 두 눈은 화면 위쪽의 프레임 밖을 향한다. 시선이 닿는 사람이나 물체는 보이지 않으며, 지문도 특정 시선 대상을 정하지 않았다. 입은 분명히 벌어져 있고 눈썹과 눈가에 고통스러운 긴장이 보인다. 무기나 방향성 있는 소품은 없다.",
        "built_space": "녹슨 청회색 컨테이너 패널 한 장의 상판, 가로 보강선들과 가까운 두꺼운 테두리가 보인다. 오른쪽 뒤 모서리에는 돌출된 모서리 부품 하나가 보인다. 미연의 머리는 가까운 테두리에 놓였고 몸통은 오른쪽 아래로 이어진다. 배경은 탁 트인 바다와 따뜻한 수평선으로, 이전 장면의 금속 재질과 새벽 자연광을 대체로 유지한다. 추가 구조물이나 중복 설비, 문제 되는 거울 반사는 없다.",
        "entities": "중년 한국인 여성으로 제시된 미연 한 명만 보인다. 검은 머리, 얼굴 윤곽, 낡고 얼룩진 갈색 작업 재킷과 회색 속옷이 참조와 대체로 일치하며, 이전 장면처럼 모자는 없다. 얼굴에는 피 묻은 긁힌 상처와 눈가 타박상이 있고 화면 아래쪽 옷에도 혈흔이 보인다. 복부 관통상 자체는 잘려 있어 확인할 수 없다. 눈은 정상적인 사람의 눈이며, 읽을 수 있는 글자나 다른 인물은 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리카락이 패널 테두리에 닿아 머리를 받치고 있다. 목과 몸통은 오른쪽 아래로 비스듬히 이어지는 기대어 누운 자세이며, 몸통 하부의 정확한 받침과 팔다리는 화면 밖이다. 보이는 부분에서 허공에 떠 있는 신체나 물체는 없다. 패널은 주변 수면에 떠 있으며, 얼굴의 움직임은 아직 살아 있는 인물이 힘겹게 눈과 입을 여는 동작으로 가능하다."
       },
       {
        "label": "A",
        "direction": "미연은 고개를 옆으로 기울인 채 카메라 가까운 프레임 밖을 흐릿하게 바라본다. 특정 시선 대상은 보이지 않는다. 눈꺼풀은 낮게 내려와 있지만 입술은 거의 붙어 있어, 애절하게 입을 벌린 순간으로는 읽히지 않는다. 방향을 확인할 무기나 사용 중인 소품은 없다.",
        "built_space": "녹슨 청회색 컨테이너 패널 한 장의 상판과 보강선, 가까운 테두리가 보이고 왼쪽 뒤에는 구멍 난 돌출 부품 하나가 있다. 미연의 머리는 가까운 패널 가장자리에 놓이고 어깨와 몸통은 오른쪽 아래로 잘린다. 바다와 수평선이 뒤에 있으며 재질과 장소는 참조에 부합한다. 다만 빛은 이전 장면보다 밝고 중성적이다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "중년 한국인 여성으로 제시된 미연 한 명만 등장한다. 검은 머리와 중년의 얼굴, 얼룩진 갈색 재킷은 참조와 잘 맞고 모자는 없다. 그러나 보이는 얼굴은 대체로 깨끗해 피투성이 상태와 구타로 부은 얼굴이 구현되지 않았다. 복부와 상처는 클로즈업 밖이므로 평가할 수 없다. 눈은 정상적인 해부학적 형태이며 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리카락이 금속 상판 및 가장자리에 눌려 있어 머리의 지지가 확인된다. 목은 어깨로 자연스럽게 이어지고 몸통의 나머지와 팔다리는 프레임 밖이다. 보이는 신체가 지지 없이 들려 있거나 떠 있는 징후는 없다. 패널은 바닷물에 떠 있는 것으로 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.054,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.054,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1054
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "피투성이가 된 얼굴, 힘겹게 뜬 눈꺼풀, 애절하게 벌린 입 등 샷 텍스트의 핵심 묘사와 새벽의 조명 분위기를 매우 충실히 구현했습니다."
   },
   {
    "label": "A",
    "score": 1054,
    "verdict_ko": "얼굴에 피가 전혀 없고 표정이 평온하여 '피투성이', '애절하게 입을 벌린 찰나'라는 핵심 지문을 완전히 누락했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S38sh6_sel.png",
    "asset_id": "f051c526-60ed-46e6-b08d-034d2e664e3f",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1387085>",
    "asset_id": "29884448-7e28-4e7c-8652-3354113e4e44",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdf9-d0a2-7888-a064-ec49d7ea0929",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S38sh6"
  }
 },
 "S63sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:52:13.049207+00:00",
  "fingerprint": "a93fe6386fcf36fd62993c91f899fd72f882d8682872f54119fb3ab2c2e2fc23",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S63sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S63sh12_sel.png",
  "source_sha256": "a029b39de6c052558c36063dcf3e7de7cbcc26c7352cc3bde2770e9e881bb07b",
  "file": "S63sh12_cine.png",
  "staged_sha256": "c86b5f2cee2cf81d6cf40bc5219acea7dd451f69660c13fab52502a5470bea0e",
  "latency_ms": 11222
 },
 "S63sh14::confined_fp_apt": {
  "applies": true,
  "reason_ko": "트럭 운전석 내부라는 제한되고 통제 장치가 중요한 공간에서 운전석에 앉은 인물을 찍는 숏입니다. 운전대와 앞창문, 좌석의 상대적인 공간 배치가 조금이라도 틀리면 서사적 몰입을 깨는 치명적인 오류가 발생하므로 평면도 레이아웃 보조가 필요합니다.",
  "input_fingerprint": "e4b09d9ae9bc985b"
 },
 "S63sh14::signage": {
  "fp": "c4f56c1ad56b450d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::1eb557c77df4": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/confinedfp_base_1eb557c77df4.png",
  "place_text": "At the wheel inside the truck's compact cab, with strong sunlight entering directly through the front windshield.",
  "input_fingerprint": "0a44e47784acf6ae"
 },
 "S63sh14::confined_fp": {
  "reads": {
   "controls": "Steering wheel located at the Driver Seat on the right side of the cab.",
   "mirrors": "No mirrors are depicted in the diagram.",
   "camera": "Positioned behind and to the right of the driver's seat, angled pointing forward and left towards the driver.",
   "occupants": "이현우 is in the Driver Seat."
  },
  "mismatches": [],
  "scene_description_en": "The camera is positioned on the right side of the truck cab, framing a close-up from slightly behind and to the right of the driver. 이현우 occupies the driver's seat, dominating the right and center foreground of the screen, facing leftward toward the front of the vehicle. The steering wheel is positioned just in front of him, appearing in the center-left midground of the shot. The front windshield spans the background across the top and left of the frame. Bright sunlight streams through this windshield from the upper background, heavily illuminating his face.",
  "fixed": true,
  "input_fingerprint": "32f37599999794b8"
 },
 "S63sh14": {
  "input_fingerprint": "033de96da6fd7a76",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앞창문으로 쏟아지는 햇살을 받은 이현우가 입술을 꽉 깨물고 결연한 눈빛으로 앞을 쏘아보는 굳은 얼굴 클로즈업.\n\nLOCATION (lock): At the wheel inside the truck's compact cab, with strong sunlight entering directly through the front windshield. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Truck windshield (Ahead of the driver) — A narrow oblique portion at the forward edge of the composition leaves the exterior unresolved; used as Marks the direction of travel and provides a small margin in front of his gaze.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Sunlight entering through the windshield brightens his face with expressive contrast while preserving detail in his set jaw and eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The operational old truck continues along the road in daylight, with sunlight entering the cab. 이현우: He remains in the driver's seat, still bearing the injuries from the village fighting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned on the right side of the truck cab, framing a close-up from slightly behind and to the right of the driver. 이현우 occupies the driver's seat, dominating the right and center foreground of the screen, facing leftward toward the front of the vehicle. The steering wheel is positioned just in front of him, appearing in the center-left midground of the shot. The front windshield spans the background across the top and left of the frame. Bright sunlight streams through this windshield from the upper background, heavily illuminating his face.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앞창문으로 쏟아지는 햇살을 받은 이현우가 입술을 꽉 깨물고 결연한 눈빛으로 앞을 쏘아보는 굳은 얼굴 클로즈업.\n\nLOCATION (lock): At the wheel inside the truck's compact cab, with strong sunlight entering directly through the front windshield. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Sunlight entering through the windshield brightens his face with expressive contrast while preserving detail in his set jaw and eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The operational old truck continues along the road in daylight, with sunlight entering the cab. 이현우: He remains in the driver's seat, still bearing the injuries from the village fighting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned on the right side of the truck cab, framing a close-up from slightly behind and to the right of the driver. 이현우 occupies the driver's seat, dominating the right and center foreground of the screen, facing leftward toward the front of the vehicle. The steering wheel is positioned just in front of him, appearing in the center-left midground of the shot. The front windshield spans the background across the top and left of the frame. Bright sunlight streams through this windshield from the upper background, heavily illuminating his face.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앞창문으로 쏟아지는 햇살을 받은 이현우가 입술을 꽉 깨물고 결연한 눈빛으로 앞을 쏘아보는 굳은 얼굴 클로즈업.\n\nLOCATION (lock): At the wheel inside the truck's compact cab, with strong sunlight entering directly through the front windshield. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Sunlight entering through the windshield brightens his face with expressive contrast while preserving detail in his set jaw and eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The operational old truck continues along the road in daylight, with sunlight entering the cab. 이현우: He remains in the driver's seat, still bearing the injuries from the village fighting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S63sh14_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S63sh14_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "앞을 쏘아보며 아랫입술을 꽉 깨물고 있음.",
    "built_space": "트럭 내부. 운전자가 평면도와 반대로 차량 좌측 좌석에 위치함.",
    "entities": "이현우의 얼굴 특징은 일치하나, 지정된 후드티 대신 반팔 티셔츠를 착용함.",
    "hard_violations": [
     "[gemini-pro] 평면도와 반대되는 위치(좌측 운전석)에 인물 배치"
    ],
    "physics": "운전대를 쥔 손과 좌석에 앉은 자세가 자연스러움."
   },
   {
    "label": "B",
    "direction": "앞을 쏘아보지만 입술을 깨물지 않고 다물고 있음.",
    "built_space": "트럭 내부. 평면도에 맞게 차량 우측 운전석에 올바르게 배치됨.",
    "entities": "이현우의 얼굴, 얼굴 상처, 낡은 후드티 모두 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "운전대를 잡은 손과 좌석에 기댄 몸이 물리적으로 안정적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "입술을 깨무는 연기는 좋으나, 평면도와 반대인 좌측 운전석 배치와 지정된 의상 누락으로 큰 감점."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "평면도의 우측 운전석 배치와 의상, 상처를 정확히 구현했으나 입술을 깨무는 표정 디테일이 부족함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앞을 쏘아보며 아랫입술을 꽉 깨물고 있음.",
        "built_space": "트럭 내부. 운전자가 평면도와 반대로 차량 좌측 좌석에 위치함.",
        "entities": "이현우의 얼굴 특징은 일치하나, 지정된 후드티 대신 반팔 티셔츠를 착용함.",
        "hard_violations": [
         "평면도와 반대되는 위치(좌측 운전석)에 인물 배치"
        ],
        "physics": "운전대를 쥔 손과 좌석에 앉은 자세가 자연스러움."
       },
       {
        "label": "B",
        "direction": "앞을 쏘아보지만 입술을 깨물지 않고 다물고 있음.",
        "built_space": "트럭 내부. 평면도에 맞게 차량 우측 운전석에 올바르게 배치됨.",
        "entities": "이현우의 얼굴, 얼굴 상처, 낡은 후드티 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "운전대를 잡은 손과 좌석에 기댄 몸이 물리적으로 안정적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "입술을 깨무는 연기는 좋으나, 평면도와 반대인 좌측 운전석 배치와 지정된 의상 누락으로 큰 감점."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "평면도의 우측 운전석 배치와 의상, 상처를 정확히 구현했으나 입술을 깨무는 표정 디테일이 부족함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앞을 쏘아보며 아랫입술을 꽉 깨물고 있음.",
        "built_space": "트럭 내부. 운전자가 평면도와 반대로 차량 좌측 좌석에 위치함.",
        "entities": "이현우의 얼굴 특징은 일치하나, 지정된 후드티 대신 반팔 티셔츠를 착용함.",
        "hard_violations": [
         "평면도와 반대되는 위치(좌측 운전석)에 인물 배치"
        ],
        "physics": "운전대를 쥔 손과 좌석에 앉은 자세가 자연스러움."
       },
       {
        "label": "B",
        "direction": "앞을 쏘아보지만 입술을 깨물지 않고 다물고 있음.",
        "built_space": "트럭 내부. 평면도에 맞게 차량 우측 운전석에 올바르게 배치됨.",
        "entities": "이현우의 얼굴, 얼굴 상처, 낡은 후드티 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "운전대를 잡은 손과 좌석에 기댄 몸이 물리적으로 안정적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "참조의 낡은 후드와 얼굴 부상은 유지했지만, 시선이 주행 전방보다 카메라 쪽으로 돌아왔고 입술을 깨무는 동작과 얼굴 중심 클로즈업이 부족하다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전방을 응시하며 아랫입술을 깨무는 핵심 연기는 더 정확하지만, 운전대와 양팔까지 드러낸 넓은 구도와 후드 대신 입은 반소매 상의가 지시에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "운전대와 몸통은 화면 왼쪽 앞유리 쪽을 향하지만, 얼굴은 카메라 쪽으로 돌아와 있다. 눈도 앞유리 너머 도로보다 카메라에 가까운 방향을 응시하여, 주행 전방을 쏘아보라는 지시와 어긋난다. 입술은 굳게 오므렸으나 깨물고 있는 모습은 명확하지 않다.",
        "built_space": "운전대 하나, 계기판 하나, 실내 거울 하나, 선바이저 하나, 앞유리와 운전자 옆창, 기둥 손잡이 하나가 보인다. 인물은 운전대 뒤 운전석 위치에 있으며 설비 중복이나 불가능한 반사는 보이지 않는다. 다만 앞유리와 도로가 화면 왼쪽의 큰 면적을 차지하고 대시보드까지 보여, 앞유리는 좁고 비스듬한 일부만 남기라는 구도와 다르다. 첨부 자료에 이전 장면의 실사 사진은 없어 마모와 설비의 정확한 연속성까지 확인할 수는 없다.",
        "entities": "인물은 한 명이며, 한국계로 설정된 참조와 부합하는 동아시아계의 젊은 남성 얼굴, 헝클어진 검은 머리, 낡은 짙은 회색 후드가 보인다. 볼과 코 주변의 상처 및 입술의 피가 전투 부상 설정을 전달한다. 정상적인 눈과 피부 질감, 낡은 트럭 실내, 낮의 햇빛이 표현되어 있다. 얼굴뿐 아니라 어깨와 팔까지 크게 포함되어 요청한 얼굴 클로즈업보다 넓다.",
        "hard_violations": [],
        "physics": "보이는 손은 운전대 왼쪽 테두리를 잡고 있으며 팔과 어깨의 연결이 자연스럽다. 몸통은 운전석에 앉은 자세이고 뒤쪽에 좌석 일부가 보인다. 골반과 발의 접촉은 프레임 밖이므로 확인할 수 없지만, 공중에 뜬 신체나 지지 없이 떠 있는 물체는 보이지 않는다. 앞유리 쪽 주광이 얼굴에 닿는 상황도 가능하다."
       },
       {
        "label": "B",
        "direction": "얼굴과 보이는 눈은 화면 왼쪽 앞유리 너머 주행 전방을 향한다. 카메라를 보지 않고 앞을 주시하며, 윗니로 아랫입술을 누르는 동작이 뚜렷하다. 양손은 전방에 놓인 운전대를 잡아 시선과 운전 동작의 방향이 일치한다.",
        "built_space": "운전대 하나, 계기판 하나, 실내 거울 하나, 넓은 앞유리, 가까운 측면 창틀과 작은 삼각창, 인물 뒤 운전석 등받이 일부 및 반대편 빈 좌석 하나가 보인다. 운전자 옆 창문에서 안쪽을 보는 시점이며 인물은 운전대 뒤에 자리한다. 중복된 설비나 불가능한 반사는 보이지 않는다. 다만 앞유리, 대시보드, 좌석까지 넓게 담아 앞유리를 좁은 가장자리로 제한한 얼굴 클로즈업 지시를 충족하지 못한다.",
        "entities": "한 명의 젊은 동아시아계 남성이 등장하며 검은 헝클어진 머리와 얼굴 윤곽은 참조 인물에 대체로 부합한다. 목과 손의 긁힌 상처, 얼굴의 오염은 부상 지속 설정에 맞는다. 그러나 참조의 낡은 후드 대신 짙은 반소매 상의를 입고 있다. 낡은 트럭과 낮의 강한 햇빛은 확인되며 눈은 정상적인 사람의 눈으로 표현되어 있다.",
        "hard_violations": [],
        "physics": "양손이 운전대 양쪽 테두리를 실제로 감싸고 있고 손목과 팔꿈치의 연결이 자연스럽다. 몸은 뒤의 운전석 등받이 앞에 앉아 있으며 하단에 허벅지 일부가 보여 착석 자세가 성립한다. 운전대는 조향 장치에 연결되어 있고 지지 없이 뜬 물체나 신체는 없다. 입술을 깨무는 동작도 실제 얼굴 근육과 치아로 가능한 형태다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "참조의 낡은 후드와 얼굴 부상은 유지했지만, 시선이 주행 전방보다 카메라 쪽으로 돌아왔고 입술을 깨무는 동작과 얼굴 중심 클로즈업이 부족하다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "전방을 응시하며 아랫입술을 깨무는 핵심 연기는 더 정확하지만, 운전대와 양팔까지 드러낸 넓은 구도와 후드 대신 입은 반소매 상의가 지시에서 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "운전대와 몸통은 화면 왼쪽 앞유리 쪽을 향하지만, 얼굴은 카메라 쪽으로 돌아와 있다. 눈도 앞유리 너머 도로보다 카메라에 가까운 방향을 응시하여, 주행 전방을 쏘아보라는 지시와 어긋난다. 입술은 굳게 오므렸으나 깨물고 있는 모습은 명확하지 않다.",
        "built_space": "운전대 하나, 계기판 하나, 실내 거울 하나, 선바이저 하나, 앞유리와 운전자 옆창, 기둥 손잡이 하나가 보인다. 인물은 운전대 뒤 운전석 위치에 있으며 설비 중복이나 불가능한 반사는 보이지 않는다. 다만 앞유리와 도로가 화면 왼쪽의 큰 면적을 차지하고 대시보드까지 보여, 앞유리는 좁고 비스듬한 일부만 남기라는 구도와 다르다. 첨부 자료에 이전 장면의 실사 사진은 없어 마모와 설비의 정확한 연속성까지 확인할 수는 없다.",
        "entities": "인물은 한 명이며, 한국계로 설정된 참조와 부합하는 동아시아계의 젊은 남성 얼굴, 헝클어진 검은 머리, 낡은 짙은 회색 후드가 보인다. 볼과 코 주변의 상처 및 입술의 피가 전투 부상 설정을 전달한다. 정상적인 눈과 피부 질감, 낡은 트럭 실내, 낮의 햇빛이 표현되어 있다. 얼굴뿐 아니라 어깨와 팔까지 크게 포함되어 요청한 얼굴 클로즈업보다 넓다.",
        "hard_violations": [],
        "physics": "보이는 손은 운전대 왼쪽 테두리를 잡고 있으며 팔과 어깨의 연결이 자연스럽다. 몸통은 운전석에 앉은 자세이고 뒤쪽에 좌석 일부가 보인다. 골반과 발의 접촉은 프레임 밖이므로 확인할 수 없지만, 공중에 뜬 신체나 지지 없이 떠 있는 물체는 보이지 않는다. 앞유리 쪽 주광이 얼굴에 닿는 상황도 가능하다."
       },
       {
        "label": "A",
        "direction": "얼굴과 보이는 눈은 화면 왼쪽 앞유리 너머 주행 전방을 향한다. 카메라를 보지 않고 앞을 주시하며, 윗니로 아랫입술을 누르는 동작이 뚜렷하다. 양손은 전방에 놓인 운전대를 잡아 시선과 운전 동작의 방향이 일치한다.",
        "built_space": "운전대 하나, 계기판 하나, 실내 거울 하나, 넓은 앞유리, 가까운 측면 창틀과 작은 삼각창, 인물 뒤 운전석 등받이 일부 및 반대편 빈 좌석 하나가 보인다. 운전자 옆 창문에서 안쪽을 보는 시점이며 인물은 운전대 뒤에 자리한다. 중복된 설비나 불가능한 반사는 보이지 않는다. 다만 앞유리, 대시보드, 좌석까지 넓게 담아 앞유리를 좁은 가장자리로 제한한 얼굴 클로즈업 지시를 충족하지 못한다.",
        "entities": "한 명의 젊은 동아시아계 남성이 등장하며 검은 헝클어진 머리와 얼굴 윤곽은 참조 인물에 대체로 부합한다. 목과 손의 긁힌 상처, 얼굴의 오염은 부상 지속 설정에 맞는다. 그러나 참조의 낡은 후드 대신 짙은 반소매 상의를 입고 있다. 낡은 트럭과 낮의 강한 햇빛은 확인되며 눈은 정상적인 사람의 눈으로 표현되어 있다.",
        "hard_violations": [],
        "physics": "양손이 운전대 양쪽 테두리를 실제로 감싸고 있고 손목과 팔꿈치의 연결이 자연스럽다. 몸은 뒤의 운전석 등받이 앞에 앉아 있으며 하단에 허벅지 일부가 보여 착석 자세가 성립한다. 운전대는 조향 장치에 연결되어 있고 지지 없이 뜬 물체나 신체는 없다. 입술을 깨무는 동작도 실제 얼굴 근육과 치아로 가능한 형태다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.667
   },
   "violations": {
    "A": [
     "[gemini-pro] 평면도와 반대되는 위치(좌측 운전석)에 인물 배치"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "A": 1179,
   "B": 1667
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "입술을 깨무는 연기는 좋으나, 평면도와 반대인 좌측 운전석 배치와 지정된 의상 누락으로 큰 감점.  ★위반: [gemini-pro] 평면도와 반대되는 위치(좌측 운전석)에 인물 배치"
   },
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "평면도의 우측 운전석 배치와 의상, 상처를 정확히 구현했으나 입술을 깨무는 표정 디테일이 부족함."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S63sh14_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "이현우",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdfa-89b2-7cf9-a126-d5477c85820b",
  "confined_fp": {
   "base_key": "confinedfp::1eb557c77df4",
   "apt_reason": "트럭 운전석 내부라는 제한되고 통제 장치가 중요한 공간에서 운전석에 앉은 인물을 찍는 숏입니다. 운전대와 앞창문, 좌석의 상대적인 공간 배치가 조금이라도 틀리면 서사적 몰입을 깨는 치명적인 오류가 발생하므로 평면도 레이아웃 보조가 필요합니다.",
   "fixed": true,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S63sh7"
  }
 },
 "S63sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:52:24.684947+00:00",
  "fingerprint": "b08daad691882609d8539e5ea39fefcea3bd5b9dc167d16f9c78edf5fb9c9463",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S63sh14_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S63sh14_sel.png",
  "source_sha256": "eb5cfee8545d65122ff184de731bb82d1fa71ac11242a33d35affac94b746d0d",
  "file": "S63sh14_cine.png",
  "staged_sha256": "608be2a35b17e523d5dde0bcc8f713d714f5c5a1386edaf8f538d00cb9fa830f",
  "latency_ms": 12900
 },
 "S64sh9::signage": {
  "fp": "16373fb83c44c785",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S64sh9": {
  "input_fingerprint": "b54828198f11c85d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200이 자신의 팔에 달린 거대한 고사포를 찰리의 가슴을 향해 맹렬히 겨눈 위압적인 구도.\n\nLOCATION (lock): In a dark recessed area inside a large abandoned machine-parts warehouse, with daylight entering through its half-open door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: B-200's arm-mounted cannon (Raised and aimed at 찰리's chest) — The barrel is seen obliquely from the side, never head-on into its bore; used as Creates a readable diagonal threat between the separated bodies without exaggerated foreground scale; Broken robots (Piled in a warehouse corner) — Overlapping bodies and parts remain behind the confrontation; used as Secondary background context connecting B-200 to the warehouse's discarded machines.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep B-200 within the warehouse darkness, using restrained tonal separation to make the cannon's aim and both bodies readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse door remains half open beside heaps of broken robots; baby birds are already concealed behind a box in B-200's dark corner. B-200 retains his modified gun arms, while Charlie carries the football and retains his dented, perforated body, damaged chest ring and faulty language system.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by B-200 right now, so B-200's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to B-200: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200이 자신의 팔에 달린 거대한 고사포를 찰리의 가슴을 향해 맹렬히 겨눈 위압적인 구도.\n\nLOCATION (lock): In a dark recessed area inside a large abandoned machine-parts warehouse, with daylight entering through its half-open door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: B-200's arm-mounted cannon (Raised and aimed at 찰리's chest) — The barrel is seen obliquely from the side, never head-on into its bore; used as Creates a readable diagonal threat between the separated bodies without exaggerated foreground scale; Broken robots (Piled in a warehouse corner) — Overlapping bodies and parts remain behind the confrontation; used as Secondary background context connecting B-200 to the warehouse's discarded machines.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep B-200 within the warehouse darkness, using restrained tonal separation to make the cannon's aim and both bodies readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse door remains half open beside heaps of broken robots; baby birds are already concealed behind a box in B-200's dark corner. B-200 retains his modified gun arms, while Charlie carries the football and retains his dented, perforated body, damaged chest ring and faulty language system.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by B-200 right now, so B-200's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to B-200: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200이 자신의 팔에 달린 거대한 고사포를 찰리의 가슴을 향해 맹렬히 겨눈 위압적인 구도.\n\nLOCATION (lock): In a dark recessed area inside a large abandoned machine-parts warehouse, with daylight entering through its half-open door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: B-200's arm-mounted cannon (Raised and aimed at 찰리's chest) — The barrel is seen obliquely from the side, never head-on into its bore; used as Creates a readable diagonal threat between the separated bodies without exaggerated foreground scale; Broken robots (Piled in a warehouse corner) — Overlapping bodies and parts remain behind the confrontation; used as Secondary background context connecting B-200 to the warehouse's discarded machines.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep B-200 within the warehouse darkness, using restrained tonal separation to make the cannon's aim and both bodies readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse door remains half open beside heaps of broken robots; baby birds are already concealed behind a box in B-200's dark corner. B-200 retains his modified gun arms, while Charlie carries the football and retains his dented, perforated body, damaged chest ring and faulty language system.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by B-200 right now, so B-200's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to B-200: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "B-200의 팔에 달린 고사포가 찰리의 가슴을 정확히 겨냥하고 있음.",
    "built_space": "창고 내부로, 중앙에 빛이 들어오는 문이 있고 주변에 부서진 로봇 부품들이 쌓여 있음.",
    "entities": "B-200은 기준 이미지와 유사함. 찰리는 고릴라 형태이나 각지지 않고 둥근 머리를 가졌으며, 타공된 몸체 대신 오돌토돌한 질감으로 표현됨. 럭비공은 들고 있으나 아기 새는 없음.",
    "hard_violations": [],
    "physics": "두 캐릭터 모두 지면에 안정적으로 서 있으며, 찰리의 손과 팔이 럭비공을 제대로 지지하고 있음."
   },
   {
    "label": "B",
    "direction": "B-200의 고사포가 찰리의 가슴을 향해 대각선 구도로 겨냥됨.",
    "built_space": "어두운 창고 내부, 우측 배경에 반쯤 열린 문이 있고 부서진 로봇 부품과 상자가 배치됨.",
    "entities": "B-200은 기준과 일치함. 찰리는 명시된 대로 각진 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입, 구멍 난 몸체를 가짐. 럭비공을 들고 있으며, 상자 위에 아기 새가 있음.",
    "hard_violations": [],
    "physics": "두 캐릭터 모두 바닥에 서 있으며, 찰리의 팔이 럭비공을 자연스럽게 받치고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리의 각진 머리와 구멍 난 몸체 등 지정된 외형 특징을 정확히 묘사하였고, 요구된 배경 요소들을 프레임 내에 충실히 배치했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리의 머리가 둥글게 표현되고 몸체의 구멍 묘사가 부족하며, 명시된 아기 새 디테일이 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "B-200의 팔에 달린 고사포가 찰리의 가슴을 정확히 겨냥하고 있음.",
        "built_space": "창고 내부로, 중앙에 빛이 들어오는 문이 있고 주변에 부서진 로봇 부품들이 쌓여 있음.",
        "entities": "B-200은 기준 이미지와 유사함. 찰리는 고릴라 형태이나 각지지 않고 둥근 머리를 가졌으며, 타공된 몸체 대신 오돌토돌한 질감으로 표현됨. 럭비공은 들고 있으나 아기 새는 없음.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 지면에 안정적으로 서 있으며, 찰리의 손과 팔이 럭비공을 제대로 지지하고 있음."
       },
       {
        "label": "B",
        "direction": "B-200의 고사포가 찰리의 가슴을 향해 대각선 구도로 겨냥됨.",
        "built_space": "어두운 창고 내부, 우측 배경에 반쯤 열린 문이 있고 부서진 로봇 부품과 상자가 배치됨.",
        "entities": "B-200은 기준과 일치함. 찰리는 명시된 대로 각진 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입, 구멍 난 몸체를 가짐. 럭비공을 들고 있으며, 상자 위에 아기 새가 있음.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 바닥에 서 있으며, 찰리의 팔이 럭비공을 자연스럽게 받치고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리의 각진 머리와 구멍 난 몸체 등 지정된 외형 특징을 정확히 묘사하였고, 요구된 배경 요소들을 프레임 내에 충실히 배치했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리의 머리가 둥글게 표현되고 몸체의 구멍 묘사가 부족하며, 명시된 아기 새 디테일이 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "B-200의 팔에 달린 고사포가 찰리의 가슴을 정확히 겨냥하고 있음.",
        "built_space": "창고 내부로, 중앙에 빛이 들어오는 문이 있고 주변에 부서진 로봇 부품들이 쌓여 있음.",
        "entities": "B-200은 기준 이미지와 유사함. 찰리는 고릴라 형태이나 각지지 않고 둥근 머리를 가졌으며, 타공된 몸체 대신 오돌토돌한 질감으로 표현됨. 럭비공은 들고 있으나 아기 새는 없음.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 지면에 안정적으로 서 있으며, 찰리의 손과 팔이 럭비공을 제대로 지지하고 있음."
       },
       {
        "label": "B",
        "direction": "B-200의 고사포가 찰리의 가슴을 향해 대각선 구도로 겨냥됨.",
        "built_space": "어두운 창고 내부, 우측 배경에 반쯤 열린 문이 있고 부서진 로봇 부품과 상자가 배치됨.",
        "entities": "B-200은 기준과 일치함. 찰리는 명시된 대로 각진 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입, 구멍 난 몸체를 가짐. 럭비공을 들고 있으며, 상자 위에 아기 새가 있음.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 바닥에 서 있으며, 찰리의 팔이 럭비공을 자연스럽게 받치고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "가슴을 향한 사선 조준과 어두운 창고는 잘 구현했지만, 무릎까지 넓어진 구도와 상자 위에 노출된 새들이 미디엄 숏 및 은폐 상태 지시에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 인물의 상체와 가슴을 겨누는 팔 장착 포에 더 집중하며 반쯤 열린 문과 후방 잔해를 유지하지만, 포신의 대각선 위협은 약하고 구도도 미디엄 숏보다 조금 넓습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 B-200의 얼굴은 오른쪽 찰리를 향하고, 찰리도 B-200 쪽을 바라본다. 포신은 화면 오른쪽 아래로 완만하게 뻗어 찰리의 가슴 고리 왼쪽 흉부를 겨눈다. 포구 내부를 정면으로 보여주지 않고 포신 측면을 보여준다.",
        "built_space": "낡은 철골 지붕, 높은 벽창 여러 개, 상부의 매달린 조명 한 개, 오른쪽 뒤의 열린 출입구 한 곳이 보인다. 문짝 하나가 출입구 일부를 가리며 낮빛이 들어온다. 두 인물 뒤에는 선반과 겹쳐 쌓인 폐로봇이 있고, 왼쪽 앞에는 나무상자가 있다. 참고 장소의 노후한 창고 재료와 적재 환경은 유사하나 정확한 동일 지점인지는 확인하기 어렵다.",
        "entities": "활동하는 인물은 기계형 B-200과 찰리 두 명이다. B-200은 금속 머리와 어두운 철제 외장, 굵은 기계식 전완 및 팔 장착 다연장 포를 갖췄다. 찰리는 각진 흰 얼굴판, 점눈, 선 모양 입, 고릴라형 머리와 패이고 뚫린 몸체를 갖췄으며 가슴 고리와 미식축구공도 보인다. 고리 자체의 파손은 뚜렷하지 않다. 왼쪽 상자 위에는 작은 새 두 마리가 드러나 있어 이미 상자 뒤에 숨겨졌다는 상태와 다르다. 언어 시스템의 고장은 정지 화면으로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "포는 B-200의 전완 장착부와 팔꿈치에 연결되어 있으며 별도로 떠 있지 않다. 포를 쥔 독립된 손가락은 보이지 않고 팔 자체가 무기로 이어진다. 찰리는 손과 굽힌 팔로 공을 몸 옆에 받친다. 두 인물의 다리는 화면 아래로 이어져 있으며 공중에 뜬 자세는 아니다. 폐로봇은 바닥과 서로의 몸체에 기대고, 새들은 상자 가장자리에 앉아 있다."
       },
       {
        "label": "B",
        "direction": "왼쪽 B-200과 오른쪽 찰리의 얼굴은 서로를 향한다. B-200의 포신은 거의 수평으로 오른쪽을 향하며 연장선이 찰리의 가슴 고리 부근 흉부에 닿는다. 조준 대상은 명확하고 포신은 측면으로 보이지만, 요구된 대각선 연결은 A보다 약하다.",
        "built_space": "뒤쪽 출입구 한 곳에서 문짝 하나가 오른쪽 부분을 가리고 왼쪽 틈으로 강한 낮빛이 들어온다. 철제 기둥과 가로보, 낡은 벽, 왼쪽 선반, 오른쪽 상자 및 중앙 뒤 폐로봇 더미가 보인다. B-200은 실내 그늘에 있고 찰리는 그 오른쪽에 서며, 잔해는 대치하는 두 몸 뒤에 놓인다. 참고의 창고 재료와 분위기를 유지하지만 참고 사진의 세부 배치를 모두 확인할 수 있는 시야는 아니다.",
        "entities": "B-200과 찰리 외에 새로운 활동 인물은 없다. B-200의 금속 머리, 어두운 중장갑과 팔에 연결된 거대한 포가 보인다. 찰리는 흰 고릴라형 얼굴판과 선형 입, 손상되고 구멍 난 금속 몸체, 원형 가슴 장치 및 들고 있는 미식축구공을 갖췄다. 다만 얼굴판은 요구한 각진 형태보다 둥글고 눈에는 점보다 복잡한 기계 테두리가 있다. 푸른 가슴 고리에는 마모가 있으나 심한 파손은 분명하지 않다. 새는 보이지 않아 은폐 상태와 충돌하지 않는다.",
        "hard_violations": [],
        "physics": "큰 포의 후방이 B-200의 전완과 팔꿈치에 결합되어 하중을 받으며, 아래쪽 보조 구조도 포 본체에 연결되어 있다. 독립된 손으로 손잡이를 쥔 모습은 확인되지 않는다. 찰리의 손가락과 전완이 공의 아래와 옆을 받친다. 두 몸은 직립한 채 다리가 프레임 아래로 이어지고, 폐로봇과 상자는 바닥에 쌓여 있어 근거 없는 부유는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "가슴을 향한 사선 조준과 어두운 창고는 잘 구현했지만, 무릎까지 넓어진 구도와 상자 위에 노출된 새들이 미디엄 숏 및 은폐 상태 지시에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 인물의 상체와 가슴을 겨누는 팔 장착 포에 더 집중하며 반쯤 열린 문과 후방 잔해를 유지하지만, 포신의 대각선 위협은 약하고 구도도 미디엄 숏보다 조금 넓습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 B-200의 얼굴은 오른쪽 찰리를 향하고, 찰리도 B-200 쪽을 바라본다. 포신은 화면 오른쪽 아래로 완만하게 뻗어 찰리의 가슴 고리 왼쪽 흉부를 겨눈다. 포구 내부를 정면으로 보여주지 않고 포신 측면을 보여준다.",
        "built_space": "낡은 철골 지붕, 높은 벽창 여러 개, 상부의 매달린 조명 한 개, 오른쪽 뒤의 열린 출입구 한 곳이 보인다. 문짝 하나가 출입구 일부를 가리며 낮빛이 들어온다. 두 인물 뒤에는 선반과 겹쳐 쌓인 폐로봇이 있고, 왼쪽 앞에는 나무상자가 있다. 참고 장소의 노후한 창고 재료와 적재 환경은 유사하나 정확한 동일 지점인지는 확인하기 어렵다.",
        "entities": "활동하는 인물은 기계형 B-200과 찰리 두 명이다. B-200은 금속 머리와 어두운 철제 외장, 굵은 기계식 전완 및 팔 장착 다연장 포를 갖췄다. 찰리는 각진 흰 얼굴판, 점눈, 선 모양 입, 고릴라형 머리와 패이고 뚫린 몸체를 갖췄으며 가슴 고리와 미식축구공도 보인다. 고리 자체의 파손은 뚜렷하지 않다. 왼쪽 상자 위에는 작은 새 두 마리가 드러나 있어 이미 상자 뒤에 숨겨졌다는 상태와 다르다. 언어 시스템의 고장은 정지 화면으로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "포는 B-200의 전완 장착부와 팔꿈치에 연결되어 있으며 별도로 떠 있지 않다. 포를 쥔 독립된 손가락은 보이지 않고 팔 자체가 무기로 이어진다. 찰리는 손과 굽힌 팔로 공을 몸 옆에 받친다. 두 인물의 다리는 화면 아래로 이어져 있으며 공중에 뜬 자세는 아니다. 폐로봇은 바닥과 서로의 몸체에 기대고, 새들은 상자 가장자리에 앉아 있다."
       },
       {
        "label": "A",
        "direction": "왼쪽 B-200과 오른쪽 찰리의 얼굴은 서로를 향한다. B-200의 포신은 거의 수평으로 오른쪽을 향하며 연장선이 찰리의 가슴 고리 부근 흉부에 닿는다. 조준 대상은 명확하고 포신은 측면으로 보이지만, 요구된 대각선 연결은 A보다 약하다.",
        "built_space": "뒤쪽 출입구 한 곳에서 문짝 하나가 오른쪽 부분을 가리고 왼쪽 틈으로 강한 낮빛이 들어온다. 철제 기둥과 가로보, 낡은 벽, 왼쪽 선반, 오른쪽 상자 및 중앙 뒤 폐로봇 더미가 보인다. B-200은 실내 그늘에 있고 찰리는 그 오른쪽에 서며, 잔해는 대치하는 두 몸 뒤에 놓인다. 참고의 창고 재료와 분위기를 유지하지만 참고 사진의 세부 배치를 모두 확인할 수 있는 시야는 아니다.",
        "entities": "B-200과 찰리 외에 새로운 활동 인물은 없다. B-200의 금속 머리, 어두운 중장갑과 팔에 연결된 거대한 포가 보인다. 찰리는 흰 고릴라형 얼굴판과 선형 입, 손상되고 구멍 난 금속 몸체, 원형 가슴 장치 및 들고 있는 미식축구공을 갖췄다. 다만 얼굴판은 요구한 각진 형태보다 둥글고 눈에는 점보다 복잡한 기계 테두리가 있다. 푸른 가슴 고리에는 마모가 있으나 심한 파손은 분명하지 않다. 새는 보이지 않아 은폐 상태와 충돌하지 않는다.",
        "hard_violations": [],
        "physics": "큰 포의 후방이 B-200의 전완과 팔꿈치에 결합되어 하중을 받으며, 아래쪽 보조 구조도 포 본체에 연결되어 있다. 독립된 손으로 손잡이를 쥔 모습은 확인되지 않는다. 찰리의 손가락과 전완이 공의 아래와 옆을 받친다. 두 몸은 직립한 채 다리가 프레임 아래로 이어지고, 폐로봇과 상자는 바닥에 쌓여 있어 근거 없는 부유는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1750,
   "A": 1571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "찰리의 각진 머리와 구멍 난 몸체 등 지정된 외형 특징을 정확히 묘사하였고, 요구된 배경 요소들을 프레임 내에 충실히 배치했습니다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "찰리의 머리가 둥글게 표현되고 몸체의 구멍 묘사가 부족하며, 명시된 아기 새 디테일이 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L98B01.png",
    "asset_id": "87c288f8-7766-4887-9147-7f2d612268cc",
    "role": "location_plate"
   },
   {
    "label": "HAND OWNER REFERENCE — B-200: the person whose hand is on the object in this shot. Use this photograph ONLY to get that hand right — its size, build, skin tone and texture, apparent age, grooming, nails, sleeve and anything worn on the wrist or fingers. Do NOT bring their face, body or clothing into the frame; the framing stays on the object and reaches no further than the wrist and forearm.",
    "path": "<bytes:1110018>",
    "asset_id": "a70896ff-b489-4df8-a200-1c23c16bdec1",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdfb-5ea0-7978-8613-c38b84845fce",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S64sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:52:37.928994+00:00",
  "fingerprint": "d1c4d2ed49fd253047c494510e6e7eedc72225d7934b59ec36f1680030eba1e5",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S64sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S64sh9_sel.png",
  "source_sha256": "b24acf62a9320b51e8bd61541799dac312ced2fc1cd6a69d8879f87d4b3206cb",
  "file": "S64sh9_cine.png",
  "staged_sha256": "d258e93e972d3c3b0ca4b8119e275c1b1aa28dbfacc404a885e056d35150ce40",
  "latency_ms": 13737
 },
 "S64sh18::signage": {
  "fp": "79e684fa9a2bf631",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S64sh18": {
  "input_fingerprint": "9533b6d431c05740",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200이 찰리의 가슴 중앙에 달린 둥근 링 부품을 향해 굵은 금속 손가락을 뻗은 채 정확히 가리키고 있는 손끝.\n\nLOCATION (lock): In the shadowed sitting area among stored machine parts inside the large warehouse, dimly lit from the half-open entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리's chest ring (Attached at the center of his chest) — Its circular front is visible obliquely on the chest rather than flattened into a straight-on product view; used as Small focal destination for B-200's pointing finger, with the surrounding torso retained.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the subdued warehouse illumination, with sufficient local contrast to separate the fingertip, its intervening gap, and the chest ring.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse door remains half open, with broken robots piled in one corner and baby birds still hidden behind a box. B-200's modified gun arms remain intact; Charlie is seated with his damaged chest ring, body dents and holes, and impaired language system.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200이 찰리의 가슴 중앙에 달린 둥근 링 부품을 향해 굵은 금속 손가락을 뻗은 채 정확히 가리키고 있는 손끝.\n\nLOCATION (lock): In the shadowed sitting area among stored machine parts inside the large warehouse, dimly lit from the half-open entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리's chest ring (Attached at the center of his chest) — Its circular front is visible obliquely on the chest rather than flattened into a straight-on product view; used as Small focal destination for B-200's pointing finger, with the surrounding torso retained.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the subdued warehouse illumination, with sufficient local contrast to separate the fingertip, its intervening gap, and the chest ring.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse door remains half open, with broken robots piled in one corner and baby birds still hidden behind a box. B-200's modified gun arms remain intact; Charlie is seated with his damaged chest ring, body dents and holes, and impaired language system.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200이 찰리의 가슴 중앙에 달린 둥근 링 부품을 향해 굵은 금속 손가락을 뻗은 채 정확히 가리키고 있는 손끝.\n\nLOCATION (lock): In the shadowed sitting area among stored machine parts inside the large warehouse, dimly lit from the half-open entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리's chest ring (Attached at the center of his chest) — Its circular front is visible obliquely on the chest rather than flattened into a straight-on product view; used as Small focal destination for B-200's pointing finger, with the surrounding torso retained.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the subdued warehouse illumination, with sufficient local contrast to separate the fingertip, its intervening gap, and the chest ring.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse door remains half open, with broken robots piled in one corner and baby birds still hidden behind a box. B-200's modified gun arms remain intact; Charlie is seated with his damaged chest ring, body dents and holes, and impaired language system.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "B-200의 굵은 금속 손가락이 찰리의 가슴 중앙 링을 향해 정확히 조준하고 있으며 중간에 틈이 존재함.",
    "built_space": "흐릿한 배경에 창고 내부의 기계 부품들이 적절한 비율로 배치되어 있음.",
    "entities": "B-200의 낡은 금속 손과 총탄 자국이 있는 찰리의 흉부 및 링 부품이 레퍼런스와 일치하게 묘사됨.",
    "hard_violations": [],
    "physics": "손과 흉부의 위치가 자연스럽게 고정되어 있으며 물리적으로 어색한 묘사가 없음."
   },
   {
    "label": "B",
    "direction": "금속 손가락이 가슴의 링 부품을 향하고 있으나, 간격 없이 표면에 닿아 있음.",
    "built_space": "아웃포커싱된 배경에 창고의 반쯤 열린 문과 내부 공간이 보임.",
    "entities": "레퍼런스에 부합하는 B-200의 기계 손과 찰리의 손상된 가슴 외장이 나타남.",
    "hard_violations": [],
    "physics": "뻗은 팔과 가슴 장갑의 배치가 안정적이며 지탱에 문제가 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "손가락 끝과 가슴 링 부품 사이의 간격을 명시한 프롬프트의 지시를 정확히 구현하여 훌륭한 구도를 보여줍니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "손가락 끝이 링 부품에 닿아 있어 프롬프트가 요구한 '손가락 끝과 링 사이의 간격'을 표현하는 데 실패했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "B-200의 굵은 금속 손가락이 찰리의 가슴 중앙 링을 향해 정확히 조준하고 있으며 중간에 틈이 존재함.",
        "built_space": "흐릿한 배경에 창고 내부의 기계 부품들이 적절한 비율로 배치되어 있음.",
        "entities": "B-200의 낡은 금속 손과 총탄 자국이 있는 찰리의 흉부 및 링 부품이 레퍼런스와 일치하게 묘사됨.",
        "hard_violations": [],
        "physics": "손과 흉부의 위치가 자연스럽게 고정되어 있으며 물리적으로 어색한 묘사가 없음."
       },
       {
        "label": "B",
        "direction": "금속 손가락이 가슴의 링 부품을 향하고 있으나, 간격 없이 표면에 닿아 있음.",
        "built_space": "아웃포커싱된 배경에 창고의 반쯤 열린 문과 내부 공간이 보임.",
        "entities": "레퍼런스에 부합하는 B-200의 기계 손과 찰리의 손상된 가슴 외장이 나타남.",
        "hard_violations": [],
        "physics": "뻗은 팔과 가슴 장갑의 배치가 안정적이며 지탱에 문제가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "손가락 끝과 가슴 링 부품 사이의 간격을 명시한 프롬프트의 지시를 정확히 구현하여 훌륭한 구도를 보여줍니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "손가락 끝이 링 부품에 닿아 있어 프롬프트가 요구한 '손가락 끝과 링 사이의 간격'을 표현하는 데 실패했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "B-200의 굵은 금속 손가락이 찰리의 가슴 중앙 링을 향해 정확히 조준하고 있으며 중간에 틈이 존재함.",
        "built_space": "흐릿한 배경에 창고 내부의 기계 부품들이 적절한 비율로 배치되어 있음.",
        "entities": "B-200의 낡은 금속 손과 총탄 자국이 있는 찰리의 흉부 및 링 부품이 레퍼런스와 일치하게 묘사됨.",
        "hard_violations": [],
        "physics": "손과 흉부의 위치가 자연스럽게 고정되어 있으며 물리적으로 어색한 묘사가 없음."
       },
       {
        "label": "B",
        "direction": "금속 손가락이 가슴의 링 부품을 향하고 있으나, 간격 없이 표면에 닿아 있음.",
        "built_space": "아웃포커싱된 배경에 창고의 반쯤 열린 문과 내부 공간이 보임.",
        "entities": "레퍼런스에 부합하는 B-200의 기계 손과 찰리의 손상된 가슴 외장이 나타남.",
        "hard_violations": [],
        "physics": "뻗은 팔과 가슴 장갑의 배치가 안정적이며 지탱에 문제가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "가슴 링을 정확히 겨누는 클로즈업이지만 손끝이 링 안쪽에 닿아 보이므로, 명시된 손끝과 링 사이의 간격을 구현하지 못했다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "굵은 금속 손끝이 비스듬히 보이는 가슴 링을 향하며, 접촉 전 간격과 주변 몸통을 함께 확보해 지시된 순간을 더 충실하게 구현했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽에서 뻗은 금속 검지가 오른쪽 찰리의 가슴 링 내부를 직접 가리킨다. 손끝은 링 왼쪽 안쪽 면과 겹쳐 닿거나 삽입된 듯 보이며, 분리된 빈 간격은 보이지 않는다. 눈과 총구는 화면 밖이다.",
        "built_space": "오른쪽에 찰리의 가슴, 왼쪽에 B-200의 손과 손목이 배치되어 있다. 배경에는 밝은 출입구 한 곳, 왼쪽 위 창 일부, 어두운 기계 잔해 더미가 보인다. 낡은 창고 재질과 낮빛이 들어오는 어두운 실내는 참조와 부합한다. 좌석과 하체는 잘려 있어 착석 여부는 확인할 수 없다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "B-200의 굵은 관절식 금속 손과 찰리의 회색 흉갑, 중앙의 원형 링 한 개, 여러 구멍과 긁힌 자국이 보인다. 링은 비스듬한 타원 형태로 투영되며 주변 몸통도 남아 있지만 화면에서 상당히 크게 강조된다. 참조의 낡고 손상된 기계 외관은 유지된다. 얼굴과 개조 총팔 전체는 프레임 밖이므로 세부 정체성과 보존 상태를 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "검지는 관절을 통해 손바닥과 손목에 연결되고 나머지 손가락은 접혀 있다. 팔은 왼쪽 화면 밖으로 이어지며, 링은 흉갑에 부착되어 있어 떠 있는 물체는 없다. 손끝이 링 안쪽에 접촉하는 자세 자체는 가능하지만, 요구된 비접촉 간격과는 다르다."
       },
       {
        "label": "B",
        "direction": "왼쪽 금속 검지가 오른쪽 아래 방향으로 뻗어 찰리의 가슴 링 안쪽을 겨눈다. 손끝과 링 왼쪽 테두리 사이에 분명한 간격이 남아 있어 가리키는 동작으로 읽힌다. 시선과 총구 방향은 프레임에 포함되지 않는다.",
        "built_space": "왼쪽 전경의 손과 오른쪽의 가슴 사이로 창고의 낡은 벽, 상부 창 세 곳, 기계 잔해 더미와 하단 보관 상자 일부가 보인다. 참조의 벽과 창, 적치된 기계 부품의 공간감이 이어진다. 출입구와 좌석은 이 클로즈업 밖이며, 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "B-200의 두꺼운 금속 손가락과 관절식 손, 찰리의 손상된 회색 흉갑과 중앙 링 한 개가 확인된다. 위쪽에는 찰리의 흰 얼굴판 하단도 조금 보인다. 링의 원형 앞면은 비스듬하게 보이고 흉갑의 구멍과 마모도 참조 외관에 부합한다. 링은 작은 배경 목적지라는 표현보다는 다소 크게 강조되어 있다. 총팔 전체와 하체는 화면 밖이며 추가 인물이나 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "뻗은 검지는 손바닥·손목·화면 밖 팔로 연속해서 연결되어 있고, 나머지 손가락을 접은 지시 자세도 기계 관절 구조상 가능하다. 링은 흉갑에 고정되어 있다. 손끝이 링에 닿지 않고 멈춘 상태이며, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "가슴 링을 정확히 겨누는 클로즈업이지만 손끝이 링 안쪽에 닿아 보이므로, 명시된 손끝과 링 사이의 간격을 구현하지 못했다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "굵은 금속 손끝이 비스듬히 보이는 가슴 링을 향하며, 접촉 전 간격과 주변 몸통을 함께 확보해 지시된 순간을 더 충실하게 구현했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽에서 뻗은 금속 검지가 오른쪽 찰리의 가슴 링 내부를 직접 가리킨다. 손끝은 링 왼쪽 안쪽 면과 겹쳐 닿거나 삽입된 듯 보이며, 분리된 빈 간격은 보이지 않는다. 눈과 총구는 화면 밖이다.",
        "built_space": "오른쪽에 찰리의 가슴, 왼쪽에 B-200의 손과 손목이 배치되어 있다. 배경에는 밝은 출입구 한 곳, 왼쪽 위 창 일부, 어두운 기계 잔해 더미가 보인다. 낡은 창고 재질과 낮빛이 들어오는 어두운 실내는 참조와 부합한다. 좌석과 하체는 잘려 있어 착석 여부는 확인할 수 없다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "B-200의 굵은 관절식 금속 손과 찰리의 회색 흉갑, 중앙의 원형 링 한 개, 여러 구멍과 긁힌 자국이 보인다. 링은 비스듬한 타원 형태로 투영되며 주변 몸통도 남아 있지만 화면에서 상당히 크게 강조된다. 참조의 낡고 손상된 기계 외관은 유지된다. 얼굴과 개조 총팔 전체는 프레임 밖이므로 세부 정체성과 보존 상태를 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "검지는 관절을 통해 손바닥과 손목에 연결되고 나머지 손가락은 접혀 있다. 팔은 왼쪽 화면 밖으로 이어지며, 링은 흉갑에 부착되어 있어 떠 있는 물체는 없다. 손끝이 링 안쪽에 접촉하는 자세 자체는 가능하지만, 요구된 비접촉 간격과는 다르다."
       },
       {
        "label": "A",
        "direction": "왼쪽 금속 검지가 오른쪽 아래 방향으로 뻗어 찰리의 가슴 링 안쪽을 겨눈다. 손끝과 링 왼쪽 테두리 사이에 분명한 간격이 남아 있어 가리키는 동작으로 읽힌다. 시선과 총구 방향은 프레임에 포함되지 않는다.",
        "built_space": "왼쪽 전경의 손과 오른쪽의 가슴 사이로 창고의 낡은 벽, 상부 창 세 곳, 기계 잔해 더미와 하단 보관 상자 일부가 보인다. 참조의 벽과 창, 적치된 기계 부품의 공간감이 이어진다. 출입구와 좌석은 이 클로즈업 밖이며, 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "B-200의 두꺼운 금속 손가락과 관절식 손, 찰리의 손상된 회색 흉갑과 중앙 링 한 개가 확인된다. 위쪽에는 찰리의 흰 얼굴판 하단도 조금 보인다. 링의 원형 앞면은 비스듬하게 보이고 흉갑의 구멍과 마모도 참조 외관에 부합한다. 링은 작은 배경 목적지라는 표현보다는 다소 크게 강조되어 있다. 총팔 전체와 하체는 화면 밖이며 추가 인물이나 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "뻗은 검지는 손바닥·손목·화면 밖 팔로 연속해서 연결되어 있고, 나머지 손가락을 접은 지시 자세도 기계 관절 구조상 가능하다. 링은 흉갑에 고정되어 있다. 손끝이 링에 닿지 않고 멈춘 상태이며, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.492
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.492
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1492
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "손가락 끝과 가슴 링 부품 사이의 간격을 명시한 프롬프트의 지시를 정확히 구현하여 훌륭한 구도를 보여줍니다."
   },
   {
    "label": "B",
    "score": 1492,
    "verdict_ko": "손가락 끝이 링 부품에 닿아 있어 프롬프트가 요구한 '손가락 끝과 링 사이의 간격'을 표현하는 데 실패했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of B-200, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S64sh9_sel.png",
    "asset_id": "bb5edb86-2e56-47cf-af9f-29855e16ae3d",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdfc-3fd0-7a99-b26d-0661f8376651",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S64sh9"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S64sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:52:51.998878+00:00",
  "fingerprint": "e59a50ab7f99b3fdc7dee2cbea76e7c11a3c8d6b1907bef4868d4b6182910a68",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S64sh18_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S64sh18_sel.png",
  "source_sha256": "d37123291903d547914bf67fcb80d791afca81ad1fd4df7cc2c6e96c4c998689",
  "file": "S64sh18_cine.png",
  "staged_sha256": "0fd22bf82725b51f638d64450169aafd02d72c514b0b37546e5ec942d86c733f",
  "latency_ms": 14008
 },
 "S64sh27::signage": {
  "fp": "6b816db4ff9f4dec",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S64sh27": {
  "input_fingerprint": "d00e3b80f9031079",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200이 둥지 안의 아기 새들을 향해 따뜻한 안광을 뿜은 채 조용히 내려다보는 평온한 구도.\n\nLOCATION (lock): At a concealed box among machine parts in the warehouse's shadowed interior, with weak daylight from the half-open entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Baby birds' nest (Holding the revealed baby birds) — Its interior is visible from the diagonal elevated viewpoint; used as Small lower-right anchor for B-200's downward attention; Moved box (Shifted aside to reveal the birds) — Seen obliquely at the frame periphery, clear of the birds; used as Retains evidence of the protective concealment without competing with the shared gaze.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: B-200's warm eye glow provides a restrained tender accent within the otherwise subdued warehouse darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The moved box exposes the baby birds in B-200's corner; the warehouse door remains half open and the broken-robot heap remains undisturbed. B-200 retains his gun arms, while Charlie has left the corner with his body and chest-ring damage unrepaired.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈); 아기 새 (어린 새의 머리, 짧은 부리, 어린 얼굴 깃털) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200이 둥지 안의 아기 새들을 향해 따뜻한 안광을 뿜은 채 조용히 내려다보는 평온한 구도.\n\nLOCATION (lock): At a concealed box among machine parts in the warehouse's shadowed interior, with weak daylight from the half-open entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Baby birds' nest (Holding the revealed baby birds) — Its interior is visible from the diagonal elevated viewpoint; used as Small lower-right anchor for B-200's downward attention; Moved box (Shifted aside to reveal the birds) — Seen obliquely at the frame periphery, clear of the birds; used as Retains evidence of the protective concealment without competing with the shared gaze.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: B-200's warm eye glow provides a restrained tender accent within the otherwise subdued warehouse darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The moved box exposes the baby birds in B-200's corner; the warehouse door remains half open and the broken-robot heap remains undisturbed. B-200 retains his gun arms, while Charlie has left the corner with his body and chest-ring damage unrepaired.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈); 아기 새 (어린 새의 머리, 짧은 부리, 어린 얼굴 깃털) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200이 둥지 안의 아기 새들을 향해 따뜻한 안광을 뿜은 채 조용히 내려다보는 평온한 구도.\n\nLOCATION (lock): At a concealed box among machine parts in the warehouse's shadowed interior, with weak daylight from the half-open entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Baby birds' nest (Holding the revealed baby birds) — Its interior is visible from the diagonal elevated viewpoint; used as Small lower-right anchor for B-200's downward attention; Moved box (Shifted aside to reveal the birds) — Seen obliquely at the frame periphery, clear of the birds; used as Retains evidence of the protective concealment without competing with the shared gaze.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: B-200's warm eye glow provides a restrained tender accent within the otherwise subdued warehouse darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The moved box exposes the baby birds in B-200's corner; the warehouse door remains half open and the broken-robot heap remains undisturbed. B-200 retains his gun arms, while Charlie has left the corner with his body and chest-ring damage unrepaired.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈); 아기 새 (어린 새의 머리, 짧은 부리, 어린 얼굴 깃털) — wearing: Non-human or non-outfit entity. Uses entity reference image as-is. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "B-200의 고개와 시선이 오른쪽 아래 둥지 안의 아기 새들을 정확히 향하고 있습니다.",
    "built_space": "어두운 창고 내부로, 배경에 넓게 열린 출입구가 보입니다. 왼쪽 가장자리에 나무 상자가 위치하며 오른쪽 금속 부품들 위에 둥지가 있습니다.",
    "entities": "B-200은 빛나는 눈을 가졌으나, 지시문과 달리 레퍼런스의 찰리(가슴 고리 및 파손 흔적) 외형을 띠고 있으며 총기 팔이 없습니다. 아기 새들의 묘사는 일치합니다.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 등장하는 찰리의 몸통(가슴 고리와 파손 흔적)을 B-200에게 적용하여 인물 외형 제한 지시(never carry their face, body or clothing)를 위반함.",
     "[gpt-high] 장면에서 떠난 찰리의 가슴 고리와 손상된 장갑 몸체를 현재 주인공의 외형으로 재사용했다. 이전 사진의 다른 인물 몸체를 옮기지 말라는 금지에 어긋난다."
    ],
    "physics": "B-200은 지면에 안정적으로 서 있으며, 둥지는 금속 구조물 위에 놓여 지탱되고 있습니다. 허공에 뜬 물체는 없습니다."
   },
   {
    "label": "B",
    "direction": "B-200의 고개와 시선이 아래쪽 종이 상자 안의 아기 새들을 향하고 있습니다.",
    "built_space": "어두운 창고 내부로, 배경에 반쯤 열린 출입구 틈새가 보입니다. 화면 오른쪽에 큰 종이 상자가 놓여 있습니다.",
    "entities": "B-200은 안광과 한쪽 총기 팔을 갖추었으나 역시 찰리의 몸통을 잘못 차용했습니다. 둥지와 아기 새들이 종이 상자 안에 배치되어 있습니다.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 등장하는 찰리의 몸통(가슴 고리와 파손 흔적)을 B-200에게 적용하여 인물 외형 제한 지시를 위반함.",
     "[gemini-pro] 상자가 새들과 분리되어 화면 주변부에 있어야 한다는 지시(clear of the birds)를 어기고 새들을 상자 안에 배치하여 무대 연출을 위반함.",
     "[gpt-high] 장면에서 제외된 찰리의 식별 가능한 가슴 고리와 탄흔이 있는 몸체를 현재 주인공에게 옮겼다. 이전 사진의 다른 인물 몸체를 이 장면에 전용하지 말라는 명시적 금지에 어긋난다."
    ],
    "physics": "B-200은 종이 상자 테두리에 양팔을 기대고 서 있으며, 둥지는 상자 내부에 안정적으로 놓여 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "레퍼런스의 찰리 외형(가슴 고리 등)을 B-200에게 잘못 적용한 치명적 오류와 총기 팔 누락이 있으나, 상자를 새들과 분리해 주변부에 배치한 구도 지시는 잘 따랐습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "반쯤 열린 출입구와 총기 팔 지시를 일부 반영했으나, 찰리 외형을 차용하는 오류를 범했고 새들을 상자 안에 배치하여 배경 구도 지시를 심각하게 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "B-200의 고개와 시선이 오른쪽 아래 둥지 안의 아기 새들을 정확히 향하고 있습니다.",
        "built_space": "어두운 창고 내부로, 배경에 넓게 열린 출입구가 보입니다. 왼쪽 가장자리에 나무 상자가 위치하며 오른쪽 금속 부품들 위에 둥지가 있습니다.",
        "entities": "B-200은 빛나는 눈을 가졌으나, 지시문과 달리 레퍼런스의 찰리(가슴 고리 및 파손 흔적) 외형을 띠고 있으며 총기 팔이 없습니다. 아기 새들의 묘사는 일치합니다.",
        "hard_violations": [
         "레퍼런스에 등장하는 찰리의 몸통(가슴 고리와 파손 흔적)을 B-200에게 적용하여 인물 외형 제한 지시(never carry their face, body or clothing)를 위반함."
        ],
        "physics": "B-200은 지면에 안정적으로 서 있으며, 둥지는 금속 구조물 위에 놓여 지탱되고 있습니다. 허공에 뜬 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "B-200의 고개와 시선이 아래쪽 종이 상자 안의 아기 새들을 향하고 있습니다.",
        "built_space": "어두운 창고 내부로, 배경에 반쯤 열린 출입구 틈새가 보입니다. 화면 오른쪽에 큰 종이 상자가 놓여 있습니다.",
        "entities": "B-200은 안광과 한쪽 총기 팔을 갖추었으나 역시 찰리의 몸통을 잘못 차용했습니다. 둥지와 아기 새들이 종이 상자 안에 배치되어 있습니다.",
        "hard_violations": [
         "레퍼런스에 등장하는 찰리의 몸통(가슴 고리와 파손 흔적)을 B-200에게 적용하여 인물 외형 제한 지시를 위반함.",
         "상자가 새들과 분리되어 화면 주변부에 있어야 한다는 지시(clear of the birds)를 어기고 새들을 상자 안에 배치하여 무대 연출을 위반함."
        ],
        "physics": "B-200은 종이 상자 테두리에 양팔을 기대고 서 있으며, 둥지는 상자 내부에 안정적으로 놓여 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "레퍼런스의 찰리 외형(가슴 고리 등)을 B-200에게 잘못 적용한 치명적 오류와 총기 팔 누락이 있으나, 상자를 새들과 분리해 주변부에 배치한 구도 지시는 잘 따랐습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "반쯤 열린 출입구와 총기 팔 지시를 일부 반영했으나, 찰리 외형을 차용하는 오류를 범했고 새들을 상자 안에 배치하여 배경 구도 지시를 심각하게 어겼습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "B-200의 고개와 시선이 오른쪽 아래 둥지 안의 아기 새들을 정확히 향하고 있습니다.",
        "built_space": "어두운 창고 내부로, 배경에 넓게 열린 출입구가 보입니다. 왼쪽 가장자리에 나무 상자가 위치하며 오른쪽 금속 부품들 위에 둥지가 있습니다.",
        "entities": "B-200은 빛나는 눈을 가졌으나, 지시문과 달리 레퍼런스의 찰리(가슴 고리 및 파손 흔적) 외형을 띠고 있으며 총기 팔이 없습니다. 아기 새들의 묘사는 일치합니다.",
        "hard_violations": [
         "레퍼런스에 등장하는 찰리의 몸통(가슴 고리와 파손 흔적)을 B-200에게 적용하여 인물 외형 제한 지시(never carry their face, body or clothing)를 위반함."
        ],
        "physics": "B-200은 지면에 안정적으로 서 있으며, 둥지는 금속 구조물 위에 놓여 지탱되고 있습니다. 허공에 뜬 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "B-200의 고개와 시선이 아래쪽 종이 상자 안의 아기 새들을 향하고 있습니다.",
        "built_space": "어두운 창고 내부로, 배경에 반쯤 열린 출입구 틈새가 보입니다. 화면 오른쪽에 큰 종이 상자가 놓여 있습니다.",
        "entities": "B-200은 안광과 한쪽 총기 팔을 갖추었으나 역시 찰리의 몸통을 잘못 차용했습니다. 둥지와 아기 새들이 종이 상자 안에 배치되어 있습니다.",
        "hard_violations": [
         "레퍼런스에 등장하는 찰리의 몸통(가슴 고리와 파손 흔적)을 B-200에게 적용하여 인물 외형 제한 지시를 위반함.",
         "상자가 새들과 분리되어 화면 주변부에 있어야 한다는 지시(clear of the birds)를 어기고 새들을 상자 안에 배치하여 무대 연출을 위반함."
        ],
        "physics": "B-200은 종이 상자 테두리에 양팔을 기대고 서 있으며, 둥지는 상자 내부에 안정적으로 놓여 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "하향 시선과 따뜻한 안광, 총기 팔은 더 충실하지만, 떠난 찰리의 가슴 고리와 손상된 몸체를 주인공에게 옮겼고 둥지도 작은 우하단 기준점보다 크다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "새를 내려다보는 동작은 맞지만 찰리의 몸체를 전용한 데다 총기 팔이 확인되지 않고, 크게 부각된 전경 둥지가 지정된 미디엄 구도의 위계를 흐린다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇은 고개를 숙여 화면 오른쪽 아래 둥지의 새끼 새 두 마리 쪽을 바라본다. 새들은 부리를 위로 들어 로봇 쪽을 향한다. 화면 왼쪽 팔에 붙은 총구 두 개는 전방 아래쪽으로 향하며 새들을 직접 겨누지는 않는다.",
        "built_space": "뒤 중앙에 부분적으로 열린 출입구 하나가 있고, 왼쪽 벽에는 높은 창들과 부품 선반·목상자들이, 오른쪽에는 폐로봇 및 기계 부품 더미가 있다. 로봇은 전경의 열린 부품 상자 뒤에 서 있고, 둥지는 그 상자의 오른쪽 가장자리 부근에 놓여 내부가 내려다보인다. 옆으로 치운 것으로 보이는 큰 상자 하나가 오른쪽 주변부에 있다. 낡은 창고 재질과 낮빛은 이어지지만, 둥지가 화면 너비의 약 3분의 1을 차지해 요구한 작은 우하단 요소보다 크다. 반사상은 없다.",
        "entities": "살아 움직이는 주체는 기계 머리의 로봇 하나와 짧은 부리·솜털을 지닌 새끼 새 두 마리다. 따뜻한 원형 안광과 화면 왼쪽의 총기 부착 팔이 보인다. 그러나 몸통은 짙은 회색 B-200 외장이라기보다 참고 이미지 오른쪽 인물의 밝은 회색 장갑이며, 원형 가슴 고리와 그 위쪽 탄흔 배치까지 반복한다. 이는 프롬프트에서 떠났다고 명시한 찰리의 가슴 손상과 외형에 해당한다. 둥지, 이동된 상자, 폐기계 더미는 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장면에서 제외된 찰리의 식별 가능한 가슴 고리와 탄흔이 있는 몸체를 현재 주인공에게 옮겼다. 이전 사진의 다른 인물 몸체를 이 장면에 전용하지 말라는 명시적 금지에 어긋난다."
        ],
        "physics": "로봇의 몸통은 골반과 아래로 이어지는 다리에 연결되어 있고, 발은 프레임 밖이다. 양손은 전경 상자 테두리에 닿아 상체를 보조한다. 총기는 팔에 고정되어 있다. 새들은 둥지 안에 앉아 있고 둥지는 부품 상자 상단의 받침에 놓여 있다. 공중에 떠 있거나 지지 없이 매달린 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "로봇은 머리를 오른쪽 아래로 기울여 전경 둥지의 새끼 새 두 마리를 바라본다. 두 새는 열린 부리를 위쪽 로봇 방향으로 든다. 팔은 몸 옆으로 내려와 있으며, 명확하게 식별되는 총구나 겨냥 동작은 없다.",
        "built_space": "뒤 왼쪽에 낮빛이 들어오는 출입구 하나, 오른쪽 벽에 선반 두 열과 벽 부착 전기함 하나, 오른쪽 중경에 부품이 든 드럼통 하나가 보인다. 왼쪽 전경에는 옆으로 치운 목상자 하나가 있고, 로봇은 중앙 통로에서 오른쪽의 열린 부품 용기를 내려다본다. 둥지는 그 용기 위에 놓여 내부가 보인다. 창고의 낡은 금속·목재와 부품 적재 환경은 비슷하지만, 둥지와 새들이 오른쪽 전경을 크게 점유하여 작은 배경 기준점이라는 지시와 맞지 않는다. 반사상은 없다.",
        "entities": "로봇 하나와 솜털·짧은 부리를 가진 새끼 새 두 마리가 보이며 다른 활동 인물은 없다. 로봇의 따뜻한 안광은 맞지만, 참고 이미지의 찰리와 같은 밝은 장갑 몸통, 원형 가슴 고리 및 주변 탄흔을 사용했다. 화면에 들어온 팔에서는 유지되어야 할 총기 구조가 확인되지 않고, 왼쪽 아래에는 일반적인 집게형 손이 보인다. 둥지와 이동된 목상자, 폐기계 부품은 존재한다. 목상자에는 운송용 화살표 표식이 있으나 읽을 수 있는 문장은 없다.",
        "hard_violations": [
         "장면에서 떠난 찰리의 가슴 고리와 손상된 장갑 몸체를 현재 주인공의 외형으로 재사용했다. 이전 사진의 다른 인물 몸체를 옮기지 말라는 금지에 어긋난다."
        ],
        "physics": "로봇은 골반과 다리가 연결된 직립 자세이며 발은 화면 밖이다. 내려뜨린 손과 팔은 관절에 정상적으로 연결되어 있다. 새들의 몸은 둥지에 받쳐지고, 둥지는 열린 부품 용기 안의 판재와 적재물 위에 놓여 있다. 목상자와 드럼통은 바닥에 놓여 있으며, 지지 없는 부유나 불가능한 동작은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "하향 시선과 따뜻한 안광, 총기 팔은 더 충실하지만, 떠난 찰리의 가슴 고리와 손상된 몸체를 주인공에게 옮겼고 둥지도 작은 우하단 기준점보다 크다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "새를 내려다보는 동작은 맞지만 찰리의 몸체를 전용한 데다 총기 팔이 확인되지 않고, 크게 부각된 전경 둥지가 지정된 미디엄 구도의 위계를 흐린다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "로봇은 고개를 숙여 화면 오른쪽 아래 둥지의 새끼 새 두 마리 쪽을 바라본다. 새들은 부리를 위로 들어 로봇 쪽을 향한다. 화면 왼쪽 팔에 붙은 총구 두 개는 전방 아래쪽으로 향하며 새들을 직접 겨누지는 않는다.",
        "built_space": "뒤 중앙에 부분적으로 열린 출입구 하나가 있고, 왼쪽 벽에는 높은 창들과 부품 선반·목상자들이, 오른쪽에는 폐로봇 및 기계 부품 더미가 있다. 로봇은 전경의 열린 부품 상자 뒤에 서 있고, 둥지는 그 상자의 오른쪽 가장자리 부근에 놓여 내부가 내려다보인다. 옆으로 치운 것으로 보이는 큰 상자 하나가 오른쪽 주변부에 있다. 낡은 창고 재질과 낮빛은 이어지지만, 둥지가 화면 너비의 약 3분의 1을 차지해 요구한 작은 우하단 요소보다 크다. 반사상은 없다.",
        "entities": "살아 움직이는 주체는 기계 머리의 로봇 하나와 짧은 부리·솜털을 지닌 새끼 새 두 마리다. 따뜻한 원형 안광과 화면 왼쪽의 총기 부착 팔이 보인다. 그러나 몸통은 짙은 회색 B-200 외장이라기보다 참고 이미지 오른쪽 인물의 밝은 회색 장갑이며, 원형 가슴 고리와 그 위쪽 탄흔 배치까지 반복한다. 이는 프롬프트에서 떠났다고 명시한 찰리의 가슴 손상과 외형에 해당한다. 둥지, 이동된 상자, 폐기계 더미는 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장면에서 제외된 찰리의 식별 가능한 가슴 고리와 탄흔이 있는 몸체를 현재 주인공에게 옮겼다. 이전 사진의 다른 인물 몸체를 이 장면에 전용하지 말라는 명시적 금지에 어긋난다."
        ],
        "physics": "로봇의 몸통은 골반과 아래로 이어지는 다리에 연결되어 있고, 발은 프레임 밖이다. 양손은 전경 상자 테두리에 닿아 상체를 보조한다. 총기는 팔에 고정되어 있다. 새들은 둥지 안에 앉아 있고 둥지는 부품 상자 상단의 받침에 놓여 있다. 공중에 떠 있거나 지지 없이 매달린 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "로봇은 머리를 오른쪽 아래로 기울여 전경 둥지의 새끼 새 두 마리를 바라본다. 두 새는 열린 부리를 위쪽 로봇 방향으로 든다. 팔은 몸 옆으로 내려와 있으며, 명확하게 식별되는 총구나 겨냥 동작은 없다.",
        "built_space": "뒤 왼쪽에 낮빛이 들어오는 출입구 하나, 오른쪽 벽에 선반 두 열과 벽 부착 전기함 하나, 오른쪽 중경에 부품이 든 드럼통 하나가 보인다. 왼쪽 전경에는 옆으로 치운 목상자 하나가 있고, 로봇은 중앙 통로에서 오른쪽의 열린 부품 용기를 내려다본다. 둥지는 그 용기 위에 놓여 내부가 보인다. 창고의 낡은 금속·목재와 부품 적재 환경은 비슷하지만, 둥지와 새들이 오른쪽 전경을 크게 점유하여 작은 배경 기준점이라는 지시와 맞지 않는다. 반사상은 없다.",
        "entities": "로봇 하나와 솜털·짧은 부리를 가진 새끼 새 두 마리가 보이며 다른 활동 인물은 없다. 로봇의 따뜻한 안광은 맞지만, 참고 이미지의 찰리와 같은 밝은 장갑 몸통, 원형 가슴 고리 및 주변 탄흔을 사용했다. 화면에 들어온 팔에서는 유지되어야 할 총기 구조가 확인되지 않고, 왼쪽 아래에는 일반적인 집게형 손이 보인다. 둥지와 이동된 목상자, 폐기계 부품은 존재한다. 목상자에는 운송용 화살표 표식이 있으나 읽을 수 있는 문장은 없다.",
        "hard_violations": [
         "장면에서 떠난 찰리의 가슴 고리와 손상된 장갑 몸체를 현재 주인공의 외형으로 재사용했다. 이전 사진의 다른 인물 몸체를 옮기지 말라는 금지에 어긋난다."
        ],
        "physics": "로봇은 골반과 다리가 연결된 직립 자세이며 발은 화면 밖이다. 내려뜨린 손과 팔은 관절에 정상적으로 연결되어 있다. 새들의 몸은 둥지에 받쳐지고, 둥지는 열린 부품 용기 안의 판재와 적재물 위에 놓여 있다. 목상자와 드럼통은 바닥에 놓여 있으며, 지지 없는 부유나 불가능한 동작은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스에 등장하는 찰리의 몸통(가슴 고리와 파손 흔적)을 B-200에게 적용하여 인물 외형 제한 지시(never carry their face, body or clothing)를 위반함.",
     "[gpt-high] 장면에서 떠난 찰리의 가슴 고리와 손상된 장갑 몸체를 현재 주인공의 외형으로 재사용했다. 이전 사진의 다른 인물 몸체를 옮기지 말라는 금지에 어긋난다."
    ],
    "B": [
     "[gemini-pro] 레퍼런스에 등장하는 찰리의 몸통(가슴 고리와 파손 흔적)을 B-200에게 적용하여 인물 외형 제한 지시를 위반함.",
     "[gemini-pro] 상자가 새들과 분리되어 화면 주변부에 있어야 한다는 지시(clear of the birds)를 어기고 새들을 상자 안에 배치하여 무대 연출을 위반함.",
     "[gpt-high] 장면에서 제외된 찰리의 식별 가능한 가슴 고리와 탄흔이 있는 몸체를 현재 주인공에게 옮겼다. 이전 사진의 다른 인물 몸체를 이 장면에 전용하지 말라는 명시적 금지에 어긋난다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1417,
   "B": 1500
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "레퍼런스의 찰리 외형(가슴 고리 등)을 B-200에게 잘못 적용한 치명적 오류와 총기 팔 누락이 있으나, 상자를 새들과 분리해 주변부에 배치한 구도 지시는 잘 따랐습니다.  ★위반: [gemini-pro] 레퍼런스에 등장하는 찰리의 몸통(가슴 고리와 파손 흔적)을 B-200에게 적용하여 인물 외형 제한 지시(never carry their face, body or clothing)를 위반함. / [gpt-high] 장면에서 떠난 찰리의 가슴 고리와 손상된 장갑 몸체를 현재 주인공의 외형으로 재사용했다. 이전 사진의 다른 인물 몸체를 옮기지 말라는 금지에 어긋난다."
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "반쯤 열린 출입구와 총기 팔 지시를 일부 반영했으나, 찰리 외형을 차용하는 오류를 범했고 새들을 상자 안에 배치하여 배경 구도 지시를 심각하게 어겼습니다.  ★위반: [gemini-pro] 레퍼런스에 등장하는 찰리의 몸통(가슴 고리와 파손 흔적)을 B-200에게 적용하여 인물 외형 제한 지시를 위반함. / [gemini-pro] 상자가 새들과 분리되어 화면 주변부에 있어야 한다는 지시(clear of the birds)를 어기고 새들을 상자 안에 배치하여 무대 연출을 위반함. / [gpt-high] 장면에서 제외된 찰리의 식별 가능한 가슴 고리와 탄흔이 있는 몸체를 현재 주인공에게 옮겼다. 이전 사진의 다른 인물 몸체를 이 장면에 전용하지 말라는 명시적 금지에 어긋난다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S64sh18_sel.png",
    "asset_id": "15e86419-dc26-48bc-8373-69c9cfdb9695",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdfd-2569-776b-926d-75a5ddf857cd",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S64sh18"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S64sh27::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:53:06.348425+00:00",
  "fingerprint": "485a0b7a127ff83fa334a8f1f0c9613dee2edcacb2e0fe17429ff793d540257d",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S64sh27_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S64sh27_sel.png",
  "source_sha256": "fc27fbc0cb17931481cae0dc68366a996632682bd71e582e5628592606078720",
  "file": "S64sh27_cine.png",
  "staged_sha256": "978625fec31435c9611444cd20a16606680f6148956768092e674f9c55b986ac",
  "latency_ms": 12010
 },
 "S65sh4::signage": {
  "fp": "1786e2d1042aea66",
  "inscriptions": [
   {
    "text_native": "0",
    "source": "scene_text_quoted",
    "reason_ko": "방사능 측정기 계기판에서 바늘이 정확히 가리키고 있는 숫자로 본문에 명시되어 있습니다.",
    "source_quote": "0"
   }
  ],
  "cues": [],
  "dropped": []
 },
 "era_assess::73f4a5af046e1573": {
  "subjects": [],
  "subject_text": "무너진 백화점 외부와 주변 수풀\n대형 상업 건물 주변을 이끼와 무성한 나무가 뒤덮은 공간. 고인 웅덩이와 꽃, 빽빽한 수풀이 건물 외벽까지 이어진다.",
  "identity": "canonical",
  "scope_id": "L100",
  "scope_role": "location_exterior",
  "scope_sha": "58473a6bc08a82e4"
 },
 "S65sh4::bgfirst_bg": {
  "input_fingerprint": "df058551ee0362c7",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수빈의 손에 들린 방사능 측정기의 바늘이 숫자 0을 정확히 가리키며 멈춘 액정 클로즈업.\n\nLOCATION (lock): On the moss-covered outdoor approach to a partially collapsed department store, among puddles and overgrown vegetation.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Radiation meter display (Needle settled at the zero reading before the instrument is tapped) — The numbered display face is angled upward toward the lens, with zero and the needle clearly visible; used as Primary detail focus, retaining the hand and surrounding ground rather than isolating an oversized instrument; Moss-covered ground (Broadly covered with moss near the collapsed department store); used as Soft surrounding context and an early visual connection between the meter reading and the recovering environment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight keeps the zero reading and supporting hand clear while preserving the location's quietly unsettling atmosphere.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수빈의 손에 들린 방사능 측정기의 바늘이 숫자 0을 정확히 가리키며 멈춘 액정 클로즈업.\n\nLOCATION (lock): On the moss-covered outdoor approach to a partially collapsed department store, among puddles and overgrown vegetation.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Radiation meter display (Needle settled at the zero reading before the instrument is tapped) — The numbered display face is angled upward toward the lens, with zero and the needle clearly visible; used as Primary detail focus, retaining the hand and surrounding ground rather than isolating an oversized instrument; Moss-covered ground (Broadly covered with moss near the collapsed department store); used as Soft surrounding context and an early visual connection between the meter reading and the recovering environment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight keeps the zero reading and supporting hand clear while preserving the location's quietly unsettling atmosphere.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S65sh4__bgfirst_bg.png",
  "asset_id": "35c04433-c677-4bca-a371-cf9cf4952d27",
  "input_asset_ids": [
   "a0f440c3-1ec2-45c0-8807-28f0c17b2ac9",
   "3ab84dc4-2692-40b2-a37a-53eff77ce828"
  ]
 },
 "S65sh4": {
  "input_fingerprint": "485a5d1c58e07e60",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈의 손에 들린 방사능 측정기의 바늘이 숫자 0을 정확히 가리키며 멈춘 액정 클로즈업.\n\nLOCATION (lock): On the moss-covered outdoor approach to a partially collapsed department store, among puddles and overgrown vegetation. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Radiation meter display (Needle settled at the zero reading before the instrument is tapped) — The numbered display face is angled upward toward the lens, with zero and the needle clearly visible; used as Primary detail focus, retaining the hand and surrounding ground rather than isolating an oversized instrument; Moss-covered ground (Broadly covered with moss near the collapsed department store); used as Soft surrounding context and an early visual connection between the meter reading and the recovering environment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight keeps the zero reading and supporting hand clear while preserving the location's quietly unsettling atmosphere.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The half-collapsed department store is covered in moss and uncontrolled vegetation, with standing pools among the ruins. The radiation meter reads close to zero, not necessarily exactly zero.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"0\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈의 손에 들린 방사능 측정기의 바늘이 숫자 0을 정확히 가리키며 멈춘 액정 클로즈업.\n\nLOCATION (lock): On the moss-covered outdoor approach to a partially collapsed department store, among puddles and overgrown vegetation. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Radiation meter display (Needle settled at the zero reading before the instrument is tapped) — The numbered display face is angled upward toward the lens, with zero and the needle clearly visible; used as Primary detail focus, retaining the hand and surrounding ground rather than isolating an oversized instrument; Moss-covered ground (Broadly covered with moss near the collapsed department store); used as Soft surrounding context and an early visual connection between the meter reading and the recovering environment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight keeps the zero reading and supporting hand clear while preserving the location's quietly unsettling atmosphere.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The half-collapsed department store is covered in moss and uncontrolled vegetation, with standing pools among the ruins. The radiation meter reads close to zero, not necessarily exactly zero.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"0\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈의 손에 들린 방사능 측정기의 바늘이 숫자 0을 정확히 가리키며 멈춘 액정 클로즈업.\n\nLOCATION (lock): On the moss-covered outdoor approach to a partially collapsed department store, among puddles and overgrown vegetation. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Radiation meter display (Needle settled at the zero reading before the instrument is tapped) — The numbered display face is angled upward toward the lens, with zero and the needle clearly visible; used as Primary detail focus, retaining the hand and surrounding ground rather than isolating an oversized instrument; Moss-covered ground (Broadly covered with moss near the collapsed department store); used as Soft surrounding context and an early visual connection between the meter reading and the recovering environment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daylight keeps the zero reading and supporting hand clear while preserving the location's quietly unsettling atmosphere.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The half-collapsed department store is covered in moss and uncontrolled vegetation, with standing pools among the ruins. The radiation meter reads close to zero, not necessarily exactly zero.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"0\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S65sh4__bgfirst_bg.png",
     "asset_id": "35c04433-c677-4bca-a371-cf9cf4952d27",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S65sh4.png",
     "asset_id": "a0f440c3-1ec2-45c0-8807-28f0c17b2ac9",
     "role": "conti_light"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L100B01.png",
     "asset_id": "3ab84dc4-2692-40b2-a37a-53eff77ce828",
     "role": "location_plate"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "측정기 화면이 렌즈를 향하고 있으며, 바늘이 숫자 0을 명확히 가리킴.",
    "built_space": "붕괴된 백화점 입구, 이끼와 물웅덩이가 레퍼런스와 동일하게 배경에 배치됨.",
    "entities": "장갑을 낀 두 손, 0이 표기된 방사능 측정기.",
    "hard_violations": [],
    "physics": "두 손이 기기의 양옆을 자연스럽게 쥐어 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "측정기 화면이 렌즈를 향하며 바늘이 0을 가리킴.",
    "built_space": "파괴된 건물 건물과 식물, 물웅덩이가 아웃포커싱되어 배경에 나타남.",
    "entities": "전술복을 입은 인물의 손, 방사능 측정기, 기기 하단의 미상 텍스트.",
    "hard_violations": [
     "[gemini-pro] 지정된 '0' 외에 측정기 우측 하단에 임의의 글자(가짜 텍스트)가 생성되어 텍스트 제약 위반."
    ],
    "physics": "손이 기기 측면과 하단을 받치고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 텍스트('0') 제약을 준수하고 요구된 클로즈업 앵글과 배경의 조화를 잘 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "측정기 표면에 프롬프트에서 금지한 임의의 텍스트가 노출되어 하드 위반에 해당함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "측정기 화면이 렌즈를 향하고 있으며, 바늘이 숫자 0을 명확히 가리킴.",
        "built_space": "붕괴된 백화점 입구, 이끼와 물웅덩이가 레퍼런스와 동일하게 배경에 배치됨.",
        "entities": "장갑을 낀 두 손, 0이 표기된 방사능 측정기.",
        "hard_violations": [],
        "physics": "두 손이 기기의 양옆을 자연스럽게 쥐어 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "측정기 화면이 렌즈를 향하며 바늘이 0을 가리킴.",
        "built_space": "파괴된 건물 건물과 식물, 물웅덩이가 아웃포커싱되어 배경에 나타남.",
        "entities": "전술복을 입은 인물의 손, 방사능 측정기, 기기 하단의 미상 텍스트.",
        "hard_violations": [
         "지정된 '0' 외에 측정기 우측 하단에 임의의 글자(가짜 텍스트)가 생성되어 텍스트 제약 위반."
        ],
        "physics": "손이 기기 측면과 하단을 받치고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 텍스트('0') 제약을 준수하고 요구된 클로즈업 앵글과 배경의 조화를 잘 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "측정기 표면에 프롬프트에서 금지한 임의의 텍스트가 노출되어 하드 위반에 해당함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "측정기 화면이 렌즈를 향하고 있으며, 바늘이 숫자 0을 명확히 가리킴.",
        "built_space": "붕괴된 백화점 입구, 이끼와 물웅덩이가 레퍼런스와 동일하게 배경에 배치됨.",
        "entities": "장갑을 낀 두 손, 0이 표기된 방사능 측정기.",
        "hard_violations": [],
        "physics": "두 손이 기기의 양옆을 자연스럽게 쥐어 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "측정기 화면이 렌즈를 향하며 바늘이 0을 가리킴.",
        "built_space": "파괴된 건물 건물과 식물, 물웅덩이가 아웃포커싱되어 배경에 나타남.",
        "entities": "전술복을 입은 인물의 손, 방사능 측정기, 기기 하단의 미상 텍스트.",
        "hard_violations": [
         "지정된 '0' 외에 측정기 우측 하단에 임의의 글자(가짜 텍스트)가 생성되어 텍스트 제약 위반."
        ],
        "physics": "손이 기기 측면과 하단을 받치고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "바늘이 단일 0 눈금에 정렬된 계기면을 손과 이끼 낀 지면 속에서 강조해, 배경과 기기 전체를 더 넓게 보여주는 B보다 요구된 클로즈업에 충실하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "장소와 양손의 지지는 잘 구현했지만, 계기면보다 기기 전체와 건물의 비중이 크고 눈금에 0이 두 번 있어 정확한 영점 표시가 덜 명료하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "바늘은 아래쪽 회전축에서 왼쪽 위로 뻗어, 유일하게 읽히는 0 옆의 눈금 시작점을 가리킨다. 계기면은 위로 기울어 렌즈에 보이며, 인물의 눈은 프레임 밖이라 시선은 확인할 수 없다.",
        "built_space": "배경에 파손된 중앙 출입구 한 곳, 양옆의 가로 창 띠, 콘크리트 외벽과 무성한 식생이 보인다. 전경에는 이끼 낀 포장과 넓은 물웅덩이가 있다. 참고 장소의 재료와 주요 구조는 맞지만, 물웅덩이는 참고 사진의 통로 양옆 배치보다 중앙에 모여 보인다. 인물은 화면 오른쪽에서 기기를 들고 있으며 구조물과 충돌하지 않는다.",
        "entities": "낡은 금속 케이스의 원형 바늘식 측정기 한 대와 이를 잡은 양손, 손목 및 소매가 보인다. 읽을 수 있는 표기는 0 하나다. 손은 밝은 피부이며 손가락이 드러나는 갈색 장갑과 회녹색 겉옷을 착용했다. 수빈의 외모 참고가 없어 손만으로 나이·성별·민족적 정체성을 검증할 수 없다. 계기는 유리 아래 물리적 바늘과 눈금판처럼 보여 액정 여부는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "앞쪽 손이 케이스의 왼쪽과 아랫부분을 감싸고, 반대 손이 오른쪽 가장자리를 받친다. 기기의 무게를 지탱하는 접촉이 보이며 떠 있는 물체는 없다. 바늘과 손에 동작 흐림이 없고 두드리는 동작도 없어, 두드리기 전 정지 순간과 부합한다."
       },
       {
        "label": "B",
        "direction": "바늘은 하단 회전축에서 왼쪽 위로 향해 왼쪽 끝의 0 표시를 가리킨다. 그러나 눈금 상단 중앙에도 별도의 0이 있어 영점 위치가 중복된다. 계기면은 렌즈 쪽으로 위를 향하고 있으며 인물의 시선은 보이지 않는다.",
        "built_space": "중앙의 파손된 출입구 한 곳, 양쪽 가로 창 띠, 오른쪽 건물 돌출부가 보인다. 접근로 양옆의 물웅덩이와 오른쪽 철근 노출 잔해는 참고 장소에 잘 대응한다. 다만 건물과 주변 폐허가 비교적 선명하고 넓게 들어와, 계기 액정 중심의 클로즈업보다 장소 설명의 비중이 높다. 물 표면의 식생 반사는 주변 배치와 모순되지 않는다.",
        "entities": "직사각형 바늘식 측정기 한 대, 조절 노브 두 개, 이를 잡은 양손과 양쪽 소매가 보인다. 읽히는 문자는 모두 0이지만 두 위치에 반복되어 있다. 밝은 피부의 손에 갈색 손가락 노출 장갑과 회녹색 소매가 보이며, 얼굴이 없어 수빈의 세부 정체성은 검증할 수 없다. 표시부는 물리적 바늘식 계기로 보이고 액정이라는 점은 확인되지 않는다. 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "왼손이 케이스 왼쪽을 감싸고 오른손이 오른쪽 아래를 받쳐 기기를 안정적으로 지지한다. 손가락과 케이스의 접촉 및 손목 연결은 자연스럽다. 기기는 공중에 떠 있지 않고, 바늘과 양손은 멈춰 있으며 두드리는 동작은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "바늘이 단일 0 눈금에 정렬된 계기면을 손과 이끼 낀 지면 속에서 강조해, 배경과 기기 전체를 더 넓게 보여주는 B보다 요구된 클로즈업에 충실하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "장소와 양손의 지지는 잘 구현했지만, 계기면보다 기기 전체와 건물의 비중이 크고 눈금에 0이 두 번 있어 정확한 영점 표시가 덜 명료하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "바늘은 아래쪽 회전축에서 왼쪽 위로 뻗어, 유일하게 읽히는 0 옆의 눈금 시작점을 가리킨다. 계기면은 위로 기울어 렌즈에 보이며, 인물의 눈은 프레임 밖이라 시선은 확인할 수 없다.",
        "built_space": "배경에 파손된 중앙 출입구 한 곳, 양옆의 가로 창 띠, 콘크리트 외벽과 무성한 식생이 보인다. 전경에는 이끼 낀 포장과 넓은 물웅덩이가 있다. 참고 장소의 재료와 주요 구조는 맞지만, 물웅덩이는 참고 사진의 통로 양옆 배치보다 중앙에 모여 보인다. 인물은 화면 오른쪽에서 기기를 들고 있으며 구조물과 충돌하지 않는다.",
        "entities": "낡은 금속 케이스의 원형 바늘식 측정기 한 대와 이를 잡은 양손, 손목 및 소매가 보인다. 읽을 수 있는 표기는 0 하나다. 손은 밝은 피부이며 손가락이 드러나는 갈색 장갑과 회녹색 겉옷을 착용했다. 수빈의 외모 참고가 없어 손만으로 나이·성별·민족적 정체성을 검증할 수 없다. 계기는 유리 아래 물리적 바늘과 눈금판처럼 보여 액정 여부는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "앞쪽 손이 케이스의 왼쪽과 아랫부분을 감싸고, 반대 손이 오른쪽 가장자리를 받친다. 기기의 무게를 지탱하는 접촉이 보이며 떠 있는 물체는 없다. 바늘과 손에 동작 흐림이 없고 두드리는 동작도 없어, 두드리기 전 정지 순간과 부합한다."
       },
       {
        "label": "A",
        "direction": "바늘은 하단 회전축에서 왼쪽 위로 향해 왼쪽 끝의 0 표시를 가리킨다. 그러나 눈금 상단 중앙에도 별도의 0이 있어 영점 위치가 중복된다. 계기면은 렌즈 쪽으로 위를 향하고 있으며 인물의 시선은 보이지 않는다.",
        "built_space": "중앙의 파손된 출입구 한 곳, 양쪽 가로 창 띠, 오른쪽 건물 돌출부가 보인다. 접근로 양옆의 물웅덩이와 오른쪽 철근 노출 잔해는 참고 장소에 잘 대응한다. 다만 건물과 주변 폐허가 비교적 선명하고 넓게 들어와, 계기 액정 중심의 클로즈업보다 장소 설명의 비중이 높다. 물 표면의 식생 반사는 주변 배치와 모순되지 않는다.",
        "entities": "직사각형 바늘식 측정기 한 대, 조절 노브 두 개, 이를 잡은 양손과 양쪽 소매가 보인다. 읽히는 문자는 모두 0이지만 두 위치에 반복되어 있다. 밝은 피부의 손에 갈색 손가락 노출 장갑과 회녹색 소매가 보이며, 얼굴이 없어 수빈의 세부 정체성은 검증할 수 없다. 표시부는 물리적 바늘식 계기로 보이고 액정이라는 점은 확인되지 않는다. 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "왼손이 케이스 왼쪽을 감싸고 오른손이 오른쪽 아래를 받쳐 기기를 안정적으로 지지한다. 손가락과 케이스의 접촉 및 손목 연결은 자연스럽다. 기기는 공중에 떠 있지 않고, 바늘과 양손은 멈춰 있으며 두드리는 동작은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.179
   },
   "violations": {
    "B": [
     "[gemini-pro] 지정된 '0' 외에 측정기 우측 하단에 임의의 글자(가짜 텍스트)가 생성되어 텍스트 제약 위반."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1875,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "지정된 텍스트('0') 제약을 준수하고 요구된 클로즈업 앵글과 배경의 조화를 잘 구현함."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "측정기 표면에 프롬프트에서 금지한 임의의 텍스트가 노출되어 하드 위반에 해당함.  ★위반: [gemini-pro] 지정된 '0' 외에 측정기 우측 하단에 임의의 글자(가짜 텍스트)가 생성되어 텍스트 제약 위반."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L100B01.png",
    "asset_id": "3ab84dc4-2692-40b2-a37a-53eff77ce828",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdfd-eb29-7d94-bf59-b8e48519d2f8",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S65sh4__bgfirst_bg.png",
   "bg_asset_id": "35c04433-c677-4bca-a371-cf9cf4952d27",
   "bg_record_key": "S65sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S65sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:53:18.787146+00:00",
  "fingerprint": "8c04a16e9402a61d303e61bd8ea23203aaf53118171b9596221031d409b03ef1",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S65sh4_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S65sh4_sel.png",
  "source_sha256": "2c30f2fb504a2b25f26b2341d3a3e91fc45f73c9729e58a634dede1b3551275c",
  "file": "S65sh4_cine.png",
  "staged_sha256": "8d4c60eae7a49e36165eb0d41178f369dcf4f217b3f8f4350725bd7cc04b7ce1",
  "latency_ms": 12644
 },
 "S65sh11::signage": {
  "fp": "8f5292b86108e1dd",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S65sh11": {
  "input_fingerprint": "3af9f515b7202dee",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 맨얼굴을 드러낸 이현우가 눈을 감고 고개를 치켜든 채 가슴을 부풀려 숨을 머금은 평온한 얼굴.\n\nLOCATION (lock): In the overgrown exterior grounds of the ruined department store, among moss, flowers and returning wildlife. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Collapsed department store (Partly collapsed) — An oblique fragment of the exterior remains behind his shoulder; used as Soft contextual background contrasting the peaceful face with structural destruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight preserves gentle facial contrast and the quiet relief of his closed-eyed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Moss, pools and uncontrolled vegetation cover the half-collapsed department store, and the radiation meter continues to read close to zero. Birds and other small wildlife occupy the regrown surroundings. 이현우: His gas mask is off, leaving his face uncovered as he breathes deeply. He retains his injuries and carries the prepared flashlight.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 맨얼굴을 드러낸 이현우가 눈을 감고 고개를 치켜든 채 가슴을 부풀려 숨을 머금은 평온한 얼굴.\n\nLOCATION (lock): In the overgrown exterior grounds of the ruined department store, among moss, flowers and returning wildlife. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Collapsed department store (Partly collapsed) — An oblique fragment of the exterior remains behind his shoulder; used as Soft contextual background contrasting the peaceful face with structural destruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight preserves gentle facial contrast and the quiet relief of his closed-eyed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Moss, pools and uncontrolled vegetation cover the half-collapsed department store, and the radiation meter continues to read close to zero. Birds and other small wildlife occupy the regrown surroundings. 이현우: His gas mask is off, leaving his face uncovered as he breathes deeply. He retains his injuries and carries the prepared flashlight.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 맨얼굴을 드러낸 이현우가 눈을 감고 고개를 치켜든 채 가슴을 부풀려 숨을 머금은 평온한 얼굴.\n\nLOCATION (lock): In the overgrown exterior grounds of the ruined department store, among moss, flowers and returning wildlife. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Collapsed department store (Partly collapsed) — An oblique fragment of the exterior remains behind his shoulder; used as Soft contextual background contrasting the peaceful face with structural destruction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight preserves gentle facial contrast and the quiet relief of his closed-eyed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Moss, pools and uncontrolled vegetation cover the half-collapsed department store, and the radiation meter continues to read close to zero. Birds and other small wildlife occupy the regrown surroundings. 이현우: His gas mask is off, leaving his face uncovered as he breathes deeply. He retains his injuries and carries the prepared flashlight.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "고개를 들고 눈을 감은 채 정면 상단을 향함.",
    "built_space": "인물 뒤편으로 이끼와 식물로 덮인 부분적으로 무너진 콘크리트 건물 외벽이 위치함.",
    "entities": "이현우(레퍼런스와 일치하는 검은 후드티, 상처 난 얼굴), 붕괴된 백화점, 이끼, 식물, 새.",
    "hard_violations": [],
    "physics": "자연스럽게 땅에 서서 가슴을 부풀려 호흡하는 자세이며, 물리적 지지 오류 없음."
   },
   {
    "label": "B",
    "direction": "고개를 들고 눈을 감은 채 정면 상단을 향함.",
    "built_space": "인물 뒤편으로 식물에 뒤덮인 붕괴된 건물이 배경으로 자리함.",
    "entities": "이현우(레퍼런스와 불일치하는 이전 샷의 녹색 재킷 및 장갑 착용), 손에 든 방독면, 붕괴된 백화점, 식물.",
    "hard_violations": [],
    "physics": "자연스럽게 서서 한 손으로 방독면을 지탱하여 쥐고 있으며, 물리적 오류 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "눈을 감고 숨을 들이마시는 평온한 표정을 잘 표현했으며, 지시된 캐릭터 레퍼런스 의상(검은 후드티)과 배경의 야생동물(새)까지 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물의 표정과 배경은 적절하나, 이전 샷의 의상을 복사하지 말라는 명시적 지시를 어기고 레퍼런스의 의상을 무시한 채 녹색 재킷과 장갑을 착용했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개를 들고 눈을 감은 채 정면 상단을 향함.",
        "built_space": "인물 뒤편으로 이끼와 식물로 덮인 부분적으로 무너진 콘크리트 건물 외벽이 위치함.",
        "entities": "이현우(레퍼런스와 일치하는 검은 후드티, 상처 난 얼굴), 붕괴된 백화점, 이끼, 식물, 새.",
        "hard_violations": [],
        "physics": "자연스럽게 땅에 서서 가슴을 부풀려 호흡하는 자세이며, 물리적 지지 오류 없음."
       },
       {
        "label": "B",
        "direction": "고개를 들고 눈을 감은 채 정면 상단을 향함.",
        "built_space": "인물 뒤편으로 식물에 뒤덮인 붕괴된 건물이 배경으로 자리함.",
        "entities": "이현우(레퍼런스와 불일치하는 이전 샷의 녹색 재킷 및 장갑 착용), 손에 든 방독면, 붕괴된 백화점, 식물.",
        "hard_violations": [],
        "physics": "자연스럽게 서서 한 손으로 방독면을 지탱하여 쥐고 있으며, 물리적 오류 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "눈을 감고 숨을 들이마시는 평온한 표정을 잘 표현했으며, 지시된 캐릭터 레퍼런스 의상(검은 후드티)과 배경의 야생동물(새)까지 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물의 표정과 배경은 적절하나, 이전 샷의 의상을 복사하지 말라는 명시적 지시를 어기고 레퍼런스의 의상을 무시한 채 녹색 재킷과 장갑을 착용했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "고개를 들고 눈을 감은 채 정면 상단을 향함.",
        "built_space": "인물 뒤편으로 이끼와 식물로 덮인 부분적으로 무너진 콘크리트 건물 외벽이 위치함.",
        "entities": "이현우(레퍼런스와 일치하는 검은 후드티, 상처 난 얼굴), 붕괴된 백화점, 이끼, 식물, 새.",
        "hard_violations": [],
        "physics": "자연스럽게 땅에 서서 가슴을 부풀려 호흡하는 자세이며, 물리적 지지 오류 없음."
       },
       {
        "label": "B",
        "direction": "고개를 들고 눈을 감은 채 정면 상단을 향함.",
        "built_space": "인물 뒤편으로 식물에 뒤덮인 붕괴된 건물이 배경으로 자리함.",
        "entities": "이현우(레퍼런스와 불일치하는 이전 샷의 녹색 재킷 및 장갑 착용), 손에 든 방독면, 붕괴된 백화점, 식물.",
        "hard_violations": [],
        "physics": "자연스럽게 서서 한 손으로 방독면을 지탱하여 쥐고 있으며, 물리적 오류 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "맨얼굴과 감은 눈, 치켜든 고개를 얼굴 중심 클로즈업으로 가장 정확히 구현했지만, 갈색 겉옷과 장갑은 인물 참조의 복장과 다르다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "검은 후드와 상처, 평온한 호흡 연기는 충실하지만, 가슴까지 넓힌 미디엄 클로즈업과 크게 드러난 폐허가 지정된 얼굴 중심 구도에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈을 감고 턱과 코를 화면 왼쪽 위로 들어 올렸다. 특정 대상을 바라보는 것이 아니라 눈을 감은 채 숨을 머금는 방향으로, 지시된 동작에 맞는다. 아래쪽 손은 벗은 방독면을 가슴 앞에서 잡고 있다.",
        "built_space": "인물 뒤 왼쪽과 상단에 회색 콘크리트 백화점 외벽 한 동의 일부가 비스듬히 보인다. 왼쪽의 파손된 큰 개구부, 중앙의 가로 창열, 덩굴과 잔해가 이전 장소의 재료와 특징을 이어간다. 아래에는 이끼 낀 돌과 꽃이 있으며, 인물은 건물 외부에 있다. 중복 설비나 부자연스러운 반사는 보이지 않는다.",
        "entities": "젊은 한국계 남성으로 보이는 인물 한 명이며 검고 헝클어진 머리와 얼굴 윤곽은 참조에 대체로 부합한다. 얼굴은 가려지지 않았고 표정은 평온하다. 다만 참조의 낡은 검은 후드 대신 갈색 계열 겉옷과 손가락 없는 장갑을 착용해 이전 장면의 복장을 가져온 듯하다. 피부의 얼룩은 보이나 뚜렷한 상처는 확인하기 어렵다. 방독면은 화면 아래 일부 보인다. 손전등과 계측기, 물웅덩이와 야생동물은 이 구도에서 확인되지 않으며, 화면 밖 소지 상태는 판단할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되어 있으며 뒤로 젖힌 정도도 가능한 범위다. 방독면은 손가락으로 잡혀 있고 끈도 목 주변으로 이어져 지지가 분명하다. 하체와 지면 접촉은 클로즈업 밖이지만 부유를 시사하는 모습은 없다. 가슴의 팽창은 하단 크롭 때문에 제한적으로만 보인다."
       },
       {
        "label": "B",
        "direction": "두 눈을 감고 얼굴을 화면 왼쪽 위로 향하게 하여 턱을 들었다. 시선의 외부 목표는 없으며 지시된 평온한 호흡 동작과 맞는다. 입은 편안하게 닫혀 있고 상체는 곧게 펴져 있다.",
        "built_space": "회색 콘크리트 백화점 외벽 한 동과 여러 어두운 개구부가 뒤에 보인다. 왼쪽에는 벽돌 속살이 드러난 커다란 사선 붕괴 부재 하나가 있고, 하단에는 이끼 낀 잔해와 꽃이 있다. 외부 폐허라는 장소는 맞지만 사선 부재의 정확한 연속성은 이전 사진으로 확인되지 않는다. 배경 파괴부가 어깨 뒤의 작은 맥락적 조각보다 넓고 두드러지게 제시된다. 중복 설비나 불가능한 반사는 없다.",
        "entities": "검고 헝클어진 머리의 젊은 한국계 남성 한 명으로, 얼굴과 체격이 참조에 대체로 부합한다. 낡고 구멍 난 검은 후드는 인물 참조와 맞으며, 코와 눈썹 주변 및 뺨의 상처도 보인다. 얼굴에 방독면은 없다. 이끼와 여러 색의 꽃, 잔해에 앉은 작은 새가 최소 다섯 마리 보인다. 손과 손전등, 방독면의 보관 위치 및 계측기는 화면 밖이므로 소지 여부를 판단할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 머리를 정상적으로 지지하고, 뒤로 약간 젖힌 머리와 펼친 상체는 깊게 들이쉰 자세로 가능하다. 가슴 부분은 보이지만 실제 호흡 팽창량은 정지 화면만으로 확정할 수 없다. 새들은 돌과 이끼 낀 잔해 위에 앉아 있으며 붕괴 부재도 아래 잔해에 닿아 있다. 하체는 화면 밖이고, 보이는 범위에 지지 없는 부유 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "맨얼굴과 감은 눈, 치켜든 고개를 얼굴 중심 클로즈업으로 가장 정확히 구현했지만, 갈색 겉옷과 장갑은 인물 참조의 복장과 다르다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "검은 후드와 상처, 평온한 호흡 연기는 충실하지만, 가슴까지 넓힌 미디엄 클로즈업과 크게 드러난 폐허가 지정된 얼굴 중심 구도에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 눈을 감고 턱과 코를 화면 왼쪽 위로 들어 올렸다. 특정 대상을 바라보는 것이 아니라 눈을 감은 채 숨을 머금는 방향으로, 지시된 동작에 맞는다. 아래쪽 손은 벗은 방독면을 가슴 앞에서 잡고 있다.",
        "built_space": "인물 뒤 왼쪽과 상단에 회색 콘크리트 백화점 외벽 한 동의 일부가 비스듬히 보인다. 왼쪽의 파손된 큰 개구부, 중앙의 가로 창열, 덩굴과 잔해가 이전 장소의 재료와 특징을 이어간다. 아래에는 이끼 낀 돌과 꽃이 있으며, 인물은 건물 외부에 있다. 중복 설비나 부자연스러운 반사는 보이지 않는다.",
        "entities": "젊은 한국계 남성으로 보이는 인물 한 명이며 검고 헝클어진 머리와 얼굴 윤곽은 참조에 대체로 부합한다. 얼굴은 가려지지 않았고 표정은 평온하다. 다만 참조의 낡은 검은 후드 대신 갈색 계열 겉옷과 손가락 없는 장갑을 착용해 이전 장면의 복장을 가져온 듯하다. 피부의 얼룩은 보이나 뚜렷한 상처는 확인하기 어렵다. 방독면은 화면 아래 일부 보인다. 손전등과 계측기, 물웅덩이와 야생동물은 이 구도에서 확인되지 않으며, 화면 밖 소지 상태는 판단할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되어 있으며 뒤로 젖힌 정도도 가능한 범위다. 방독면은 손가락으로 잡혀 있고 끈도 목 주변으로 이어져 지지가 분명하다. 하체와 지면 접촉은 클로즈업 밖이지만 부유를 시사하는 모습은 없다. 가슴의 팽창은 하단 크롭 때문에 제한적으로만 보인다."
       },
       {
        "label": "A",
        "direction": "두 눈을 감고 얼굴을 화면 왼쪽 위로 향하게 하여 턱을 들었다. 시선의 외부 목표는 없으며 지시된 평온한 호흡 동작과 맞는다. 입은 편안하게 닫혀 있고 상체는 곧게 펴져 있다.",
        "built_space": "회색 콘크리트 백화점 외벽 한 동과 여러 어두운 개구부가 뒤에 보인다. 왼쪽에는 벽돌 속살이 드러난 커다란 사선 붕괴 부재 하나가 있고, 하단에는 이끼 낀 잔해와 꽃이 있다. 외부 폐허라는 장소는 맞지만 사선 부재의 정확한 연속성은 이전 사진으로 확인되지 않는다. 배경 파괴부가 어깨 뒤의 작은 맥락적 조각보다 넓고 두드러지게 제시된다. 중복 설비나 불가능한 반사는 없다.",
        "entities": "검고 헝클어진 머리의 젊은 한국계 남성 한 명으로, 얼굴과 체격이 참조에 대체로 부합한다. 낡고 구멍 난 검은 후드는 인물 참조와 맞으며, 코와 눈썹 주변 및 뺨의 상처도 보인다. 얼굴에 방독면은 없다. 이끼와 여러 색의 꽃, 잔해에 앉은 작은 새가 최소 다섯 마리 보인다. 손과 손전등, 방독면의 보관 위치 및 계측기는 화면 밖이므로 소지 여부를 판단할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 머리를 정상적으로 지지하고, 뒤로 약간 젖힌 머리와 펼친 상체는 깊게 들이쉰 자세로 가능하다. 가슴 부분은 보이지만 실제 호흡 팽창량은 정지 화면만으로 확정할 수 없다. 새들은 돌과 이끼 낀 잔해 위에 앉아 있으며 붕괴 부재도 아래 잔해에 닿아 있다. 하체는 화면 밖이고, 보이는 범위에 지지 없는 부유 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.571
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1875,
   "B": 1571
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "눈을 감고 숨을 들이마시는 평온한 표정을 잘 표현했으며, 지시된 캐릭터 레퍼런스 의상(검은 후드티)과 배경의 야생동물(새)까지 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "인물의 표정과 배경은 적절하나, 이전 샷의 의상을 복사하지 말라는 명시적 지시를 어기고 레퍼런스의 의상을 무시한 채 녹색 재킷과 장갑을 착용했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S65sh4_sel.png",
    "asset_id": "001f7631-5505-4f63-b493-d674d86521fd",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdfe-bc53-7eb3-b9d6-5f53822b11e3",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S65sh4"
  }
 },
 "S65sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:53:31.781299+00:00",
  "fingerprint": "6529e1b667eed79c63d3cf7f5b6b1e6f19f8327c6fe99bbf6451f490d89f3af3",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S65sh11_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S65sh11_sel.png",
  "source_sha256": "103e0fc5ba256e42116615245a180e91cb2a359736d3172ac54f1df451c5167c",
  "file": "S65sh11_cine.png",
  "staged_sha256": "f2dea701d3468eb3d39f86a39a8f1af9e7350e7335b4d2e86d6705232bf91f9b",
  "latency_ms": 10226
 },
 "S65sh17::signage": {
  "fp": "5b5df94c432bcca6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S65sh17::bgfirst_bg": {
  "input_fingerprint": "ef16cb1ec1a43259",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 높은 잔해 위 어둠 속, 우뚝 선 아시아계 남성의 짙은 실루엣 뒷모습 너머로 아래쪽의 이현우와 수빈이 백화점 틈새 안으로 몸이 반쯤 가려진 조감 구도.\n\nLOCATION (lock): On elevated rubble outside the collapsed department store, overlooking the breach through which the visitors enter.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Department-store opening concealing the entering pair in the lower-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Department-store opening (Open passage through the collapsed structure) — Viewed obliquely from above; its near edge conceals the entering pair's lower bodies; used as Lower-right destination and shared depth anchor linking the observer to the pair; Elevated collapsed remains (Collapsed) — The upper surface supports the observer in the foreground; used as Establishes the observer's physical elevation without an exaggerated foreground mass.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the observer in deep shadow against the more readable daytime opening below, without revealing his facial identity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 높은 잔해 위 어둠 속, 우뚝 선 아시아계 남성의 짙은 실루엣 뒷모습 너머로 아래쪽의 이현우와 수빈이 백화점 틈새 안으로 몸이 반쯤 가려진 조감 구도.\n\nLOCATION (lock): On elevated rubble outside the collapsed department store, overlooking the breach through which the visitors enter.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Department-store opening concealing the entering pair in the lower-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Department-store opening (Open passage through the collapsed structure) — Viewed obliquely from above; its near edge conceals the entering pair's lower bodies; used as Lower-right destination and shared depth anchor linking the observer to the pair; Elevated collapsed remains (Collapsed) — The upper surface supports the observer in the foreground; used as Establishes the observer's physical elevation without an exaggerated foreground mass.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the observer in deep shadow against the more readable daytime opening below, without revealing his facial identity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S65sh17__bgfirst_bg.png",
  "asset_id": "934a963a-1e3d-4b39-89f4-7038e0daab7d",
  "input_asset_ids": [
   "c606afa2-ca2e-490c-88a0-db4e0d779212",
   "3ab84dc4-2692-40b2-a37a-53eff77ce828",
   "9dcd2529-ce3b-4608-b19e-97c9ddf73437"
  ]
 },
 "S65sh17": {
  "input_fingerprint": "0492676bdcaa5836",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 높은 잔해 위 어둠 속, 우뚝 선 아시아계 남성의 짙은 실루엣 뒷모습 너머로 아래쪽의 이현우와 수빈이 백화점 틈새 안으로 몸이 반쯤 가려진 조감 구도.\n\nLOCATION (lock): On elevated rubble outside the collapsed department store, overlooking the breach through which the visitors enter. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Department-store opening concealing the entering pair in the lower-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Department-store opening (Open passage through the collapsed structure) — Viewed obliquely from above; its near edge conceals the entering pair's lower bodies; used as Lower-right destination and shared depth anchor linking the observer to the pair; Elevated collapsed remains (Collapsed) — The upper surface supports the observer in the foreground; used as Establishes the observer's physical elevation without an exaggerated foreground mass.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the observer in deep shadow against the more readable daytime opening below, without revealing his facial identity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The moss-covered, half-collapsed department store has an accessible gap through its ruins, with pools and vegetation around it. The discarded gas mask lies outside, and the radiation meter's established reading remains close to zero. 이현우: He enters the ruined building with his face uncovered and a flashlight carried for the dark interior. His existing injuries persist. 수빈: She enters with her gas mask removed and discarded, retaining the radiation meter and a prepared flashlight. Her facial wounds and radiation-damaged abdominal skin remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 높은 잔해 위 어둠 속, 우뚝 선 아시아계 남성의 짙은 실루엣 뒷모습 너머로 아래쪽의 이현우와 수빈이 백화점 틈새 안으로 몸이 반쯤 가려진 조감 구도.\n\nLOCATION (lock): On elevated rubble outside the collapsed department store, overlooking the breach through which the visitors enter. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Department-store opening concealing the entering pair in the lower-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Department-store opening (Open passage through the collapsed structure) — Viewed obliquely from above; its near edge conceals the entering pair's lower bodies; used as Lower-right destination and shared depth anchor linking the observer to the pair; Elevated collapsed remains (Collapsed) — The upper surface supports the observer in the foreground; used as Establishes the observer's physical elevation without an exaggerated foreground mass.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the observer in deep shadow against the more readable daytime opening below, without revealing his facial identity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The moss-covered, half-collapsed department store has an accessible gap through its ruins, with pools and vegetation around it. The discarded gas mask lies outside, and the radiation meter's established reading remains close to zero. 이현우: He enters the ruined building with his face uncovered and a flashlight carried for the dark interior. His existing injuries persist. 수빈: She enters with her gas mask removed and discarded, retaining the radiation meter and a prepared flashlight. Her facial wounds and radiation-damaged abdominal skin remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 높은 잔해 위 어둠 속, 우뚝 선 아시아계 남성의 짙은 실루엣 뒷모습 너머로 아래쪽의 이현우와 수빈이 백화점 틈새 안으로 몸이 반쯤 가려진 조감 구도.\n\nLOCATION (lock): On elevated rubble outside the collapsed department store, overlooking the breach through which the visitors enter. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Department-store opening concealing the entering pair in the lower-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Department-store opening (Open passage through the collapsed structure) — Viewed obliquely from above; its near edge conceals the entering pair's lower bodies; used as Lower-right destination and shared depth anchor linking the observer to the pair; Elevated collapsed remains (Collapsed) — The upper surface supports the observer in the foreground; used as Establishes the observer's physical elevation without an exaggerated foreground mass.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the observer in deep shadow against the more readable daytime opening below, without revealing his facial identity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The moss-covered, half-collapsed department store has an accessible gap through its ruins, with pools and vegetation around it. The discarded gas mask lies outside, and the radiation meter's established reading remains close to zero. 이현우: He enters the ruined building with his face uncovered and a flashlight carried for the dark interior. His existing injuries persist. 수빈: She enters with her gas mask removed and discarded, retaining the radiation meter and a prepared flashlight. Her facial wounds and radiation-damaged abdominal skin remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S65sh17__bgfirst_bg.png",
     "asset_id": "934a963a-1e3d-4b39-89f4-7038e0daab7d",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S65sh17.png",
     "asset_id": "c606afa2-ca2e-490c-88a0-db4e0d779212",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904512>",
     "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L100B01.png",
     "asset_id": "3ab84dc4-2692-40b2-a37a-53eff77ce828",
     "role": "location_plate"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_collapsed_department_store_sel.png",
     "asset_id": "9dcd2529-ce3b-4608-b19e-97c9ddf73437",
     "role": "structure_seed_look"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904512>",
     "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "관찰자는 원경의 두 인물을 향하고 있으며, 인물들은 정면으로 손전등을 비추고 있습니다.",
    "built_space": "위치 참조와 유사한 폐허이나, 지정된 삼풍백화점 외관(색상, 간판 등)이 반영되지 않았습니다. 인물들은 틈새에 가려지지 않고 완전히 노출되어 있습니다.",
    "entities": "전경에 남성 실루엣이 있고, 원경에 이현우와 수빈이 확인되나 수빈의 방사능 측정기는 보이지 않습니다.",
    "hard_violations": [],
    "physics": "모든 인물이 지면과 잔해 위에 서 있어 중력과 지지 상태가 정상적입니다."
   },
   {
    "label": "B",
    "direction": "관찰자는 우측 아래의 인물들을 내려다보고 있으며, 두 인물은 앞을 향해 손전등을 비추고 있습니다.",
    "built_space": "배경에 구조 참조 사진과 일치하는 핑크색 백화점 외관과 간판이 렌더링되었고, 우측 하단 잔해 틈새에 인물들이 반쯤 가려진 채 진입하는 구도가 형성되었습니다.",
    "entities": "전경에 남성의 짙은 실루엣이 있으며, 아래쪽의 이현우와 수빈은 참조된 외모와 일치합니다. 수빈은 손전등과 노란색 방사능 측정기를 들고 있고 곁에 방독면이 있습니다.",
    "hard_violations": [
     "[gpt-high] 오른쪽 위 배너의 영문이 판독 가능하여 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
    ],
    "physics": "관찰자와 두 인물 모두 잔해와 바닥 위에 안정적으로 서 있으며, 방독면 등 사물도 바닥에 정상적으로 놓여 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 백화점 외관을 배경에 잘 재현했으며, 하단 틈새에 반쯤 가려진 인물들의 구도와 수빈의 방사능 측정기 등 세부 지침을 충실히 반영했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "필수적인 백화점 외관 참조를 완전히 누락했으며, 인물들이 틈새에 가려지지 않고 노출되어 구도 지침을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "관찰자는 우측 아래의 인물들을 내려다보고 있으며, 두 인물은 앞을 향해 손전등을 비추고 있습니다.",
        "built_space": "배경에 구조 참조 사진과 일치하는 핑크색 백화점 외관과 간판이 렌더링되었고, 우측 하단 잔해 틈새에 인물들이 반쯤 가려진 채 진입하는 구도가 형성되었습니다.",
        "entities": "전경에 남성의 짙은 실루엣이 있으며, 아래쪽의 이현우와 수빈은 참조된 외모와 일치합니다. 수빈은 손전등과 노란색 방사능 측정기를 들고 있고 곁에 방독면이 있습니다.",
        "hard_violations": [],
        "physics": "관찰자와 두 인물 모두 잔해와 바닥 위에 안정적으로 서 있으며, 방독면 등 사물도 바닥에 정상적으로 놓여 있습니다."
       },
       {
        "label": "A",
        "direction": "관찰자는 원경의 두 인물을 향하고 있으며, 인물들은 정면으로 손전등을 비추고 있습니다.",
        "built_space": "위치 참조와 유사한 폐허이나, 지정된 삼풍백화점 외관(색상, 간판 등)이 반영되지 않았습니다. 인물들은 틈새에 가려지지 않고 완전히 노출되어 있습니다.",
        "entities": "전경에 남성 실루엣이 있고, 원경에 이현우와 수빈이 확인되나 수빈의 방사능 측정기는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "모든 인물이 지면과 잔해 위에 서 있어 중력과 지지 상태가 정상적입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 백화점 외관을 배경에 잘 재현했으며, 하단 틈새에 반쯤 가려진 인물들의 구도와 수빈의 방사능 측정기 등 세부 지침을 충실히 반영했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "필수적인 백화점 외관 참조를 완전히 누락했으며, 인물들이 틈새에 가려지지 않고 노출되어 구도 지침을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "관찰자는 우측 아래의 인물들을 내려다보고 있으며, 두 인물은 앞을 향해 손전등을 비추고 있습니다.",
        "built_space": "배경에 구조 참조 사진과 일치하는 핑크색 백화점 외관과 간판이 렌더링되었고, 우측 하단 잔해 틈새에 인물들이 반쯤 가려진 채 진입하는 구도가 형성되었습니다.",
        "entities": "전경에 남성의 짙은 실루엣이 있으며, 아래쪽의 이현우와 수빈은 참조된 외모와 일치합니다. 수빈은 손전등과 노란색 방사능 측정기를 들고 있고 곁에 방독면이 있습니다.",
        "hard_violations": [],
        "physics": "관찰자와 두 인물 모두 잔해와 바닥 위에 안정적으로 서 있으며, 방독면 등 사물도 바닥에 정상적으로 놓여 있습니다."
       },
       {
        "label": "A",
        "direction": "관찰자는 원경의 두 인물을 향하고 있으며, 인물들은 정면으로 손전등을 비추고 있습니다.",
        "built_space": "위치 참조와 유사한 폐허이나, 지정된 삼풍백화점 외관(색상, 간판 등)이 반영되지 않았습니다. 인물들은 틈새에 가려지지 않고 완전히 노출되어 있습니다.",
        "entities": "전경에 남성 실루엣이 있고, 원경에 이현우와 수빈이 확인되나 수빈의 방사능 측정기는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "모든 인물이 지면과 잔해 위에 서 있어 중력과 지지 상태가 정상적입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "판독 가능한 배너 글자가 금지 조건을 위반하며, 두 사람이 틈새로 들어가기보다 관찰자를 향해 올라오는 모습이고 목적지도 배경보다 지나치게 가깝다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "높은 잔해의 어두운 관찰자 너머 오른쪽 아래 출입구를 내려다보는 공간 관계는 우세하지만, 두 사람이 밖으로 나오며 하반신도 충분히 가려지지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경 남성은 등을 보이고 오른쪽 아래 두 사람 쪽으로 고개를 돌린다. 이현우와 수빈은 건물 내부가 아니라 위쪽 관찰자를 바라보며 그쪽으로 향한다. 두 손전등도 어두운 내부보다 카메라와 관찰자 쪽을 향한다. 요구된 진입 방향과 반대다.",
        "built_space": "왼쪽 전경의 관찰자는 이끼 낀 높은 콘크리트 판 위에 서 있다. 오른쪽에는 큰 경사 붕괴판 아래 틈새 하나가 있고, 두 사람은 그 안쪽 아래에 있다. 가까운 잔해가 다리 일부를 가리지만 두 사람과 틈새가 배경이라기보다 바로 인접한 공간으로 보인다. 중앙 후방에는 별도의 직립 출입구가 보인다. 분홍 외장, 파란 띠, 노출된 붕괴층과 식생은 구조 참고의 특징을 반영하나, 장소 참고의 진입 동선보다 오른쪽 경사판 틈새가 주된 목적지가 되었다.",
        "entities": "총 세 사람이 보이며, 전경의 얼굴을 숨긴 검은 머리 남성과 아래쪽 젊은 동아시아계 남녀로 읽힌다. 전경 남성의 정확한 민족 정체성은 확인할 수 없다. 이현우의 헝클어진 검은 머리·짙은 후드·찢어진 청바지와 수빈의 짧은 검은 머리·갈색 재킷·목도리는 참고와 대체로 맞는다. 정확한 나이와 얼굴 동일성은 이 거리에서 확정하기 어렵다. 두 사람 모두 방독면을 쓰지 않았고 손전등을 하나씩 들었다. 수빈의 다른 손에는 노란 계측기가 있으나 수치는 확인되지 않는다. 방독면 하나가 전경 판 위에 놓여 있다. 얼굴 상처의 일치 여부는 불분명하며 복부는 옷으로 가려져 있다. 오른쪽 위 배너에는 읽을 수 있는 영문이 있다.",
        "hard_violations": [
         "오른쪽 위 배너의 영문이 판독 가능하여 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "관찰자의 두 신발은 전경 콘크리트 판에 닿아 체중을 지탱한다. 두 방문자의 발은 잔해에 가려져 직접 확인되지 않지만 몸은 아래 통로에 서 있는 위치로 이어지며 공중에 뜬 증거는 없다. 손전등과 계측기는 각각 손으로 잡고 있고 방독면은 판 위에 놓여 있다. 경사 붕괴판은 주변 잔해에 걸쳐 지지된다."
       },
       {
        "label": "B",
        "direction": "전경 남성의 몸과 고개는 오른쪽 아래 출입구와 두 사람을 향한다. 이현우는 출입구에서 바깥 포장로 쪽으로 발을 내딛고, 수빈도 그 뒤에서 바깥을 향한다. 이현우의 손전등은 바깥쪽 아래 바닥을 비춘다. 관찰 대상은 맞지만 방문자들의 이동은 요구된 건물 안 진입과 반대다.",
        "built_space": "왼쪽 전경의 높은 잔해 위에 관찰자가 있고, 오른쪽 아래 중경에 큰 파손 출입구 하나가 있다. 출입구 양쪽에는 낮은 파손 벽체가 하나씩 남아 있으며 내부 수직 기둥과 바닥이 보인다. 이현우는 문턱 바깥으로 나와 다리와 발까지 드러나고 수빈도 하반신 대부분이 보여, 가까운 개구부 가장자리가 두 사람을 반쯤 가리는 구도는 아니다. 외벽의 분홍·회색 패널, 가로 창 띠, 중앙 접근로와 양쪽 물웅덩이 및 식생은 장소 참고에 가깝다. 구조 참고의 큰 붕괴부는 이 화면에서 확인되지 않는다.",
        "entities": "총 세 사람이 있으며 전경은 검은 머리의 남성 뒷모습이다. 얼굴이 어두워 정확한 민족 정체성은 확인할 수 없다. 출입구의 두 사람은 젊은 동아시아계 남녀로 읽히며, 이현우의 어두운 후드와 청바지, 수빈의 짧은 머리와 재킷·목도리는 참고와 대체로 맞는다. 작은 화면 크기 때문에 정확한 나이와 얼굴 동일성은 확인하기 어렵다. 두 사람의 얼굴은 방독면으로 가리지 않았고 손전등을 각각 들고 있다. 문턱 앞 바닥에 버려진 방독면으로 보이는 물체 하나가 있다. 방사선 계측기와 수치, 얼굴 상처는 식별하기 어렵고 복부 피부는 옷에 가려져 있다. 판독 가능한 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "관찰자의 신발은 잔해 상면에 닿아 있다. 이현우는 한 발로 문턱 부근을 딛고 다른 발을 옮기는 보행 자세이며, 수빈도 출입구 바닥 위에 서 있다. 두 손전등은 손에 잡혀 있고 방독면으로 보이는 물체는 포장로 위에 놓여 있다. 지지 없는 인체나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "판독 가능한 배너 글자가 금지 조건을 위반하며, 두 사람이 틈새로 들어가기보다 관찰자를 향해 올라오는 모습이고 목적지도 배경보다 지나치게 가깝다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "높은 잔해의 어두운 관찰자 너머 오른쪽 아래 출입구를 내려다보는 공간 관계는 우세하지만, 두 사람이 밖으로 나오며 하반신도 충분히 가려지지 않는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "전경 남성은 등을 보이고 오른쪽 아래 두 사람 쪽으로 고개를 돌린다. 이현우와 수빈은 건물 내부가 아니라 위쪽 관찰자를 바라보며 그쪽으로 향한다. 두 손전등도 어두운 내부보다 카메라와 관찰자 쪽을 향한다. 요구된 진입 방향과 반대다.",
        "built_space": "왼쪽 전경의 관찰자는 이끼 낀 높은 콘크리트 판 위에 서 있다. 오른쪽에는 큰 경사 붕괴판 아래 틈새 하나가 있고, 두 사람은 그 안쪽 아래에 있다. 가까운 잔해가 다리 일부를 가리지만 두 사람과 틈새가 배경이라기보다 바로 인접한 공간으로 보인다. 중앙 후방에는 별도의 직립 출입구가 보인다. 분홍 외장, 파란 띠, 노출된 붕괴층과 식생은 구조 참고의 특징을 반영하나, 장소 참고의 진입 동선보다 오른쪽 경사판 틈새가 주된 목적지가 되었다.",
        "entities": "총 세 사람이 보이며, 전경의 얼굴을 숨긴 검은 머리 남성과 아래쪽 젊은 동아시아계 남녀로 읽힌다. 전경 남성의 정확한 민족 정체성은 확인할 수 없다. 이현우의 헝클어진 검은 머리·짙은 후드·찢어진 청바지와 수빈의 짧은 검은 머리·갈색 재킷·목도리는 참고와 대체로 맞는다. 정확한 나이와 얼굴 동일성은 이 거리에서 확정하기 어렵다. 두 사람 모두 방독면을 쓰지 않았고 손전등을 하나씩 들었다. 수빈의 다른 손에는 노란 계측기가 있으나 수치는 확인되지 않는다. 방독면 하나가 전경 판 위에 놓여 있다. 얼굴 상처의 일치 여부는 불분명하며 복부는 옷으로 가려져 있다. 오른쪽 위 배너에는 읽을 수 있는 영문이 있다.",
        "hard_violations": [
         "오른쪽 위 배너의 영문이 판독 가능하여 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
        ],
        "physics": "관찰자의 두 신발은 전경 콘크리트 판에 닿아 체중을 지탱한다. 두 방문자의 발은 잔해에 가려져 직접 확인되지 않지만 몸은 아래 통로에 서 있는 위치로 이어지며 공중에 뜬 증거는 없다. 손전등과 계측기는 각각 손으로 잡고 있고 방독면은 판 위에 놓여 있다. 경사 붕괴판은 주변 잔해에 걸쳐 지지된다."
       },
       {
        "label": "A",
        "direction": "전경 남성의 몸과 고개는 오른쪽 아래 출입구와 두 사람을 향한다. 이현우는 출입구에서 바깥 포장로 쪽으로 발을 내딛고, 수빈도 그 뒤에서 바깥을 향한다. 이현우의 손전등은 바깥쪽 아래 바닥을 비춘다. 관찰 대상은 맞지만 방문자들의 이동은 요구된 건물 안 진입과 반대다.",
        "built_space": "왼쪽 전경의 높은 잔해 위에 관찰자가 있고, 오른쪽 아래 중경에 큰 파손 출입구 하나가 있다. 출입구 양쪽에는 낮은 파손 벽체가 하나씩 남아 있으며 내부 수직 기둥과 바닥이 보인다. 이현우는 문턱 바깥으로 나와 다리와 발까지 드러나고 수빈도 하반신 대부분이 보여, 가까운 개구부 가장자리가 두 사람을 반쯤 가리는 구도는 아니다. 외벽의 분홍·회색 패널, 가로 창 띠, 중앙 접근로와 양쪽 물웅덩이 및 식생은 장소 참고에 가깝다. 구조 참고의 큰 붕괴부는 이 화면에서 확인되지 않는다.",
        "entities": "총 세 사람이 있으며 전경은 검은 머리의 남성 뒷모습이다. 얼굴이 어두워 정확한 민족 정체성은 확인할 수 없다. 출입구의 두 사람은 젊은 동아시아계 남녀로 읽히며, 이현우의 어두운 후드와 청바지, 수빈의 짧은 머리와 재킷·목도리는 참고와 대체로 맞는다. 작은 화면 크기 때문에 정확한 나이와 얼굴 동일성은 확인하기 어렵다. 두 사람의 얼굴은 방독면으로 가리지 않았고 손전등을 각각 들고 있다. 문턱 앞 바닥에 버려진 방독면으로 보이는 물체 하나가 있다. 방사선 계측기와 수치, 얼굴 상처는 식별하기 어렵고 복부 피부는 옷에 가려져 있다. 판독 가능한 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "관찰자의 신발은 잔해 상면에 닿아 있다. 이현우는 한 발로 문턱 부근을 딛고 다른 발을 옮기는 보행 자세이며, 수빈도 출입구 바닥 위에 서 있다. 두 손전등은 손에 잡혀 있고 방독면으로 보이는 물체는 포장로 위에 놓여 있다. 지지 없는 인체나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.6
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.35
   },
   "violations": {
    "B": [
     "[gpt-high] 오른쪽 위 배너의 영문이 판독 가능하여 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1350,
   "A": 1571
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1350,
    "verdict_ko": "지정된 백화점 외관을 배경에 잘 재현했으며, 하단 틈새에 반쯤 가려진 인물들의 구도와 수빈의 방사능 측정기 등 세부 지침을 충실히 반영했습니다.  ★위반: [gpt-high] 오른쪽 위 배너의 영문이 판독 가능하여 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "필수적인 백화점 외관 참조를 완전히 누락했으며, 인물들이 틈새에 가려지지 않고 노출되어 구도 지침을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L100B01.png",
    "asset_id": "3ab84dc4-2692-40b2-a37a-53eff77ce828",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_collapsed_department_store_sel.png",
    "asset_id": "9dcd2529-ce3b-4608-b19e-97c9ddf73437",
    "role": "structure_seed_look"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:904512>",
    "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aacdff-6572-7346-a729-dd6440b2b309",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S65sh17__bgfirst_bg.png",
   "bg_asset_id": "934a963a-1e3d-4b39-89f4-7038e0daab7d",
   "bg_record_key": "S65sh17::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S65sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:53:42.437560+00:00",
  "fingerprint": "b73348c9d34b28ebb9ba597e48546310da0dd9f59f71bee7ce21f79e3445af84",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S65sh17_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S65sh17_sel.png",
  "source_sha256": "e6cd776d6ce92f9e27733ed244ee3ded498d2569e993bcc43b3362477173c9f4",
  "file": "S65sh17_cine.png",
  "staged_sha256": "a657b18d7d9e1f01b7bbc957fdbc3ee249915d14e377f7c94317e12017345291",
  "latency_ms": 25749
 },
 "S66sh9::signage": {
  "fp": "9ab2a7e3815338c9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::39066a096593e7f3": {
  "subjects": [],
  "subject_text": "무너진 백화점 1층과 붕괴된 바닥 주변\n먼지 쌓인 안내판과 지하로 내려가는 에스컬레이터가 있는 어두운 대형 매장층. 두꺼운 벽체 사이로 좁은 통로가 이어진다.",
  "identity": "canonical",
  "scope_id": "L101",
  "scope_role": "location_interior",
  "scope_sha": "50aa63082d5beb11"
 },
 "S66sh9::bgfirst_bg": {
  "input_fingerprint": "a585993f18179480",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 구멍 가장자리로 몸을 날려 엎드린 수빈이 허공에 매달린 이현우의 손목을 두 손으로 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): At the collapsing floor edge near a descending escalator inside the ruined department store, lit by the explorers' flashlights.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Newly broken floor edge (Broken open above the lower floor) — The upper surface beneath 수빈 and the exposed edge above 이현우 are both visible; used as A narrow diagonal divider establishes the load-bearing edge and the difference in body levels; Opening beneath the hands (Open after the sudden floor collapse) — Seen downward from beside the rim; used as Negative space below the wrist conveys the unsupported drop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the department store's dim ambient illumination with enough local contrast to separate both gripping hands and the suspended wrist.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 구멍 가장자리로 몸을 날려 엎드린 수빈이 허공에 매달린 이현우의 손목을 두 손으로 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): At the collapsing floor edge near a descending escalator inside the ruined department store, lit by the explorers' flashlights.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Newly broken floor edge (Broken open above the lower floor) — The upper surface beneath 수빈 and the exposed edge above 이현우 are both visible; used as A narrow diagonal divider establishes the load-bearing edge and the difference in body levels; Opening beneath the hands (Open after the sudden floor collapse) — Seen downward from beside the rim; used as Negative space below the wrist conveys the unsupported drop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the department store's dim ambient illumination with enough local contrast to separate both gripping hands and the suspended wrist.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S66sh9__bgfirst_bg.png",
  "asset_id": "1bad1e80-714e-4266-89b0-17fd4934cc35",
  "input_asset_ids": [
   "2eb97da9-f2ec-49fc-b344-a9e2342f515e",
   "ee9186a1-023a-4aaa-a042-70e8ff736bb2"
  ]
 },
 "S66sh9": {
  "input_fingerprint": "94c5e687df1fdc08",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 구멍 가장자리로 몸을 날려 엎드린 수빈이 허공에 매달린 이현우의 손목을 두 손으로 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): At the collapsing floor edge near a descending escalator inside the ruined department store, lit by the explorers' flashlights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Newly broken floor edge (Broken open above the lower floor) — The upper surface beneath 수빈 and the exposed edge above 이현우 are both visible; used as A narrow diagonal divider establishes the load-bearing edge and the difference in body levels; Opening beneath the hands (Open after the sudden floor collapse) — Seen downward from beside the rim; used as Negative space below the wrist conveys the unsupported drop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the department store's dim ambient illumination with enough local contrast to separate both gripping hands and the suspended wrist.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A newly opened sinkhole breaks the unstable floor near the barely accessible descending escalator. The dust-cleared directory identifies the food sections from basement level one through the third floor. 이현우: He hangs over the new opening, unmasked and still bearing his earlier injuries. His flashlight remains in his possession, without a specified grip during the fall. 수빈: She braces at the opening with her arms extended, her face uncovered. She retains her flashlight and radiation meter, along with her existing facial and abdominal skin damage.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 구멍 가장자리로 몸을 날려 엎드린 수빈이 허공에 매달린 이현우의 손목을 두 손으로 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): At the collapsing floor edge near a descending escalator inside the ruined department store, lit by the explorers' flashlights. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Newly broken floor edge (Broken open above the lower floor) — The upper surface beneath 수빈 and the exposed edge above 이현우 are both visible; used as A narrow diagonal divider establishes the load-bearing edge and the difference in body levels; Opening beneath the hands (Open after the sudden floor collapse) — Seen downward from beside the rim; used as Negative space below the wrist conveys the unsupported drop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the department store's dim ambient illumination with enough local contrast to separate both gripping hands and the suspended wrist.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A newly opened sinkhole breaks the unstable floor near the barely accessible descending escalator. The dust-cleared directory identifies the food sections from basement level one through the third floor. 이현우: He hangs over the new opening, unmasked and still bearing his earlier injuries. His flashlight remains in his possession, without a specified grip during the fall. 수빈: She braces at the opening with her arms extended, her face uncovered. She retains her flashlight and radiation meter, along with her existing facial and abdominal skin damage.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 구멍 가장자리로 몸을 날려 엎드린 수빈이 허공에 매달린 이현우의 손목을 두 손으로 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): At the collapsing floor edge near a descending escalator inside the ruined department store, lit by the explorers' flashlights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Newly broken floor edge (Broken open above the lower floor) — The upper surface beneath 수빈 and the exposed edge above 이현우 are both visible; used as A narrow diagonal divider establishes the load-bearing edge and the difference in body levels; Opening beneath the hands (Open after the sudden floor collapse) — Seen downward from beside the rim; used as Negative space below the wrist conveys the unsupported drop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the department store's dim ambient illumination with enough local contrast to separate both gripping hands and the suspended wrist.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A newly opened sinkhole breaks the unstable floor near the barely accessible descending escalator. The dust-cleared directory identifies the food sections from basement level one through the third floor. 이현우: He hangs over the new opening, unmasked and still bearing his earlier injuries. His flashlight remains in his possession, without a specified grip during the fall. 수빈: She braces at the opening with her arms extended, her face uncovered. She retains her flashlight and radiation meter, along with her existing facial and abdominal skin damage.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 수빈 (한국인 여성 얼굴, 20세, 검은 머리카락, 짧고 날렵한 단발머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S66sh9__bgfirst_bg.png",
     "asset_id": "1bad1e80-714e-4266-89b0-17fd4934cc35",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S66sh9.png",
     "asset_id": "2eb97da9-f2ec-49fc-b344-a9e2342f515e",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904512>",
     "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L101B02.png",
     "asset_id": "ee9186a1-023a-4aaa-a042-70e8ff736bb2",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904512>",
     "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "수빈의 시선은 아래쪽에 있는 이현우를 향하고, 이현우는 위쪽 수빈을 바라봅니다. 이현우가 쥐고 있는 손전등의 빛은 화면 왼쪽을 비추고 있습니다.",
    "built_space": "무너진 백화점 내부로, 배경에 에스컬레이터가 확인되며 전경에는 지시된 대로 부서진 콘크리트 바닥 가장자리와 그 아래의 열린 공간이 잘 배치되어 있습니다.",
    "entities": "수빈은 단발머리, 가죽 재킷, 방사선 측정기 등 레퍼런스의 특징을 정확히 반영하고 있으며, 이현우 역시 헝클어진 머리와 어두운 후드티, 상처 난 얼굴을 잘 유지하고 있습니다.",
    "hard_violations": [],
    "physics": "수빈의 몸은 부서진 바닥 면에 엎드린 채 지탱되고 있으며, 이현우는 수빈이 붙잡은 두 손의 완력에 의지해 허공에 매달린 상태가 물리적으로 타당하게 묘사되었습니다."
   },
   {
    "label": "B",
    "direction": "수빈의 시선은 구멍 아래를 향하고 있으며, 우측 상단에 위치한 손전등이 아래쪽을 비추고 있습니다.",
    "built_space": "부서진 콘크리트 바닥 가장자리가 있는 폐허 공간이나, 배경의 디테일(에스컬레이터 등)은 명확하게 식별되지 않습니다.",
    "entities": "수빈이 지시된 가죽 재킷을 입지 않고 반소매 티셔츠 차림이며, 이현우는 팔 일부만 묘사되었습니다. 가장 큰 문제는 화면 우측 상단에 프롬프트에 없는 다른 인물의 다리와 손전등을 쥔 손이 등장한다는 점입니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 추가 인물(화면 우측 상단의 다리와 손) 등장",
     "[gpt-high] 우측 상단의 손전등을 든 손과 무릎은 엎드린 수빈이나 구멍 아래 이현우에게 연결될 수 없는 제3자의 신체로, 허용된 두 사람 외 인물을 추가했다."
    ],
    "physics": "수빈은 바닥에 엎드려 체중을 싣고 있으며, 아래쪽의 팔은 수빈의 양손에 의해 붙잡혀 허공에 지탱되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스와 일치하는 인물 외양과 복장을 유지하며, 두 인물의 긴박한 위치 관계와 그립을 물리적으로 자연스럽게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에 명시되지 않은 제3자가 등장하는 치명적인 오류가 있으며, 수빈의 복장(가죽 재킷 누락) 또한 레퍼런스와 일치하지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수빈의 시선은 아래쪽에 있는 이현우를 향하고, 이현우는 위쪽 수빈을 바라봅니다. 이현우가 쥐고 있는 손전등의 빛은 화면 왼쪽을 비추고 있습니다.",
        "built_space": "무너진 백화점 내부로, 배경에 에스컬레이터가 확인되며 전경에는 지시된 대로 부서진 콘크리트 바닥 가장자리와 그 아래의 열린 공간이 잘 배치되어 있습니다.",
        "entities": "수빈은 단발머리, 가죽 재킷, 방사선 측정기 등 레퍼런스의 특징을 정확히 반영하고 있으며, 이현우 역시 헝클어진 머리와 어두운 후드티, 상처 난 얼굴을 잘 유지하고 있습니다.",
        "hard_violations": [],
        "physics": "수빈의 몸은 부서진 바닥 면에 엎드린 채 지탱되고 있으며, 이현우는 수빈이 붙잡은 두 손의 완력에 의지해 허공에 매달린 상태가 물리적으로 타당하게 묘사되었습니다."
       },
       {
        "label": "B",
        "direction": "수빈의 시선은 구멍 아래를 향하고 있으며, 우측 상단에 위치한 손전등이 아래쪽을 비추고 있습니다.",
        "built_space": "부서진 콘크리트 바닥 가장자리가 있는 폐허 공간이나, 배경의 디테일(에스컬레이터 등)은 명확하게 식별되지 않습니다.",
        "entities": "수빈이 지시된 가죽 재킷을 입지 않고 반소매 티셔츠 차림이며, 이현우는 팔 일부만 묘사되었습니다. 가장 큰 문제는 화면 우측 상단에 프롬프트에 없는 다른 인물의 다리와 손전등을 쥔 손이 등장한다는 점입니다.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 추가 인물(화면 우측 상단의 다리와 손) 등장"
        ],
        "physics": "수빈은 바닥에 엎드려 체중을 싣고 있으며, 아래쪽의 팔은 수빈의 양손에 의해 붙잡혀 허공에 지탱되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스와 일치하는 인물 외양과 복장을 유지하며, 두 인물의 긴박한 위치 관계와 그립을 물리적으로 자연스럽게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에 명시되지 않은 제3자가 등장하는 치명적인 오류가 있으며, 수빈의 복장(가죽 재킷 누락) 또한 레퍼런스와 일치하지 않습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "수빈의 시선은 아래쪽에 있는 이현우를 향하고, 이현우는 위쪽 수빈을 바라봅니다. 이현우가 쥐고 있는 손전등의 빛은 화면 왼쪽을 비추고 있습니다.",
        "built_space": "무너진 백화점 내부로, 배경에 에스컬레이터가 확인되며 전경에는 지시된 대로 부서진 콘크리트 바닥 가장자리와 그 아래의 열린 공간이 잘 배치되어 있습니다.",
        "entities": "수빈은 단발머리, 가죽 재킷, 방사선 측정기 등 레퍼런스의 특징을 정확히 반영하고 있으며, 이현우 역시 헝클어진 머리와 어두운 후드티, 상처 난 얼굴을 잘 유지하고 있습니다.",
        "hard_violations": [],
        "physics": "수빈의 몸은 부서진 바닥 면에 엎드린 채 지탱되고 있으며, 이현우는 수빈이 붙잡은 두 손의 완력에 의지해 허공에 매달린 상태가 물리적으로 타당하게 묘사되었습니다."
       },
       {
        "label": "B",
        "direction": "수빈의 시선은 구멍 아래를 향하고 있으며, 우측 상단에 위치한 손전등이 아래쪽을 비추고 있습니다.",
        "built_space": "부서진 콘크리트 바닥 가장자리가 있는 폐허 공간이나, 배경의 디테일(에스컬레이터 등)은 명확하게 식별되지 않습니다.",
        "entities": "수빈이 지시된 가죽 재킷을 입지 않고 반소매 티셔츠 차림이며, 이현우는 팔 일부만 묘사되었습니다. 가장 큰 문제는 화면 우측 상단에 프롬프트에 없는 다른 인물의 다리와 손전등을 쥔 손이 등장한다는 점입니다.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 추가 인물(화면 우측 상단의 다리와 손) 등장"
        ],
        "physics": "수빈은 바닥에 엎드려 체중을 싣고 있으며, 아래쪽의 팔은 수빈의 양손에 의해 붙잡혀 허공에 지탱되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "두 손으로 손목을 붙들고 상층 바닥에 엎드린 동작은 분명하지만, 우측 상단에 허용되지 않은 제3자의 신체가 등장하고 수빈의 의상도 참조와 다르다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "수빈의 바닥 지지와 이현우의 매달림, 대각선 파단면과 손 아래 추락 공간을 충실히 구현했지만, 손목만 빈틈없이 조이는 대신 손과 손목을 함께 감싼 악력이 표현된다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수빈은 아래로 뻗은 자신의 두 손과 붙잡힌 이현우의 손목 쪽을 본다. 이현우의 팔은 구멍 아래에서 위로 뻗어 수빈에게 연결되며 얼굴과 시선은 화면 밖이다. 우측 상단의 다른 사람이 든 손전등은 왼쪽 아래의 구멍 쪽을 향한다.",
        "built_space": "수빈이 엎드린 상층 바닥, 그 아래 노출된 콘크리트 파단면, 하층 기둥과 잔해가 보인다. 구멍 하나의 양쪽 가장자리가 넓게 드러나며 우측 바닥이 전경을 크게 차지한다. 노출 철근들이 구멍을 가로지르고 에스컬레이터는 보이지 않는다. 상하층 관계는 성립하지만, 요구한 좁은 대각선 경계보다 공간 전체를 넓게 설명하는 구도다.",
        "entities": "수빈은 검은 단발의 젊은 동아시아계 여성으로 보이지만 참조의 가죽 재킷과 목도리 대신 회색 반소매 상의를 입었다. 얼굴에는 오염과 작은 상처가 보인다. 이현우는 검은 소매와 젊은 남성의 팔만 보여 얼굴 정체성은 판별할 수 없다. 그의 얼굴과 휴대품, 수빈의 복부 및 장비는 가려져 확인할 수 없다. 우측 상단에는 손전등을 잡은 손과 검은 바지를 입은 무릎이 추가로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "우측 상단의 손전등을 든 손과 무릎은 엎드린 수빈이나 구멍 아래 이현우에게 연결될 수 없는 제3자의 신체로, 허용된 두 사람 외 인물을 추가했다."
        ],
        "physics": "수빈의 몸통과 하체는 상층 바닥에 닿아 지지되고, 두 팔은 가장자리 밖으로 뻗어 이현우의 손목을 감싼다. 이현우의 보이는 팔에는 수빈의 두 손이라는 명확한 지지가 있으므로 근거 없이 떠 있는 자세는 아니다. 추가 인물의 무릎은 우측 바닥에 놓이고 손전등은 손에 잡혀 있다. 철근은 파단면 쪽으로 이어진다."
       },
       {
        "label": "B",
        "direction": "수빈은 아래쪽 이현우와 맞잡은 손 방향을 응시하고, 이현우는 위쪽 수빈과 손의 연결부를 바라본다. 이현우의 한 팔은 위로 뻗어 수빈에게 연결되며 다른 손의 손전등은 왼쪽 위 파단면을 향한다. 파단면에 생긴 밝은 빛도 이 방향과 일치한다.",
        "built_space": "하나의 붕괴 구멍을 사이에 두고 수빈이 놓인 상층 바닥과 이현우 뒤쪽의 하층 기둥·잔해가 구분된다. 노출된 바닥 단면은 왼쪽 아래에서 오른쪽 위로 이어져 두 사람의 높이를 나눈다. 배경 중앙에는 에스컬레이터 한 대가 작게 보이고, 콘크리트 기둥과 어두운 점포 개구부가 참조 장소의 재료와 구성을 따른다. 손 아래의 빈 공간도 보인다. 다만 파단면과 배경의 비중이 커서 손 중심의 타격감 있는 클로즈업보다는 조금 넓다.",
        "entities": "두 사람만 등장한다. 수빈은 검은 짧은 머리의 젊은 동아시아계 여성으로, 참조의 어두운 가죽 재킷·후드·목도리를 유지한다. 옆구리에는 끈으로 부착된 방사선 계측기로 읽히는 장비가 있다. 이현우는 헝클어진 검은 머리와 젊은 동아시아계 남성 얼굴, 어두운 후드 상의로 참조와 대체로 맞는다. 두 얼굴 모두 가려지지 않고 상처가 보인다. 이현우는 손전등을 쥐고 있으며 수빈의 손전등과 복부 피부는 화면에서 확인되지 않는다. 읽을 수 있는 문구는 없다.",
        "hard_violations": [],
        "physics": "수빈의 몸통과 팔꿈치 부근은 상층 바닥에 닿아 하중을 지탱한다. 이현우의 몸은 아래로 처지고 들어 올린 팔이 수빈의 양손에 연결되어 매달림의 지지 경로가 명확하다. 다만 수빈의 손은 이현우의 손목만 두 겹으로 조이기보다 손과 손목을 함께 감싸며 이현우도 손가락을 걸고 있다. 손전등은 이현우의 다른 손에 잡혀 있고 계측기는 수빈의 몸에 부착되어 있어 무지지 부유물은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "두 손으로 손목을 붙들고 상층 바닥에 엎드린 동작은 분명하지만, 우측 상단에 허용되지 않은 제3자의 신체가 등장하고 수빈의 의상도 참조와 다르다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "수빈의 바닥 지지와 이현우의 매달림, 대각선 파단면과 손 아래 추락 공간을 충실히 구현했지만, 손목만 빈틈없이 조이는 대신 손과 손목을 함께 감싼 악력이 표현된다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "수빈은 아래로 뻗은 자신의 두 손과 붙잡힌 이현우의 손목 쪽을 본다. 이현우의 팔은 구멍 아래에서 위로 뻗어 수빈에게 연결되며 얼굴과 시선은 화면 밖이다. 우측 상단의 다른 사람이 든 손전등은 왼쪽 아래의 구멍 쪽을 향한다.",
        "built_space": "수빈이 엎드린 상층 바닥, 그 아래 노출된 콘크리트 파단면, 하층 기둥과 잔해가 보인다. 구멍 하나의 양쪽 가장자리가 넓게 드러나며 우측 바닥이 전경을 크게 차지한다. 노출 철근들이 구멍을 가로지르고 에스컬레이터는 보이지 않는다. 상하층 관계는 성립하지만, 요구한 좁은 대각선 경계보다 공간 전체를 넓게 설명하는 구도다.",
        "entities": "수빈은 검은 단발의 젊은 동아시아계 여성으로 보이지만 참조의 가죽 재킷과 목도리 대신 회색 반소매 상의를 입었다. 얼굴에는 오염과 작은 상처가 보인다. 이현우는 검은 소매와 젊은 남성의 팔만 보여 얼굴 정체성은 판별할 수 없다. 그의 얼굴과 휴대품, 수빈의 복부 및 장비는 가려져 확인할 수 없다. 우측 상단에는 손전등을 잡은 손과 검은 바지를 입은 무릎이 추가로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "우측 상단의 손전등을 든 손과 무릎은 엎드린 수빈이나 구멍 아래 이현우에게 연결될 수 없는 제3자의 신체로, 허용된 두 사람 외 인물을 추가했다."
        ],
        "physics": "수빈의 몸통과 하체는 상층 바닥에 닿아 지지되고, 두 팔은 가장자리 밖으로 뻗어 이현우의 손목을 감싼다. 이현우의 보이는 팔에는 수빈의 두 손이라는 명확한 지지가 있으므로 근거 없이 떠 있는 자세는 아니다. 추가 인물의 무릎은 우측 바닥에 놓이고 손전등은 손에 잡혀 있다. 철근은 파단면 쪽으로 이어진다."
       },
       {
        "label": "A",
        "direction": "수빈은 아래쪽 이현우와 맞잡은 손 방향을 응시하고, 이현우는 위쪽 수빈과 손의 연결부를 바라본다. 이현우의 한 팔은 위로 뻗어 수빈에게 연결되며 다른 손의 손전등은 왼쪽 위 파단면을 향한다. 파단면에 생긴 밝은 빛도 이 방향과 일치한다.",
        "built_space": "하나의 붕괴 구멍을 사이에 두고 수빈이 놓인 상층 바닥과 이현우 뒤쪽의 하층 기둥·잔해가 구분된다. 노출된 바닥 단면은 왼쪽 아래에서 오른쪽 위로 이어져 두 사람의 높이를 나눈다. 배경 중앙에는 에스컬레이터 한 대가 작게 보이고, 콘크리트 기둥과 어두운 점포 개구부가 참조 장소의 재료와 구성을 따른다. 손 아래의 빈 공간도 보인다. 다만 파단면과 배경의 비중이 커서 손 중심의 타격감 있는 클로즈업보다는 조금 넓다.",
        "entities": "두 사람만 등장한다. 수빈은 검은 짧은 머리의 젊은 동아시아계 여성으로, 참조의 어두운 가죽 재킷·후드·목도리를 유지한다. 옆구리에는 끈으로 부착된 방사선 계측기로 읽히는 장비가 있다. 이현우는 헝클어진 검은 머리와 젊은 동아시아계 남성 얼굴, 어두운 후드 상의로 참조와 대체로 맞는다. 두 얼굴 모두 가려지지 않고 상처가 보인다. 이현우는 손전등을 쥐고 있으며 수빈의 손전등과 복부 피부는 화면에서 확인되지 않는다. 읽을 수 있는 문구는 없다.",
        "hard_violations": [],
        "physics": "수빈의 몸통과 팔꿈치 부근은 상층 바닥에 닿아 하중을 지탱한다. 이현우의 몸은 아래로 처지고 들어 올린 팔이 수빈의 양손에 연결되어 매달림의 지지 경로가 명확하다. 다만 수빈의 손은 이현우의 손목만 두 겹으로 조이기보다 손과 손목을 함께 감싸며 이현우도 손가락을 걸고 있다. 손전등은 이현우의 다른 손에 잡혀 있고 계측기는 수빈의 몸에 부착되어 있어 무지지 부유물은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.679
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.429
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트에 명시되지 않은 추가 인물(화면 우측 상단의 다리와 손) 등장",
     "[gpt-high] 우측 상단의 손전등을 든 손과 무릎은 엎드린 수빈이나 구멍 아래 이현우에게 연결될 수 없는 제3자의 신체로, 허용된 두 사람 외 인물을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스와 일치하는 인물 외양과 복장을 유지하며, 두 인물의 긴박한 위치 관계와 그립을 물리적으로 자연스럽게 구현했습니다."
   },
   {
    "label": "B",
    "score": 429,
    "verdict_ko": "프롬프트에 명시되지 않은 제3자가 등장하는 치명적인 오류가 있으며, 수빈의 복장(가죽 재킷 누락) 또한 레퍼런스와 일치하지 않습니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 추가 인물(화면 우측 상단의 다리와 손) 등장 / [gpt-high] 우측 상단의 손전등을 든 손과 무릎은 엎드린 수빈이나 구멍 아래 이현우에게 연결될 수 없는 제3자의 신체로, 허용된 두 사람 외 인물을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L101B02.png",
    "asset_id": "ee9186a1-023a-4aaa-a042-70e8ff736bb2",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:904512>",
    "asset_id": "3fcfbef1-5495-46ee-b675-b82808ddfcda",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aace01-083d-70d0-b903-1708b597f3b8",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S66sh9__bgfirst_bg.png",
   "bg_asset_id": "1bad1e80-714e-4266-89b0-17fd4934cc35",
   "bg_record_key": "S66sh9::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S66sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:54:08.605627+00:00",
  "fingerprint": "72d12c80c6f3cf905c00d2b7d2debe94a47352d1bfb8ea569fb80a20832948a1",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S66sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S66sh9_sel.png",
  "source_sha256": "042a685d45826a110deb0485c1569d7392b6ac3acdf6a8b679f8d5bb3ac06014",
  "file": "S66sh9_cine.png",
  "staged_sha256": "4a2e12a9a5cbf081187b17698af534f46cf7b2f7aceeadeb09b644e394e9f94b",
  "latency_ms": 10517
 },
 "S66sh31::signage": {
  "fp": "4b34d73683f8a794",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S66sh31": {
  "input_fingerprint": "2279806660a0db22",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아득히 먼 천장 구멍 틈새에서 굵고 튼튼한 밧줄 하나가 바닥을 향해 맹렬히 떨어져 내리고 있는 역동적인 구도의 한 찰나.\n\nLOCATION (lock): Inside the department store's buried lower-level food section, beneath the distant collapse opening, with flashlights and faint light from above. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant ceiling opening in the upper-center of the frame, background; Descending rope with falling free end in the middle-center of the frame, midground, moves toward Lower-right frame toward the lower floor.\n- KEY BACKGROUND ELEMENTS: Ceiling opening (Open where the pair previously fell) — Its underside and irregular rim are visible at an oblique upward angle; used as Small upper-frame origin for the descending rope, establishing the remote upper floor; Descending rope (Falling from the ceiling opening toward the lower floor) — Extends from upper center toward lower right, with its free end still descending; used as A narrow dynamic line through the open space, occupying only a small fraction of the image; Department-store interior walls (Partly collapsed) — Receding interior faces border the upward view; used as Peripheral scale references prevent the rope from reading as an isolated close-up.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the dim interior and use restrained tonal separation to keep the descending rope readable against the distant ceiling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The collapsed floor leaves a high opening above the blocked lower food section, where dusty food stocks, opened cans and an unfolded parasol remain. A rope now descends through that opening, and rain drips into the space.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아득히 먼 천장 구멍 틈새에서 굵고 튼튼한 밧줄 하나가 바닥을 향해 맹렬히 떨어져 내리고 있는 역동적인 구도의 한 찰나.\n\nLOCATION (lock): Inside the department store's buried lower-level food section, beneath the distant collapse opening, with flashlights and faint light from above. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant ceiling opening in the upper-center of the frame, background; Descending rope with falling free end in the middle-center of the frame, midground, moves toward Lower-right frame toward the lower floor.\n- KEY BACKGROUND ELEMENTS: Ceiling opening (Open where the pair previously fell) — Its underside and irregular rim are visible at an oblique upward angle; used as Small upper-frame origin for the descending rope, establishing the remote upper floor; Descending rope (Falling from the ceiling opening toward the lower floor) — Extends from upper center toward lower right, with its free end still descending; used as A narrow dynamic line through the open space, occupying only a small fraction of the image; Department-store interior walls (Partly collapsed) — Receding interior faces border the upward view; used as Peripheral scale references prevent the rope from reading as an isolated close-up.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the dim interior and use restrained tonal separation to keep the descending rope readable against the distant ceiling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The collapsed floor leaves a high opening above the blocked lower food section, where dusty food stocks, opened cans and an unfolded parasol remain. A rope now descends through that opening, and rain drips into the space.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아득히 먼 천장 구멍 틈새에서 굵고 튼튼한 밧줄 하나가 바닥을 향해 맹렬히 떨어져 내리고 있는 역동적인 구도의 한 찰나.\n\nLOCATION (lock): Inside the department store's buried lower-level food section, beneath the distant collapse opening, with flashlights and faint light from above. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant ceiling opening in the upper-center of the frame, background; Descending rope with falling free end in the middle-center of the frame, midground, moves toward Lower-right frame toward the lower floor.\n- KEY BACKGROUND ELEMENTS: Ceiling opening (Open where the pair previously fell) — Its underside and irregular rim are visible at an oblique upward angle; used as Small upper-frame origin for the descending rope, establishing the remote upper floor; Descending rope (Falling from the ceiling opening toward the lower floor) — Extends from upper center toward lower right, with its free end still descending; used as A narrow dynamic line through the open space, occupying only a small fraction of the image; Department-store interior walls (Partly collapsed) — Receding interior faces border the upward view; used as Peripheral scale references prevent the rope from reading as an isolated close-up.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the dim interior and use restrained tonal separation to keep the descending rope readable against the distant ceiling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The collapsed floor leaves a high opening above the blocked lower food section, where dusty food stocks, opened cans and an unfolded parasol remain. A rope now descends through that opening, and rain drips into the space.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "밧줄이 천장 구멍에서 시작해 우측 하단으로 향하고 있음.",
    "built_space": "상단 중앙에 무너진 구멍이 위치하며, 주변에 기둥과 식품 매대 구조물이 배치됨.",
    "entities": "굵은 밧줄, 펼쳐진 파라솔, 흩어진 통조림, 켜진 손전등이 모두 존재함. 사람은 없음.",
    "hard_violations": [
     "[gpt-high] 오른쪽 전경 통조림 라벨에 읽을 수 있는 한글이 있어, 이미지 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반합니다."
    ],
    "physics": "밧줄의 끝이 바닥에 닿기 전 떨어지는 상태로 허공에 떠 있고, 다른 사물들은 바닥이나 잔해 위에 지지되어 있음."
   },
   {
    "label": "B",
    "direction": "밧줄이 상단에서 내려와 전경을 향해 복잡하게 얽히고 있음.",
    "built_space": "무너진 천장이 상단 중앙에 국한되지 않고 화면 상단 전체를 가로지르며 과도하게 넓게 잡힘.",
    "entities": "밧줄, 파라솔, 흩어진 캔들이 확인됨. 사람은 없음.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 밧줄 형태 (지지대나 당기는 힘 없이 전경에서 둥글게 꼬인 채 허공에 떠 있음)"
    ],
    "physics": "밧줄이 자유낙하하는 끝부분을 보여주지 않고, 화면 앞에서 물리적으로 설명할 수 없는 꼬임 상태로 공중에 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 상단 중앙의 천장 구멍과 우측 하단으로 떨어지는 밧줄의 역동적인 구도를 정확히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지정된 프레이밍을 벗어났으며, 전경의 밧줄이 물리법칙을 무시하고 허공에 떠 있어 치명적인 오류임."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "밧줄이 천장 구멍에서 시작해 우측 하단으로 향하고 있음.",
        "built_space": "상단 중앙에 무너진 구멍이 위치하며, 주변에 기둥과 식품 매대 구조물이 배치됨.",
        "entities": "굵은 밧줄, 펼쳐진 파라솔, 흩어진 통조림, 켜진 손전등이 모두 존재함. 사람은 없음.",
        "hard_violations": [],
        "physics": "밧줄의 끝이 바닥에 닿기 전 떨어지는 상태로 허공에 떠 있고, 다른 사물들은 바닥이나 잔해 위에 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "밧줄이 상단에서 내려와 전경을 향해 복잡하게 얽히고 있음.",
        "built_space": "무너진 천장이 상단 중앙에 국한되지 않고 화면 상단 전체를 가로지르며 과도하게 넓게 잡힘.",
        "entities": "밧줄, 파라솔, 흩어진 캔들이 확인됨. 사람은 없음.",
        "hard_violations": [
         "물리적으로 불가능한 밧줄 형태 (지지대나 당기는 힘 없이 전경에서 둥글게 꼬인 채 허공에 떠 있음)"
        ],
        "physics": "밧줄이 자유낙하하는 끝부분을 보여주지 않고, 화면 앞에서 물리적으로 설명할 수 없는 꼬임 상태로 공중에 떠 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 상단 중앙의 천장 구멍과 우측 하단으로 떨어지는 밧줄의 역동적인 구도를 정확히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지정된 프레이밍을 벗어났으며, 전경의 밧줄이 물리법칙을 무시하고 허공에 떠 있어 치명적인 오류임."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "밧줄이 천장 구멍에서 시작해 우측 하단으로 향하고 있음.",
        "built_space": "상단 중앙에 무너진 구멍이 위치하며, 주변에 기둥과 식품 매대 구조물이 배치됨.",
        "entities": "굵은 밧줄, 펼쳐진 파라솔, 흩어진 통조림, 켜진 손전등이 모두 존재함. 사람은 없음.",
        "hard_violations": [],
        "physics": "밧줄의 끝이 바닥에 닿기 전 떨어지는 상태로 허공에 떠 있고, 다른 사물들은 바닥이나 잔해 위에 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "밧줄이 상단에서 내려와 전경을 향해 복잡하게 얽히고 있음.",
        "built_space": "무너진 천장이 상단 중앙에 국한되지 않고 화면 상단 전체를 가로지르며 과도하게 넓게 잡힘.",
        "entities": "밧줄, 파라솔, 흩어진 캔들이 확인됨. 사람은 없음.",
        "hard_violations": [
         "물리적으로 불가능한 밧줄 형태 (지지대나 당기는 힘 없이 전경에서 둥글게 꼬인 채 허공에 떠 있음)"
        ],
        "physics": "밧줄이 자유낙하하는 끝부분을 보여주지 않고, 화면 앞에서 물리적으로 설명할 수 없는 꼬임 상태로 공중에 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "인물과 판독 가능한 글자 없이 장소의 붕괴 상태는 비교적 충실하지만, 하향 구도와 크게 부각된 밧줄 뭉치가 먼 천장 틈에서 가느다랗게 내려오는 구도에 어긋납니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "위를 올려다보는 시점과 우하향하는 자유단은 더 정확하지만, 오른쪽 통조림의 읽을 수 있는 한글이 금지 조건을 위반하며 구멍과 밧줄도 지나치게 큽니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "밧줄은 상단 중앙의 무너진 바닥 가장자리에서 중앙을 지나 오른쪽 아래로 이어집니다. 우하단의 둥근 밧줄 뭉치는 하부 바닥 쪽으로 내려오는 경로에 있지만, 실제 자유단은 화면 아래로 잘려 보이지 않습니다. 사람의 시선이나 무기는 없습니다.",
        "built_space": "상단에 넓게 열린 붕괴부 하나와 그 너머 에스컬레이터 한 대가 보입니다. 전경 왼쪽과 오른쪽의 큰 콘크리트 기둥, 뒤쪽의 여러 기둥, 노출된 보와 철근은 참조의 구조적 특징을 이어갑니다. 다만 카메라는 하부 바닥을 많이 내려다보며, 작은 먼 천장 구멍의 밑면을 비스듬히 올려다보는 구도가 아닙니다. 펼친 파라솔 하나는 왼쪽 아래에 있습니다.",
        "entities": "사람과 얼굴은 없습니다. 굵은 섬유 밧줄 하나, 펼친 줄무늬 파라솔 하나, 바닥의 식품 용기와 열린 통조림, 낙수와 고인 물이 보입니다. 파손된 콘크리트와 어두운 하부 공간은 참조와 유사합니다. 판독 가능한 글자는 보이지 않습니다. 밧줄 뭉치가 전경에서 커져, 화면의 작은 비율만 차지해야 한다는 조건은 충족하지 못합니다.",
        "hard_violations": [],
        "physics": "밧줄은 상단 붕괴 가장자리까지 연속해서 이어지고, 아래쪽 고리는 던져 내린 밧줄이 풀리며 낙하하는 순간으로 가능한 형태입니다. 공중에 있다는 이유만으로 무지지 정지 물체라고 볼 근거는 없습니다. 다만 자유단이 잘려 낙하 순간의 확인은 약합니다. 파라솔은 기울어진 자루가 잔해 쪽에 닿으며, 통조림과 상자는 바닥이나 잔해 위에 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "밧줄은 상단 중앙에서 내려와 중앙에서 굽은 뒤 오른쪽 아래 바닥을 향합니다. 풀어진 자유단이 화면 안에 보여 요청한 이동 방향과 낙하 대상을 명확하게 드러냅니다. 사람의 시선이나 무기는 없습니다.",
        "built_space": "상단 중앙에 불규칙한 붕괴 구멍 하나가 있고, 아래에서 올려다본 슬래브 밑면과 철근이 보입니다. 좌우 콘크리트 기둥과 보가 공간을 둘러싸며, 오른쪽에는 긴 진열대, 뒤쪽에는 기울어진 진열대와 세워진 진열대가 있습니다. 왼쪽 바닥에는 펼친 파라솔 하나가 있습니다. 올려다보는 관계는 맞지만 구멍이 상부 화면을 크게 차지해 아득히 먼 작은 기점으로 읽히지 않습니다. 참조보다 식품 진열대가 훨씬 두드러집니다.",
        "entities": "사람과 얼굴은 없습니다. 굵은 밧줄 하나와 풀어진 끝, 펼친 파라솔 하나, 식품이 남은 진열대, 바닥의 통조림, 손전등과 낙수가 보입니다. 낮의 외광과 어두운 내부도 구현되어 있습니다. 그러나 오른쪽 전경 통조림 라벨에 판독 가능한 한글 상품명이 남아 있습니다. 밧줄 역시 중경의 좁은 선보다는 크게 확대된 전경 물체로 보입니다.",
        "hard_violations": [
         "오른쪽 전경 통조림 라벨에 읽을 수 있는 한글이 있어, 이미지 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반합니다."
        ],
        "physics": "밧줄 윗부분은 천장 구멍 너머로 이어지고, 굽은 몸통과 흐려진 자유단은 위에서 풀어 내린 밧줄의 낙하로 가능한 모습입니다. 자유단 아래에는 하부 바닥이 있어 낙하 경로도 성립합니다. 식품과 통조림은 선반이나 바닥에 지지되어 있고, 파라솔은 아래쪽이 바닥 잔해에 걸쳐 있습니다. 빗방울은 위쪽 개구부에서 아래로 떨어집니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물과 판독 가능한 글자 없이 장소의 붕괴 상태는 비교적 충실하지만, 하향 구도와 크게 부각된 밧줄 뭉치가 먼 천장 틈에서 가느다랗게 내려오는 구도에 어긋납니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "위를 올려다보는 시점과 우하향하는 자유단은 더 정확하지만, 오른쪽 통조림의 읽을 수 있는 한글이 금지 조건을 위반하며 구멍과 밧줄도 지나치게 큽니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "밧줄은 상단 중앙의 무너진 바닥 가장자리에서 중앙을 지나 오른쪽 아래로 이어집니다. 우하단의 둥근 밧줄 뭉치는 하부 바닥 쪽으로 내려오는 경로에 있지만, 실제 자유단은 화면 아래로 잘려 보이지 않습니다. 사람의 시선이나 무기는 없습니다.",
        "built_space": "상단에 넓게 열린 붕괴부 하나와 그 너머 에스컬레이터 한 대가 보입니다. 전경 왼쪽과 오른쪽의 큰 콘크리트 기둥, 뒤쪽의 여러 기둥, 노출된 보와 철근은 참조의 구조적 특징을 이어갑니다. 다만 카메라는 하부 바닥을 많이 내려다보며, 작은 먼 천장 구멍의 밑면을 비스듬히 올려다보는 구도가 아닙니다. 펼친 파라솔 하나는 왼쪽 아래에 있습니다.",
        "entities": "사람과 얼굴은 없습니다. 굵은 섬유 밧줄 하나, 펼친 줄무늬 파라솔 하나, 바닥의 식품 용기와 열린 통조림, 낙수와 고인 물이 보입니다. 파손된 콘크리트와 어두운 하부 공간은 참조와 유사합니다. 판독 가능한 글자는 보이지 않습니다. 밧줄 뭉치가 전경에서 커져, 화면의 작은 비율만 차지해야 한다는 조건은 충족하지 못합니다.",
        "hard_violations": [],
        "physics": "밧줄은 상단 붕괴 가장자리까지 연속해서 이어지고, 아래쪽 고리는 던져 내린 밧줄이 풀리며 낙하하는 순간으로 가능한 형태입니다. 공중에 있다는 이유만으로 무지지 정지 물체라고 볼 근거는 없습니다. 다만 자유단이 잘려 낙하 순간의 확인은 약합니다. 파라솔은 기울어진 자루가 잔해 쪽에 닿으며, 통조림과 상자는 바닥이나 잔해 위에 놓여 있습니다."
       },
       {
        "label": "A",
        "direction": "밧줄은 상단 중앙에서 내려와 중앙에서 굽은 뒤 오른쪽 아래 바닥을 향합니다. 풀어진 자유단이 화면 안에 보여 요청한 이동 방향과 낙하 대상을 명확하게 드러냅니다. 사람의 시선이나 무기는 없습니다.",
        "built_space": "상단 중앙에 불규칙한 붕괴 구멍 하나가 있고, 아래에서 올려다본 슬래브 밑면과 철근이 보입니다. 좌우 콘크리트 기둥과 보가 공간을 둘러싸며, 오른쪽에는 긴 진열대, 뒤쪽에는 기울어진 진열대와 세워진 진열대가 있습니다. 왼쪽 바닥에는 펼친 파라솔 하나가 있습니다. 올려다보는 관계는 맞지만 구멍이 상부 화면을 크게 차지해 아득히 먼 작은 기점으로 읽히지 않습니다. 참조보다 식품 진열대가 훨씬 두드러집니다.",
        "entities": "사람과 얼굴은 없습니다. 굵은 밧줄 하나와 풀어진 끝, 펼친 파라솔 하나, 식품이 남은 진열대, 바닥의 통조림, 손전등과 낙수가 보입니다. 낮의 외광과 어두운 내부도 구현되어 있습니다. 그러나 오른쪽 전경 통조림 라벨에 판독 가능한 한글 상품명이 남아 있습니다. 밧줄 역시 중경의 좁은 선보다는 크게 확대된 전경 물체로 보입니다.",
        "hard_violations": [
         "오른쪽 전경 통조림 라벨에 읽을 수 있는 한글이 있어, 이미지 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반합니다."
        ],
        "physics": "밧줄 윗부분은 천장 구멍 너머로 이어지고, 굽은 몸통과 흐려진 자유단은 위에서 풀어 내린 밧줄의 낙하로 가능한 모습입니다. 자유단 아래에는 하부 바닥이 있어 낙하 경로도 성립합니다. 식품과 통조림은 선반이나 바닥에 지지되어 있고, 파라솔은 아래쪽이 바닥 잔해에 걸쳐 있습니다. 빗방울은 위쪽 개구부에서 아래로 떨어집니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.35,
    "B": 1.179
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 밧줄 형태 (지지대나 당기는 힘 없이 전경에서 둥글게 꼬인 채 허공에 떠 있음)"
    ],
    "A": [
     "[gpt-high] 오른쪽 전경 통조림 라벨에 읽을 수 있는 한글이 있어, 이미지 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1350,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "프롬프트가 요구한 상단 중앙의 천장 구멍과 우측 하단으로 떨어지는 밧줄의 역동적인 구도를 정확히 구현함.  ★위반: [gpt-high] 오른쪽 전경 통조림 라벨에 읽을 수 있는 한글이 있어, 이미지 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반합니다."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "지정된 프레이밍을 벗어났으며, 전경의 밧줄이 물리법칙을 무시하고 허공에 떠 있어 치명적인 오류임.  ★위반: [gemini-pro] 물리적으로 불가능한 밧줄 형태 (지지대나 당기는 힘 없이 전경에서 둥글게 꼬인 채 허공에 떠 있음)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L101B01.png",
    "asset_id": "ccec8321-7807-4baa-b0bf-ef4f2dba7377",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aace01-b752-7c27-b7d8-2fc9ec4cdccd",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S66sh31::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:54:19.470122+00:00",
  "fingerprint": "ae46d82a70da27d5039a425c64301d270a5981c1c0e744dbd39737da69ea29e1",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S66sh31_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S66sh31_sel.png",
  "source_sha256": "189b1d878470e78c0bb75ef476b8ddea83150dc131022360b769f2b3c2838a41",
  "file": "S66sh31_cine.png",
  "staged_sha256": "ba5a04f8ca80eb013e5d7e3e99b0d092b6e906ef9cb31bfea3f202b732d9f30a",
  "latency_ms": 8957
 },
 "S66sh40::signage": {
  "fp": "94f9b2e8001e7d8d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S66sh40": {
  "input_fingerprint": "3c6f90aa783f6cc7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 짧은 머리의 정한수가 무너진 잔해 위에서 이현우를 향해 날카로운 눈빛을 쏘아보는 자세로 우뚝 선 위압적인 전신.\n\nLOCATION (lock): On the upper-floor rubble beside the collapse opening inside the ruined department store, where light filters through structural breaches. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Collapsed remains beneath 정한수 (Collapsed) — An oblique side and supporting upper surfaces are visible beneath his feet; used as Establishes his elevated footing while preserving natural figure scale; Upper-floor opening (Open) — A partial rim remains beside 이현우 at the lower-left edge; used as Connects the completed climb to the confrontation; Rescue rope (Held by 이현우 after his ascent) — A short visible section descends beside his cropped foreground body; used as Small continuity detail rather than a competing focal object.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim interior illumination and controlled contrast keep 정한수's short hair, eyes, and full-body outline readable without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rope remains hanging through the collapsed floor opening into the food section, whose other exits remain blocked. The unfolded parasol and opened food cans remain below. 정한수: He stands in the upper ruins with short hair and a sharp gaze.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 정한수 (한국인 남성 얼굴, 성인, 짧은 머리) — wearing: 수납용 주머니가 여러 개 달린 카고 바지와 튼튼하고 낡은 야상 점퍼. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 짧은 머리의 정한수가 무너진 잔해 위에서 이현우를 향해 날카로운 눈빛을 쏘아보는 자세로 우뚝 선 위압적인 전신.\n\nLOCATION (lock): On the upper-floor rubble beside the collapse opening inside the ruined department store, where light filters through structural breaches. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Collapsed remains beneath 정한수 (Collapsed) — An oblique side and supporting upper surfaces are visible beneath his feet; used as Establishes his elevated footing while preserving natural figure scale; Upper-floor opening (Open) — A partial rim remains beside 이현우 at the lower-left edge; used as Connects the completed climb to the confrontation; Rescue rope (Held by 이현우 after his ascent) — A short visible section descends beside his cropped foreground body; used as Small continuity detail rather than a competing focal object.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim interior illumination and controlled contrast keep 정한수's short hair, eyes, and full-body outline readable without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rope remains hanging through the collapsed floor opening into the food section, whose other exits remain blocked. The unfolded parasol and opened food cans remain below. 정한수: He stands in the upper ruins with short hair and a sharp gaze.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 정한수 (한국인 남성 얼굴, 성인, 짧은 머리) — wearing: 수납용 주머니가 여러 개 달린 카고 바지와 튼튼하고 낡은 야상 점퍼. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 짧은 머리의 정한수가 무너진 잔해 위에서 이현우를 향해 날카로운 눈빛을 쏘아보는 자세로 우뚝 선 위압적인 전신.\n\nLOCATION (lock): On the upper-floor rubble beside the collapse opening inside the ruined department store, where light filters through structural breaches. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Collapsed remains beneath 정한수 (Collapsed) — An oblique side and supporting upper surfaces are visible beneath his feet; used as Establishes his elevated footing while preserving natural figure scale; Upper-floor opening (Open) — A partial rim remains beside 이현우 at the lower-left edge; used as Connects the completed climb to the confrontation; Rescue rope (Held by 이현우 after his ascent) — A short visible section descends beside his cropped foreground body; used as Small continuity detail rather than a competing focal object.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim interior illumination and controlled contrast keep 정한수's short hair, eyes, and full-body outline readable without inventing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rope remains hanging through the collapsed floor opening into the food section, whose other exits remain blocked. The unfolded parasol and opened food cans remain below. 정한수: He stands in the upper ruins with short hair and a sharp gaze.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 정한수 (한국인 남성 얼굴, 성인, 짧은 머리) — wearing: 수납용 주머니가 여러 개 달린 카고 바지와 튼튼하고 낡은 야상 점퍼. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "정한수의 시선이 이현우를 직접 향하지 않고 약간 빗겨 허공을 향해 있으며, 이현우는 정한수를 올려다봄.",
    "built_space": "기둥과 유리 진열장이 남은 폐허 내부. 평평한 바닥이 붕괴되어 형성된 사각형 형태의 구멍 가장자리에 인물들이 배치됨.",
    "entities": "정한수는 짧은 머리, 야상 점퍼, 카고 바지 차림으로 지시에 부합함. 이현우는 왼쪽 하단에 크롭되어 나타남.",
    "hard_violations": [],
    "physics": "정한수는 평평한 바닥 끝에 안정적으로 서 있고, 이현우는 한 손으로 밧줄을 쥐고 있음. 중력을 거스르거나 지지대가 없는 물체는 없음."
   },
   {
    "label": "B",
    "direction": "정한수의 날카로운 시선이 이현우를 향해 정확히 고정되어 있으며, 이현우 역시 위를 주시함.",
    "built_space": "천장 채광창이 뚫려 빛이 들어오는 붕괴된 상업 시설. 거대한 경사진 콘크리트 판이 층과 층 사이를 잇고 있음.",
    "entities": "두 인물 모두 지시된 복장과 신체적 특징, 프레이밍 요구사항을 훌륭하게 충족함.",
    "hard_violations": [
     "[gemini-pro] 화면 하단 상점 구조물에 'FOOD'라는 텍스트가 명확하게 읽히는 상태로 렌더링되어 'No readable writing' 제약을 위반함.",
     "[gpt-high] 왼쪽 아래 매장 간판에 읽을 수 있는 영문 단어가 있어, 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반합니다.",
     "[gpt-high] 등반을 마친 이현우를 상층 개구부 가장자리의 잘린 전경 인물로 두어야 하는데, 정한수가 선 상층 슬래브 아래쪽 층에 서 있는 배치로 표현했습니다."
    ],
    "physics": "정한수는 경사진 잔해 위에서 균형을 잡고 서 있고, 이현우는 밧줄을 잡고 낭떠러지 가장자리에 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "정한수의 시선 방향과 딛고 있는 잔해의 형태가 지시문과 다소 다르나, 금지된 텍스트나 물리적 오류 없이 주요 샷 구성을 충실히 따랐습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "경사진 잔해와 시선 처리는 훌륭하나, 하단에 'FOOD'라는 명확한 텍스트가 생성되어 텍스트 렌더링 금지 제약을 치명적으로 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "정한수의 시선이 이현우를 직접 향하지 않고 약간 빗겨 허공을 향해 있으며, 이현우는 정한수를 올려다봄.",
        "built_space": "기둥과 유리 진열장이 남은 폐허 내부. 평평한 바닥이 붕괴되어 형성된 사각형 형태의 구멍 가장자리에 인물들이 배치됨.",
        "entities": "정한수는 짧은 머리, 야상 점퍼, 카고 바지 차림으로 지시에 부합함. 이현우는 왼쪽 하단에 크롭되어 나타남.",
        "hard_violations": [],
        "physics": "정한수는 평평한 바닥 끝에 안정적으로 서 있고, 이현우는 한 손으로 밧줄을 쥐고 있음. 중력을 거스르거나 지지대가 없는 물체는 없음."
       },
       {
        "label": "B",
        "direction": "정한수의 날카로운 시선이 이현우를 향해 정확히 고정되어 있으며, 이현우 역시 위를 주시함.",
        "built_space": "천장 채광창이 뚫려 빛이 들어오는 붕괴된 상업 시설. 거대한 경사진 콘크리트 판이 층과 층 사이를 잇고 있음.",
        "entities": "두 인물 모두 지시된 복장과 신체적 특징, 프레이밍 요구사항을 훌륭하게 충족함.",
        "hard_violations": [
         "화면 하단 상점 구조물에 'FOOD'라는 텍스트가 명확하게 읽히는 상태로 렌더링되어 'No readable writing' 제약을 위반함."
        ],
        "physics": "정한수는 경사진 잔해 위에서 균형을 잡고 서 있고, 이현우는 밧줄을 잡고 낭떠러지 가장자리에 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "정한수의 시선 방향과 딛고 있는 잔해의 형태가 지시문과 다소 다르나, 금지된 텍스트나 물리적 오류 없이 주요 샷 구성을 충실히 따랐습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "경사진 잔해와 시선 처리는 훌륭하나, 하단에 'FOOD'라는 명확한 텍스트가 생성되어 텍스트 렌더링 금지 제약을 치명적으로 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "정한수의 시선이 이현우를 직접 향하지 않고 약간 빗겨 허공을 향해 있으며, 이현우는 정한수를 올려다봄.",
        "built_space": "기둥과 유리 진열장이 남은 폐허 내부. 평평한 바닥이 붕괴되어 형성된 사각형 형태의 구멍 가장자리에 인물들이 배치됨.",
        "entities": "정한수는 짧은 머리, 야상 점퍼, 카고 바지 차림으로 지시에 부합함. 이현우는 왼쪽 하단에 크롭되어 나타남.",
        "hard_violations": [],
        "physics": "정한수는 평평한 바닥 끝에 안정적으로 서 있고, 이현우는 한 손으로 밧줄을 쥐고 있음. 중력을 거스르거나 지지대가 없는 물체는 없음."
       },
       {
        "label": "B",
        "direction": "정한수의 날카로운 시선이 이현우를 향해 정확히 고정되어 있으며, 이현우 역시 위를 주시함.",
        "built_space": "천장 채광창이 뚫려 빛이 들어오는 붕괴된 상업 시설. 거대한 경사진 콘크리트 판이 층과 층 사이를 잇고 있음.",
        "entities": "두 인물 모두 지시된 복장과 신체적 특징, 프레이밍 요구사항을 훌륭하게 충족함.",
        "hard_violations": [
         "화면 하단 상점 구조물에 'FOOD'라는 텍스트가 명확하게 읽히는 상태로 렌더링되어 'No readable writing' 제약을 위반함."
        ],
        "physics": "정한수는 경사진 잔해 위에서 균형을 잡고 서 있고, 이현우는 밧줄을 잡고 낭떠러지 가장자리에 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "전신과 서로 마주 보는 시선은 맞지만, 읽을 수 있는 간판이 금지 조건을 위반하고 이현우를 등반 완료 지점보다 아래층에 둔 구도가 지정된 대치 배치와 어긋납니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "잔해 위 정한수의 전신, 왼쪽 아래에 잘린 이현우와 손에 쥔 짧은 로프, 상층 붕괴구 가장자리를 함께 구현해 지정된 대치 구도와 장소 연속성에 더 충실합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "정한수는 고개와 눈을 왼쪽 아래 이현우 쪽으로 향하고, 이현우는 오른쪽 위 정한수를 올려다봅니다. 상호 대치의 시선 방향은 맞습니다. 무기나 이동 중인 인물은 없습니다.",
        "built_space": "정한수 아래에는 경사진 대형 콘크리트 판과 발을 받치는 파손 슬래브가 있고, 뒤에는 여러 사각 기둥과 상층 난간, 위에는 큰 채광창 하나가 보입니다. 왼쪽 아래 이현우는 상층 개구부 가장자리에 올라온 전경 인물보다는 슬래브 아래층에 선 인물로 배치되어 있습니다. 참조의 어두운 상층 바닥보다 천창과 수직 층간 높이가 강조되었습니다.",
        "entities": "두 성인 남성이 보입니다. 정한수는 한국인 남성 설정에 부합하는 외관, 짧은 머리, 낡은 야상과 여러 주머니가 있는 카고 바지를 갖췄습니다. 왼쪽 인물은 이현우 역할로 보이며 검은 곱슬머리와 어두운 후드 차림입니다. 손에 로프가 있고 콘크리트 잔해와 철근도 보입니다. 아래층 간판에는 읽을 수 있는 영문 음식 매장 표기가 있습니다. 파라솔과 식품 캔은 프레임에 보이지 않습니다.",
        "hard_violations": [
         "왼쪽 아래 매장 간판에 읽을 수 있는 영문 단어가 있어, 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반합니다.",
         "등반을 마친 이현우를 상층 개구부 가장자리의 잘린 전경 인물로 두어야 하는데, 정한수가 선 상층 슬래브 아래쪽 층에 서 있는 배치로 표현했습니다."
        ],
        "physics": "정한수의 양발은 파손 슬래브 윗면에 닿아 있어 체중 지지가 보입니다. 큰 콘크리트 판은 주변 잔해에 기대어 있습니다. 이현우는 손으로 로프를 잡고 있으며 로프는 손 아래로 처집니다. 이현우의 발과 바닥 접점은 화면 밖이므로 실제 발판은 확인할 수 없지만, 공중에 떠 있다고 단정할 근거는 없습니다."
       },
       {
        "label": "B",
        "direction": "정한수의 얼굴과 눈은 왼쪽 전경 이현우를 향하고, 이현우 역시 오른쪽의 정한수를 바라봅니다. 정한수의 날카로운 대치 시선이 요구된 상대에게 향합니다. 무기나 이동 중인 인물은 없습니다.",
        "built_space": "큰 바닥 붕괴구 하나, 그 안으로 기울어진 대형 슬래브 하나, 정한수를 받치는 오른쪽 잔해 윗면과 수직·사선 측면이 보입니다. 이현우는 왼쪽 아래 전경에 잘려 있고 바로 옆으로 개구부 가장자리가 이어집니다. 뒤에는 여러 사각 기둥, 왼쪽의 두 매대 구역, 오른쪽 위 에스컬레이터 일부가 보입니다. 참조의 콘크리트 기둥, 파손 바닥, 어두운 백화점 내부를 비교적 잘 유지합니다.",
        "entities": "두 성인 남성이 보이며 추가 인물은 없습니다. 정한수는 한국인 남성 설정에 부합하는 외관과 짧은 머리를 갖추고, 낡은 야상과 다중 주머니 카고 바지, 부츠를 착용했습니다. 이현우 역할의 왼쪽 전경 인물은 곱슬머리와 어두운 후드 차림이고 손에 구조 로프를 쥐고 있습니다. 아래층 파라솔과 캔은 보이지 않지만 이 구도의 필수 가시 대상은 아닙니다. 판독 가능한 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "정한수의 두 부츠는 잔해의 평평한 윗면에 확실히 닿아 있고, 그 아래에는 두꺼운 콘크리트 구조가 이어집니다. 기울어진 대형 슬래브는 하단 잔해에 걸쳐 있습니다. 이현우의 손가락은 로프를 감싸며 로프는 중력 방향으로 화면 아래까지 처집니다. 이현우의 하체는 전경 크롭 밖이어서 발판은 확인할 수 없으나, 떠 있는 자세로 묘사되지는 않았습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "전신과 서로 마주 보는 시선은 맞지만, 읽을 수 있는 간판이 금지 조건을 위반하고 이현우를 등반 완료 지점보다 아래층에 둔 구도가 지정된 대치 배치와 어긋납니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "잔해 위 정한수의 전신, 왼쪽 아래에 잘린 이현우와 손에 쥔 짧은 로프, 상층 붕괴구 가장자리를 함께 구현해 지정된 대치 구도와 장소 연속성에 더 충실합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "정한수는 고개와 눈을 왼쪽 아래 이현우 쪽으로 향하고, 이현우는 오른쪽 위 정한수를 올려다봅니다. 상호 대치의 시선 방향은 맞습니다. 무기나 이동 중인 인물은 없습니다.",
        "built_space": "정한수 아래에는 경사진 대형 콘크리트 판과 발을 받치는 파손 슬래브가 있고, 뒤에는 여러 사각 기둥과 상층 난간, 위에는 큰 채광창 하나가 보입니다. 왼쪽 아래 이현우는 상층 개구부 가장자리에 올라온 전경 인물보다는 슬래브 아래층에 선 인물로 배치되어 있습니다. 참조의 어두운 상층 바닥보다 천창과 수직 층간 높이가 강조되었습니다.",
        "entities": "두 성인 남성이 보입니다. 정한수는 한국인 남성 설정에 부합하는 외관, 짧은 머리, 낡은 야상과 여러 주머니가 있는 카고 바지를 갖췄습니다. 왼쪽 인물은 이현우 역할로 보이며 검은 곱슬머리와 어두운 후드 차림입니다. 손에 로프가 있고 콘크리트 잔해와 철근도 보입니다. 아래층 간판에는 읽을 수 있는 영문 음식 매장 표기가 있습니다. 파라솔과 식품 캔은 프레임에 보이지 않습니다.",
        "hard_violations": [
         "왼쪽 아래 매장 간판에 읽을 수 있는 영문 단어가 있어, 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반합니다.",
         "등반을 마친 이현우를 상층 개구부 가장자리의 잘린 전경 인물로 두어야 하는데, 정한수가 선 상층 슬래브 아래쪽 층에 서 있는 배치로 표현했습니다."
        ],
        "physics": "정한수의 양발은 파손 슬래브 윗면에 닿아 있어 체중 지지가 보입니다. 큰 콘크리트 판은 주변 잔해에 기대어 있습니다. 이현우는 손으로 로프를 잡고 있으며 로프는 손 아래로 처집니다. 이현우의 발과 바닥 접점은 화면 밖이므로 실제 발판은 확인할 수 없지만, 공중에 떠 있다고 단정할 근거는 없습니다."
       },
       {
        "label": "A",
        "direction": "정한수의 얼굴과 눈은 왼쪽 전경 이현우를 향하고, 이현우 역시 오른쪽의 정한수를 바라봅니다. 정한수의 날카로운 대치 시선이 요구된 상대에게 향합니다. 무기나 이동 중인 인물은 없습니다.",
        "built_space": "큰 바닥 붕괴구 하나, 그 안으로 기울어진 대형 슬래브 하나, 정한수를 받치는 오른쪽 잔해 윗면과 수직·사선 측면이 보입니다. 이현우는 왼쪽 아래 전경에 잘려 있고 바로 옆으로 개구부 가장자리가 이어집니다. 뒤에는 여러 사각 기둥, 왼쪽의 두 매대 구역, 오른쪽 위 에스컬레이터 일부가 보입니다. 참조의 콘크리트 기둥, 파손 바닥, 어두운 백화점 내부를 비교적 잘 유지합니다.",
        "entities": "두 성인 남성이 보이며 추가 인물은 없습니다. 정한수는 한국인 남성 설정에 부합하는 외관과 짧은 머리를 갖추고, 낡은 야상과 다중 주머니 카고 바지, 부츠를 착용했습니다. 이현우 역할의 왼쪽 전경 인물은 곱슬머리와 어두운 후드 차림이고 손에 구조 로프를 쥐고 있습니다. 아래층 파라솔과 캔은 보이지 않지만 이 구도의 필수 가시 대상은 아닙니다. 판독 가능한 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "정한수의 두 부츠는 잔해의 평평한 윗면에 확실히 닿아 있고, 그 아래에는 두꺼운 콘크리트 구조가 이어집니다. 기울어진 대형 슬래브는 하단 잔해에 걸쳐 있습니다. 이현우의 손가락은 로프를 감싸며 로프는 중력 방향으로 화면 아래까지 처집니다. 이현우의 하체는 전경 크롭 밖이어서 발판은 확인할 수 없으나, 떠 있는 자세로 묘사되지는 않았습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.708
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.458
   },
   "violations": {
    "B": [
     "[gemini-pro] 화면 하단 상점 구조물에 'FOOD'라는 텍스트가 명확하게 읽히는 상태로 렌더링되어 'No readable writing' 제약을 위반함.",
     "[gpt-high] 왼쪽 아래 매장 간판에 읽을 수 있는 영문 단어가 있어, 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반합니다.",
     "[gpt-high] 등반을 마친 이현우를 상층 개구부 가장자리의 잘린 전경 인물로 두어야 하는데, 정한수가 선 상층 슬래브 아래쪽 층에 서 있는 배치로 표현했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 458
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "정한수의 시선 방향과 딛고 있는 잔해의 형태가 지시문과 다소 다르나, 금지된 텍스트나 물리적 오류 없이 주요 샷 구성을 충실히 따랐습니다."
   },
   {
    "label": "B",
    "score": 458,
    "verdict_ko": "경사진 잔해와 시선 처리는 훌륭하나, 하단에 'FOOD'라는 명확한 텍스트가 생성되어 텍스트 렌더링 금지 제약을 치명적으로 위반했습니다.  ★위반: [gemini-pro] 화면 하단 상점 구조물에 'FOOD'라는 텍스트가 명확하게 읽히는 상태로 렌더링되어 'No readable writing' 제약을 위반함. / [gpt-high] 왼쪽 아래 매장 간판에 읽을 수 있는 영문 단어가 있어, 어디에도 판독 가능한 글자를 두지 말라는 조건을 위반합니다. / [gpt-high] 등반을 마친 이현우를 상층 개구부 가장자리의 잘린 전경 인물로 두어야 하는데, 정한수가 선 상층 슬래브 아래쪽 층에 서 있는 배치로 표현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S66sh9_sel.png",
    "asset_id": "a814044f-d8ab-44ab-b805-098653591bec",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aace02-4c2c-7a36-8752-fd2ccd46da64",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S66sh9"
  }
 },
 "S66sh40::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:54:28.831543+00:00",
  "fingerprint": "c26b2a321424eafe0d32018596ec87a494fad2ce6839596e0efd327c00923d06",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S66sh40_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S66sh40_sel.png",
  "source_sha256": "85a2344a65e2b93ab4caa3dcc55bf61ca64ef876504e6623536ae75eb48300a8",
  "file": "S66sh40_cine.png",
  "staged_sha256": "fd1281b8e2997b4613c0c24b119d8a2cfe58a316389a97895ed53c153e1f0659",
  "latency_ms": 13017
 },
 "S67sh80": {
  "input_fingerprint": "e548cb306a1efe33",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 찰리가 부품을 쥔 채 두 눈의 푸른빛이 완전히 꺼지고 고개를 푹 떨군 깊은 슬픔의 자태.\n\nLOCATION (lock): At the smoke-filled blast site on the village departure road, amid shattered robot parts and military wreckage. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: B-200's remaining chest component (Recovered and held in 찰리's hands) — Visible obliquely between both palms without specifying an unsupported face or marking; used as Small lower-center emotional anchor linking the bowed face to the loss; Blast site (After the explosion, with smoke clearing); used as Subdued background context that does not compete with the face and hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Under the established night illumination, the complete loss of 찰리's blue eye light leaves restrained facial contrast and a solemn, quiet mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Black explosion smoke is clearing over overturned trucks, the opened cage and broken restraints. B-200 is shattered, with only his chest component remaining, now held by the heavily damaged Charlie; Charlie closes his eyes, without an established permanent loss of power.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 찰리가 부품을 쥔 채 두 눈의 푸른빛이 완전히 꺼지고 고개를 푹 떨군 깊은 슬픔의 자태.\n\nLOCATION (lock): At the smoke-filled blast site on the village departure road, amid shattered robot parts and military wreckage. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: B-200's remaining chest component (Recovered and held in 찰리's hands) — Visible obliquely between both palms without specifying an unsupported face or marking; used as Small lower-center emotional anchor linking the bowed face to the loss; Blast site (After the explosion, with smoke clearing); used as Subdued background context that does not compete with the face and hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Under the established night illumination, the complete loss of 찰리's blue eye light leaves restrained facial contrast and a solemn, quiet mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Black explosion smoke is clearing over overturned trucks, the opened cage and broken restraints. B-200 is shattered, with only his chest component remaining, now held by the heavily damaged Charlie; Charlie closes his eyes, without an established permanent loss of power.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 찰리가 부품을 쥔 채 두 눈의 푸른빛이 완전히 꺼지고 고개를 푹 떨군 깊은 슬픔의 자태.\n\nLOCATION (lock): At the smoke-filled blast site on the village departure road, amid shattered robot parts and military wreckage. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: B-200's remaining chest component (Recovered and held in 찰리's hands) — Visible obliquely between both palms without specifying an unsupported face or marking; used as Small lower-center emotional anchor linking the bowed face to the loss; Blast site (After the explosion, with smoke clearing); used as Subdued background context that does not compete with the face and hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Under the established night illumination, the complete loss of 찰리's blue eye light leaves restrained facial contrast and a solemn, quiet mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Black explosion smoke is clearing over overturned trucks, the opened cage and broken restraints. B-200 is shattered, with only his chest component remaining, now held by the heavily damaged Charlie; Charlie closes his eyes, without an established permanent loss of power.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 시선이 양손에 든 기계 부품을 향해 아래로 내려다보고 있음.",
    "built_space": "야간의 마을 도로로, 왼쪽의 붉은 벽돌 건물과 트럭, 오른쪽의 전통 가옥과 뒤집힌 트럭, 하늘의 보름달 등 레퍼런스의 공간 배치가 정확히 일치함.",
    "entities": "고릴라 형태의 기계 머리, 각진 외장, '흰 얼굴판'을 가진 찰리의 외형이 텍스트와 일치하며, 눈의 푸른빛이 완전히 꺼져 있음. 손에는 부품이 들려 있음.",
    "hard_violations": [],
    "physics": "찰리가 두 발로 서서 양손으로 부품의 무게를 지탱하며 안정적으로 들고 있음."
   },
   {
    "label": "B",
    "direction": "찰리의 시선이 양손에 든 부품을 향해 아래로 향함.",
    "built_space": "레퍼런스와 동일하게 트럭, 건물, 달, 연기 등이 배치된 야간의 마을 도로 구조를 따름.",
    "entities": "찰리의 얼굴판이 프롬프트가 요구한 '흰색'이 아닌 어두운 금속으로 묘사됨. 좌측 하단에 지시되지 않은 작은 로봇 형태가 등장함.",
    "hard_violations": [
     "[gemini-pro] Invented people or objects (좌측 하단에 프롬프트에 명시되지 않은 형태의 로봇 캐릭터/객체 추가)",
     "[gpt-high] 왼쪽 하단에 가슴 부품 외의 로봇 몸통과 연결된 팔이 추가되어, B-200은 손에 든 가슴 부품만 남았다는 상태 및 추가 몸체 금지를 위반한다."
    ],
    "physics": "부품을 양손으로 들고 서 있는 자세가 물리적으로 가능하게 표현됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리의 '흰 얼굴판'과 꺼진 눈빛을 정확히 묘사하여 프롬프트의 지시를 충실히 따랐으며, 배경의 연속성과 슬픈 분위기도 잘 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "명시된 '흰 얼굴판' 지시를 어기고 어두운 금속 질감으로 렌더링했으며, 좌측 하단에 프롬프트에 없는 로봇 형태를 임의로 추가하여 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선이 양손에 든 기계 부품을 향해 아래로 내려다보고 있음.",
        "built_space": "야간의 마을 도로로, 왼쪽의 붉은 벽돌 건물과 트럭, 오른쪽의 전통 가옥과 뒤집힌 트럭, 하늘의 보름달 등 레퍼런스의 공간 배치가 정확히 일치함.",
        "entities": "고릴라 형태의 기계 머리, 각진 외장, '흰 얼굴판'을 가진 찰리의 외형이 텍스트와 일치하며, 눈의 푸른빛이 완전히 꺼져 있음. 손에는 부품이 들려 있음.",
        "hard_violations": [],
        "physics": "찰리가 두 발로 서서 양손으로 부품의 무게를 지탱하며 안정적으로 들고 있음."
       },
       {
        "label": "B",
        "direction": "찰리의 시선이 양손에 든 부품을 향해 아래로 향함.",
        "built_space": "레퍼런스와 동일하게 트럭, 건물, 달, 연기 등이 배치된 야간의 마을 도로 구조를 따름.",
        "entities": "찰리의 얼굴판이 프롬프트가 요구한 '흰색'이 아닌 어두운 금속으로 묘사됨. 좌측 하단에 지시되지 않은 작은 로봇 형태가 등장함.",
        "hard_violations": [
         "Invented people or objects (좌측 하단에 프롬프트에 명시되지 않은 형태의 로봇 캐릭터/객체 추가)"
        ],
        "physics": "부품을 양손으로 들고 서 있는 자세가 물리적으로 가능하게 표현됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리의 '흰 얼굴판'과 꺼진 눈빛을 정확히 묘사하여 프롬프트의 지시를 충실히 따랐으며, 배경의 연속성과 슬픈 분위기도 잘 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "명시된 '흰 얼굴판' 지시를 어기고 어두운 금속 질감으로 렌더링했으며, 좌측 하단에 프롬프트에 없는 로봇 형태를 임의로 추가하여 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선이 양손에 든 기계 부품을 향해 아래로 내려다보고 있음.",
        "built_space": "야간의 마을 도로로, 왼쪽의 붉은 벽돌 건물과 트럭, 오른쪽의 전통 가옥과 뒤집힌 트럭, 하늘의 보름달 등 레퍼런스의 공간 배치가 정확히 일치함.",
        "entities": "고릴라 형태의 기계 머리, 각진 외장, '흰 얼굴판'을 가진 찰리의 외형이 텍스트와 일치하며, 눈의 푸른빛이 완전히 꺼져 있음. 손에는 부품이 들려 있음.",
        "hard_violations": [],
        "physics": "찰리가 두 발로 서서 양손으로 부품의 무게를 지탱하며 안정적으로 들고 있음."
       },
       {
        "label": "B",
        "direction": "찰리의 시선이 양손에 든 부품을 향해 아래로 향함.",
        "built_space": "레퍼런스와 동일하게 트럭, 건물, 달, 연기 등이 배치된 야간의 마을 도로 구조를 따름.",
        "entities": "찰리의 얼굴판이 프롬프트가 요구한 '흰색'이 아닌 어두운 금속으로 묘사됨. 좌측 하단에 지시되지 않은 작은 로봇 형태가 등장함.",
        "hard_violations": [
         "Invented people or objects (좌측 하단에 프롬프트에 명시되지 않은 형태의 로봇 캐릭터/객체 추가)"
        ],
        "physics": "부품을 양손으로 들고 서 있는 자세가 물리적으로 가능하게 표현됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "미디엄 숏과 양손의 부품 지지는 맞지만, 별도의 로봇 몸통·팔 잔해가 ‘가슴 부품만 남음’에 어긋나며 흰 얼굴판과 깊이 떨군 고개도 구현하지 못했다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "흰 얼굴판, 소등된 눈, 양손에 든 가슴 부품과 잦아드는 연기 배경이 더 충실하지만, 고개 숙임이 얕고 부품이 작고 비스듬한 하단 중심점보다 크고 정면으로 제시된다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴은 조금 아래로 기울었지만 여전히 카메라 쪽을 향한다. 검은 눈은 열린 듯 보이며 시선이 손의 부품에 확실히 닿지는 않는다. 부품은 양손 사이에서 대각선으로 놓이고 원통형 끝은 화면 오른쪽 위를 향한다. 겨누는 무기나 이동 중인 인물은 없다.",
        "built_space": "왼쪽 벽돌 건물, 중앙의 높은 건물, 오른쪽 기와 건물이 이전 장소와 대응한다. 왼쪽에는 철제 적재함 틀이 있는 트럭 한 대와 열린 문이, 오른쪽에는 뒤집힌 차량 한 대가 보인다. 하단에는 낮은 콘크리트 경계가 이어진다. 찰리는 도로 중앙 전경에 허리 부근까지 잡혀 미디엄 숏에 해당하지만, 짙은 검은 연기와 선명한 잔해가 배경의 존재감을 높인다.",
        "entities": "중앙에는 손상된 고릴라형 기계 찰리 한 명이 있고 두 눈에 푸른 발광은 없다. 그러나 얼굴은 흰 얼굴판보다 어두운 금속의 입체적인 고릴라 얼굴이며 눈도 작은 점 형태와 다르다. 양손은 관절식 금속 손이지만 참고의 손보다 육중하고 손목에 참고의 색색 소매는 없다. 가슴 부품으로 읽히는 기계 상자를 들고 있으며, 왼쪽 하단에는 별도의 로봇 몸통과 연결된 팔이 보인다. 보름달, 차량 잔해와 연기는 있으나 끊어진 구속구는 분명하지 않다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "왼쪽 하단에 가슴 부품 외의 로봇 몸통과 연결된 팔이 추가되어, B-200은 손에 든 가슴 부품만 남았다는 상태 및 추가 몸체 금지를 위반한다."
        ],
        "physics": "부품의 양옆과 아래쪽을 두 금속 손이 감싸 지지하고, 손목과 팔꿈치가 몸통으로 이어져 무게를 받는 구조가 보인다. 하체는 프레임 밖이므로 발의 접지는 확인할 수 없지만 공중에 떠 있는 자세는 아니다. 왼쪽 로봇 잔해와 양쪽 차량은 도로 위에 놓여 있다. 지지 없는 부유 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리는 턱을 약간 내리고 얼굴을 정면 아래로 향하지만, 고개를 푹 떨군 정도는 아니다. 눈 부분은 빛 없는 좁은 검은 틈으로 보이며 완전히 감았는지는 명확하지 않다. 손의 가슴 부품은 넓은 면이 거의 카메라를 향하고 약간만 기울어져 있어, 양손 사이에 비스듬히 보이게 하라는 지시에는 덜 맞는다.",
        "built_space": "왼쪽 벽돌 건물 한 동, 중앙의 높은 건물, 오른쪽 기와 건물군이 이전 장소의 배치를 유지한다. 왼쪽 철제 적재함 트럭 한 대와 오른쪽 뒤집힌 차량 한 대가 보이고, 열린 차량 문은 왼쪽 트럭 옆에 일부 드러난다. 찰리는 도로 중앙 전경에서 허리 부근까지 보인다. 연무와 얕은 배경 초점이 건물과 잔해를 얼굴·손보다 덜 두드러지게 한다. 중복된 고정 시설이나 반사는 없다.",
        "entities": "찰리 한 명만 보이며 각진 기계 머리, 흰 얼굴판, 짧은 선 형태의 입과 푸른빛 없는 눈이 명시된 정체성에 더 가깝다. 외장에는 그을음과 긁힘이 있으나 심하게 파손된 정도는 약하다. 관절식 금속 손은 참고와 유형이 맞지만 손목의 색색 소매는 반영되지 않았다. 양손에는 찢긴 가장자리의 가슴 외장 부품이 있고 추가 로봇 몸체는 없다. 부품은 화면 하단 중앙에 있지만 작은 감정적 중심점이라기에는 크다. 보름달과 흩어지는 연기, 차량 잔해는 보이며 구속구는 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양손의 손가락이 가슴 부품의 좌우 가장자리를 감싸고 아래쪽을 받쳐 부품이 실제로 지지된다. 팔과 손목의 연결 및 구부린 팔 자세도 유지 가능하다. 하체는 프레임 밖이며 떠 있는 몸으로 보이지 않는다. 배경 차량들은 지면에 놓여 있고 공중에 떠 있는 잔해는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미디엄 숏과 양손의 부품 지지는 맞지만, 별도의 로봇 몸통·팔 잔해가 ‘가슴 부품만 남음’에 어긋나며 흰 얼굴판과 깊이 떨군 고개도 구현하지 못했다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "흰 얼굴판, 소등된 눈, 양손에 든 가슴 부품과 잦아드는 연기 배경이 더 충실하지만, 고개 숙임이 얕고 부품이 작고 비스듬한 하단 중심점보다 크고 정면으로 제시된다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴은 조금 아래로 기울었지만 여전히 카메라 쪽을 향한다. 검은 눈은 열린 듯 보이며 시선이 손의 부품에 확실히 닿지는 않는다. 부품은 양손 사이에서 대각선으로 놓이고 원통형 끝은 화면 오른쪽 위를 향한다. 겨누는 무기나 이동 중인 인물은 없다.",
        "built_space": "왼쪽 벽돌 건물, 중앙의 높은 건물, 오른쪽 기와 건물이 이전 장소와 대응한다. 왼쪽에는 철제 적재함 틀이 있는 트럭 한 대와 열린 문이, 오른쪽에는 뒤집힌 차량 한 대가 보인다. 하단에는 낮은 콘크리트 경계가 이어진다. 찰리는 도로 중앙 전경에 허리 부근까지 잡혀 미디엄 숏에 해당하지만, 짙은 검은 연기와 선명한 잔해가 배경의 존재감을 높인다.",
        "entities": "중앙에는 손상된 고릴라형 기계 찰리 한 명이 있고 두 눈에 푸른 발광은 없다. 그러나 얼굴은 흰 얼굴판보다 어두운 금속의 입체적인 고릴라 얼굴이며 눈도 작은 점 형태와 다르다. 양손은 관절식 금속 손이지만 참고의 손보다 육중하고 손목에 참고의 색색 소매는 없다. 가슴 부품으로 읽히는 기계 상자를 들고 있으며, 왼쪽 하단에는 별도의 로봇 몸통과 연결된 팔이 보인다. 보름달, 차량 잔해와 연기는 있으나 끊어진 구속구는 분명하지 않다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "왼쪽 하단에 가슴 부품 외의 로봇 몸통과 연결된 팔이 추가되어, B-200은 손에 든 가슴 부품만 남았다는 상태 및 추가 몸체 금지를 위반한다."
        ],
        "physics": "부품의 양옆과 아래쪽을 두 금속 손이 감싸 지지하고, 손목과 팔꿈치가 몸통으로 이어져 무게를 받는 구조가 보인다. 하체는 프레임 밖이므로 발의 접지는 확인할 수 없지만 공중에 떠 있는 자세는 아니다. 왼쪽 로봇 잔해와 양쪽 차량은 도로 위에 놓여 있다. 지지 없는 부유 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리는 턱을 약간 내리고 얼굴을 정면 아래로 향하지만, 고개를 푹 떨군 정도는 아니다. 눈 부분은 빛 없는 좁은 검은 틈으로 보이며 완전히 감았는지는 명확하지 않다. 손의 가슴 부품은 넓은 면이 거의 카메라를 향하고 약간만 기울어져 있어, 양손 사이에 비스듬히 보이게 하라는 지시에는 덜 맞는다.",
        "built_space": "왼쪽 벽돌 건물 한 동, 중앙의 높은 건물, 오른쪽 기와 건물군이 이전 장소의 배치를 유지한다. 왼쪽 철제 적재함 트럭 한 대와 오른쪽 뒤집힌 차량 한 대가 보이고, 열린 차량 문은 왼쪽 트럭 옆에 일부 드러난다. 찰리는 도로 중앙 전경에서 허리 부근까지 보인다. 연무와 얕은 배경 초점이 건물과 잔해를 얼굴·손보다 덜 두드러지게 한다. 중복된 고정 시설이나 반사는 없다.",
        "entities": "찰리 한 명만 보이며 각진 기계 머리, 흰 얼굴판, 짧은 선 형태의 입과 푸른빛 없는 눈이 명시된 정체성에 더 가깝다. 외장에는 그을음과 긁힘이 있으나 심하게 파손된 정도는 약하다. 관절식 금속 손은 참고와 유형이 맞지만 손목의 색색 소매는 반영되지 않았다. 양손에는 찢긴 가장자리의 가슴 외장 부품이 있고 추가 로봇 몸체는 없다. 부품은 화면 하단 중앙에 있지만 작은 감정적 중심점이라기에는 크다. 보름달과 흩어지는 연기, 차량 잔해는 보이며 구속구는 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양손의 손가락이 가슴 부품의 좌우 가장자리를 감싸고 아래쪽을 받쳐 부품이 실제로 지지된다. 팔과 손목의 연결 및 구부린 팔 자세도 유지 가능하다. 하체는 프레임 밖이며 떠 있는 몸으로 보이지 않는다. 배경 차량들은 지면에 놓여 있고 공중에 떠 있는 잔해는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.0
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.75
   },
   "violations": {
    "B": [
     "[gemini-pro] Invented people or objects (좌측 하단에 프롬프트에 명시되지 않은 형태의 로봇 캐릭터/객체 추가)",
     "[gpt-high] 왼쪽 하단에 가슴 부품 외의 로봇 몸통과 연결된 팔이 추가되어, B-200은 손에 든 가슴 부품만 남았다는 상태 및 추가 몸체 금지를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "찰리의 '흰 얼굴판'과 꺼진 눈빛을 정확히 묘사하여 프롬프트의 지시를 충실히 따랐으며, 배경의 연속성과 슬픈 분위기도 잘 구현했습니다."
   },
   {
    "label": "B",
    "score": 750,
    "verdict_ko": "명시된 '흰 얼굴판' 지시를 어기고 어두운 금속 질감으로 렌더링했으며, 좌측 하단에 프롬프트에 없는 로봇 형태를 임의로 추가하여 감점되었습니다.  ★위반: [gemini-pro] Invented people or objects (좌측 하단에 프롬프트에 명시되지 않은 형태의 로봇 캐릭터/객체 추가) / [gpt-high] 왼쪽 하단에 가슴 부품 외의 로봇 몸통과 연결된 팔이 추가되어, B-200은 손에 든 가슴 부품만 남았다는 상태 및 추가 몸체 금지를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S67sh75_sel.png",
    "asset_id": "9c85090d-269c-4523-98db-58c20d4c6ddf",
    "role": "prev_still"
   },
   {
    "label": "HAND OWNER REFERENCE — 찰리: the person whose hand is on the object in this shot. Use this photograph ONLY to get that hand right — its size, build, skin tone and texture, apparent age, grooming, nails, sleeve and anything worn on the wrist or fingers. Do NOT bring their face, body or clothing into the frame; the framing stays on the object and reaches no further than the wrist and forearm.",
    "path": "<bytes:885644>",
    "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace16-62a6-74b2-88f4-22b3088fd901",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S67sh75"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S68sh4": {
  "input_fingerprint": "83ccb99d3834382e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갈림길에서 서로 다른 방향을 향해 꺾여 달리고 있는 두 대의 트럭의 mid-action 구도.\n\nLOCATION (lock): At an open countryside road fork where two groups of old trucks separate in daylight. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이현우's turning truck in the middle-left of the frame, midground, moves toward Upper-left road branch; 수빈's turning truck and following route in the middle-right of the frame, midground, moves toward Upper-right road branch.\n- KEY BACKGROUND ELEMENTS: 이현우's truck (Old, roofless, turning into one branch) — Seen from above and behind, angled toward the upper-left road; used as Left-moving half of the separation, held at a natural scale within the road layout; 수빈's truck (Old, roofless, turning into the other branch) — Seen from above and behind, angled toward the upper-right road; used as Right-moving counterpart that makes the divergence legible; Road fork (Splitting into two routes) — The common approach enters from below and separates toward the upper corners; used as Primary spatial structure distinguishing the two destinations; Following vehicles (Following 수빈's truck) — Rear-oblique views align with the upper-right branch at different stages of the turn; used as Establishes the convoy's chosen route without duplicating vehicle spacing or orientation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and balanced contrast keep both turning vehicles and the road's branching geometry equally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Two old roofless trucks separate at the fork, with additional vehicles following one branch. Charlie sits in a cargo bed carrying B-200's surviving chest component, his holes, dents and fresh combat damage still unrepaired.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갈림길에서 서로 다른 방향을 향해 꺾여 달리고 있는 두 대의 트럭의 mid-action 구도.\n\nLOCATION (lock): At an open countryside road fork where two groups of old trucks separate in daylight. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이현우's turning truck in the middle-left of the frame, midground, moves toward Upper-left road branch; 수빈's turning truck and following route in the middle-right of the frame, midground, moves toward Upper-right road branch.\n- KEY BACKGROUND ELEMENTS: 이현우's truck (Old, roofless, turning into one branch) — Seen from above and behind, angled toward the upper-left road; used as Left-moving half of the separation, held at a natural scale within the road layout; 수빈's truck (Old, roofless, turning into the other branch) — Seen from above and behind, angled toward the upper-right road; used as Right-moving counterpart that makes the divergence legible; Road fork (Splitting into two routes) — The common approach enters from below and separates toward the upper corners; used as Primary spatial structure distinguishing the two destinations; Following vehicles (Following 수빈's truck) — Rear-oblique views align with the upper-right branch at different stages of the turn; used as Establishes the convoy's chosen route without duplicating vehicle spacing or orientation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and balanced contrast keep both turning vehicles and the road's branching geometry equally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Two old roofless trucks separate at the fork, with additional vehicles following one branch. Charlie sits in a cargo bed carrying B-200's surviving chest component, his holes, dents and fresh combat damage still unrepaired.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갈림길에서 서로 다른 방향을 향해 꺾여 달리고 있는 두 대의 트럭의 mid-action 구도.\n\nLOCATION (lock): At an open countryside road fork where two groups of old trucks separate in daylight. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이현우's turning truck in the middle-left of the frame, midground, moves toward Upper-left road branch; 수빈's turning truck and following route in the middle-right of the frame, midground, moves toward Upper-right road branch.\n- KEY BACKGROUND ELEMENTS: 이현우's truck (Old, roofless, turning into one branch) — Seen from above and behind, angled toward the upper-left road; used as Left-moving half of the separation, held at a natural scale within the road layout; 수빈's truck (Old, roofless, turning into the other branch) — Seen from above and behind, angled toward the upper-right road; used as Right-moving counterpart that makes the divergence legible; Road fork (Splitting into two routes) — The common approach enters from below and separates toward the upper corners; used as Primary spatial structure distinguishing the two destinations; Following vehicles (Following 수빈's truck) — Rear-oblique views align with the upper-right branch at different stages of the turn; used as Establishes the convoy's chosen route without duplicating vehicle spacing or orientation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and balanced contrast keep both turning vehicles and the road's branching geometry equally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Two old roofless trucks separate at the fork, with additional vehicles following one branch. Charlie sits in a cargo bed carrying B-200's surviving chest component, his holes, dents and fresh combat damage still unrepaired.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "왼쪽 트럭은 왼쪽 길을 향하지만, 오른쪽 트럭과 뒤따르는 차량들은 카메라 쪽으로 다가오며 역방향을 향함.",
    "built_space": "참조 사진에 있는 갈림길, 중앙의 큰 나무, 정자, 표지판의 위치와 형태가 정확히 묘사됨.",
    "entities": "오른쪽 트럭에 찰리가 손상된 기계 부품을 들고 있으나, 난민 그룹의 묘사가 거의 없음.",
    "hard_violations": [
     "[gemini-pro] 오른쪽 트럭과 후발 차량의 진행 방향 및 시점 위반 (뒤에서 멀어지는 구도가 아님)",
     "[gpt-high] 후속 트럭들에 샷에서 허용하지 않은 여러 탑승자를 추가했다."
    ],
    "physics": "트럭의 접지 상태나 찰리가 부품을 들고 있는 자세 등은 물리적으로 자연스러움."
   },
   {
    "label": "B",
    "direction": "두 트럭 모두 카메라를 등지고 각각 좌측 상단과 우측 상단을 향해 올바르게 갈라져 이동함.",
    "built_space": "Y자 갈림길, 정자, 표지판, 나무, 우측 철로 등 참조 사진의 모든 환경 구조가 정확한 시점에서 묘사됨.",
    "entities": "지붕 없는 두 트럭 모두에 다국적 난민들이 탑승해 있으며, 오른쪽 트럭에 찰리가 손상된 B-200 흉부 부품을 안고 있음.",
    "hard_violations": [
     "[gpt-high] 샷에서 허용하지 않은 다수의 적재함 승객과 오토바이 탑승자를 추가했다."
    ],
    "physics": "차량들의 바닥 밀착 상태가 정상적이며, 다수의 인물과 찰리가 짐칸에 안정적으로 안착해 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "멀어지며 갈라지는 두 트럭의 구도, 난민 그룹, 찰리의 외형 및 파손된 부품, 완벽한 배경 구현까지 모든 프롬프트 요구사항을 정확히 충족함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오른쪽 트럭과 후발 차량이 뒤에서 바라보는 구도가 아니라 카메라를 향해 정면으로 다가오도록 묘사되어 치명적인 연출 위반이 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 트럭은 왼쪽 길을 향하지만, 오른쪽 트럭과 뒤따르는 차량들은 카메라 쪽으로 다가오며 역방향을 향함.",
        "built_space": "참조 사진에 있는 갈림길, 중앙의 큰 나무, 정자, 표지판의 위치와 형태가 정확히 묘사됨.",
        "entities": "오른쪽 트럭에 찰리가 손상된 기계 부품을 들고 있으나, 난민 그룹의 묘사가 거의 없음.",
        "hard_violations": [
         "오른쪽 트럭과 후발 차량의 진행 방향 및 시점 위반 (뒤에서 멀어지는 구도가 아님)"
        ],
        "physics": "트럭의 접지 상태나 찰리가 부품을 들고 있는 자세 등은 물리적으로 자연스러움."
       },
       {
        "label": "B",
        "direction": "두 트럭 모두 카메라를 등지고 각각 좌측 상단과 우측 상단을 향해 올바르게 갈라져 이동함.",
        "built_space": "Y자 갈림길, 정자, 표지판, 나무, 우측 철로 등 참조 사진의 모든 환경 구조가 정확한 시점에서 묘사됨.",
        "entities": "지붕 없는 두 트럭 모두에 다국적 난민들이 탑승해 있으며, 오른쪽 트럭에 찰리가 손상된 B-200 흉부 부품을 안고 있음.",
        "hard_violations": [],
        "physics": "차량들의 바닥 밀착 상태가 정상적이며, 다수의 인물과 찰리가 짐칸에 안정적으로 안착해 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "멀어지며 갈라지는 두 트럭의 구도, 난민 그룹, 찰리의 외형 및 파손된 부품, 완벽한 배경 구현까지 모든 프롬프트 요구사항을 정확히 충족함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오른쪽 트럭과 후발 차량이 뒤에서 바라보는 구도가 아니라 카메라를 향해 정면으로 다가오도록 묘사되어 치명적인 연출 위반이 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 트럭은 왼쪽 길을 향하지만, 오른쪽 트럭과 뒤따르는 차량들은 카메라 쪽으로 다가오며 역방향을 향함.",
        "built_space": "참조 사진에 있는 갈림길, 중앙의 큰 나무, 정자, 표지판의 위치와 형태가 정확히 묘사됨.",
        "entities": "오른쪽 트럭에 찰리가 손상된 기계 부품을 들고 있으나, 난민 그룹의 묘사가 거의 없음.",
        "hard_violations": [
         "오른쪽 트럭과 후발 차량의 진행 방향 및 시점 위반 (뒤에서 멀어지는 구도가 아님)"
        ],
        "physics": "트럭의 접지 상태나 찰리가 부품을 들고 있는 자세 등은 물리적으로 자연스러움."
       },
       {
        "label": "B",
        "direction": "두 트럭 모두 카메라를 등지고 각각 좌측 상단과 우측 상단을 향해 올바르게 갈라져 이동함.",
        "built_space": "Y자 갈림길, 정자, 표지판, 나무, 우측 철로 등 참조 사진의 모든 환경 구조가 정확한 시점에서 묘사됨.",
        "entities": "지붕 없는 두 트럭 모두에 다국적 난민들이 탑승해 있으며, 오른쪽 트럭에 찰리가 손상된 B-200 흉부 부품을 안고 있음.",
        "hard_violations": [],
        "physics": "차량들의 바닥 밀착 상태가 정상적이며, 다수의 인물과 찰리가 짐칸에 안정적으로 안착해 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "주요 트럭 두 대의 후방 사선과 좌우 분기 방향은 맞지만, 허용되지 않은 다수의 승객을 추가했고 오른쪽 후속 차량들이 반대 방향을 향한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "장소와 왼쪽 트럭의 방향은 맞지만, 오른쪽 트럭을 정면에서 보여 아래쪽으로 달리게 해 핵심 분리 동작을 어겼으며 추가 인물도 보인다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 트럭은 적재함 뒤쪽이 카메라를 향하고 운전석 쪽이 왼쪽 위 도로를 향한다. 오른쪽 주요 트럭도 뒤쪽을 보이며 앞부분이 오른쪽 위 갈래를 향해 두 대의 분리는 읽힌다. 그러나 오른쪽 먼 도로의 자동차와 오토바이들은 대부분 전면을 카메라 쪽으로 드러내며 아래쪽 갈림길로 접근한다. 따라서 수빈을 따라 오른쪽 위로 나가는 후속 행렬이 아니다. 승객들의 시선은 차 안과 옆쪽으로 흩어져 있으며, 로봇은 아래쪽의 들고 있는 부품을 보는 듯하다.",
        "built_space": "아래에서 들어와 좌우 위로 나뉘는 포장도로, 중앙 섬의 큰 낙엽수 한 그루, 그 왼쪽 목조 정자 한 채, 나무 앞 삼각 경고표지 한 개, 오른쪽 단선 철로가 보인다. 정자 난간과 기둥, 도로변 볼라드와 전신주, 뒤쪽 밭과 마을도 참고 장소에 대체로 대응한다. 높은 후방 시점은 허용된 구도지만, 주요 트럭들이 중경의 자연스러운 작은 비중보다 화면 아래쪽을 크게 차지한다.",
        "entities": "오래되고 녹슨 지붕 없는 주요 트럭 두 대와 추가 차량들이 보인다. 두 적재함에는 샷에서 요구하지 않은 다수의 성인 남녀 승객이 있으며, 여러 오토바이 운전자도 추가되었다. 이현우와 수빈의 신원은 이 거리와 가림으로 확인하기 어렵다. 오른쪽에는 고릴라형 머리의 대형 기계 찰리가 있고, 손 사이에 구멍과 손상이 있는 별도 흉부 외장처럼 보이는 물체가 있다. 다만 찰리의 얼굴은 지정된 흰 얼굴판과 점눈·선입보다 어두운 입체 고릴라 얼굴에 가깝다. 낮의 실제 농촌 재질 표현은 부합한다.",
        "hard_violations": [
         "샷에서 허용하지 않은 다수의 적재함 승객과 오토바이 탑승자를 추가했다."
        ],
        "physics": "트럭과 후속 차량은 타이어로 노면에 지지된다. 승객들은 적재함 바닥이나 가장자리에 앉거나 기대고 있고, 찰리의 하체도 적재함 안에 놓여 있다. 흉부 부품은 찰리의 양손과 몸 앞쪽에 받쳐져 있어 공중에 떠 있지 않다. 두 주요 트럭의 차체 각도는 각각의 도로로 선회할 수 있는 자세지만, 후속 차량들의 진행 방향은 행렬 설정과 충돌한다."
       },
       {
        "label": "B",
        "direction": "왼쪽 트럭은 뒤 적재함을 보이며 왼쪽 위 갈래로 향하고 운전자도 그쪽을 본다. 오른쪽 주요 트럭은 전면 그릴과 보닛을 카메라에 보이며 왼쪽 아래의 공통 진입로 쪽으로 향한다. 운전자 역시 그 전방을 보고 있어 오른쪽 위로 갈라져 나가는 동작이 아니다. 뒤의 차량들도 전면을 보이며 아래쪽으로 접근한다. 찰리의 얼굴도 대체로 카메라 쪽을 향한다.",
        "built_space": "중앙의 큰 나무 한 그루, 왼쪽 정자 한 채, 삼각 경고표지 한 개, 오른쪽 단선 철로와 전신주·볼라드, 배후 농경지와 마을이 유지된다. 공통 도로가 아래에서 좌우 위로 갈라지는 장소 구조는 맞고, 정자의 바닥과 난간도 자연스럽다. 다만 오른쪽 주요 트럭이 화면 오른쪽 아래에 크게 놓여 중경 후방 사선이라는 지정 배치를 벗어난다.",
        "entities": "주요 차량은 녹슨 지붕 없는 트럭 두 대이며 왼쪽 적재함은 비어 있고 오른쪽 적재함에는 회색 로봇이 있다. 두 운전자는 성인 남성으로 보이지만 명명된 인물의 신원을 확인할 구체적 근거는 없다. 찰리는 고릴라형 흰 얼굴판보다는 둥근 회색 인간형 기계 머리에 가깝다. 별도로 들고 있는 B-200의 생존 흉부 부품은 식별되지 않으며, 보이는 구멍들은 찰리 자신의 몸통에 난 것으로 읽힌다. 후속 트럭들에는 샷에서 요구하지 않은 여러 사람이 추가되어 있다.",
        "hard_violations": [
         "후속 트럭들에 샷에서 허용하지 않은 여러 탑승자를 추가했다."
        ],
        "physics": "주요 트럭들은 바퀴로 도로에 지지되고 운전자들은 좌석에 앉아 조향 장치를 잡고 있다. 찰리의 하체는 적재함 안에 놓여 있어 떠 있는 자세는 아니다. 후속 탑승자들은 차체 뒤쪽에 서 있거나 기대며 하체 일부가 가려져 있다. 오른쪽 트럭의 주행 자세 자체는 가능하지만, 그 자세가 만드는 운동은 요구한 오른쪽 위 분기 주행과 반대다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "주요 트럭 두 대의 후방 사선과 좌우 분기 방향은 맞지만, 허용되지 않은 다수의 승객을 추가했고 오른쪽 후속 차량들이 반대 방향을 향한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "장소와 왼쪽 트럭의 방향은 맞지만, 오른쪽 트럭을 정면에서 보여 아래쪽으로 달리게 해 핵심 분리 동작을 어겼으며 추가 인물도 보인다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 트럭은 적재함 뒤쪽이 카메라를 향하고 운전석 쪽이 왼쪽 위 도로를 향한다. 오른쪽 주요 트럭도 뒤쪽을 보이며 앞부분이 오른쪽 위 갈래를 향해 두 대의 분리는 읽힌다. 그러나 오른쪽 먼 도로의 자동차와 오토바이들은 대부분 전면을 카메라 쪽으로 드러내며 아래쪽 갈림길로 접근한다. 따라서 수빈을 따라 오른쪽 위로 나가는 후속 행렬이 아니다. 승객들의 시선은 차 안과 옆쪽으로 흩어져 있으며, 로봇은 아래쪽의 들고 있는 부품을 보는 듯하다.",
        "built_space": "아래에서 들어와 좌우 위로 나뉘는 포장도로, 중앙 섬의 큰 낙엽수 한 그루, 그 왼쪽 목조 정자 한 채, 나무 앞 삼각 경고표지 한 개, 오른쪽 단선 철로가 보인다. 정자 난간과 기둥, 도로변 볼라드와 전신주, 뒤쪽 밭과 마을도 참고 장소에 대체로 대응한다. 높은 후방 시점은 허용된 구도지만, 주요 트럭들이 중경의 자연스러운 작은 비중보다 화면 아래쪽을 크게 차지한다.",
        "entities": "오래되고 녹슨 지붕 없는 주요 트럭 두 대와 추가 차량들이 보인다. 두 적재함에는 샷에서 요구하지 않은 다수의 성인 남녀 승객이 있으며, 여러 오토바이 운전자도 추가되었다. 이현우와 수빈의 신원은 이 거리와 가림으로 확인하기 어렵다. 오른쪽에는 고릴라형 머리의 대형 기계 찰리가 있고, 손 사이에 구멍과 손상이 있는 별도 흉부 외장처럼 보이는 물체가 있다. 다만 찰리의 얼굴은 지정된 흰 얼굴판과 점눈·선입보다 어두운 입체 고릴라 얼굴에 가깝다. 낮의 실제 농촌 재질 표현은 부합한다.",
        "hard_violations": [
         "샷에서 허용하지 않은 다수의 적재함 승객과 오토바이 탑승자를 추가했다."
        ],
        "physics": "트럭과 후속 차량은 타이어로 노면에 지지된다. 승객들은 적재함 바닥이나 가장자리에 앉거나 기대고 있고, 찰리의 하체도 적재함 안에 놓여 있다. 흉부 부품은 찰리의 양손과 몸 앞쪽에 받쳐져 있어 공중에 떠 있지 않다. 두 주요 트럭의 차체 각도는 각각의 도로로 선회할 수 있는 자세지만, 후속 차량들의 진행 방향은 행렬 설정과 충돌한다."
       },
       {
        "label": "A",
        "direction": "왼쪽 트럭은 뒤 적재함을 보이며 왼쪽 위 갈래로 향하고 운전자도 그쪽을 본다. 오른쪽 주요 트럭은 전면 그릴과 보닛을 카메라에 보이며 왼쪽 아래의 공통 진입로 쪽으로 향한다. 운전자 역시 그 전방을 보고 있어 오른쪽 위로 갈라져 나가는 동작이 아니다. 뒤의 차량들도 전면을 보이며 아래쪽으로 접근한다. 찰리의 얼굴도 대체로 카메라 쪽을 향한다.",
        "built_space": "중앙의 큰 나무 한 그루, 왼쪽 정자 한 채, 삼각 경고표지 한 개, 오른쪽 단선 철로와 전신주·볼라드, 배후 농경지와 마을이 유지된다. 공통 도로가 아래에서 좌우 위로 갈라지는 장소 구조는 맞고, 정자의 바닥과 난간도 자연스럽다. 다만 오른쪽 주요 트럭이 화면 오른쪽 아래에 크게 놓여 중경 후방 사선이라는 지정 배치를 벗어난다.",
        "entities": "주요 차량은 녹슨 지붕 없는 트럭 두 대이며 왼쪽 적재함은 비어 있고 오른쪽 적재함에는 회색 로봇이 있다. 두 운전자는 성인 남성으로 보이지만 명명된 인물의 신원을 확인할 구체적 근거는 없다. 찰리는 고릴라형 흰 얼굴판보다는 둥근 회색 인간형 기계 머리에 가깝다. 별도로 들고 있는 B-200의 생존 흉부 부품은 식별되지 않으며, 보이는 구멍들은 찰리 자신의 몸통에 난 것으로 읽힌다. 후속 트럭들에는 샷에서 요구하지 않은 여러 사람이 추가되어 있다.",
        "hard_violations": [
         "후속 트럭들에 샷에서 허용하지 않은 여러 탑승자를 추가했다."
        ],
        "physics": "주요 트럭들은 바퀴로 도로에 지지되고 운전자들은 좌석에 앉아 조향 장치를 잡고 있다. 찰리의 하체는 적재함 안에 놓여 있어 떠 있는 자세는 아니다. 후속 탑승자들은 차체 뒤쪽에 서 있거나 기대며 하체 일부가 가려져 있다. 오른쪽 트럭의 주행 자세 자체는 가능하지만, 그 자세가 만드는 운동은 요구한 오른쪽 위 분기 주행과 반대다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.095,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.845,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 오른쪽 트럭과 후발 차량의 진행 방향 및 시점 위반 (뒤에서 멀어지는 구도가 아님)",
     "[gpt-high] 후속 트럭들에 샷에서 허용하지 않은 여러 탑승자를 추가했다."
    ],
    "B": [
     "[gpt-high] 샷에서 허용하지 않은 다수의 적재함 승객과 오토바이 탑승자를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 845
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "멀어지며 갈라지는 두 트럭의 구도, 난민 그룹, 찰리의 외형 및 파손된 부품, 완벽한 배경 구현까지 모든 프롬프트 요구사항을 정확히 충족함.  ★위반: [gpt-high] 샷에서 허용하지 않은 다수의 적재함 승객과 오토바이 탑승자를 추가했다."
   },
   {
    "label": "A",
    "score": 845,
    "verdict_ko": "오른쪽 트럭과 후발 차량이 뒤에서 바라보는 구도가 아니라 카메라를 향해 정면으로 다가오도록 묘사되어 치명적인 연출 위반이 발생함.  ★위반: [gemini-pro] 오른쪽 트럭과 후발 차량의 진행 방향 및 시점 위반 (뒤에서 멀어지는 구도가 아님) / [gpt-high] 후속 트럭들에 샷에서 허용하지 않은 여러 탑승자를 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_food_transport_fork_sel.png",
    "asset_id": "3353276e-8593-44a0-8507-e9efa6680c2a",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace1a-8648-77b9-80fc-84db1da16bf6",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S68sh10": {
  "input_fingerprint": "788848714f13285e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈의 안광이 꺼진 채 고개를 푹 숙이고 있는 찰리의 평온한 자세.\n\nLOCATION (lock): In the open rear cargo bed of an old truck traveling away from the village in daylight. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Truck bed (Carrying the seated 찰리 during travel) — A partial interior edge is visible obliquely behind his shoulder; used as Soft, narrow continuity cue retaining the road setting before the dissolve.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly controlled daylight keeps his unlit eyes and bowed facial contours readable without introducing any imagined-field color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The roofless truck continues along the road, with Charlie seated in its cargo bed holding B-200's chest component and a baby bird perched on his finger. Charlie's accumulated body damage remains as he closes his eyes and enters the dream.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈의 안광이 꺼진 채 고개를 푹 숙이고 있는 찰리의 평온한 자세.\n\nLOCATION (lock): In the open rear cargo bed of an old truck traveling away from the village in daylight. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Truck bed (Carrying the seated 찰리 during travel) — A partial interior edge is visible obliquely behind his shoulder; used as Soft, narrow continuity cue retaining the road setting before the dissolve.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly controlled daylight keeps his unlit eyes and bowed facial contours readable without introducing any imagined-field color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The roofless truck continues along the road, with Charlie seated in its cargo bed holding B-200's chest component and a baby bird perched on his finger. Charlie's accumulated body damage remains as he closes his eyes and enters the dream.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈의 안광이 꺼진 채 고개를 푹 숙이고 있는 찰리의 평온한 자세.\n\nLOCATION (lock): In the open rear cargo bed of an old truck traveling away from the village in daylight. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Truck bed (Carrying the seated 찰리 during travel) — A partial interior edge is visible obliquely behind his shoulder; used as Soft, narrow continuity cue retaining the road setting before the dissolve.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly controlled daylight keeps his unlit eyes and bowed facial contours readable without introducing any imagined-field color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The roofless truck continues along the road, with Charlie seated in its cargo bed holding B-200's chest component and a baby bird perched on his finger. Charlie's accumulated body damage remains as he closes his eyes and enters the dream.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "고개를 숙여 손 위에 있는 아기 새를 향해 시선이 향함.",
    "built_space": "주행 중인 트럭의 짐칸 내부에 앉아 있으며, 등 뒤로 트럭의 측면 패널과 배경의 도로가 보임.",
    "entities": "찰리의 지정된 디자인(고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입)이 정확히 일치함. 손에 상자(가슴 부품으로 추정)와 아기 새가 있음.",
    "hard_violations": [
     "[gpt-high] 장소와 소품 지시에 없는 별도의 천 자루류 화물을 적재함 왼쪽에 추가했다."
    ],
    "physics": "트럭 바닥에 안정적으로 앉아 있으며, 손에 상자와 새를 무리 없이 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "고개를 깊게 숙여 품에 안은 부품 쪽을 향함.",
    "built_space": "주행 중인 트럭의 짐칸 내부에 앉아 있으며, 뒤로 트럭 벽면과 도로가 보임.",
    "entities": "기계 인간이 등장하나, 지문이 요구한 찰리의 외형(고릴라형 머리 등)이 아닌 검은 바이저 형태의 얼굴을 하고 있음. 기계 부품과 아기 새가 있음.",
    "hard_violations": [
     "[gemini-pro] 지정된 인물(찰리)의 형태학적 묘사 완전 불일치",
     "[gemini-pro] 안광이 꺼져야 한다는 지문과 달리 눈(바이저)에 빛이 남"
    ],
    "physics": "트럭 짐칸에 앉아 기계 부품과 새를 지탱하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 찰리의 외형(고릴라형 기계 머리, 점 형태의 눈 등)과 안광이 꺼진 상태를 훌륭하게 구현했으나, 들고 있는 부품이 상자처럼 보임."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 외형 묘사를 완전히 무시한 다른 형태의 로봇을 생성했으며, 눈의 안광이 꺼져야 한다는 지문도 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개를 숙여 손 위에 있는 아기 새를 향해 시선이 향함.",
        "built_space": "주행 중인 트럭의 짐칸 내부에 앉아 있으며, 등 뒤로 트럭의 측면 패널과 배경의 도로가 보임.",
        "entities": "찰리의 지정된 디자인(고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입)이 정확히 일치함. 손에 상자(가슴 부품으로 추정)와 아기 새가 있음.",
        "hard_violations": [],
        "physics": "트럭 바닥에 안정적으로 앉아 있으며, 손에 상자와 새를 무리 없이 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "고개를 깊게 숙여 품에 안은 부품 쪽을 향함.",
        "built_space": "주행 중인 트럭의 짐칸 내부에 앉아 있으며, 뒤로 트럭 벽면과 도로가 보임.",
        "entities": "기계 인간이 등장하나, 지문이 요구한 찰리의 외형(고릴라형 머리 등)이 아닌 검은 바이저 형태의 얼굴을 하고 있음. 기계 부품과 아기 새가 있음.",
        "hard_violations": [
         "지정된 인물(찰리)의 형태학적 묘사 완전 불일치",
         "안광이 꺼져야 한다는 지문과 달리 눈(바이저)에 빛이 남"
        ],
        "physics": "트럭 짐칸에 앉아 기계 부품과 새를 지탱하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 찰리의 외형(고릴라형 기계 머리, 점 형태의 눈 등)과 안광이 꺼진 상태를 훌륭하게 구현했으나, 들고 있는 부품이 상자처럼 보임."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 외형 묘사를 완전히 무시한 다른 형태의 로봇을 생성했으며, 눈의 안광이 꺼져야 한다는 지문도 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "고개를 숙여 손 위에 있는 아기 새를 향해 시선이 향함.",
        "built_space": "주행 중인 트럭의 짐칸 내부에 앉아 있으며, 등 뒤로 트럭의 측면 패널과 배경의 도로가 보임.",
        "entities": "찰리의 지정된 디자인(고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입)이 정확히 일치함. 손에 상자(가슴 부품으로 추정)와 아기 새가 있음.",
        "hard_violations": [],
        "physics": "트럭 바닥에 안정적으로 앉아 있으며, 손에 상자와 새를 무리 없이 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "고개를 깊게 숙여 품에 안은 부품 쪽을 향함.",
        "built_space": "주행 중인 트럭의 짐칸 내부에 앉아 있으며, 뒤로 트럭 벽면과 도로가 보임.",
        "entities": "기계 인간이 등장하나, 지문이 요구한 찰리의 외형(고릴라형 머리 등)이 아닌 검은 바이저 형태의 얼굴을 하고 있음. 기계 부품과 아기 새가 있음.",
        "hard_violations": [
         "지정된 인물(찰리)의 형태학적 묘사 완전 불일치",
         "안광이 꺼져야 한다는 지문과 달리 눈(바이저)에 빛이 남"
        ],
        "physics": "트럭 짐칸에 앉아 기계 부품과 새를 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "추가 화물 없이 이동 중인 적재함의 착석 상태는 구현했지만, 켜진 눈과 검은 얼굴판이 핵심 묘사에 어긋나며 클로즈업보다 지나치게 넓다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "고릴라형 흰 얼굴판과 꺼진 점형 눈은 더 정확하지만, 지시되지 않은 화물을 추가했으며 적재함과 무릎까지 드러낸 구도도 클로즈업 요구를 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 머리는 아래쪽으로 기울어 얼굴이 품에 든 부품과 손가락 위 새를 향한다. 두 눈에는 밝은 발광점이 보여 ‘안광이 꺼진’ 상태와 반대다. 뒤로 길게 이어지는 도로와 먼 마을은 마을을 떠나는 이동 설정에 부합한다.",
        "built_space": "지붕 없는 낡은 금속 적재함 안에 찰리 한 명이 앉아 있다. 화면 왼쪽의 측벽 하나와 오른쪽 뒤의 가로 벽 하나, 하단 바닥 일부가 보이며 중복된 설비는 없다. 그러나 벽면과 도로가 넓게 노출되어, 어깨 뒤에 좁고 부드러운 적재함 가장자리만 남기는 클로즈업과 다르다.",
        "entities": "찰리 한 명, 작은 새 한 마리, 두 손으로 안은 기계 흉부 부품 하나가 보인다. 외장과 옷에는 긁힘과 손상이 있다. 머리는 기계식이지만 둥근 흰 두개 외장과 검은 안면창으로 표현되어, 각진 고릴라형 머리·흰 얼굴판·선형 입과는 차이가 크다. 흉부 부품이 B-200의 것인지는 외관만으로 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몸은 적재함 안에 앉아 벽 가까이에 기대어 있으며, 화면 아래로 잘린 하체가 공중에 뜬 징후는 없다. 흉부 부품은 양손과 몸통 사이에 받쳐져 있다. 새의 발은 내민 기계 손가락에 닿아 있다. 고개를 숙인 목과 팔의 연결도 물리적으로 성립한다."
       },
       {
        "label": "B",
        "direction": "찰리는 고개를 앞으로 숙이고 얼굴을 손가락 위 새와 품의 부품 쪽으로 향한다. 두 원형 눈은 검게 보이며 빛나지 않는다. 다만 눈을 감았다기보다 열린 비발광 렌즈처럼 보인다. 새는 찰리 쪽을 향한다. 도로는 오른쪽에 보이지만 마을에서 멀어지는 진행 방향은 확정하기 어렵다.",
        "built_space": "찰리 뒤에 운전실 후면과 창 하나, 왼쪽 적재함 측벽, 오른쪽 세로 기둥, 하단 금속 가장자리와 바닥이 보인다. 착석 위치와 구조물 사이에는 명백한 공간 모순이 없다. 그러나 운전실과 적재함 전체가 큰 비중을 차지하고 무릎까지 포함되어, 좁은 배경 단서만 두는 클로즈업 요구와 다르다. 왼쪽 바닥에는 별도의 천 화물 더미도 놓여 있다.",
        "entities": "찰리 한 명과 작은 새 한 마리가 보인다. 각진 기계 머리, 고릴라형 코와 주둥이, 흰 얼굴판, 검은 점형 눈, 선형 입은 지정된 정체성에 잘 맞는다. 손과 의복에 누적 마모가 있다. 안고 있는 물체는 직사각형 금속 케이스에 가까워 흉부 부품인지 불분명하다. 왼쪽의 천 자루류는 지시되지 않은 추가 물체다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장소와 소품 지시에 없는 별도의 천 자루류 화물을 적재함 왼쪽에 추가했다."
        ],
        "physics": "굽힌 무릎과 착석한 몸통은 적재함 바닥에 지지되는 자세로 읽힌다. 금속 물체는 무릎과 몸통 앞에 놓이고 양손으로 지지된다. 새는 손가락 위에 발을 딛고 있으며, 추가 화물도 바닥에 놓여 있다. 지지 없이 떠 있는 물체나 불가능한 관절 연결은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "추가 화물 없이 이동 중인 적재함의 착석 상태는 구현했지만, 켜진 눈과 검은 얼굴판이 핵심 묘사에 어긋나며 클로즈업보다 지나치게 넓다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "고릴라형 흰 얼굴판과 꺼진 점형 눈은 더 정확하지만, 지시되지 않은 화물을 추가했으며 적재함과 무릎까지 드러낸 구도도 클로즈업 요구를 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 머리는 아래쪽으로 기울어 얼굴이 품에 든 부품과 손가락 위 새를 향한다. 두 눈에는 밝은 발광점이 보여 ‘안광이 꺼진’ 상태와 반대다. 뒤로 길게 이어지는 도로와 먼 마을은 마을을 떠나는 이동 설정에 부합한다.",
        "built_space": "지붕 없는 낡은 금속 적재함 안에 찰리 한 명이 앉아 있다. 화면 왼쪽의 측벽 하나와 오른쪽 뒤의 가로 벽 하나, 하단 바닥 일부가 보이며 중복된 설비는 없다. 그러나 벽면과 도로가 넓게 노출되어, 어깨 뒤에 좁고 부드러운 적재함 가장자리만 남기는 클로즈업과 다르다.",
        "entities": "찰리 한 명, 작은 새 한 마리, 두 손으로 안은 기계 흉부 부품 하나가 보인다. 외장과 옷에는 긁힘과 손상이 있다. 머리는 기계식이지만 둥근 흰 두개 외장과 검은 안면창으로 표현되어, 각진 고릴라형 머리·흰 얼굴판·선형 입과는 차이가 크다. 흉부 부품이 B-200의 것인지는 외관만으로 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몸은 적재함 안에 앉아 벽 가까이에 기대어 있으며, 화면 아래로 잘린 하체가 공중에 뜬 징후는 없다. 흉부 부품은 양손과 몸통 사이에 받쳐져 있다. 새의 발은 내민 기계 손가락에 닿아 있다. 고개를 숙인 목과 팔의 연결도 물리적으로 성립한다."
       },
       {
        "label": "A",
        "direction": "찰리는 고개를 앞으로 숙이고 얼굴을 손가락 위 새와 품의 부품 쪽으로 향한다. 두 원형 눈은 검게 보이며 빛나지 않는다. 다만 눈을 감았다기보다 열린 비발광 렌즈처럼 보인다. 새는 찰리 쪽을 향한다. 도로는 오른쪽에 보이지만 마을에서 멀어지는 진행 방향은 확정하기 어렵다.",
        "built_space": "찰리 뒤에 운전실 후면과 창 하나, 왼쪽 적재함 측벽, 오른쪽 세로 기둥, 하단 금속 가장자리와 바닥이 보인다. 착석 위치와 구조물 사이에는 명백한 공간 모순이 없다. 그러나 운전실과 적재함 전체가 큰 비중을 차지하고 무릎까지 포함되어, 좁은 배경 단서만 두는 클로즈업 요구와 다르다. 왼쪽 바닥에는 별도의 천 화물 더미도 놓여 있다.",
        "entities": "찰리 한 명과 작은 새 한 마리가 보인다. 각진 기계 머리, 고릴라형 코와 주둥이, 흰 얼굴판, 검은 점형 눈, 선형 입은 지정된 정체성에 잘 맞는다. 손과 의복에 누적 마모가 있다. 안고 있는 물체는 직사각형 금속 케이스에 가까워 흉부 부품인지 불분명하다. 왼쪽의 천 자루류는 지시되지 않은 추가 물체다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장소와 소품 지시에 없는 별도의 천 자루류 화물을 적재함 왼쪽에 추가했다."
        ],
        "physics": "굽힌 무릎과 착석한 몸통은 적재함 바닥에 지지되는 자세로 읽힌다. 금속 물체는 무릎과 몸통 앞에 놓이고 양손으로 지지된다. 새는 손가락 위에 발을 딛고 있으며, 추가 화물도 바닥에 놓여 있다. 지지 없이 떠 있는 물체나 불가능한 관절 연결은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.179
   },
   "violations": {
    "B": [
     "[gemini-pro] 지정된 인물(찰리)의 형태학적 묘사 완전 불일치",
     "[gemini-pro] 안광이 꺼져야 한다는 지문과 달리 눈(바이저)에 빛이 남"
    ],
    "A": [
     "[gpt-high] 장소와 소품 지시에 없는 별도의 천 자루류 화물을 적재함 왼쪽에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "지정된 찰리의 외형(고릴라형 기계 머리, 점 형태의 눈 등)과 안광이 꺼진 상태를 훌륭하게 구현했으나, 들고 있는 부품이 상자처럼 보임.  ★위반: [gpt-high] 장소와 소품 지시에 없는 별도의 천 자루류 화물을 적재함 왼쪽에 추가했다."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "찰리의 외형 묘사를 완전히 무시한 다른 형태의 로봇을 생성했으며, 눈의 안광이 꺼져야 한다는 지문도 위반함.  ★위반: [gemini-pro] 지정된 인물(찰리)의 형태학적 묘사 완전 불일치 / [gemini-pro] 안광이 꺼져야 한다는 지문과 달리 눈(바이저)에 빛이 남"
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace1e-e14f-7f97-af7a-e0aa4317d9a9",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "era_fail::b3e5ca9e4df6bba2": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "era_fail::60132f46c1f2b04a": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S69sh4": {
  "input_fingerprint": "115d72d5623f0bdf",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 그늘막 아래 잠든 아이들을 보며 찰리가 입가 금속 부품을 위로 올린 채 환하게 미소 짓고 있는 얼굴.\n\nLOCATION (lock): Beside the sleeping children in the parked truck's open cargo bed, under a shade cloth beside a rocky outcrop. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Shade (Repositioned to shield the sleeping companions) — Only an edge of its sheltering surface enters the upper margin; used as Retains evidence of the caring action without obscuring the smile; Truck-bed edge (Partially visible behind 찰리) — An oblique inner edge remains low in the background; used as Maintains location continuity with the sleeping group.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight moderated by the repositioned shade preserves gentle facial contrast and readable metal contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the parked truck bed, nearby rock shelter, and daytime surface colors. EXCLUDE direct sunlight over the sleeping area, since the shade cover has been moved to block it.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's cargo bed with his head gathered together with Amber's and Raul's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's cargo bed with her head gathered together with Hyunwoo's and Raul's beneath the shade. Her exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's cargo bed with his head gathered together with Hyunwoo's and Amber's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck remains parked beside the rocks. The repositioned shade cover now blocks the sunlight over the cargo-bed sleeping area. 찰리: He stands beside the shaded sleeping area with a gentle smile. His battle-damaged body and exposed chest opening remain, and he retains the salvaged chest component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 그늘막 아래 잠든 아이들을 보며 찰리가 입가 금속 부품을 위로 올린 채 환하게 미소 짓고 있는 얼굴.\n\nLOCATION (lock): Beside the sleeping children in the parked truck's open cargo bed, under a shade cloth beside a rocky outcrop. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Shade (Repositioned to shield the sleeping companions) — Only an edge of its sheltering surface enters the upper margin; used as Retains evidence of the caring action without obscuring the smile; Truck-bed edge (Partially visible behind 찰리) — An oblique inner edge remains low in the background; used as Maintains location continuity with the sleeping group.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight moderated by the repositioned shade preserves gentle facial contrast and readable metal contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the parked truck bed, nearby rock shelter, and daytime surface colors. EXCLUDE direct sunlight over the sleeping area, since the shade cover has been moved to block it.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's cargo bed with his head gathered together with Amber's and Raul's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's cargo bed with her head gathered together with Hyunwoo's and Raul's beneath the shade. Her exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's cargo bed with his head gathered together with Hyunwoo's and Amber's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck remains parked beside the rocks. The repositioned shade cover now blocks the sunlight over the cargo-bed sleeping area. 찰리: He stands beside the shaded sleeping area with a gentle smile. His battle-damaged body and exposed chest opening remain, and he retains the salvaged chest component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 그늘막 아래 잠든 아이들을 보며 찰리가 입가 금속 부품을 위로 올린 채 환하게 미소 짓고 있는 얼굴.\n\nLOCATION (lock): Beside the sleeping children in the parked truck's open cargo bed, under a shade cloth beside a rocky outcrop. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Shade (Repositioned to shield the sleeping companions) — Only an edge of its sheltering surface enters the upper margin; used as Retains evidence of the caring action without obscuring the smile; Truck-bed edge (Partially visible behind 찰리) — An oblique inner edge remains low in the background; used as Maintains location continuity with the sleeping group.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight moderated by the repositioned shade preserves gentle facial contrast and readable metal contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the parked truck bed, nearby rock shelter, and daytime surface colors. EXCLUDE direct sunlight over the sleeping area, since the shade cover has been moved to block it.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's cargo bed with his head gathered together with Amber's and Raul's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's cargo bed with her head gathered together with Hyunwoo's and Raul's beneath the shade. Her exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's cargo bed with his head gathered together with Hyunwoo's and Amber's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck remains parked beside the rocks. The repositioned shade cover now blocks the sunlight over the cargo-bed sleeping area. 찰리: He stands beside the shaded sleeping area with a gentle smile. His battle-damaged body and exposed chest opening remain, and he retains the salvaged chest component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 시선이 화면 밖 아래쪽(잠든 아이들이 있는 방향)을 향하고 있음.",
    "built_space": "상단 가장자리에 그늘막 모서리가 걸쳐 있고, 찰리 뒤로 트럭 적재함의 비스듬한 선이 낮게 깔려 있으며 배경에 바위가 보임.",
    "entities": "찰리는 지시된 대로 가슴 장갑이 열려 손상된 내부 부품이 드러난 상태임. 입가의 금속 부품이 들린 묘사 대신 치아가 보이는 형태로 미소 짓고 있음.",
    "hard_violations": [],
    "physics": "찰리의 서 있는 자세와 무게 중심이 안정적으로 유지됨."
   },
   {
    "label": "B",
    "direction": "찰리의 시선이 화면 좌측 하단에 있는 아이들을 향하고 있음.",
    "built_space": "상단에 그늘막이 보이고 뒤로 트럭 가장자리와 배경의 바위가 적절히 배치됨.",
    "entities": "찰리가 손상된 가슴 상태 대신 레퍼런스의 비옷과 밀짚모자를 착용하고 있음. 프레임 좌측 하단에 지시문이 배제하도록 한 이전 샷의 아이들이 등장함.",
    "hard_violations": [
     "[gemini-pro] 지시문에서 명시적으로 프레임 배제를 요구한 이전 샷의 인물(잠든 아이들)이 화면에 포함됨",
     "[gpt-high] 이전 숏의 인물들을 이번 화면에 등장시키지 말라는 명시적 제한에도 잠든 아이 세 명의 얼굴과 몸을 배경에 재현했습니다."
    ],
    "physics": "찰리의 자세 및 사물들의 중력 작용에 어색함이 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 프레이밍과 찰리의 손상된 가슴 상태를 잘 구현했으며, 명시된 대로 아이들을 프레임에서 정확히 배제했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "등장해서는 안 되는 이전 샷의 아이들을 프레임에 포함시켰고, 지시된 가슴 파손 상태 대신 레퍼런스의 옷을 입고 있어 지시를 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선이 화면 밖 아래쪽(잠든 아이들이 있는 방향)을 향하고 있음.",
        "built_space": "상단 가장자리에 그늘막 모서리가 걸쳐 있고, 찰리 뒤로 트럭 적재함의 비스듬한 선이 낮게 깔려 있으며 배경에 바위가 보임.",
        "entities": "찰리는 지시된 대로 가슴 장갑이 열려 손상된 내부 부품이 드러난 상태임. 입가의 금속 부품이 들린 묘사 대신 치아가 보이는 형태로 미소 짓고 있음.",
        "hard_violations": [],
        "physics": "찰리의 서 있는 자세와 무게 중심이 안정적으로 유지됨."
       },
       {
        "label": "B",
        "direction": "찰리의 시선이 화면 좌측 하단에 있는 아이들을 향하고 있음.",
        "built_space": "상단에 그늘막이 보이고 뒤로 트럭 가장자리와 배경의 바위가 적절히 배치됨.",
        "entities": "찰리가 손상된 가슴 상태 대신 레퍼런스의 비옷과 밀짚모자를 착용하고 있음. 프레임 좌측 하단에 지시문이 배제하도록 한 이전 샷의 아이들이 등장함.",
        "hard_violations": [
         "지시문에서 명시적으로 프레임 배제를 요구한 이전 샷의 인물(잠든 아이들)이 화면에 포함됨"
        ],
        "physics": "찰리의 자세 및 사물들의 중력 작용에 어색함이 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 프레이밍과 찰리의 손상된 가슴 상태를 잘 구현했으며, 명시된 대로 아이들을 프레임에서 정확히 배제했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "등장해서는 안 되는 이전 샷의 아이들을 프레임에 포함시켰고, 지시된 가슴 파손 상태 대신 레퍼런스의 옷을 입고 있어 지시를 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선이 화면 밖 아래쪽(잠든 아이들이 있는 방향)을 향하고 있음.",
        "built_space": "상단 가장자리에 그늘막 모서리가 걸쳐 있고, 찰리 뒤로 트럭 적재함의 비스듬한 선이 낮게 깔려 있으며 배경에 바위가 보임.",
        "entities": "찰리는 지시된 대로 가슴 장갑이 열려 손상된 내부 부품이 드러난 상태임. 입가의 금속 부품이 들린 묘사 대신 치아가 보이는 형태로 미소 짓고 있음.",
        "hard_violations": [],
        "physics": "찰리의 서 있는 자세와 무게 중심이 안정적으로 유지됨."
       },
       {
        "label": "B",
        "direction": "찰리의 시선이 화면 좌측 하단에 있는 아이들을 향하고 있음.",
        "built_space": "상단에 그늘막이 보이고 뒤로 트럭 가장자리와 배경의 바위가 적절히 배치됨.",
        "entities": "찰리가 손상된 가슴 상태 대신 레퍼런스의 비옷과 밀짚모자를 착용하고 있음. 프레임 좌측 하단에 지시문이 배제하도록 한 이전 샷의 아이들이 등장함.",
        "hard_violations": [
         "지시문에서 명시적으로 프레임 배제를 요구한 이전 샷의 인물(잠든 아이들)이 화면에 포함됨"
        ],
        "physics": "찰리의 자세 및 사물들의 중력 작용에 어색함이 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "얼굴 클로즈업과 의상은 더 가깝지만, 이번 화면에서 제외하라고 명시한 이전 숏의 아이들 세 명을 다시 등장시켰고 시선도 그들에게 향하지 않습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "찰리만 등장하고 장소와 손상된 가슴은 유지하지만, 얼굴 클로즈업 대신 상반신으로 넓어졌으며 모자·의상이 사라지고 아이들을 보는 방향도 확인되지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴은 화면 왼쪽 앞을 향해 조금 숙여져 있습니다. 잠든 세 아이는 찰리의 왼쪽 뒤에 있어, 보이는 눈과 얼굴의 방향이 실제 아이들의 위치로 향하지 않습니다. 입선의 양 끝은 올라가 미소로 읽힙니다.",
        "built_space": "회색 적재함 측판 하나가 뒤를 가로지르고 오른쪽에는 차량의 창 달린 부분이 보입니다. 상단에는 검은 그늘막 한 장, 바깥에는 바위들이 있습니다. 장소의 재질과 낮빛은 이어지지만, 그늘막이 상단 가장자리보다 넓게 들어오고 적재함과 아이들도 상당 부분 노출됩니다.",
        "entities": "찰리의 각진 기계 외장, 흰 고릴라형 얼굴판, 점눈, 선형 입, 밀짚모자와 다색 외투는 참조와 대체로 일치합니다. 그러나 배경에 검은 머리 남자아이, 금발 여자아이, 어두운 피부의 땋은 머리 아이까지 세 명이 나타나 이전 숏의 인물들을 제외하라는 지시를 어깁니다. 가슴 아래와 손에 든 부품은 프레임 밖이므로 판단하지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "이전 숏의 인물들을 이번 화면에 등장시키지 말라는 명시적 제한에도 잠든 아이 세 명의 얼굴과 몸을 배경에 재현했습니다."
        ],
        "physics": "찰리의 머리는 기계 목과 몸통에 연결되어 있고 모자는 머리에 얹혀 있습니다. 아이들의 머리는 기울어진 몸이나 서로 맞댄 부분에 기대어 보이며, 하단이 잘려 정확한 바닥 접촉은 확인되지 않습니다. 명백하게 공중에 뜬 신체나 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "찰리는 화면 오른쪽 앞을 향해 약간 아래를 보고 있습니다. 적재함은 뒤에 있고 아이들은 화면 밖이므로, 그 시선이 잠든 아이들에게 도달한다고 확인할 수 없습니다. 입 양끝이 위로 올라간 넓은 웃음은 분명합니다.",
        "built_space": "흰 적재함의 가까운 측판과 뒤쪽 측판이 각각 보이고, 좌우에 그늘막 지지대가 하나씩 있습니다. 검은 그늘막 한 장이 상단을 덮고 바위가 뒤를 채웁니다. 적재함 안 오른쪽의 녹색 상자와 밝은 상자는 이전 장소와 연결됩니다. 찰리는 가까운 측판 바깥에 서 있지만, 얼굴뿐 아니라 가슴과 양팔까지 보여 요구한 클로즈업보다 훨씬 넓으며 적재함도 낮은 배경의 일부 이상을 차지합니다.",
        "entities": "등장인물은 찰리 하나입니다. 흰 고릴라형 얼굴판과 점눈은 유지되지만 머리 외장이 더 둥글고, 선형 입에 치아 같은 구조가 추가되었습니다. 참조의 밀짚모자와 다색 외투는 보이는 머리와 상반신에서 사라졌습니다. 손상된 금속 몸체와 열린 가슴 내부는 나타납니다. 손과 회수 부품의 휴대 상태는 프레임 밖이라 평가하지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "머리와 팔은 목·어깨 관절로 몸통에 연결되고, 열린 가슴 덮개는 가장자리 연결부에 붙어 있습니다. 그늘막은 양쪽 지지대에 걸려 처지고 상자들은 적재함 안에 놓여 있습니다. 하체는 잘렸지만 상체에 부유를 의심할 자세는 없으며, 지지 없이 떠 있는 물체도 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "얼굴 클로즈업과 의상은 더 가깝지만, 이번 화면에서 제외하라고 명시한 이전 숏의 아이들 세 명을 다시 등장시켰고 시선도 그들에게 향하지 않습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "찰리만 등장하고 장소와 손상된 가슴은 유지하지만, 얼굴 클로즈업 대신 상반신으로 넓어졌으며 모자·의상이 사라지고 아이들을 보는 방향도 확인되지 않습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴은 화면 왼쪽 앞을 향해 조금 숙여져 있습니다. 잠든 세 아이는 찰리의 왼쪽 뒤에 있어, 보이는 눈과 얼굴의 방향이 실제 아이들의 위치로 향하지 않습니다. 입선의 양 끝은 올라가 미소로 읽힙니다.",
        "built_space": "회색 적재함 측판 하나가 뒤를 가로지르고 오른쪽에는 차량의 창 달린 부분이 보입니다. 상단에는 검은 그늘막 한 장, 바깥에는 바위들이 있습니다. 장소의 재질과 낮빛은 이어지지만, 그늘막이 상단 가장자리보다 넓게 들어오고 적재함과 아이들도 상당 부분 노출됩니다.",
        "entities": "찰리의 각진 기계 외장, 흰 고릴라형 얼굴판, 점눈, 선형 입, 밀짚모자와 다색 외투는 참조와 대체로 일치합니다. 그러나 배경에 검은 머리 남자아이, 금발 여자아이, 어두운 피부의 땋은 머리 아이까지 세 명이 나타나 이전 숏의 인물들을 제외하라는 지시를 어깁니다. 가슴 아래와 손에 든 부품은 프레임 밖이므로 판단하지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "이전 숏의 인물들을 이번 화면에 등장시키지 말라는 명시적 제한에도 잠든 아이 세 명의 얼굴과 몸을 배경에 재현했습니다."
        ],
        "physics": "찰리의 머리는 기계 목과 몸통에 연결되어 있고 모자는 머리에 얹혀 있습니다. 아이들의 머리는 기울어진 몸이나 서로 맞댄 부분에 기대어 보이며, 하단이 잘려 정확한 바닥 접촉은 확인되지 않습니다. 명백하게 공중에 뜬 신체나 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "찰리는 화면 오른쪽 앞을 향해 약간 아래를 보고 있습니다. 적재함은 뒤에 있고 아이들은 화면 밖이므로, 그 시선이 잠든 아이들에게 도달한다고 확인할 수 없습니다. 입 양끝이 위로 올라간 넓은 웃음은 분명합니다.",
        "built_space": "흰 적재함의 가까운 측판과 뒤쪽 측판이 각각 보이고, 좌우에 그늘막 지지대가 하나씩 있습니다. 검은 그늘막 한 장이 상단을 덮고 바위가 뒤를 채웁니다. 적재함 안 오른쪽의 녹색 상자와 밝은 상자는 이전 장소와 연결됩니다. 찰리는 가까운 측판 바깥에 서 있지만, 얼굴뿐 아니라 가슴과 양팔까지 보여 요구한 클로즈업보다 훨씬 넓으며 적재함도 낮은 배경의 일부 이상을 차지합니다.",
        "entities": "등장인물은 찰리 하나입니다. 흰 고릴라형 얼굴판과 점눈은 유지되지만 머리 외장이 더 둥글고, 선형 입에 치아 같은 구조가 추가되었습니다. 참조의 밀짚모자와 다색 외투는 보이는 머리와 상반신에서 사라졌습니다. 손상된 금속 몸체와 열린 가슴 내부는 나타납니다. 손과 회수 부품의 휴대 상태는 프레임 밖이라 평가하지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "머리와 팔은 목·어깨 관절로 몸통에 연결되고, 열린 가슴 덮개는 가장자리 연결부에 붙어 있습니다. 그늘막은 양쪽 지지대에 걸려 처지고 상자들은 적재함 안에 놓여 있습니다. 하체는 잘렸지만 상체에 부유를 의심할 자세는 없으며, 지지 없이 떠 있는 물체도 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.829
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.579
   },
   "violations": {
    "B": [
     "[gemini-pro] 지시문에서 명시적으로 프레임 배제를 요구한 이전 샷의 인물(잠든 아이들)이 화면에 포함됨",
     "[gpt-high] 이전 숏의 인물들을 이번 화면에 등장시키지 말라는 명시적 제한에도 잠든 아이 세 명의 얼굴과 몸을 배경에 재현했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 579
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 클로즈업 프레이밍과 찰리의 손상된 가슴 상태를 잘 구현했으며, 명시된 대로 아이들을 프레임에서 정확히 배제했습니다."
   },
   {
    "label": "B",
    "score": 579,
    "verdict_ko": "등장해서는 안 되는 이전 샷의 아이들을 프레임에 포함시켰고, 지시된 가슴 파손 상태 대신 레퍼런스의 옷을 입고 있어 지시를 위반했습니다.  ★위반: [gemini-pro] 지시문에서 명시적으로 프레임 배제를 요구한 이전 샷의 인물(잠든 아이들)이 화면에 포함됨 / [gpt-high] 이전 숏의 인물들을 이번 화면에 등장시키지 말라는 명시적 제한에도 잠든 아이 세 명의 얼굴과 몸을 배경에 재현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S69sh2_sel.png",
    "asset_id": "1c4b19cd-b4e0-4483-a834-ffc684ee2786",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:885644>",
    "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace33-89bf-7748-8eb6-8dd76557f05e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S69sh2"
  }
 },
 "era_fail::c136306f7326421e": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S70sh10": {
  "input_fingerprint": "7737833c8f08972a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 깁스를 한 박철진의 붉게 충혈된 남은 한쪽 눈에서 분노의 눈물이 뺨에 맺힌 얼굴 클로즈업.\n\nLOCATION (lock): At the injured commander's bed inside a military hospital room, under ordinary nighttime clinical lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Treatment covering around the injured eye (The injured eye remains covered during treatment) — Its edge is visible beside the uncovered eye from the oblique bedside angle; used as Preserves the physical context of his injury within the facial close-up; Hospital bed (Supporting the reclined patient) — A small area beside the patient's head remains visible; used as Provides minimal spatial context behind the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding hospital illumination, allowing the bloodshot red and the tear's small highlight to carry the emotional contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the hospital bed, bedside surroundings, and established nighttime interior lighting. EXCLUDE the minister and accompanying officer as they leave the room.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Cheoljin is reclining on a hospital bed, severely injured, with one eye covered and an arm and a leg immobilized in casts; his uncovered eye looks upward. His exact head support, torso inclination, limb placement, and the sides of the covered eye and cast limbs are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military-hospital bed and patient-room setting remain unchanged. 박철진: He remains bedridden with one eye covered and an arm and a leg in casts. His uncovered eye is bloodshot and filling with tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 깁스를 한 박철진의 붉게 충혈된 남은 한쪽 눈에서 분노의 눈물이 뺨에 맺힌 얼굴 클로즈업.\n\nLOCATION (lock): At the injured commander's bed inside a military hospital room, under ordinary nighttime clinical lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Treatment covering around the injured eye (The injured eye remains covered during treatment) — Its edge is visible beside the uncovered eye from the oblique bedside angle; used as Preserves the physical context of his injury within the facial close-up; Hospital bed (Supporting the reclined patient) — A small area beside the patient's head remains visible; used as Provides minimal spatial context behind the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding hospital illumination, allowing the bloodshot red and the tear's small highlight to carry the emotional contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the hospital bed, bedside surroundings, and established nighttime interior lighting. EXCLUDE the minister and accompanying officer as they leave the room.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Cheoljin is reclining on a hospital bed, severely injured, with one eye covered and an arm and a leg immobilized in casts; his uncovered eye looks upward. His exact head support, torso inclination, limb placement, and the sides of the covered eye and cast limbs are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military-hospital bed and patient-room setting remain unchanged. 박철진: He remains bedridden with one eye covered and an arm and a leg in casts. His uncovered eye is bloodshot and filling with tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 깁스를 한 박철진의 붉게 충혈된 남은 한쪽 눈에서 분노의 눈물이 뺨에 맺힌 얼굴 클로즈업.\n\nLOCATION (lock): At the injured commander's bed inside a military hospital room, under ordinary nighttime clinical lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Treatment covering around the injured eye (The injured eye remains covered during treatment) — Its edge is visible beside the uncovered eye from the oblique bedside angle; used as Preserves the physical context of his injury within the facial close-up; Hospital bed (Supporting the reclined patient) — A small area beside the patient's head remains visible; used as Provides minimal spatial context behind the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding hospital illumination, allowing the bloodshot red and the tear's small highlight to carry the emotional contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the hospital bed, bedside surroundings, and established nighttime interior lighting. EXCLUDE the minister and accompanying officer as they leave the room.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Cheoljin is reclining on a hospital bed, severely injured, with one eye covered and an arm and a leg immobilized in casts; his uncovered eye looks upward. His exact head support, torso inclination, limb placement, and the sides of the covered eye and cast limbs are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military-hospital bed and patient-room setting remain unchanged. 박철진: He remains bedridden with one eye covered and an arm and a leg in casts. His uncovered eye is bloodshot and filling with tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "가려지지 않은 오른쪽 눈은 위쪽 약간 우측을 향해 시선을 두고 있습니다.",
    "built_space": "환자는 병원 침대 베개에 누워 있으며, 뒤쪽으로 병상 헤드보드와 의료 기기들이 위치해 있습니다.",
    "entities": "박철진(레퍼런스와 일치), 회색 환자복 상의, 왼쪽 눈을 덮은 안대, 오른쪽 팔의 흰색 깁스, 충혈된 오른쪽 눈과 뺨에 맺힌 눈방울이 정확히 묘사되었습니다.",
    "hard_violations": [
     "[gpt-high] 참고 장면에서 복부 옆 침상에 놓여 있던 깁스 팔과 손을 상부 가슴 위로 옮겼다. 카메라만 바뀌고 움직일 수 없는 인물의 자세와 접촉점은 유지되어야 한다는 지시를 위반한다."
    ],
    "physics": "몸이 침대에 눕혀져 있고, 깁스를 한 팔 역시 가슴/배 위에 자연스럽게 얹혀 있어 중력에 맞게 지지되고 있습니다."
   },
   {
    "label": "B",
    "direction": "가려지지 않은 오른쪽 눈은 위쪽을 향해 시선을 두고 있습니다.",
    "built_space": "환자는 병원 침대 베개에 누워 있으며, 배경에는 레퍼런스 이미지와 동일한 형태의 벽면 조명 패널과 의료용 모니터가 아웃포커싱되어 보입니다.",
    "entities": "박철진(레퍼런스와 일치하는 중년 남성), 회색 환자복 상의, 왼쪽 눈의 하얀색 안대(거즈), 심하게 충혈된 오른쪽 눈, 뺨에 흐르는 눈물이 모두 정확하게 묘사되었습니다.",
    "hard_violations": [],
    "physics": "머리와 상체가 침대 베개에 자연스럽게 기대어 중력의 영향을 받아 안정적으로 지지되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지정된 '얼굴 클로즈업' 프레이밍을 가장 정확하게 구현했으며, 레퍼런스의 배경에 있는 특징적인 조명 패널까지 완벽하게 재현하면서 충혈된 눈과 눈물의 감정선을 훌륭하게 표현했습니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "오른팔의 깁스까지 레퍼런스에 맞게 정확히 구현하고 감정 표현도 훌륭하나, 깁스를 보여주기 위해 프레임이 다소 넓어져 엄격한 '얼굴 클로즈업' 기준에서는 후보 B에 비해 우선순위가 약간 밀립니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "가려지지 않은 오른쪽 눈은 위쪽을 향해 시선을 두고 있습니다.",
        "built_space": "환자는 병원 침대 베개에 누워 있으며, 배경에는 레퍼런스 이미지와 동일한 형태의 벽면 조명 패널과 의료용 모니터가 아웃포커싱되어 보입니다.",
        "entities": "박철진(레퍼런스와 일치하는 중년 남성), 회색 환자복 상의, 왼쪽 눈의 하얀색 안대(거즈), 심하게 충혈된 오른쪽 눈, 뺨에 흐르는 눈물이 모두 정확하게 묘사되었습니다.",
        "hard_violations": [],
        "physics": "머리와 상체가 침대 베개에 자연스럽게 기대어 중력의 영향을 받아 안정적으로 지지되고 있습니다."
       },
       {
        "label": "A",
        "direction": "가려지지 않은 오른쪽 눈은 위쪽 약간 우측을 향해 시선을 두고 있습니다.",
        "built_space": "환자는 병원 침대 베개에 누워 있으며, 뒤쪽으로 병상 헤드보드와 의료 기기들이 위치해 있습니다.",
        "entities": "박철진(레퍼런스와 일치), 회색 환자복 상의, 왼쪽 눈을 덮은 안대, 오른쪽 팔의 흰색 깁스, 충혈된 오른쪽 눈과 뺨에 맺힌 눈방울이 정확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "몸이 침대에 눕혀져 있고, 깁스를 한 팔 역시 가슴/배 위에 자연스럽게 얹혀 있어 중력에 맞게 지지되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지정된 '얼굴 클로즈업' 프레이밍을 가장 정확하게 구현했으며, 레퍼런스의 배경에 있는 특징적인 조명 패널까지 완벽하게 재현하면서 충혈된 눈과 눈물의 감정선을 훌륭하게 표현했습니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "오른팔의 깁스까지 레퍼런스에 맞게 정확히 구현하고 감정 표현도 훌륭하나, 깁스를 보여주기 위해 프레임이 다소 넓어져 엄격한 '얼굴 클로즈업' 기준에서는 후보 B에 비해 우선순위가 약간 밀립니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "가려지지 않은 오른쪽 눈은 위쪽을 향해 시선을 두고 있습니다.",
        "built_space": "환자는 병원 침대 베개에 누워 있으며, 배경에는 레퍼런스 이미지와 동일한 형태의 벽면 조명 패널과 의료용 모니터가 아웃포커싱되어 보입니다.",
        "entities": "박철진(레퍼런스와 일치하는 중년 남성), 회색 환자복 상의, 왼쪽 눈의 하얀색 안대(거즈), 심하게 충혈된 오른쪽 눈, 뺨에 흐르는 눈물이 모두 정확하게 묘사되었습니다.",
        "hard_violations": [],
        "physics": "머리와 상체가 침대 베개에 자연스럽게 기대어 중력의 영향을 받아 안정적으로 지지되고 있습니다."
       },
       {
        "label": "A",
        "direction": "가려지지 않은 오른쪽 눈은 위쪽 약간 우측을 향해 시선을 두고 있습니다.",
        "built_space": "환자는 병원 침대 베개에 누워 있으며, 뒤쪽으로 병상 헤드보드와 의료 기기들이 위치해 있습니다.",
        "entities": "박철진(레퍼런스와 일치), 회색 환자복 상의, 왼쪽 눈을 덮은 안대, 오른쪽 팔의 흰색 깁스, 충혈된 오른쪽 눈과 뺨에 맺힌 눈방울이 정확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "몸이 침대에 눕혀져 있고, 깁스를 한 팔 역시 가슴/배 위에 자연스럽게 얹혀 있어 중력에 맞게 지지되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "얼굴에 밀착한 클로즈업으로 위를 보는 충혈된 눈과 뺨의 눈물을 강조하고 침상 지지를 유지한다. 다만 분노의 표정은 다소 약하다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "충혈과 눈물은 표현했지만, 고정되어야 할 깁스 팔과 손을 기존의 복부 옆 침상에서 가슴 위로 옮겼으며 얼굴 주변의 불필요한 노출도 늘었다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "드러난 눈의 동공은 카메라 렌즈가 아니라 얼굴 위쪽의 화면 밖을 향한다. 특정 대상은 보이지 않으며, 누운 상태에서 위를 바라보라는 지시와 맞는다.",
        "built_space": "머리 뒤의 베개 하나, 오른쪽 침대 난간 일부 하나, 뒤쪽의 흐릿한 모니터 하나, 상단의 병상 조명 일부와 커튼이 보인다. 참고 병실의 침대 주변 구성과 야간 실내 조명이 유지되며, 얼굴이 화면 대부분을 차지한다. 중복 설비나 반사는 없다.",
        "entities": "중년 동아시아계 남성 한 명만 보이며 44세 박철진의 인물 조건에 부합한다. 한쪽 눈은 흰 거즈와 테이프로 덮였고, 반대쪽 눈에는 정상적인 홍채·동공과 붉은 충혈이 보인다. 뺨에는 반사광을 띤 눈물 자국과 방울이 있으며 회색 상의 일부가 보인다. 팔과 다리의 깁스는 적절한 얼굴 클로즈업 밖이라 확인할 수 없다. 다른 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 목은 베개에 기대고 몸통은 기울어진 침상에 지지된다. 눈물은 피부에 붙어 뺨을 따라 흐르며 작은 방울로 맺혀 있다. 공중에 떠 있거나 지지 없이 들린 신체와 물체는 없다."
       },
       {
        "label": "B",
        "direction": "노출된 눈은 화면 위쪽의 먼 지점을 바라보며 렌즈를 직접 응시하지 않는다. 위를 보는 시선 지시는 충족한다.",
        "built_space": "베개 하나, 뒤쪽 침대 머리판 일부 하나, 오른쪽 난간 일부 하나와 오른쪽 위 모니터 하나가 보인다. 병실 재질과 낮은 야간 조명은 참고와 유사하지만, 얼굴 외에 베개와 가슴·손이 차지하는 면적이 A보다 크다. 설비 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "중년 동아시아계 남성 한 명으로 인물 조건에 부합한다. 한쪽 눈의 거즈 덮개, 반대쪽 눈의 충혈, 뺨을 타고 맺힌 눈물, 회색 상의와 팔 깁스가 보인다. 다리는 프레임 밖이다. 눈의 해부학적 형태는 자연스럽고 다른 인물이나 읽을 수 있는 글자는 없다. 표정은 분노보다 슬픔과 긴장에 가깝다.",
        "hard_violations": [
         "참고 장면에서 복부 옆 침상에 놓여 있던 깁스 팔과 손을 상부 가슴 위로 옮겼다. 카메라만 바뀌고 움직일 수 없는 인물의 자세와 접촉점은 유지되어야 한다는 지시를 위반한다."
        ],
        "physics": "머리와 목은 베개에 지지되고 몸통은 침상에 기대어 있다. 깁스 팔과 손은 가슴 위에 얹혀 있어 물리적으로 떠 있지는 않지만, 기존 침상 접촉 위치는 유지하지 않았다. 눈물은 뺨 표면을 따라 아래로 흐르는 형태다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "얼굴에 밀착한 클로즈업으로 위를 보는 충혈된 눈과 뺨의 눈물을 강조하고 침상 지지를 유지한다. 다만 분노의 표정은 다소 약하다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "충혈과 눈물은 표현했지만, 고정되어야 할 깁스 팔과 손을 기존의 복부 옆 침상에서 가슴 위로 옮겼으며 얼굴 주변의 불필요한 노출도 늘었다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "드러난 눈의 동공은 카메라 렌즈가 아니라 얼굴 위쪽의 화면 밖을 향한다. 특정 대상은 보이지 않으며, 누운 상태에서 위를 바라보라는 지시와 맞는다.",
        "built_space": "머리 뒤의 베개 하나, 오른쪽 침대 난간 일부 하나, 뒤쪽의 흐릿한 모니터 하나, 상단의 병상 조명 일부와 커튼이 보인다. 참고 병실의 침대 주변 구성과 야간 실내 조명이 유지되며, 얼굴이 화면 대부분을 차지한다. 중복 설비나 반사는 없다.",
        "entities": "중년 동아시아계 남성 한 명만 보이며 44세 박철진의 인물 조건에 부합한다. 한쪽 눈은 흰 거즈와 테이프로 덮였고, 반대쪽 눈에는 정상적인 홍채·동공과 붉은 충혈이 보인다. 뺨에는 반사광을 띤 눈물 자국과 방울이 있으며 회색 상의 일부가 보인다. 팔과 다리의 깁스는 적절한 얼굴 클로즈업 밖이라 확인할 수 없다. 다른 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 목은 베개에 기대고 몸통은 기울어진 침상에 지지된다. 눈물은 피부에 붙어 뺨을 따라 흐르며 작은 방울로 맺혀 있다. 공중에 떠 있거나 지지 없이 들린 신체와 물체는 없다."
       },
       {
        "label": "A",
        "direction": "노출된 눈은 화면 위쪽의 먼 지점을 바라보며 렌즈를 직접 응시하지 않는다. 위를 보는 시선 지시는 충족한다.",
        "built_space": "베개 하나, 뒤쪽 침대 머리판 일부 하나, 오른쪽 난간 일부 하나와 오른쪽 위 모니터 하나가 보인다. 병실 재질과 낮은 야간 조명은 참고와 유사하지만, 얼굴 외에 베개와 가슴·손이 차지하는 면적이 A보다 크다. 설비 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "중년 동아시아계 남성 한 명으로 인물 조건에 부합한다. 한쪽 눈의 거즈 덮개, 반대쪽 눈의 충혈, 뺨을 타고 맺힌 눈물, 회색 상의와 팔 깁스가 보인다. 다리는 프레임 밖이다. 눈의 해부학적 형태는 자연스럽고 다른 인물이나 읽을 수 있는 글자는 없다. 표정은 분노보다 슬픔과 긴장에 가깝다.",
        "hard_violations": [
         "참고 장면에서 복부 옆 침상에 놓여 있던 깁스 팔과 손을 상부 가슴 위로 옮겼다. 카메라만 바뀌고 움직일 수 없는 인물의 자세와 접촉점은 유지되어야 한다는 지시를 위반한다."
        ],
        "physics": "머리와 목은 베개에 지지되고 몸통은 침상에 기대어 있다. 깁스 팔과 손은 가슴 위에 얹혀 있어 물리적으로 떠 있지는 않지만, 기존 침상 접촉 위치는 유지하지 않았다. 눈물은 뺨 표면을 따라 아래로 흐르는 형태다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.344,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.094,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gpt-high] 참고 장면에서 복부 옆 침상에 놓여 있던 깁스 팔과 손을 상부 가슴 위로 옮겼다. 카메라만 바뀌고 움직일 수 없는 인물의 자세와 접촉점은 유지되어야 한다는 지시를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1094
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 '얼굴 클로즈업' 프레이밍을 가장 정확하게 구현했으며, 레퍼런스의 배경에 있는 특징적인 조명 패널까지 완벽하게 재현하면서 충혈된 눈과 눈물의 감정선을 훌륭하게 표현했습니다."
   },
   {
    "label": "A",
    "score": 1094,
    "verdict_ko": "오른팔의 깁스까지 레퍼런스에 맞게 정확히 구현하고 감정 표현도 훌륭하나, 깁스를 보여주기 위해 프레임이 다소 넓어져 엄격한 '얼굴 클로즈업' 기준에서는 후보 B에 비해 우선순위가 약간 밀립니다.  ★위반: [gpt-high] 참고 장면에서 복부 옆 침상에 놓여 있던 깁스 팔과 손을 상부 가슴 위로 옮겼다. 카메라만 바뀌고 움직일 수 없는 인물의 자세와 접촉점은 유지되어야 한다는 지시를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S70sh8_sel.png",
    "asset_id": "ca875c40-ef5f-4915-b449-e8ea0f2a1b99",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace40-0fa4-78a5-bf87-601fed6c0bd2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S70sh8"
  },
  "locked_char_refs_excluded": [
   "박철진(C17)"
  ]
 },
 "S71sh4": {
  "input_fingerprint": "30dbe12c34d2cb1d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 찰리가 낡은 종이 뒷면에 펜촉을 댄 채 보라색 꽃을 정성껏 그리고 있는 한 순간의 손끝 클로즈업.\n\nLOCATION (lock): In the open rear cargo bed of a southbound truck before dawn, where the robot draws a flower on a sheet of paper. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Old paper (A flower drawing is in progress on its reverse) — The reverse faces the camera at a slant and carries the developing flower drawing; used as Provides the readable action surface without filling the frame; Pen (Its tip is touching the paper) — The shaft leads diagonally from 찰리's fingers to the developing drawing; used as Marks the exact moment of careful inscription; Truck-bed interior (Visible around 찰리's seated lower body) — Seen steeply from above beside his forearm; used as Maintains the drawing's physical scale and location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient dawn light keeps the fingertips and drawing legible without introducing a separate task light or an unsupported ink color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old open-topped truck is traveling south at dawn. A flower drawing is taking shape on the reverse of a sheet of paper; the text does not establish colored pigment.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 찰리가 낡은 종이 뒷면에 펜촉을 댄 채 보라색 꽃을 정성껏 그리고 있는 한 순간의 손끝 클로즈업.\n\nLOCATION (lock): In the open rear cargo bed of a southbound truck before dawn, where the robot draws a flower on a sheet of paper. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Old paper (A flower drawing is in progress on its reverse) — The reverse faces the camera at a slant and carries the developing flower drawing; used as Provides the readable action surface without filling the frame; Pen (Its tip is touching the paper) — The shaft leads diagonally from 찰리's fingers to the developing drawing; used as Marks the exact moment of careful inscription; Truck-bed interior (Visible around 찰리's seated lower body) — Seen steeply from above beside his forearm; used as Maintains the drawing's physical scale and location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient dawn light keeps the fingertips and drawing legible without introducing a separate task light or an unsupported ink color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old open-topped truck is traveling south at dawn. A flower drawing is taking shape on the reverse of a sheet of paper; the text does not establish colored pigment.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 찰리가 낡은 종이 뒷면에 펜촉을 댄 채 보라색 꽃을 정성껏 그리고 있는 한 순간의 손끝 클로즈업.\n\nLOCATION (lock): In the open rear cargo bed of a southbound truck before dawn, where the robot draws a flower on a sheet of paper. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Old paper (A flower drawing is in progress on its reverse) — The reverse faces the camera at a slant and carries the developing flower drawing; used as Provides the readable action surface without filling the frame; Pen (Its tip is touching the paper) — The shaft leads diagonally from 찰리's fingers to the developing drawing; used as Marks the exact moment of careful inscription; Truck-bed interior (Visible around 찰리's seated lower body) — Seen steeply from above beside his forearm; used as Maintains the drawing's physical scale and location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient dawn light keeps the fingertips and drawing legible without introducing a separate task light or an unsupported ink color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old open-topped truck is traveling south at dawn. A flower drawing is taking shape on the reverse of a sheet of paper; the text does not establish colored pigment.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "펜이 종이 중앙에 그려진 흑백 꽃을 향하고 있음.",
    "built_space": "트럭 화물칸 내부. 정면에서 피사체를 바라보는 구도로, 요구된 시점과 다름.",
    "entities": "레퍼런스와 전혀 다른 로봇 갑옷 외형. 보라색이 아닌 흑백 꽃. 펜촉(nib)이 아닌 일반 펜.",
    "hard_violations": [
     "[gemini-pro] 지지대나 잡고 있는 손 없이 공중에 떠 있는 종이",
     "[gpt-high] 종이가 적재함 문턱 바깥 공중에 펼쳐져 있는데 받치는 판, 무릎, 종이를 잡은 손이 전혀 보이지 않는다. 펜촉의 접촉만으로는 이 상태를 지탱할 수 없어 지지 없는 부유 물체가 된다."
    ],
    "physics": "종이를 붙잡거나 받치는 지지대가 화면에 없어 종이가 허공에 떠 있는 상태."
   },
   {
    "label": "B",
    "direction": "펜촉이 종이 위에 그려지는 보라색 꽃을 정확히 향하고 있음.",
    "built_space": "트럭 화물칸 내부. 팔뚝 옆에서 가파르게 내려다보는(도로가 보이는) 구도를 정확히 구현함.",
    "entities": "레퍼런스와 일치하는 기계 얼굴과 녹색 바지. 펜촉. 보라색 꽃 그림. 낡은 종이.",
    "hard_violations": [],
    "physics": "왼손으로 종이 가장자리를 잡고 있으며, 다리 위에서 안정적으로 드로잉 동작을 수행 중."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "허공에 뜬 종이로 인한 물리적 위반(Hard Violation)이 있으며, 요구된 보라색 꽃과 레퍼런스의 인물 외형(의상 등)을 전혀 반영하지 못했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "팔뚝 옆에서 가파르게 내려다보는 요구 구도를 완벽히 따랐으며, 보라색 꽃, 펜촉, 레퍼런스와 일치하는 인물 외형(녹색 바지, 얼굴)을 충실히 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "펜이 종이 중앙에 그려진 흑백 꽃을 향하고 있음.",
        "built_space": "트럭 화물칸 내부. 정면에서 피사체를 바라보는 구도로, 요구된 시점과 다름.",
        "entities": "레퍼런스와 전혀 다른 로봇 갑옷 외형. 보라색이 아닌 흑백 꽃. 펜촉(nib)이 아닌 일반 펜.",
        "hard_violations": [
         "지지대나 잡고 있는 손 없이 공중에 떠 있는 종이"
        ],
        "physics": "종이를 붙잡거나 받치는 지지대가 화면에 없어 종이가 허공에 떠 있는 상태."
       },
       {
        "label": "B",
        "direction": "펜촉이 종이 위에 그려지는 보라색 꽃을 정확히 향하고 있음.",
        "built_space": "트럭 화물칸 내부. 팔뚝 옆에서 가파르게 내려다보는(도로가 보이는) 구도를 정확히 구현함.",
        "entities": "레퍼런스와 일치하는 기계 얼굴과 녹색 바지. 펜촉. 보라색 꽃 그림. 낡은 종이.",
        "hard_violations": [],
        "physics": "왼손으로 종이 가장자리를 잡고 있으며, 다리 위에서 안정적으로 드로잉 동작을 수행 중."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "허공에 뜬 종이로 인한 물리적 위반(Hard Violation)이 있으며, 요구된 보라색 꽃과 레퍼런스의 인물 외형(의상 등)을 전혀 반영하지 못했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "팔뚝 옆에서 가파르게 내려다보는 요구 구도를 완벽히 따랐으며, 보라색 꽃, 펜촉, 레퍼런스와 일치하는 인물 외형(녹색 바지, 얼굴)을 충실히 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "펜이 종이 중앙에 그려진 흑백 꽃을 향하고 있음.",
        "built_space": "트럭 화물칸 내부. 정면에서 피사체를 바라보는 구도로, 요구된 시점과 다름.",
        "entities": "레퍼런스와 전혀 다른 로봇 갑옷 외형. 보라색이 아닌 흑백 꽃. 펜촉(nib)이 아닌 일반 펜.",
        "hard_violations": [
         "지지대나 잡고 있는 손 없이 공중에 떠 있는 종이"
        ],
        "physics": "종이를 붙잡거나 받치는 지지대가 화면에 없어 종이가 허공에 떠 있는 상태."
       },
       {
        "label": "B",
        "direction": "펜촉이 종이 위에 그려지는 보라색 꽃을 정확히 향하고 있음.",
        "built_space": "트럭 화물칸 내부. 팔뚝 옆에서 가파르게 내려다보는(도로가 보이는) 구도를 정확히 구현함.",
        "entities": "레퍼런스와 일치하는 기계 얼굴과 녹색 바지. 펜촉. 보라색 꽃 그림. 낡은 종이.",
        "hard_violations": [],
        "physics": "왼손으로 종이 가장자리를 잡고 있으며, 다리 위에서 안정적으로 드로잉 동작을 수행 중."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "위에서 내려다본 적재함과 손에 받쳐진 종이, 꽃에 닿는 펜촉은 맞지만, 손끝을 넘어 얼굴·몸까지 담았고 근거 없는 보라색 잉크와 참조에 없는 맨금속 팔을 보여준다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "무채색 꽃 선묘는 색소 제한에 부합하지만, 종이가 지지 없이 떠 있으며 적재함 밖 정면 시점과 과도한 신체 노출이 요구된 손끝 클로즈업을 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른손에서 뻗은 펜은 왼쪽 아래로 기울어 꽃의 오른쪽 꽃잎 가장자리에 펜촉을 대고 있다. 종이의 그림 면은 찰리와 위쪽 카메라 양쪽에서 볼 수 있는 방향이다. 얼굴 일부는 종이 쪽으로 향하지만 눈의 정확한 시선은 확인하기 어렵다. 도로는 보이지만 남행 여부는 판별할 수 없다.",
        "built_space": "골이 진 금속 적재함 바닥 하나, 왼쪽 측벽 하나, 화면 위쪽 가로 벽 하나가 보인다. 찰리는 적재함 안에 앉아 무릎 위에서 종이를 받치고 있으며, 카메라는 팔 옆 위쪽에서 가파르게 내려다본다. 위치 관계는 요구와 가깝지만 얼굴, 어깨와 하체까지 들어와 손·손목·전완으로 제한한 구도를 넘는다. 중복 설비나 반사는 없다.",
        "entities": "낡고 찢어진 종이 한 장, 펜 하나, 진행 중인 꽃 그림 하나와 찰리의 기계 손 두 개가 보인다. 금속 분절 손가락은 참조의 로봇 손과 대체로 맞지만, 참조의 색색 소매 대신 맨금속 전완이 노출되어 있다. 종이에는 보라색 선이 명백히 있어, 보라색 꽃이라는 소재는 전달하지만 별도 색소를 설정하지 말라는 조건에는 어긋난다. 읽을 수 있는 글자는 없다. 종이가 뒷면인지는 이 면만으로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "왼손이 종이의 왼쪽 아래 가장자리를 받치고 종이 아래에는 무릎이 있어 지지가 성립한다. 오른손 손가락들이 펜 몸통을 쥐고 펜촉은 종이 표면에 닿아 있다. 앉은 몸은 적재함 내부에 자리하며, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "펜은 손에서 오른쪽 아래로 기울어 꽃 오른쪽의 미완성 윤곽선에 닿아 있다. 얼굴은 화면 밖이라 시선은 확인할 수 없다. 그림 면은 낮은 카메라를 향해 넓게 드러나며 찰리에게도 보일 수 있는 각도다. 차량의 남행 여부는 확인할 수 없다.",
        "built_space": "좌우 적재함 측벽 각 하나, 내부 바닥 하나와 뒤쪽 문턱 및 아래 금속 패널이 보인다. 찰리는 뒤쪽 개구부에 걸터앉은 듯하고 한쪽 다리가 문턱 밖으로 내려온다. 카메라는 적재함 내부를 위에서 보는 대신 차량 뒤 외부에서 거의 정면으로 바라본다. 손뿐 아니라 몸통과 다리까지 크게 포함한다. 내부에는 요청되지 않은 상자와 천 같은 물품도 보인다.",
        "entities": "낡은 종이 한 장, 펜 하나, 무채색 꽃 선묘 하나와 로봇 손 두 개가 보인다. 무채색 선은 색소를 추가하지 말라는 조건과 맞는다. 손은 참조보다 두껍고 장갑형 외장이 강하며 참조의 소매가 없다. 참조에서 가져오지 말라고 한 몸통과 하체까지 노출되어 있다. 읽을 수 있는 글자는 없고 종이의 앞뒷면은 확인되지 않는다.",
        "hard_violations": [
         "종이가 적재함 문턱 바깥 공중에 펼쳐져 있는데 받치는 판, 무릎, 종이를 잡은 손이 전혀 보이지 않는다. 펜촉의 접촉만으로는 이 상태를 지탱할 수 없어 지지 없는 부유 물체가 된다."
        ],
        "physics": "그리는 손은 펜을 쥐고 다른 손은 허벅지 위에 놓여 있다. 몸은 적재함 문턱에 앉아 지지받는 것으로 보인다. 그러나 종이는 두 손 어느 쪽에도 잡혀 있지 않고 아래에 받침도 없으며, 문턱 밖으로 길게 떠 있다. 종이를 지지하는 것은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "위에서 내려다본 적재함과 손에 받쳐진 종이, 꽃에 닿는 펜촉은 맞지만, 손끝을 넘어 얼굴·몸까지 담았고 근거 없는 보라색 잉크와 참조에 없는 맨금속 팔을 보여준다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "무채색 꽃 선묘는 색소 제한에 부합하지만, 종이가 지지 없이 떠 있으며 적재함 밖 정면 시점과 과도한 신체 노출이 요구된 손끝 클로즈업을 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "오른손에서 뻗은 펜은 왼쪽 아래로 기울어 꽃의 오른쪽 꽃잎 가장자리에 펜촉을 대고 있다. 종이의 그림 면은 찰리와 위쪽 카메라 양쪽에서 볼 수 있는 방향이다. 얼굴 일부는 종이 쪽으로 향하지만 눈의 정확한 시선은 확인하기 어렵다. 도로는 보이지만 남행 여부는 판별할 수 없다.",
        "built_space": "골이 진 금속 적재함 바닥 하나, 왼쪽 측벽 하나, 화면 위쪽 가로 벽 하나가 보인다. 찰리는 적재함 안에 앉아 무릎 위에서 종이를 받치고 있으며, 카메라는 팔 옆 위쪽에서 가파르게 내려다본다. 위치 관계는 요구와 가깝지만 얼굴, 어깨와 하체까지 들어와 손·손목·전완으로 제한한 구도를 넘는다. 중복 설비나 반사는 없다.",
        "entities": "낡고 찢어진 종이 한 장, 펜 하나, 진행 중인 꽃 그림 하나와 찰리의 기계 손 두 개가 보인다. 금속 분절 손가락은 참조의 로봇 손과 대체로 맞지만, 참조의 색색 소매 대신 맨금속 전완이 노출되어 있다. 종이에는 보라색 선이 명백히 있어, 보라색 꽃이라는 소재는 전달하지만 별도 색소를 설정하지 말라는 조건에는 어긋난다. 읽을 수 있는 글자는 없다. 종이가 뒷면인지는 이 면만으로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "왼손이 종이의 왼쪽 아래 가장자리를 받치고 종이 아래에는 무릎이 있어 지지가 성립한다. 오른손 손가락들이 펜 몸통을 쥐고 펜촉은 종이 표면에 닿아 있다. 앉은 몸은 적재함 내부에 자리하며, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "펜은 손에서 오른쪽 아래로 기울어 꽃 오른쪽의 미완성 윤곽선에 닿아 있다. 얼굴은 화면 밖이라 시선은 확인할 수 없다. 그림 면은 낮은 카메라를 향해 넓게 드러나며 찰리에게도 보일 수 있는 각도다. 차량의 남행 여부는 확인할 수 없다.",
        "built_space": "좌우 적재함 측벽 각 하나, 내부 바닥 하나와 뒤쪽 문턱 및 아래 금속 패널이 보인다. 찰리는 뒤쪽 개구부에 걸터앉은 듯하고 한쪽 다리가 문턱 밖으로 내려온다. 카메라는 적재함 내부를 위에서 보는 대신 차량 뒤 외부에서 거의 정면으로 바라본다. 손뿐 아니라 몸통과 다리까지 크게 포함한다. 내부에는 요청되지 않은 상자와 천 같은 물품도 보인다.",
        "entities": "낡은 종이 한 장, 펜 하나, 무채색 꽃 선묘 하나와 로봇 손 두 개가 보인다. 무채색 선은 색소를 추가하지 말라는 조건과 맞는다. 손은 참조보다 두껍고 장갑형 외장이 강하며 참조의 소매가 없다. 참조에서 가져오지 말라고 한 몸통과 하체까지 노출되어 있다. 읽을 수 있는 글자는 없고 종이의 앞뒷면은 확인되지 않는다.",
        "hard_violations": [
         "종이가 적재함 문턱 바깥 공중에 펼쳐져 있는데 받치는 판, 무릎, 종이를 잡은 손이 전혀 보이지 않는다. 펜촉의 접촉만으로는 이 상태를 지탱할 수 없어 지지 없는 부유 물체가 된다."
        ],
        "physics": "그리는 손은 펜을 쥐고 다른 손은 허벅지 위에 놓여 있다. 몸은 적재함 문턱에 앉아 지지받는 것으로 보인다. 그러나 종이는 두 손 어느 쪽에도 잡혀 있지 않고 아래에 받침도 없으며, 문턱 밖으로 길게 떠 있다. 종이를 지지하는 것은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.629,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.379,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지지대나 잡고 있는 손 없이 공중에 떠 있는 종이",
     "[gpt-high] 종이가 적재함 문턱 바깥 공중에 펼쳐져 있는데 받치는 판, 무릎, 종이를 잡은 손이 전혀 보이지 않는다. 펜촉의 접촉만으로는 이 상태를 지탱할 수 없어 지지 없는 부유 물체가 된다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 379,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 379,
    "verdict_ko": "허공에 뜬 종이로 인한 물리적 위반(Hard Violation)이 있으며, 요구된 보라색 꽃과 레퍼런스의 인물 외형(의상 등)을 전혀 반영하지 못했습니다.  ★위반: [gemini-pro] 지지대나 잡고 있는 손 없이 공중에 떠 있는 종이 / [gpt-high] 종이가 적재함 문턱 바깥 공중에 펼쳐져 있는데 받치는 판, 무릎, 종이를 잡은 손이 전혀 보이지 않는다. 펜촉의 접촉만으로는 이 상태를 지탱할 수 없어 지지 없는 부유 물체가 된다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "팔뚝 옆에서 가파르게 내려다보는 요구 구도를 완벽히 따랐으며, 보라색 꽃, 펜촉, 레퍼런스와 일치하는 인물 외형(녹색 바지, 얼굴)을 충실히 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "HAND OWNER REFERENCE — 찰리: the person whose hand is on the object in this shot. Use this photograph ONLY to get that hand right — its size, build, skin tone and texture, apparent age, grooming, nails, sleeve and anything worn on the wrist or fingers. Do NOT bring their face, body or clothing into the frame; the framing stays on the object and reaches no further than the wrist and forearm.",
    "path": "<bytes:885644>",
    "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace44-5466-76ec-9e88-ac800d13d4bf",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "era_fail::e5474b9c2b17cdaf": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "era_fail::b3b950dfa6afdda8": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "era_fail::14712d4fd10d329f": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S73sh8": {
  "input_fingerprint": "cc5834b731cfb59a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우와 쿠마가 서로의 어깨를 두 팔로 꽉 껴안은 채 미소 짓는 역동적인 찰나.\n\nLOCATION (lock): In the open bedside area of the harbor-town clinic room, under nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Health-center entrance (The entrance through which 쿠마 has arrived) — Seen obliquely behind the embracing pair; used as Retains the destination of the preceding track without competing with the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the health center's restrained ambient illumination, allowing the men's smiles and physical contact to provide the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed and nighttime treatment-room setting remain unchanged. 이현우: He stands with both arms raised in an embrace and a relieved smile. His skin and clothing remain grimy and his earlier injuries remain. 쿠마: He stands inside the clinic with both arms raised in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우와 쿠마가 서로의 어깨를 두 팔로 꽉 껴안은 채 미소 짓는 역동적인 찰나.\n\nLOCATION (lock): In the open bedside area of the harbor-town clinic room, under nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Health-center entrance (The entrance through which 쿠마 has arrived) — Seen obliquely behind the embracing pair; used as Retains the destination of the preceding track without competing with the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the health center's restrained ambient illumination, allowing the men's smiles and physical contact to provide the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed and nighttime treatment-room setting remain unchanged. 이현우: He stands with both arms raised in an embrace and a relieved smile. His skin and clothing remain grimy and his earlier injuries remain. 쿠마: He stands inside the clinic with both arms raised in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우와 쿠마가 서로의 어깨를 두 팔로 꽉 껴안은 채 미소 짓는 역동적인 찰나.\n\nLOCATION (lock): In the open bedside area of the harbor-town clinic room, under nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Health-center entrance (The entrance through which 쿠마 has arrived) — Seen obliquely behind the embracing pair; used as Retains the destination of the preceding track without competing with the embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the health center's restrained ambient illumination, allowing the men's smiles and physical contact to provide the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed and nighttime treatment-room setting remain unchanged. 이현우: He stands with both arms raised in an embrace and a relieved smile. His skin and clothing remain grimy and his earlier injuries remain. 쿠마: He stands inside the clinic with both arms raised in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "두 인물이 서로를 마주보며 미소 짓고 있음.",
    "built_space": "병실 내부. 왼쪽 전경에 크게 열린 문이 있고 오른쪽에 침대가 위치함. 문이 프레임 전경을 가려 배경 요소 지시와 어긋남.",
    "entities": "이현우(회색 후드, 얼굴 상처)와 쿠마(가죽 모자와 조끼, 패턴 셔츠)의 인상과 복장은 레퍼런스와 일치함.",
    "hard_violations": [
     "[gemini-pro] 이현우의 오른쪽 팔이 두 개로 그려지는 물리적으로 불가능한 해부학적 오류 (하나는 쿠마의 허리를 감싸고, 다른 하나는 쿠마의 어깨에 올려져 있음).",
     "[gpt-high] 창가에 고정되어 있던 침대를 의료 패널 아래 다른 방향으로 배치했다.",
     "[gpt-high] 이전 장면에서 움직일 수 없는 환자가 놓였던 병상 전체를 비어 있는 상태로 보여, 환자의 위치·자세·접촉점을 유지하라는 조건을 위반한다."
    ],
    "physics": "바닥에 서서 포옹하는 자세이나, 추가로 그려진 팔이 구조적으로 불가능한 위치에 있음."
   },
   {
    "label": "B",
    "direction": "두 인물이 서로를 마주보며 미소 짓고 있음.",
    "built_space": "병실 내부. 카메라 앵글을 틀어 왼쪽에 창문, 오른쪽에 침대가 있으며 인물들 뒤편 배경에 출입문이 비스듬히 보이도록 정확히 구성됨.",
    "entities": "이현우와 쿠마 모두 레퍼런스 이미지의 인물 특징과 복장을 정확히 반영함.",
    "hard_violations": [
     "[gpt-high] 고정된 병실 구조에서 출입문·창문·침대의 상대 배치를 바꾸어, 이전 장면과 같은 장소의 연속 촬영으로 성립하지 않는다."
    ],
    "physics": "바닥에 안정적으로 서서 서로의 체중을 지탱하며 두 팔로 자연스럽게 껴안고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "해부학적 오류 없이 두 사람의 포옹을 잘 묘사했으며, 배경 요소인 출입문을 지시대로 인물들 뒤에 적절히 배치하여 프롬프트 충실도가 높습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이현우의 팔이 3개로 묘사되는 치명적인 해부학적 오류가 있으며, 출입문이 배경이 아닌 전경을 차지하여 지시사항을 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 서로를 마주보며 미소 짓고 있음.",
        "built_space": "병실 내부. 왼쪽 전경에 크게 열린 문이 있고 오른쪽에 침대가 위치함. 문이 프레임 전경을 가려 배경 요소 지시와 어긋남.",
        "entities": "이현우(회색 후드, 얼굴 상처)와 쿠마(가죽 모자와 조끼, 패턴 셔츠)의 인상과 복장은 레퍼런스와 일치함.",
        "hard_violations": [
         "이현우의 오른쪽 팔이 두 개로 그려지는 물리적으로 불가능한 해부학적 오류 (하나는 쿠마의 허리를 감싸고, 다른 하나는 쿠마의 어깨에 올려져 있음)."
        ],
        "physics": "바닥에 서서 포옹하는 자세이나, 추가로 그려진 팔이 구조적으로 불가능한 위치에 있음."
       },
       {
        "label": "B",
        "direction": "두 인물이 서로를 마주보며 미소 짓고 있음.",
        "built_space": "병실 내부. 카메라 앵글을 틀어 왼쪽에 창문, 오른쪽에 침대가 있으며 인물들 뒤편 배경에 출입문이 비스듬히 보이도록 정확히 구성됨.",
        "entities": "이현우와 쿠마 모두 레퍼런스 이미지의 인물 특징과 복장을 정확히 반영함.",
        "hard_violations": [],
        "physics": "바닥에 안정적으로 서서 서로의 체중을 지탱하며 두 팔로 자연스럽게 껴안고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "해부학적 오류 없이 두 사람의 포옹을 잘 묘사했으며, 배경 요소인 출입문을 지시대로 인물들 뒤에 적절히 배치하여 프롬프트 충실도가 높습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이현우의 팔이 3개로 묘사되는 치명적인 해부학적 오류가 있으며, 출입문이 배경이 아닌 전경을 차지하여 지시사항을 어겼습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 서로를 마주보며 미소 짓고 있음.",
        "built_space": "병실 내부. 왼쪽 전경에 크게 열린 문이 있고 오른쪽에 침대가 위치함. 문이 프레임 전경을 가려 배경 요소 지시와 어긋남.",
        "entities": "이현우(회색 후드, 얼굴 상처)와 쿠마(가죽 모자와 조끼, 패턴 셔츠)의 인상과 복장은 레퍼런스와 일치함.",
        "hard_violations": [
         "이현우의 오른쪽 팔이 두 개로 그려지는 물리적으로 불가능한 해부학적 오류 (하나는 쿠마의 허리를 감싸고, 다른 하나는 쿠마의 어깨에 올려져 있음)."
        ],
        "physics": "바닥에 서서 포옹하는 자세이나, 추가로 그려진 팔이 구조적으로 불가능한 위치에 있음."
       },
       {
        "label": "B",
        "direction": "두 인물이 서로를 마주보며 미소 짓고 있음.",
        "built_space": "병실 내부. 카메라 앵글을 틀어 왼쪽에 창문, 오른쪽에 침대가 있으며 인물들 뒤편 배경에 출입문이 비스듬히 보이도록 정확히 구성됨.",
        "entities": "이현우와 쿠마 모두 레퍼런스 이미지의 인물 특징과 복장을 정확히 반영함.",
        "hard_violations": [],
        "physics": "바닥에 안정적으로 서서 서로의 체중을 지탱하며 두 팔로 자연스럽게 껴안고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "미소와 포옹은 구현했지만 허벅지까지 담은 넓은 구도이며, 출입문·창문·침대의 고정 배치가 이전 장면과 달라진다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "허리 부근에서 자른 미디엄 숏과 서로를 향한 표정은 더 정확하지만, 침대의 위치와 비어 있는 상태가 장소·부동 상태 유지 조건을 어긴다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 얼굴을 오른쪽 쿠마 쪽으로 돌리고 눈을 가늘게 뜬 채 웃는다. 쿠마는 이현우와 두 사람의 접촉 부근을 내려다보며 웃는다. 두 사람 모두 상대를 향해 팔을 두르고 있으나 쿠마의 아래쪽 팔은 어깨가 아니라 이현우의 허리와 등을 감싼다. 무기나 방향을 확인할 휴대 도구는 없다.",
        "built_space": "왼쪽에 큰 창문 한 조와 커튼, 중앙 오른쪽 뒤에 열린 목재 출입문 하나, 오른쪽에 병상 하나가 보인다. 병상에는 끝판 하나와 양쪽 난간이 보이며 두 사람은 그 옆 빈 바닥 공간에 서 있다. 출입구는 인물 뒤에 비스듬히 보인다. 그러나 이전 장면에서 서로 마주 보는 벽에 있던 출입문과 창문이 여기서는 인접한 벽에 놓인 것으로 보이고, 침대 끝판도 출입문 쪽에 배치되어 원래 병실의 고정 관계가 유지되지 않는다. 창유리에 실내 반사가 있으나 별도 인물이라고 확정할 만큼 선명하지 않다.",
        "entities": "젊은 남성 두 명만 직접 등장한다. 이현우의 헝클어진 검은 머리, 어린 얼굴, 얼굴의 오염과 상처, 낡은 회갈색 후드와 청바지가 참고 이미지에 대체로 맞는다. 쿠마는 참고 인물과 유사한 젊은 성인 얼굴이며 검은 장식 모자, 금속 장식 가죽 조끼, 다색 무늬 셔츠와 허리 주머니가 유지된다. 한국계 및 중국계 혼혈이라는 구체적 배경 자체는 외형만으로 확정할 수 없다. 판독 가능한 문구는 보이지 않는다.",
        "hard_violations": [
         "고정된 병실 구조에서 출입문·창문·침대의 상대 배치를 바꾸어, 이전 장면과 같은 장소의 연속 촬영으로 성립하지 않는다."
        ],
        "physics": "두 사람의 몸통과 화면에 보이는 다리는 서 있는 자세로 이어지며 공중에 뜬 모습은 없다. 발은 화면 밖이므로 바닥 접촉 자체는 확인할 수 없다. 손들은 상대의 어깨와 등에 닿아 있고 팔의 굽힘도 포옹으로 가능한 범위다. 침대는 하부 프레임으로 지지되며 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "쿠마는 왼쪽 이현우의 얼굴을 바라보며 웃고, 이현우도 쿠마 쪽으로 얼굴을 기울인 채 웃는다. 시선과 몸의 방향이 상대에게 모여 있다. 이현우의 두 손은 쿠마의 어깨 부근을 잡지만 쿠마의 한 팔은 이현우의 어깨, 다른 팔은 아래쪽 등을 감싸므로 두 사람 모두 두 팔로 어깨만 껴안는 지시와는 차이가 있다.",
        "built_space": "왼쪽 뒤에 열린 목재 출입문 하나와 복도 너머 문 하나가 보인다. 오른쪽 벽에는 가로 벽등 하나, 의료 연결 패널 한 줄, 아래쪽 수납장 일부가 있고, 오른쪽 뒤 병상 하나에는 머리판·발판·양쪽 난간과 베개가 보인다. 두 사람은 병상 옆 공간에 서 있고 출입구는 비스듬한 배경으로 남는다. 다만 이전 장면의 창가 침대가 여기서는 의료 패널 아래 벽을 향해 놓였으며, 거의 전체가 드러난 침대는 비어 있다. 이전 장면의 움직일 수 없는 환자를 단순히 화면 밖으로 제외한 구도가 아니다.",
        "entities": "이현우와 쿠마에 해당하는 젊은 남성 두 명이 보인다. 이현우의 검은 헝클어진 머리, 어린 얼굴, 볼의 상처와 때, 해진 후드가 참고 이미지와 잘 연결된다. 쿠마의 얼굴과 체격도 참고 이미지에 대체로 가깝고 장식 모자, 스터드 조끼, 다색 셔츠, 허리 주머니가 유지된다. 구체적인 민족적 배경은 외형만으로 확정할 수 없다. 침대 표찰과 복도 표지는 흐려 읽히지 않으며 명확한 자막이나 오버레이는 없다.",
        "hard_violations": [
         "창가에 고정되어 있던 침대를 의료 패널 아래 다른 방향으로 배치했다.",
         "이전 장면에서 움직일 수 없는 환자가 놓였던 병상 전체를 비어 있는 상태로 보여, 환자의 위치·자세·접촉점을 유지하라는 조건을 위반한다."
        ],
        "physics": "두 사람은 몸통을 서로 기울여 접촉하고 있으며, 보이는 손과 팔은 상대의 어깨와 등을 실제로 잡고 있다. 하체와 발 대부분은 미디엄 숏 밖이지만 공중 부양이나 불가능한 체중 지지는 보이지 않는다. 베개와 침구는 매트리스 위에 놓여 있고 침대는 프레임으로 지지된다. 포옹의 팔 교차도 해부학적으로 가능한 모습이다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "미소와 포옹은 구현했지만 허벅지까지 담은 넓은 구도이며, 출입문·창문·침대의 고정 배치가 이전 장면과 달라진다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "허리 부근에서 자른 미디엄 숏과 서로를 향한 표정은 더 정확하지만, 침대의 위치와 비어 있는 상태가 장소·부동 상태 유지 조건을 어긴다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 얼굴을 오른쪽 쿠마 쪽으로 돌리고 눈을 가늘게 뜬 채 웃는다. 쿠마는 이현우와 두 사람의 접촉 부근을 내려다보며 웃는다. 두 사람 모두 상대를 향해 팔을 두르고 있으나 쿠마의 아래쪽 팔은 어깨가 아니라 이현우의 허리와 등을 감싼다. 무기나 방향을 확인할 휴대 도구는 없다.",
        "built_space": "왼쪽에 큰 창문 한 조와 커튼, 중앙 오른쪽 뒤에 열린 목재 출입문 하나, 오른쪽에 병상 하나가 보인다. 병상에는 끝판 하나와 양쪽 난간이 보이며 두 사람은 그 옆 빈 바닥 공간에 서 있다. 출입구는 인물 뒤에 비스듬히 보인다. 그러나 이전 장면에서 서로 마주 보는 벽에 있던 출입문과 창문이 여기서는 인접한 벽에 놓인 것으로 보이고, 침대 끝판도 출입문 쪽에 배치되어 원래 병실의 고정 관계가 유지되지 않는다. 창유리에 실내 반사가 있으나 별도 인물이라고 확정할 만큼 선명하지 않다.",
        "entities": "젊은 남성 두 명만 직접 등장한다. 이현우의 헝클어진 검은 머리, 어린 얼굴, 얼굴의 오염과 상처, 낡은 회갈색 후드와 청바지가 참고 이미지에 대체로 맞는다. 쿠마는 참고 인물과 유사한 젊은 성인 얼굴이며 검은 장식 모자, 금속 장식 가죽 조끼, 다색 무늬 셔츠와 허리 주머니가 유지된다. 한국계 및 중국계 혼혈이라는 구체적 배경 자체는 외형만으로 확정할 수 없다. 판독 가능한 문구는 보이지 않는다.",
        "hard_violations": [
         "고정된 병실 구조에서 출입문·창문·침대의 상대 배치를 바꾸어, 이전 장면과 같은 장소의 연속 촬영으로 성립하지 않는다."
        ],
        "physics": "두 사람의 몸통과 화면에 보이는 다리는 서 있는 자세로 이어지며 공중에 뜬 모습은 없다. 발은 화면 밖이므로 바닥 접촉 자체는 확인할 수 없다. 손들은 상대의 어깨와 등에 닿아 있고 팔의 굽힘도 포옹으로 가능한 범위다. 침대는 하부 프레임으로 지지되며 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "쿠마는 왼쪽 이현우의 얼굴을 바라보며 웃고, 이현우도 쿠마 쪽으로 얼굴을 기울인 채 웃는다. 시선과 몸의 방향이 상대에게 모여 있다. 이현우의 두 손은 쿠마의 어깨 부근을 잡지만 쿠마의 한 팔은 이현우의 어깨, 다른 팔은 아래쪽 등을 감싸므로 두 사람 모두 두 팔로 어깨만 껴안는 지시와는 차이가 있다.",
        "built_space": "왼쪽 뒤에 열린 목재 출입문 하나와 복도 너머 문 하나가 보인다. 오른쪽 벽에는 가로 벽등 하나, 의료 연결 패널 한 줄, 아래쪽 수납장 일부가 있고, 오른쪽 뒤 병상 하나에는 머리판·발판·양쪽 난간과 베개가 보인다. 두 사람은 병상 옆 공간에 서 있고 출입구는 비스듬한 배경으로 남는다. 다만 이전 장면의 창가 침대가 여기서는 의료 패널 아래 벽을 향해 놓였으며, 거의 전체가 드러난 침대는 비어 있다. 이전 장면의 움직일 수 없는 환자를 단순히 화면 밖으로 제외한 구도가 아니다.",
        "entities": "이현우와 쿠마에 해당하는 젊은 남성 두 명이 보인다. 이현우의 검은 헝클어진 머리, 어린 얼굴, 볼의 상처와 때, 해진 후드가 참고 이미지와 잘 연결된다. 쿠마의 얼굴과 체격도 참고 이미지에 대체로 가깝고 장식 모자, 스터드 조끼, 다색 셔츠, 허리 주머니가 유지된다. 구체적인 민족적 배경은 외형만으로 확정할 수 없다. 침대 표찰과 복도 표지는 흐려 읽히지 않으며 명확한 자막이나 오버레이는 없다.",
        "hard_violations": [
         "창가에 고정되어 있던 침대를 의료 패널 아래 다른 방향으로 배치했다.",
         "이전 장면에서 움직일 수 없는 환자가 놓였던 병상 전체를 비어 있는 상태로 보여, 환자의 위치·자세·접촉점을 유지하라는 조건을 위반한다."
        ],
        "physics": "두 사람은 몸통을 서로 기울여 접촉하고 있으며, 보이는 손과 팔은 상대의 어깨와 등을 실제로 잡고 있다. 하체와 발 대부분은 미디엄 숏 밖이지만 공중 부양이나 불가능한 체중 지지는 보이지 않는다. 베개와 침구는 매트리스 위에 놓여 있고 침대는 프레임으로 지지된다. 포옹의 팔 교차도 해부학적으로 가능한 모습이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.8
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.55
   },
   "violations": {
    "A": [
     "[gemini-pro] 이현우의 오른쪽 팔이 두 개로 그려지는 물리적으로 불가능한 해부학적 오류 (하나는 쿠마의 허리를 감싸고, 다른 하나는 쿠마의 어깨에 올려져 있음).",
     "[gpt-high] 창가에 고정되어 있던 침대를 의료 패널 아래 다른 방향으로 배치했다.",
     "[gpt-high] 이전 장면에서 움직일 수 없는 환자가 놓였던 병상 전체를 비어 있는 상태로 보여, 환자의 위치·자세·접촉점을 유지하라는 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 고정된 병실 구조에서 출입문·창문·침대의 상대 배치를 바꾸어, 이전 장면과 같은 장소의 연속 촬영으로 성립하지 않는다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1550,
   "A": 1179
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1550,
    "verdict_ko": "해부학적 오류 없이 두 사람의 포옹을 잘 묘사했으며, 배경 요소인 출입문을 지시대로 인물들 뒤에 적절히 배치하여 프롬프트 충실도가 높습니다.  ★위반: [gpt-high] 고정된 병실 구조에서 출입문·창문·침대의 상대 배치를 바꾸어, 이전 장면과 같은 장소의 연속 촬영으로 성립하지 않는다."
   },
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "이현우의 팔이 3개로 묘사되는 치명적인 해부학적 오류가 있으며, 출입문이 배경이 아닌 전경을 차지하여 지시사항을 어겼습니다.  ★위반: [gemini-pro] 이현우의 오른쪽 팔이 두 개로 그려지는 물리적으로 불가능한 해부학적 오류 (하나는 쿠마의 허리를 감싸고, 다른 하나는 쿠마의 어깨에 올려져 있음). / [gpt-high] 창가에 고정되어 있던 침대를 의료 패널 아래 다른 방향으로 배치했다. / [gpt-high] 이전 장면에서 움직일 수 없는 환자가 놓였던 병상 전체를 비어 있는 상태로 보여, 환자의 위치·자세·접촉점을 유지하라는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S73sh2_sel.png",
    "asset_id": "7c758b18-ac14-481b-af61-1f3558ffcc4a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:904970>",
    "asset_id": "afafdeae-9f05-477d-8a3f-6cedcc1096d3",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace7f-8df3-7ca3-9c24-5387989db01b",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S73sh2"
  }
 },
 "era_fail::3d37fa42e862f1ba": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S74sh6": {
  "input_fingerprint": "1d33459ef4f10754",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 쟁반 위에 놓인 생선회 표면에 흉측한 돌연변이 종양들이 불거져 나온 클로즈업.\n\nLOCATION (lock): On the serving table of an outdoor harbor-market eatery at night, where a fish vendor has delivered a platter. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sliced mutant fish (Served with grotesque abnormal growths protruding from its surface) — The upper surfaces and raised profiles of the growths are visible from a steep oblique angle; used as Primary focus reveals why the offered meal is disturbing; Serving tray (Resting on the table with the sliced fish) — Its serving side and boundary are visible beneath the fish; used as Contains the food and supplies a clear size reference without filling the image; Restaurant tabletop (Supporting the newly delivered tray) — Seen from above with the near edge retained; used as Preserves the same side of the conversation space and prevents a contextless macro image.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination gives the abnormal growths readable relief without inventing luminous contamination or an unnatural color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A serving of raw fish has been set down at the roadside eating place. Its visibly mutated appearance indicates radioactive contamination, without establishing any specific tumor pattern.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 쟁반 위에 놓인 생선회 표면에 흉측한 돌연변이 종양들이 불거져 나온 클로즈업.\n\nLOCATION (lock): On the serving table of an outdoor harbor-market eatery at night, where a fish vendor has delivered a platter. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sliced mutant fish (Served with grotesque abnormal growths protruding from its surface) — The upper surfaces and raised profiles of the growths are visible from a steep oblique angle; used as Primary focus reveals why the offered meal is disturbing; Serving tray (Resting on the table with the sliced fish) — Its serving side and boundary are visible beneath the fish; used as Contains the food and supplies a clear size reference without filling the image; Restaurant tabletop (Supporting the newly delivered tray) — Seen from above with the near edge retained; used as Preserves the same side of the conversation space and prevents a contextless macro image.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination gives the abnormal growths readable relief without inventing luminous contamination or an unnatural color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A serving of raw fish has been set down at the roadside eating place. Its visibly mutated appearance indicates radioactive contamination, without establishing any specific tumor pattern.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 쟁반 위에 놓인 생선회 표면에 흉측한 돌연변이 종양들이 불거져 나온 클로즈업.\n\nLOCATION (lock): On the serving table of an outdoor harbor-market eatery at night, where a fish vendor has delivered a platter. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sliced mutant fish (Served with grotesque abnormal growths protruding from its surface) — The upper surfaces and raised profiles of the growths are visible from a steep oblique angle; used as Primary focus reveals why the offered meal is disturbing; Serving tray (Resting on the table with the sliced fish) — Its serving side and boundary are visible beneath the fish; used as Contains the food and supplies a clear size reference without filling the image; Restaurant tabletop (Supporting the newly delivered tray) — Seen from above with the near edge retained; used as Preserves the same side of the conversation space and prevents a contextless macro image.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination gives the abnormal growths readable relief without inventing luminous contamination or an unnatural color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A serving of raw fish has been set down at the roadside eating place. Its visibly mutated appearance indicates radioactive contamination, without establishing any specific tumor pattern.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 테이블 위 쟁반과 생선회를 뚜렷한 사각으로 내려다보며 샷의 목표를 정확히 향함.",
    "built_space": "항구 바닥 배경에 파란색 플라스틱 테이블이 제자리에 위치하나, 수저통이 두 개로 늘어남.",
    "entities": "기괴한 종양이 있는 생선회, 흰색 쟁반, 파란 테이블, 수저통. 인물 없음.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 하나뿐인 수저통을 두 개로 복제하여 배치함 (중복된 구조물/사물).",
     "[gpt-high] 참조에 하나인 금속 젓가락 통을 왼쪽 앞뒤에 두 개로 중복 배치하여, 이어지는 장면에 없는 소품을 추가했다."
    ],
    "physics": "생선회는 쟁반 위에, 쟁반과 수저통은 테이블 위에 물리적으로 자연스럽게 지탱됨."
   },
   {
    "label": "B",
    "direction": "카메라가 테이블 위 금속 쟁반과 생선회를 비스듬히 바라봄.",
    "built_space": "항구 배경 앞에 테이블이 위치하나, 고정된 세트인 파란색 테이블 대신 나무 판자 테이블이 배치됨.",
    "entities": "살점 형태의 종양이 있는 생선회, 금속 쟁반, 나무 테이블. 인물 없음.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 지정된 파란색 플라스틱 테이블을 완전히 다른 재질의 나무 테이블로 임의 생성함 (발명된 사물 및 장소 위반)."
    ],
    "physics": "생선회는 쟁반 위에, 쟁반은 테이블 위에 안정적으로 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "레퍼런스의 파란색 플라스틱 테이블과 조명 분위기를 잘 유지하며 요구된 생선회 클로즈업을 구현했으나, 수저통이 중복 생성된 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 생선회의 형태는 묘사했으나, 지정된 장소의 핵심인 파란색 플라스틱 테이블을 임의로 나무 테이블로 변경하여 장소 일관성을 심각하게 훼손했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 테이블 위 쟁반과 생선회를 뚜렷한 사각으로 내려다보며 샷의 목표를 정확히 향함.",
        "built_space": "항구 바닥 배경에 파란색 플라스틱 테이블이 제자리에 위치하나, 수저통이 두 개로 늘어남.",
        "entities": "기괴한 종양이 있는 생선회, 흰색 쟁반, 파란 테이블, 수저통. 인물 없음.",
        "hard_violations": [
         "레퍼런스에 하나뿐인 수저통을 두 개로 복제하여 배치함 (중복된 구조물/사물)."
        ],
        "physics": "생선회는 쟁반 위에, 쟁반과 수저통은 테이블 위에 물리적으로 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "카메라가 테이블 위 금속 쟁반과 생선회를 비스듬히 바라봄.",
        "built_space": "항구 배경 앞에 테이블이 위치하나, 고정된 세트인 파란색 테이블 대신 나무 판자 테이블이 배치됨.",
        "entities": "살점 형태의 종양이 있는 생선회, 금속 쟁반, 나무 테이블. 인물 없음.",
        "hard_violations": [
         "레퍼런스에 지정된 파란색 플라스틱 테이블을 완전히 다른 재질의 나무 테이블로 임의 생성함 (발명된 사물 및 장소 위반)."
        ],
        "physics": "생선회는 쟁반 위에, 쟁반은 테이블 위에 안정적으로 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "레퍼런스의 파란색 플라스틱 테이블과 조명 분위기를 잘 유지하며 요구된 생선회 클로즈업을 구현했으나, 수저통이 중복 생성된 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 생선회의 형태는 묘사했으나, 지정된 장소의 핵심인 파란색 플라스틱 테이블을 임의로 나무 테이블로 변경하여 장소 일관성을 심각하게 훼손했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 테이블 위 쟁반과 생선회를 뚜렷한 사각으로 내려다보며 샷의 목표를 정확히 향함.",
        "built_space": "항구 바닥 배경에 파란색 플라스틱 테이블이 제자리에 위치하나, 수저통이 두 개로 늘어남.",
        "entities": "기괴한 종양이 있는 생선회, 흰색 쟁반, 파란 테이블, 수저통. 인물 없음.",
        "hard_violations": [
         "레퍼런스에 하나뿐인 수저통을 두 개로 복제하여 배치함 (중복된 구조물/사물)."
        ],
        "physics": "생선회는 쟁반 위에, 쟁반과 수저통은 테이블 위에 물리적으로 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "카메라가 테이블 위 금속 쟁반과 생선회를 비스듬히 바라봄.",
        "built_space": "항구 배경 앞에 테이블이 위치하나, 고정된 세트인 파란색 테이블 대신 나무 판자 테이블이 배치됨.",
        "entities": "살점 형태의 종양이 있는 생선회, 금속 쟁반, 나무 테이블. 인물 없음.",
        "hard_violations": [
         "레퍼런스에 지정된 파란색 플라스틱 테이블을 완전히 다른 재질의 나무 테이블로 임의 생성함 (발명된 사물 및 장소 위반)."
        ],
        "physics": "생선회는 쟁반 위에, 쟁반은 테이블 위에 안정적으로 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "회 표면의 종양과 물리적 접촉은 명확하고 소품 중복은 없지만, 나무 탁자로 장소가 바뀌고 시점이 낮아 요청한 내려다보는 클로즈업에는 미달한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "종양을 내려다보는 클로즈업과 파란 상판·가까운 가장자리는 더 충실하지만, 참조의 젓가락 통 하나를 두 개로 늘린 명백한 소품 중복이 결정적인 위반이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "사람이나 시선, 무기, 이동체는 없다. 카메라는 쟁반을 낮은 사선으로 바라보며 종양의 윗면과 옆면을 보여 준다. 요청한 가파른 내려다보기보다 수평에 가까워 항구 배경이 크게 들어온다.",
        "built_space": "나무 판재 상판 하나와 금속 쟁반 하나가 보인다. 뒤에는 시장 가판, 천막 기둥, 조명과 항구가 이어진다. 참조의 마모된 파란 원형 탁자가 나무 탁자로 바뀌었으며, 상판의 가까운 바깥 가장자리도 프레임 안에서 확인되지 않는다. 쟁반은 오른쪽과 아래쪽에서 잘린다.",
        "entities": "껍질이 붙은 여러 생선회 조각 위에 살색의 울퉁불퉁한 덩어리가 솟아 있어 요청한 돌연변이 종양으로 읽힌다. 금속 쟁반은 음식을 담는 면과 테두리가 보인다. 사람이나 읽을 수 있는 글자는 없고 야간 항구 분위기는 유지된다. 상판의 재질과 색은 참조와 다르다.",
        "hard_violations": [],
        "physics": "회는 쟁반 바닥에 놓이고 쟁반은 상판에 닿아 있다. 종양은 회 표면에 붙어 있으며 공중에 뜬 물체는 없다. 젖은 살과 금속의 반사는 주변 조명에 따른 것으로 보이고, 스스로 빛나는 오염 표현은 없다."
       },
       {
        "label": "B",
        "direction": "사람이나 시선, 무기, 이동체는 없다. 카메라가 위쪽 사선에서 회를 내려다보므로 종양의 윗면과 솟은 윤곽이 함께 보인다. 쟁반의 음식을 담는 면도 자연스럽게 카메라 쪽으로 드러난다.",
        "built_space": "파란 상판 하나에 흰 직사각형 쟁반 하나가 놓여 있고 가까운 상판 가장자리도 남아 있다. 왼쪽에는 젓가락이 꽂힌 금속 통이 앞뒤로 두 개 보이지만, 참조에는 하나만 있다. 파란 상판의 마모와 야간의 젖은 바닥은 참조와 가깝지만 탁자 윤곽은 참조의 원형보다 사각형에 가깝다.",
        "entities": "여러 생선회 조각에서 회색·녹색·자주색의 불규칙한 종양이 솟아 있다. 회와 종양은 젖은 입체 물질로 표현되며 발광하지 않는다. 쟁반과 파란 식당 탁자가 있고 사람이나 읽을 수 있는 글자는 없다. 왼쪽의 젓가락 통은 참조보다 하나 많다.",
        "hard_violations": [
         "참조에 하나인 금속 젓가락 통을 왼쪽 앞뒤에 두 개로 중복 배치하여, 이어지는 장면에 없는 소품을 추가했다."
        ],
        "physics": "회 조각은 쟁반에, 쟁반은 상판에 안정적으로 놓여 있다. 종양은 회에 붙어 있고 젓가락은 각각 금속 통 안에 지지되어 있다. 상판 아래에는 다리 일부가 보이며, 지지 없이 떠 있는 물체나 불가능한 반사는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "회 표면의 종양과 물리적 접촉은 명확하고 소품 중복은 없지만, 나무 탁자로 장소가 바뀌고 시점이 낮아 요청한 내려다보는 클로즈업에는 미달한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "종양을 내려다보는 클로즈업과 파란 상판·가까운 가장자리는 더 충실하지만, 참조의 젓가락 통 하나를 두 개로 늘린 명백한 소품 중복이 결정적인 위반이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "사람이나 시선, 무기, 이동체는 없다. 카메라는 쟁반을 낮은 사선으로 바라보며 종양의 윗면과 옆면을 보여 준다. 요청한 가파른 내려다보기보다 수평에 가까워 항구 배경이 크게 들어온다.",
        "built_space": "나무 판재 상판 하나와 금속 쟁반 하나가 보인다. 뒤에는 시장 가판, 천막 기둥, 조명과 항구가 이어진다. 참조의 마모된 파란 원형 탁자가 나무 탁자로 바뀌었으며, 상판의 가까운 바깥 가장자리도 프레임 안에서 확인되지 않는다. 쟁반은 오른쪽과 아래쪽에서 잘린다.",
        "entities": "껍질이 붙은 여러 생선회 조각 위에 살색의 울퉁불퉁한 덩어리가 솟아 있어 요청한 돌연변이 종양으로 읽힌다. 금속 쟁반은 음식을 담는 면과 테두리가 보인다. 사람이나 읽을 수 있는 글자는 없고 야간 항구 분위기는 유지된다. 상판의 재질과 색은 참조와 다르다.",
        "hard_violations": [],
        "physics": "회는 쟁반 바닥에 놓이고 쟁반은 상판에 닿아 있다. 종양은 회 표면에 붙어 있으며 공중에 뜬 물체는 없다. 젖은 살과 금속의 반사는 주변 조명에 따른 것으로 보이고, 스스로 빛나는 오염 표현은 없다."
       },
       {
        "label": "A",
        "direction": "사람이나 시선, 무기, 이동체는 없다. 카메라가 위쪽 사선에서 회를 내려다보므로 종양의 윗면과 솟은 윤곽이 함께 보인다. 쟁반의 음식을 담는 면도 자연스럽게 카메라 쪽으로 드러난다.",
        "built_space": "파란 상판 하나에 흰 직사각형 쟁반 하나가 놓여 있고 가까운 상판 가장자리도 남아 있다. 왼쪽에는 젓가락이 꽂힌 금속 통이 앞뒤로 두 개 보이지만, 참조에는 하나만 있다. 파란 상판의 마모와 야간의 젖은 바닥은 참조와 가깝지만 탁자 윤곽은 참조의 원형보다 사각형에 가깝다.",
        "entities": "여러 생선회 조각에서 회색·녹색·자주색의 불규칙한 종양이 솟아 있다. 회와 종양은 젖은 입체 물질로 표현되며 발광하지 않는다. 쟁반과 파란 식당 탁자가 있고 사람이나 읽을 수 있는 글자는 없다. 왼쪽의 젓가락 통은 참조보다 하나 많다.",
        "hard_violations": [
         "참조에 하나인 금속 젓가락 통을 왼쪽 앞뒤에 두 개로 중복 배치하여, 이어지는 장면에 없는 소품을 추가했다."
        ],
        "physics": "회 조각은 쟁반에, 쟁반은 상판에 안정적으로 놓여 있다. 종양은 회에 붙어 있고 젓가락은 각각 금속 통 안에 지지되어 있다. 상판 아래에는 다리 일부가 보이며, 지지 없이 떠 있는 물체나 불가능한 반사는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.35,
    "B": 1.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스에 하나뿐인 수저통을 두 개로 복제하여 배치함 (중복된 구조물/사물).",
     "[gpt-high] 참조에 하나인 금속 젓가락 통을 왼쪽 앞뒤에 두 개로 중복 배치하여, 이어지는 장면에 없는 소품을 추가했다."
    ],
    "B": [
     "[gemini-pro] 레퍼런스에 지정된 파란색 플라스틱 테이블을 완전히 다른 재질의 나무 테이블로 임의 생성함 (발명된 사물 및 장소 위반)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1350,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "레퍼런스의 파란색 플라스틱 테이블과 조명 분위기를 잘 유지하며 요구된 생선회 클로즈업을 구현했으나, 수저통이 중복 생성된 점이 감점 요인입니다.  ★위반: [gemini-pro] 레퍼런스에 하나뿐인 수저통을 두 개로 복제하여 배치함 (중복된 구조물/사물). / [gpt-high] 참조에 하나인 금속 젓가락 통을 왼쪽 앞뒤에 두 개로 중복 배치하여, 이어지는 장면에 없는 소품을 추가했다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "요구된 생선회의 형태는 묘사했으나, 지정된 장소의 핵심인 파란색 플라스틱 테이블을 임의로 나무 테이블로 변경하여 장소 일관성을 심각하게 훼손했습니다.  ★위반: [gemini-pro] 레퍼런스에 지정된 파란색 플라스틱 테이블을 완전히 다른 재질의 나무 테이블로 임의 생성함 (발명된 사물 및 장소 위반)."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S74sh3_sel.png",
    "asset_id": "83fc650f-ef96-419c-92d6-d9fca3698b09",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace8f-e663-7443-9a40-de6c97b9d158",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S74sh3"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S75sh4": {
  "input_fingerprint": "9120dd7378ac11b9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 창고 안, 날개가 부서지고 녹슨 낡은 경비행기 여러 대가 덩그러니 방치된 전경.\n\nLOCATION (lock): Inside a dark hangar at an abandoned airfield, among rows of derelict light aircraft. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Abandoned light aircraft (Multiple old, nonoperational aircraft with rust and broken wings) — Staggered oblique body and wing views avoid a uniform frontal lineup; used as Distributed across depth as the main environmental reveal, with no single plane overwhelming the frame; Warehouse doorway (Opened for the men's entry) — A narrow interior-side edge remains visible beside the camera's route; used as Anchors the inward movement and distinguishes this interior reveal from the airport exterior; Warehouse interior (Housing the abandoned aircraft) — Viewed diagonally into its depth; used as Spaces between the aircraft preserve environmental scale and solemn negative space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low ambient illumination in the dark nighttime warehouse leaves the aircraft readable while preserving the weight of their abandonment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large hangar door is open, revealing numerous old, abandoned light aircraft that do not currently operate. Outside, the derelict airport retains moss, broken pavement and a partly fallen sign, with a campfire burning.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 창고 안, 날개가 부서지고 녹슨 낡은 경비행기 여러 대가 덩그러니 방치된 전경.\n\nLOCATION (lock): Inside a dark hangar at an abandoned airfield, among rows of derelict light aircraft. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Abandoned light aircraft (Multiple old, nonoperational aircraft with rust and broken wings) — Staggered oblique body and wing views avoid a uniform frontal lineup; used as Distributed across depth as the main environmental reveal, with no single plane overwhelming the frame; Warehouse doorway (Opened for the men's entry) — A narrow interior-side edge remains visible beside the camera's route; used as Anchors the inward movement and distinguishes this interior reveal from the airport exterior; Warehouse interior (Housing the abandoned aircraft) — Viewed diagonally into its depth; used as Spaces between the aircraft preserve environmental scale and solemn negative space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low ambient illumination in the dark nighttime warehouse leaves the aircraft readable while preserving the weight of their abandonment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large hangar door is open, revealing numerous old, abandoned light aircraft that do not currently operate. Outside, the derelict airport retains moss, broken pavement and a partly fallen sign, with a campfire burning.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 창고 안, 날개가 부서지고 녹슨 낡은 경비행기 여러 대가 덩그러니 방치된 전경.\n\nLOCATION (lock): Inside a dark hangar at an abandoned airfield, among rows of derelict light aircraft. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Abandoned light aircraft (Multiple old, nonoperational aircraft with rust and broken wings) — Staggered oblique body and wing views avoid a uniform frontal lineup; used as Distributed across depth as the main environmental reveal, with no single plane overwhelming the frame; Warehouse doorway (Opened for the men's entry) — A narrow interior-side edge remains visible beside the camera's route; used as Anchors the inward movement and distinguishes this interior reveal from the airport exterior; Warehouse interior (Housing the abandoned aircraft) — Viewed diagonally into its depth; used as Spaces between the aircraft preserve environmental scale and solemn negative space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low ambient illumination in the dark nighttime warehouse leaves the aircraft readable while preserving the weight of their abandonment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large hangar door is open, revealing numerous old, abandoned light aircraft that do not currently operate. Outside, the derelict airport retains moss, broken pavement and a partly fallen sign, with a campfire burning.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 격납고 내부 깊숙한 곳을 정면으로 향하며, 비행기들은 렌즈 쪽이나 약간 측면을 향해 배치되어 있습니다.",
    "built_space": "원본의 대칭형 박공지붕 트러스 구조를 잘 유지하고 있으며, 좌측 가장자리의 출입구 프레임을 통해 야외 공간과 모닥불이 자연스럽게 노출됩니다.",
    "entities": "녹슬고 부서진 경비행기 여러 대가 중경에 분산 배치되어 있으며, 좌측 바깥에 모닥불이 명확히 보입니다.",
    "hard_violations": [],
    "physics": "모든 비행기 기체는 랜딩 기어를 통해 콘크리트 바닥에 안정적이고 정상적으로 지지되어 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 격납고 내부에서 우측에 위치한 거대한 출입구를 향해 대각선으로 시선을 향하고 있습니다.",
    "built_space": "단일 전면 출입구를 가진 레퍼런스의 구조를 무시하고 우측 벽면 전체가 개방된 형태로 건축 구조를 왜곡했으며, 내부 바닥에 이끼가 과도하게 자라 있습니다.",
    "entities": "여러 대의 경비행기가 있으나 좌측 전경의 기체는 심하게 뭉개져 있으며, 우측 출입구 밖으로 모닥불이 노출됩니다.",
    "hard_violations": [
     "[gemini-pro] 배경 요소(비행기)를 전경 요소로 거대하게 확대하지 말라는 프레이밍 제한을 정면으로 위반함.",
     "[gemini-pro] 좌측 전경 비행기의 동체 구조가 성립하지 않고 바닥에 무너진 형태(physically impossible anatomy)."
    ],
    "physics": "좌측 앞경에 놓인 비행기 잔해는 동체와의 정상적인 연결 부위 없이 날개와 꼬리가 바닥에 납작하게 눌려 있어 물리적 구조가 불가능합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "대각선 뷰 지침 대신 레퍼런스와 유사한 정면 구도를 취한 점은 아쉬우나, 배경 피사체의 스케일 제한, 조명, 구조적 일관성 등 프롬프트의 핵심 의도를 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배경 요소인 비행기를 전경에 거대하게 배치하지 말라는 지침을 심각하게 위반했으며, 비행기의 물리적 형태와 격납고의 건축 구조가 심하게 왜곡되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 격납고 내부 깊숙한 곳을 정면으로 향하며, 비행기들은 렌즈 쪽이나 약간 측면을 향해 배치되어 있습니다.",
        "built_space": "원본의 대칭형 박공지붕 트러스 구조를 잘 유지하고 있으며, 좌측 가장자리의 출입구 프레임을 통해 야외 공간과 모닥불이 자연스럽게 노출됩니다.",
        "entities": "녹슬고 부서진 경비행기 여러 대가 중경에 분산 배치되어 있으며, 좌측 바깥에 모닥불이 명확히 보입니다.",
        "hard_violations": [],
        "physics": "모든 비행기 기체는 랜딩 기어를 통해 콘크리트 바닥에 안정적이고 정상적으로 지지되어 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 격납고 내부에서 우측에 위치한 거대한 출입구를 향해 대각선으로 시선을 향하고 있습니다.",
        "built_space": "단일 전면 출입구를 가진 레퍼런스의 구조를 무시하고 우측 벽면 전체가 개방된 형태로 건축 구조를 왜곡했으며, 내부 바닥에 이끼가 과도하게 자라 있습니다.",
        "entities": "여러 대의 경비행기가 있으나 좌측 전경의 기체는 심하게 뭉개져 있으며, 우측 출입구 밖으로 모닥불이 노출됩니다.",
        "hard_violations": [
         "배경 요소(비행기)를 전경 요소로 거대하게 확대하지 말라는 프레이밍 제한을 정면으로 위반함.",
         "좌측 전경 비행기의 동체 구조가 성립하지 않고 바닥에 무너진 형태(physically impossible anatomy)."
        ],
        "physics": "좌측 앞경에 놓인 비행기 잔해는 동체와의 정상적인 연결 부위 없이 날개와 꼬리가 바닥에 납작하게 눌려 있어 물리적 구조가 불가능합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "대각선 뷰 지침 대신 레퍼런스와 유사한 정면 구도를 취한 점은 아쉬우나, 배경 피사체의 스케일 제한, 조명, 구조적 일관성 등 프롬프트의 핵심 의도를 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배경 요소인 비행기를 전경에 거대하게 배치하지 말라는 지침을 심각하게 위반했으며, 비행기의 물리적 형태와 격납고의 건축 구조가 심하게 왜곡되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 격납고 내부 깊숙한 곳을 정면으로 향하며, 비행기들은 렌즈 쪽이나 약간 측면을 향해 배치되어 있습니다.",
        "built_space": "원본의 대칭형 박공지붕 트러스 구조를 잘 유지하고 있으며, 좌측 가장자리의 출입구 프레임을 통해 야외 공간과 모닥불이 자연스럽게 노출됩니다.",
        "entities": "녹슬고 부서진 경비행기 여러 대가 중경에 분산 배치되어 있으며, 좌측 바깥에 모닥불이 명확히 보입니다.",
        "hard_violations": [],
        "physics": "모든 비행기 기체는 랜딩 기어를 통해 콘크리트 바닥에 안정적이고 정상적으로 지지되어 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 격납고 내부에서 우측에 위치한 거대한 출입구를 향해 대각선으로 시선을 향하고 있습니다.",
        "built_space": "단일 전면 출입구를 가진 레퍼런스의 구조를 무시하고 우측 벽면 전체가 개방된 형태로 건축 구조를 왜곡했으며, 내부 바닥에 이끼가 과도하게 자라 있습니다.",
        "entities": "여러 대의 경비행기가 있으나 좌측 전경의 기체는 심하게 뭉개져 있으며, 우측 출입구 밖으로 모닥불이 노출됩니다.",
        "hard_violations": [
         "배경 요소(비행기)를 전경 요소로 거대하게 확대하지 말라는 프레이밍 제한을 정면으로 위반함.",
         "좌측 전경 비행기의 동체 구조가 성립하지 않고 바닥에 무너진 형태(physically impossible anatomy)."
        ],
        "physics": "좌측 앞경에 놓인 비행기 잔해는 동체와의 정상적인 연결 부위 없이 날개와 꼬리가 바닥에 납작하게 눌려 있어 물리적 구조가 불가능합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "격납고 내부의 사선 시점과 서로 다른 방향으로 방치된 파손 기체들이 핵심 연출에 더 충실하지만, 전경 잔해가 크고 출입구의 외부 노출이 다소 넓다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "참조의 격납고 구조와 야간 분위기는 잘 유지했으나, 양쪽 문틀을 모두 담은 정면 입구 구도여서 내부를 대각선으로 드러내라는 구도 지시와 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 대형 기체의 기수는 카메라 쪽으로 비스듬히 향하고, 뒤쪽 중앙 기체는 화면 오른쪽을 향한다. 오른쪽 뒤 기체는 카메라 쪽을 향하며, 바닥 중앙의 파손 동체는 오른쪽 뒤로 뻗는다. 왼쪽 아래에는 다른 기체의 꼬리와 후방 동체가 보인다. 동일한 정면 방향의 일렬 배치가 아니며, 카메라는 내부를 사선으로 바라본다. 사람의 시선이나 무기는 없다.",
        "built_space": "골강판 벽과 지붕, 반복되는 철골 기둥·보, 콘크리트 바닥이 보인다. 식별 가능한 매달린 조명은 두 개이고, 뒤쪽 벽에는 작은 상부 창 하나가 보인다. 오른쪽에 열린 대형 출입구 하나와 그 문설주가 있으며, 기체 사이로 바닥 공간이 이어진다. 참조의 재료와 구조 유형은 유지하지만 문 가장자리만 좁게 남기기보다는 외부 공간을 상당 부분 보여 준다. 반사면이나 인물 배치는 없다.",
        "entities": "경비행기 또는 그 잔해가 다섯 대가량 식별된다. 왼쪽의 흰색·청색 줄무늬 기체와 골조가 드러난 날개는 참조의 특징을 반영한다. 중앙과 전경에는 심하게 파손된 동체·날개가 있고, 여러 표면에 녹과 도장 박리가 보인다. 문밖에는 모닥불, 어두운 야외 지면과 이끼가 보인다. 사람과 읽을 수 있는 글자는 없다. 넘어진 표지판은 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "뒤쪽 기체들은 보이는 착륙장치와 바퀴로 바닥에 지지된다. 중앙의 파손 기체는 낮게 주저앉아 날개 끝과 하부가 바닥에 닿는 배치이고, 전경 잔해도 바닥에 놓여 있다. 왼쪽 큰 기체의 하부 지지는 전경 잔해에 일부 가려져 있지만 공중에 떠 있다는 증거는 없다. 천장등은 매달린 연결부가 있으며, 모닥불은 지면 위에서 탄다. 운동 중인 인물이나 비행 중인 물체는 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽과 오른쪽 전경 기체 및 오른쪽 뒤 기체는 대체로 출입구와 카메라 쪽을 향한다. 중앙의 작은 기체만 왼쪽 앞을 향해 비스듬히 놓여 있다. 카메라는 격납고의 중심축을 따라 거의 정면으로 안쪽을 바라보므로, 요청한 사선 깊이감과 다양한 동체 방향은 A보다 약하다. 사람의 시선이나 무기는 없다.",
        "built_space": "참조와 유사한 박공지붕 철골, 골강판 벽, 양쪽 상부 작은 창, 콘크리트 바닥이 보인다. 천장에는 가까운 두 개, 중간 두 개, 중앙 깊숙한 한 개의 조명이 식별된다. 하나의 넓은 출입구 양쪽 문틀과 바닥 문턱을 모두 담고 있으며, 왼쪽에는 열린 문짝 가장자리와 외부가 보인다. 중앙 통로의 빈 공간은 잘 확보했지만, 좁은 내부 측 문 가장자리를 남긴 사선 시점이 아니라 입구에서 정면으로 들여다보는 구도다. 불가능한 반사는 없다.",
        "entities": "낡은 경비행기 네 대가 식별된다. 왼쪽 흰색·청색 줄무늬 기체, 노출된 날개 골조, 오른쪽 뒤의 프로펠러 기체는 참조와 연결된다. 기체들에는 부식과 도장 박리가 있으며, 왼쪽 날개와 중앙 기체에 파손이 보인다. 왼쪽 문밖에는 모닥불과 이끼 낀 지면이 보인다. 인물이나 판독 가능한 글자는 없고, 넘어진 표지판은 확인되지 않는다.",
        "hard_violations": [],
        "physics": "전경 양쪽 기체는 보이는 착륙장치와 바퀴로 콘크리트 바닥에 서 있다. 중앙 기체도 하부 바퀴가 보이며, 오른쪽 뒤 기체의 지지부는 다른 날개에 일부 가려진다. 길게 뻗은 오른쪽 전경 날개는 동체에 연결되어 있고, 왼쪽의 노출된 날개 골조도 기체에 붙어 있다. 천장등은 매달려 있고 모닥불은 지면에 놓여 있다. 지지 없이 떠 있는 물체나 불가능한 동작은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "격납고 내부의 사선 시점과 서로 다른 방향으로 방치된 파손 기체들이 핵심 연출에 더 충실하지만, 전경 잔해가 크고 출입구의 외부 노출이 다소 넓다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "참조의 격납고 구조와 야간 분위기는 잘 유지했으나, 양쪽 문틀을 모두 담은 정면 입구 구도여서 내부를 대각선으로 드러내라는 구도 지시와 어긋난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 대형 기체의 기수는 카메라 쪽으로 비스듬히 향하고, 뒤쪽 중앙 기체는 화면 오른쪽을 향한다. 오른쪽 뒤 기체는 카메라 쪽을 향하며, 바닥 중앙의 파손 동체는 오른쪽 뒤로 뻗는다. 왼쪽 아래에는 다른 기체의 꼬리와 후방 동체가 보인다. 동일한 정면 방향의 일렬 배치가 아니며, 카메라는 내부를 사선으로 바라본다. 사람의 시선이나 무기는 없다.",
        "built_space": "골강판 벽과 지붕, 반복되는 철골 기둥·보, 콘크리트 바닥이 보인다. 식별 가능한 매달린 조명은 두 개이고, 뒤쪽 벽에는 작은 상부 창 하나가 보인다. 오른쪽에 열린 대형 출입구 하나와 그 문설주가 있으며, 기체 사이로 바닥 공간이 이어진다. 참조의 재료와 구조 유형은 유지하지만 문 가장자리만 좁게 남기기보다는 외부 공간을 상당 부분 보여 준다. 반사면이나 인물 배치는 없다.",
        "entities": "경비행기 또는 그 잔해가 다섯 대가량 식별된다. 왼쪽의 흰색·청색 줄무늬 기체와 골조가 드러난 날개는 참조의 특징을 반영한다. 중앙과 전경에는 심하게 파손된 동체·날개가 있고, 여러 표면에 녹과 도장 박리가 보인다. 문밖에는 모닥불, 어두운 야외 지면과 이끼가 보인다. 사람과 읽을 수 있는 글자는 없다. 넘어진 표지판은 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "뒤쪽 기체들은 보이는 착륙장치와 바퀴로 바닥에 지지된다. 중앙의 파손 기체는 낮게 주저앉아 날개 끝과 하부가 바닥에 닿는 배치이고, 전경 잔해도 바닥에 놓여 있다. 왼쪽 큰 기체의 하부 지지는 전경 잔해에 일부 가려져 있지만 공중에 떠 있다는 증거는 없다. 천장등은 매달린 연결부가 있으며, 모닥불은 지면 위에서 탄다. 운동 중인 인물이나 비행 중인 물체는 없다."
       },
       {
        "label": "A",
        "direction": "왼쪽과 오른쪽 전경 기체 및 오른쪽 뒤 기체는 대체로 출입구와 카메라 쪽을 향한다. 중앙의 작은 기체만 왼쪽 앞을 향해 비스듬히 놓여 있다. 카메라는 격납고의 중심축을 따라 거의 정면으로 안쪽을 바라보므로, 요청한 사선 깊이감과 다양한 동체 방향은 A보다 약하다. 사람의 시선이나 무기는 없다.",
        "built_space": "참조와 유사한 박공지붕 철골, 골강판 벽, 양쪽 상부 작은 창, 콘크리트 바닥이 보인다. 천장에는 가까운 두 개, 중간 두 개, 중앙 깊숙한 한 개의 조명이 식별된다. 하나의 넓은 출입구 양쪽 문틀과 바닥 문턱을 모두 담고 있으며, 왼쪽에는 열린 문짝 가장자리와 외부가 보인다. 중앙 통로의 빈 공간은 잘 확보했지만, 좁은 내부 측 문 가장자리를 남긴 사선 시점이 아니라 입구에서 정면으로 들여다보는 구도다. 불가능한 반사는 없다.",
        "entities": "낡은 경비행기 네 대가 식별된다. 왼쪽 흰색·청색 줄무늬 기체, 노출된 날개 골조, 오른쪽 뒤의 프로펠러 기체는 참조와 연결된다. 기체들에는 부식과 도장 박리가 있으며, 왼쪽 날개와 중앙 기체에 파손이 보인다. 왼쪽 문밖에는 모닥불과 이끼 낀 지면이 보인다. 인물이나 판독 가능한 글자는 없고, 넘어진 표지판은 확인되지 않는다.",
        "hard_violations": [],
        "physics": "전경 양쪽 기체는 보이는 착륙장치와 바퀴로 콘크리트 바닥에 서 있다. 중앙 기체도 하부 바퀴가 보이며, 오른쪽 뒤 기체의 지지부는 다른 날개에 일부 가려진다. 길게 뻗은 오른쪽 전경 날개는 동체에 연결되어 있고, 왼쪽의 노출된 날개 골조도 기체에 붙어 있다. 천장등은 매달려 있고 모닥불은 지면에 놓여 있다. 지지 없이 떠 있는 물체나 불가능한 동작은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.179
   },
   "violations": {
    "B": [
     "[gemini-pro] 배경 요소(비행기)를 전경 요소로 거대하게 확대하지 말라는 프레이밍 제한을 정면으로 위반함.",
     "[gemini-pro] 좌측 전경 비행기의 동체 구조가 성립하지 않고 바닥에 무너진 형태(physically impossible anatomy)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1750,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "대각선 뷰 지침 대신 레퍼런스와 유사한 정면 구도를 취한 점은 아쉬우나, 배경 피사체의 스케일 제한, 조명, 구조적 일관성 등 프롬프트의 핵심 의도를 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "배경 요소인 비행기를 전경에 거대하게 배치하지 말라는 지침을 심각하게 위반했으며, 비행기의 물리적 형태와 격납고의 건축 구조가 심하게 왜곡되었습니다.  ★위반: [gemini-pro] 배경 요소(비행기)를 전경 요소로 거대하게 확대하지 말라는 프레이밍 제한을 정면으로 위반함. / [gemini-pro] 좌측 전경 비행기의 동체 구조가 성립하지 않고 바닥에 무너진 형태(physically impossible anatomy)."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L119B05.png",
    "asset_id": "8760f640-2995-47fc-8ff2-a9b14d437e93",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace93-5fc4-776f-9c7d-7d80a2fd173b",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "era_fail::9f50c0c873dea96b": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S76sh4": {
  "input_fingerprint": "429fade5b907b7e7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 의사의 팔목을 두 손으로 꽉 움켜쥔 채 절박하게 매달리는 자세.\n\nLOCATION (lock): At the feverish child's bedside inside the harbor-town health clinic, under nighttime clinical lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient illumination appropriate to the nighttime health center keeps the strained face and overlapping hands clearly readable with sober contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic bed remains occupied for continuing treatment. 이현우: He stands in the bedside area, visibly desperate. His grimy appearance and accumulated injuries remain. 보건소 의사: He is in the bedside consultation area.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 보건소 의사 (성인 얼굴, 얼굴형 미상, 머리 모양 미상) — wearing: 오래 빨지 않아 핏자국과 오염이 남아 있는 흰색 의사 가운. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 의사의 팔목을 두 손으로 꽉 움켜쥔 채 절박하게 매달리는 자세.\n\nLOCATION (lock): At the feverish child's bedside inside the harbor-town health clinic, under nighttime clinical lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient illumination appropriate to the nighttime health center keeps the strained face and overlapping hands clearly readable with sober contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic bed remains occupied for continuing treatment. 이현우: He stands in the bedside area, visibly desperate. His grimy appearance and accumulated injuries remain. 보건소 의사: He is in the bedside consultation area.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 보건소 의사 (성인 얼굴, 얼굴형 미상, 머리 모양 미상) — wearing: 오래 빨지 않아 핏자국과 오염이 남아 있는 흰색 의사 가운. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 의사의 팔목을 두 손으로 꽉 움켜쥔 채 절박하게 매달리는 자세.\n\nLOCATION (lock): At the feverish child's bedside inside the harbor-town health clinic, under nighttime clinical lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient illumination appropriate to the nighttime health center keeps the strained face and overlapping hands clearly readable with sober contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic bed remains occupied for continuing treatment. 이현우: He stands in the bedside area, visibly desperate. His grimy appearance and accumulated injuries remain. 보건소 의사: He is in the bedside consultation area.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 보건소 의사 (성인 얼굴, 얼굴형 미상, 머리 모양 미상) — wearing: 오래 빨지 않아 핏자국과 오염이 남아 있는 흰색 의사 가운. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "두 인물이 서로의 얼굴을 마주보며 시선이 일치함.",
    "built_space": "참조 이미지의 병실 구조(창문, 침대, 벽면 콘센트)와 잘 일치함.",
    "entities": "현우의 외모와 의상은 참조와 일치. 의사의 피 묻은 가운 표현됨. 배경의 아이는 자세가 정면으로 바뀌고 팔이 이불 속에 가려져 참조 요건에 어긋남.",
    "hard_violations": [
     "[gemini-pro] 읽을 수 있는 텍스트 존재 ('오진소 의사')",
     "[gemini-pro] 기절한 인물(아이)의 고정된 자세 임의 변경",
     "[gpt-high] 숏 텍스트가 허용한 두 남자 외에 아이의 얼굴과 상체를 노출했다.",
     "[gpt-high] 의사 가운의 가슴에 판독 가능한 한글을 넣어 읽을 수 있는 글자 금지 조건을 위반했다."
    ],
    "physics": "현우가 두 손으로 의사의 팔목을 잡고 있으며 물리적 지지가 자연스러움."
   },
   {
    "label": "B",
    "direction": "두 인물이 서로의 눈을 강하게 응시함.",
    "built_space": "병실 내부이나, 참조에 없던 천장 조명과 수액 걸이가 추가됨.",
    "entities": "현우의 외모는 일치함. 배경에 있는 아이의 크기가 비정상적으로 작아지고 침대 하단으로 위치가 변경됨.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (손가락 융합 및 기형)",
     "[gemini-pro] 기절한 인물(아이)의 크기 및 위치 임의 변경",
     "[gpt-high] 숏 텍스트가 허용한 이현우와 의사 외에 아이의 얼굴과 몸을 노출했다.",
     "[gpt-high] 고정되어야 하는 침대와 움직일 수 없는 아이의 위치·방향을 참고의 창가 머리맡 배치에서 뒤쪽 벽 방향으로 바꿨다."
    ],
    "physics": "서로 맞잡은 손이 얽히면서 손가락이 살덩이처럼 융합되어 물리적인 파지가 불가능한 형태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "가운에 읽을 수 있는 텍스트('오진소 의사')가 포함되었고, 고정되어야 할 아이의 자세가 임의로 변경되어 실격입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "두 인물이 맞잡은 손가락이 융합된 물리적으로 불가능한 해부학적 오류가 발생했으며, 아이의 크기가 비정상적으로 축소되어 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 서로의 얼굴을 마주보며 시선이 일치함.",
        "built_space": "참조 이미지의 병실 구조(창문, 침대, 벽면 콘센트)와 잘 일치함.",
        "entities": "현우의 외모와 의상은 참조와 일치. 의사의 피 묻은 가운 표현됨. 배경의 아이는 자세가 정면으로 바뀌고 팔이 이불 속에 가려져 참조 요건에 어긋남.",
        "hard_violations": [
         "읽을 수 있는 텍스트 존재 ('오진소 의사')",
         "기절한 인물(아이)의 고정된 자세 임의 변경"
        ],
        "physics": "현우가 두 손으로 의사의 팔목을 잡고 있으며 물리적 지지가 자연스러움."
       },
       {
        "label": "B",
        "direction": "두 인물이 서로의 눈을 강하게 응시함.",
        "built_space": "병실 내부이나, 참조에 없던 천장 조명과 수액 걸이가 추가됨.",
        "entities": "현우의 외모는 일치함. 배경에 있는 아이의 크기가 비정상적으로 작아지고 침대 하단으로 위치가 변경됨.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (손가락 융합 및 기형)",
         "기절한 인물(아이)의 크기 및 위치 임의 변경"
        ],
        "physics": "서로 맞잡은 손이 얽히면서 손가락이 살덩이처럼 융합되어 물리적인 파지가 불가능한 형태임."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "가운에 읽을 수 있는 텍스트('오진소 의사')가 포함되었고, 고정되어야 할 아이의 자세가 임의로 변경되어 실격입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "두 인물이 맞잡은 손가락이 융합된 물리적으로 불가능한 해부학적 오류가 발생했으며, 아이의 크기가 비정상적으로 축소되어 실격입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 서로의 얼굴을 마주보며 시선이 일치함.",
        "built_space": "참조 이미지의 병실 구조(창문, 침대, 벽면 콘센트)와 잘 일치함.",
        "entities": "현우의 외모와 의상은 참조와 일치. 의사의 피 묻은 가운 표현됨. 배경의 아이는 자세가 정면으로 바뀌고 팔이 이불 속에 가려져 참조 요건에 어긋남.",
        "hard_violations": [
         "읽을 수 있는 텍스트 존재 ('오진소 의사')",
         "기절한 인물(아이)의 고정된 자세 임의 변경"
        ],
        "physics": "현우가 두 손으로 의사의 팔목을 잡고 있으며 물리적 지지가 자연스러움."
       },
       {
        "label": "B",
        "direction": "두 인물이 서로의 눈을 강하게 응시함.",
        "built_space": "병실 내부이나, 참조에 없던 천장 조명과 수액 걸이가 추가됨.",
        "entities": "현우의 외모는 일치함. 배경에 있는 아이의 크기가 비정상적으로 작아지고 침대 하단으로 위치가 변경됨.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (손가락 융합 및 기형)",
         "기절한 인물(아이)의 크기 및 위치 임의 변경"
        ],
        "physics": "서로 맞잡은 손이 얽히면서 손가락이 살덩이처럼 융합되어 물리적인 파지가 불가능한 형태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "미디엄 숏과 외형은 대체로 맞지만, 손목에 매달리기보다 서로 손을 맞잡는 동작이며 금지된 아이를 노출하고 침상 배치도 바꿨다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "두 손으로 의사의 한쪽 손목을 붙들고 절박하게 기울이는 핵심 동작은 더 정확하지만, 아이의 노출과 읽히는 가운 글자로 최종 사용에는 부적합하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 이현우는 오른쪽 의사의 얼굴을 바라보고 의사도 그를 마주 본다. 손들은 두 사람 사이 가슴 높이에 모여 있지만, 이현우가 의사의 한쪽 손목을 두 손으로 붙드는 모습보다는 의사가 이현우의 손목을 감싸며 서로 손을 맞잡는 모습이다.",
        "built_space": "벽등 하나, 벽면 의료 연결 패널 하나, 벽 게시물 두 장, 창과 커튼, 서랍장 하나, 침대 하나, 수액대 하나가 보인다. 두 남자는 침대 앞에 서 있다. 벽의 투톤 도장과 야간 창은 유사하지만, 침대 머리가 참고의 창 아래가 아니라 뒤쪽 벽과 서랍장 쪽으로 옮겨진 배치다. 창에 비친 천장등은 현재 시점에서 가능한 반사다.",
        "entities": "이현우는 검은 헝클어진 머리의 젊은 동아시아계 남성으로, 참고와 유사한 얼굴과 오염된 짙은 후드티, 얼굴 상처를 유지한다. 의사는 성인 남성이며 핏자국과 때가 묻은 흰 가운을 입었다. 그러나 두 사람 외에 금발의 누운 아이가 분명히 등장한다. 판독 가능한 문구는 뚜렷하지 않다.",
        "hard_violations": [
         "숏 텍스트가 허용한 이현우와 의사 외에 아이의 얼굴과 몸을 노출했다.",
         "고정되어야 하는 침대와 움직일 수 없는 아이의 위치·방향을 참고의 창가 머리맡 배치에서 뒤쪽 벽 방향으로 바꿨다."
        ],
        "physics": "두 사람의 몸통은 서 있는 하체로 이어지며 발은 프레임 밖이다. 공중에 뜬 몸은 없고, 맞잡은 손에는 실제 접촉이 보인다. 다만 의사의 손이 이현우를 붙드는 상호 지지 동작이 강해 요청한 일방적인 매달림과 다르다. 아이는 베개와 매트리스에 지지되고 수액 주머니는 수액대 고리에 걸려 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 몸을 앞으로 숙여 의사의 얼굴을 올려다보고, 의사는 이현우 쪽으로 시선을 내린다. 이현우의 두 손은 아래로 뻗은 의사의 한쪽 팔목과 소매 끝을 함께 감싼다. 붙드는 대상과 간청하는 방향이 요청에 맞는다.",
        "built_space": "창과 커튼, 투톤 벽, 침대 하나의 머리판과 측면 난간, 오른쪽 서랍장 하나가 보인다. 침대 머리는 참고처럼 창 아래에 있으며 두 남자는 그 옆 상담 공간에 서 있다. 서랍장은 참고보다 화면 오른쪽에 놓여 연속성이 다소 약하지만, 침대 주변 공간과 인물 크기는 자연스럽다. 불가능한 반사는 보이지 않는다.",
        "entities": "이현우는 참고와 유사한 젊은 동아시아계 남성 얼굴, 검은 헝클어진 머리, 상처와 때, 낡은 짙은 후드티와 청바지를 유지한다. 의사는 성인 남성이고 오염과 핏자국이 남은 흰 가운을 입었다. 금발의 누운 아이가 추가로 보인다. 의사의 가슴에는 '의사' 등을 읽을 수 있는 글자가 있다.",
        "hard_violations": [
         "숏 텍스트가 허용한 두 남자 외에 아이의 얼굴과 상체를 노출했다.",
         "의사 가운의 가슴에 판독 가능한 한글을 넣어 읽을 수 있는 글자 금지 조건을 위반했다."
        ],
        "physics": "이현우는 하체를 뒤에 둔 채 상체를 기울이고 두 손으로 의사의 팔목을 잡는다. 발은 잘렸지만 몸이 떠 있다는 징후는 없으며, 팔과 손의 접촉이 당기는 동작을 설득력 있게 지지한다. 의사는 선 자세로 팔을 내주고, 아이의 머리와 몸은 베개와 침대에 받쳐져 있다. 서랍장 위 물건도 표면에 놓여 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "미디엄 숏과 외형은 대체로 맞지만, 손목에 매달리기보다 서로 손을 맞잡는 동작이며 금지된 아이를 노출하고 침상 배치도 바꿨다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "두 손으로 의사의 한쪽 손목을 붙들고 절박하게 기울이는 핵심 동작은 더 정확하지만, 아이의 노출과 읽히는 가운 글자로 최종 사용에는 부적합하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 이현우는 오른쪽 의사의 얼굴을 바라보고 의사도 그를 마주 본다. 손들은 두 사람 사이 가슴 높이에 모여 있지만, 이현우가 의사의 한쪽 손목을 두 손으로 붙드는 모습보다는 의사가 이현우의 손목을 감싸며 서로 손을 맞잡는 모습이다.",
        "built_space": "벽등 하나, 벽면 의료 연결 패널 하나, 벽 게시물 두 장, 창과 커튼, 서랍장 하나, 침대 하나, 수액대 하나가 보인다. 두 남자는 침대 앞에 서 있다. 벽의 투톤 도장과 야간 창은 유사하지만, 침대 머리가 참고의 창 아래가 아니라 뒤쪽 벽과 서랍장 쪽으로 옮겨진 배치다. 창에 비친 천장등은 현재 시점에서 가능한 반사다.",
        "entities": "이현우는 검은 헝클어진 머리의 젊은 동아시아계 남성으로, 참고와 유사한 얼굴과 오염된 짙은 후드티, 얼굴 상처를 유지한다. 의사는 성인 남성이며 핏자국과 때가 묻은 흰 가운을 입었다. 그러나 두 사람 외에 금발의 누운 아이가 분명히 등장한다. 판독 가능한 문구는 뚜렷하지 않다.",
        "hard_violations": [
         "숏 텍스트가 허용한 이현우와 의사 외에 아이의 얼굴과 몸을 노출했다.",
         "고정되어야 하는 침대와 움직일 수 없는 아이의 위치·방향을 참고의 창가 머리맡 배치에서 뒤쪽 벽 방향으로 바꿨다."
        ],
        "physics": "두 사람의 몸통은 서 있는 하체로 이어지며 발은 프레임 밖이다. 공중에 뜬 몸은 없고, 맞잡은 손에는 실제 접촉이 보인다. 다만 의사의 손이 이현우를 붙드는 상호 지지 동작이 강해 요청한 일방적인 매달림과 다르다. 아이는 베개와 매트리스에 지지되고 수액 주머니는 수액대 고리에 걸려 있다."
       },
       {
        "label": "A",
        "direction": "이현우는 몸을 앞으로 숙여 의사의 얼굴을 올려다보고, 의사는 이현우 쪽으로 시선을 내린다. 이현우의 두 손은 아래로 뻗은 의사의 한쪽 팔목과 소매 끝을 함께 감싼다. 붙드는 대상과 간청하는 방향이 요청에 맞는다.",
        "built_space": "창과 커튼, 투톤 벽, 침대 하나의 머리판과 측면 난간, 오른쪽 서랍장 하나가 보인다. 침대 머리는 참고처럼 창 아래에 있으며 두 남자는 그 옆 상담 공간에 서 있다. 서랍장은 참고보다 화면 오른쪽에 놓여 연속성이 다소 약하지만, 침대 주변 공간과 인물 크기는 자연스럽다. 불가능한 반사는 보이지 않는다.",
        "entities": "이현우는 참고와 유사한 젊은 동아시아계 남성 얼굴, 검은 헝클어진 머리, 상처와 때, 낡은 짙은 후드티와 청바지를 유지한다. 의사는 성인 남성이고 오염과 핏자국이 남은 흰 가운을 입었다. 금발의 누운 아이가 추가로 보인다. 의사의 가슴에는 '의사' 등을 읽을 수 있는 글자가 있다.",
        "hard_violations": [
         "숏 텍스트가 허용한 두 남자 외에 아이의 얼굴과 상체를 노출했다.",
         "의사 가운의 가슴에 판독 가능한 한글을 넣어 읽을 수 있는 글자 금지 조건을 위반했다."
        ],
        "physics": "이현우는 하체를 뒤에 둔 채 상체를 기울이고 두 손으로 의사의 팔목을 잡는다. 발은 잘렸지만 몸이 떠 있다는 징후는 없으며, 팔과 손의 접촉이 당기는 동작을 설득력 있게 지지한다. 의사는 선 자세로 팔을 내주고, 아이의 머리와 몸은 베개와 침대에 받쳐져 있다. 서랍장 위 물건도 표면에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.333
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.083
   },
   "violations": {
    "A": [
     "[gemini-pro] 읽을 수 있는 텍스트 존재 ('오진소 의사')",
     "[gemini-pro] 기절한 인물(아이)의 고정된 자세 임의 변경",
     "[gpt-high] 숏 텍스트가 허용한 두 남자 외에 아이의 얼굴과 상체를 노출했다.",
     "[gpt-high] 의사 가운의 가슴에 판독 가능한 한글을 넣어 읽을 수 있는 글자 금지 조건을 위반했다."
    ],
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (손가락 융합 및 기형)",
     "[gemini-pro] 기절한 인물(아이)의 크기 및 위치 임의 변경",
     "[gpt-high] 숏 텍스트가 허용한 이현우와 의사 외에 아이의 얼굴과 몸을 노출했다.",
     "[gpt-high] 고정되어야 하는 침대와 움직일 수 없는 아이의 위치·방향을 참고의 창가 머리맡 배치에서 뒤쪽 벽 방향으로 바꿨다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1083
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "가운에 읽을 수 있는 텍스트('오진소 의사')가 포함되었고, 고정되어야 할 아이의 자세가 임의로 변경되어 실격입니다.  ★위반: [gemini-pro] 읽을 수 있는 텍스트 존재 ('오진소 의사') / [gemini-pro] 기절한 인물(아이)의 고정된 자세 임의 변경 / [gpt-high] 숏 텍스트가 허용한 두 남자 외에 아이의 얼굴과 상체를 노출했다. / [gpt-high] 의사 가운의 가슴에 판독 가능한 한글을 넣어 읽을 수 있는 글자 금지 조건을 위반했다."
   },
   {
    "label": "B",
    "score": 1083,
    "verdict_ko": "두 인물이 맞잡은 손가락이 융합된 물리적으로 불가능한 해부학적 오류가 발생했으며, 아이의 크기가 비정상적으로 축소되어 실격입니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (손가락 융합 및 기형) / [gemini-pro] 기절한 인물(아이)의 크기 및 위치 임의 변경 / [gpt-high] 숏 텍스트가 허용한 이현우와 의사 외에 아이의 얼굴과 몸을 노출했다. / [gpt-high] 고정되어야 하는 침대와 움직일 수 없는 아이의 위치·방향을 참고의 창가 머리맡 배치에서 뒤쪽 벽 방향으로 바꿨다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S73sh2_sel.png",
    "asset_id": "7c758b18-ac14-481b-af61-1f3558ffcc4a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacea3-b68d-7d93-b4b6-7d256903bc28",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S73sh2"
  }
 },
 "S76sh9": {
  "input_fingerprint": "a1f414c416635f7b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리가 두 주먹을 가슴 앞에 불끈 쥐고 두 눈에서 푸른빛이 강하게 빛나는 결연한 자태.\n\nLOCATION (lock): In the bedside gathering area of the harbor-town clinic room, lit by ordinary nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The strong blue light in 찰리's eyes provides the selective color accent against restrained health-center ambient illumination, with no HUD or subjective overlay.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed and nighttime treatment setting remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리가 두 주먹을 가슴 앞에 불끈 쥐고 두 눈에서 푸른빛이 강하게 빛나는 결연한 자태.\n\nLOCATION (lock): In the bedside gathering area of the harbor-town clinic room, lit by ordinary nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The strong blue light in 찰리's eyes provides the selective color accent against restrained health-center ambient illumination, with no HUD or subjective overlay.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed and nighttime treatment setting remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리가 두 주먹을 가슴 앞에 불끈 쥐고 두 눈에서 푸른빛이 강하게 빛나는 결연한 자태.\n\nLOCATION (lock): In the bedside gathering area of the harbor-town clinic room, lit by ordinary nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The strong blue light in 찰리's eyes provides the selective color accent against restrained health-center ambient illumination, with no HUD or subjective overlay.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed and nighttime treatment setting remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "로봇이 카메라를 정면으로 응시하며 가슴 앞에서 두 주먹을 쥐고 있음.",
    "built_space": "병실 내부. 오른쪽에 야경이 보이는 창문, 왼쪽에 의료용 가스 패널과 수납장이 있고, 배경 중앙에 빈 병원 침대가 배치됨.",
    "entities": "찰리는 둥글고 각진 외장의 기계 머리, 흰 얼굴판, 점 형태의 푸른 눈, 선 형태의 입을 가짐. 다른 인물은 없음.",
    "hard_violations": [],
    "physics": "로봇이 바닥에 안정적으로 서 있으며, 팔의 관절이 꽉 쥔 주먹의 자세를 자연스럽게 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "로봇이 정면을 향해 서서 가슴 앞에 두 주먹을 쥐고 카메라를 응시함.",
    "built_space": "병실 내부. 배경에 야경이 보이는 창문이 넓게 있고, 전경 좌측에 병원 침대가 위치함.",
    "entities": "찰리는 고릴라형 기계 머리와 흰 얼굴판, 푸른 눈을 가졌으나 레퍼런스 이미지의 인물과 동일한 후드티를 입고 있음. 침대 위에는 머리 없이 팔만 보임.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스 이미지 인물의 의상(후드티)을 그대로 복사하여 로봇에게 입힘",
     "[gemini-pro] 침대 위에 머리 없이 팔만 존재하는 물리적으로 불가능한 신체가 나타남",
     "[gpt-high] 찰리 외에는 등장할 수 없는데도 침대 위에 다른 사람의 맨팔과 손이 보여, 이전 장면의 인체 일부가 남았다."
    ],
    "physics": "로봇은 안정적으로 서 있으나, 침대 위의 팔은 몸통과 연결되지 않은 채 허공이나 이불 밑에서 부자연스럽게 나와 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 의상을 배제하는 규칙을 지켰으며, 찰리의 기계적 외형과 결연한 자세를 프롬프트에 맞게 잘 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 머리 조형은 우수하나, 레퍼런스의 의상(후드티)을 그대로 복사해 온 점과 침대 위의 기형적 신체가 치명적인 규칙 위반임."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇이 카메라를 정면으로 응시하며 가슴 앞에서 두 주먹을 쥐고 있음.",
        "built_space": "병실 내부. 오른쪽에 야경이 보이는 창문, 왼쪽에 의료용 가스 패널과 수납장이 있고, 배경 중앙에 빈 병원 침대가 배치됨.",
        "entities": "찰리는 둥글고 각진 외장의 기계 머리, 흰 얼굴판, 점 형태의 푸른 눈, 선 형태의 입을 가짐. 다른 인물은 없음.",
        "hard_violations": [],
        "physics": "로봇이 바닥에 안정적으로 서 있으며, 팔의 관절이 꽉 쥔 주먹의 자세를 자연스럽게 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "로봇이 정면을 향해 서서 가슴 앞에 두 주먹을 쥐고 카메라를 응시함.",
        "built_space": "병실 내부. 배경에 야경이 보이는 창문이 넓게 있고, 전경 좌측에 병원 침대가 위치함.",
        "entities": "찰리는 고릴라형 기계 머리와 흰 얼굴판, 푸른 눈을 가졌으나 레퍼런스 이미지의 인물과 동일한 후드티를 입고 있음. 침대 위에는 머리 없이 팔만 보임.",
        "hard_violations": [
         "레퍼런스 이미지 인물의 의상(후드티)을 그대로 복사하여 로봇에게 입힘",
         "침대 위에 머리 없이 팔만 존재하는 물리적으로 불가능한 신체가 나타남"
        ],
        "physics": "로봇은 안정적으로 서 있으나, 침대 위의 팔은 몸통과 연결되지 않은 채 허공이나 이불 밑에서 부자연스럽게 나와 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 의상을 배제하는 규칙을 지켰으며, 찰리의 기계적 외형과 결연한 자세를 프롬프트에 맞게 잘 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 머리 조형은 우수하나, 레퍼런스의 의상(후드티)을 그대로 복사해 온 점과 침대 위의 기형적 신체가 치명적인 규칙 위반임."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "로봇이 카메라를 정면으로 응시하며 가슴 앞에서 두 주먹을 쥐고 있음.",
        "built_space": "병실 내부. 오른쪽에 야경이 보이는 창문, 왼쪽에 의료용 가스 패널과 수납장이 있고, 배경 중앙에 빈 병원 침대가 배치됨.",
        "entities": "찰리는 둥글고 각진 외장의 기계 머리, 흰 얼굴판, 점 형태의 푸른 눈, 선 형태의 입을 가짐. 다른 인물은 없음.",
        "hard_violations": [],
        "physics": "로봇이 바닥에 안정적으로 서 있으며, 팔의 관절이 꽉 쥔 주먹의 자세를 자연스럽게 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "로봇이 정면을 향해 서서 가슴 앞에 두 주먹을 쥐고 카메라를 응시함.",
        "built_space": "병실 내부. 배경에 야경이 보이는 창문이 넓게 있고, 전경 좌측에 병원 침대가 위치함.",
        "entities": "찰리는 고릴라형 기계 머리와 흰 얼굴판, 푸른 눈을 가졌으나 레퍼런스 이미지의 인물과 동일한 후드티를 입고 있음. 침대 위에는 머리 없이 팔만 보임.",
        "hard_violations": [
         "레퍼런스 이미지 인물의 의상(후드티)을 그대로 복사하여 로봇에게 입힘",
         "침대 위에 머리 없이 팔만 존재하는 물리적으로 불가능한 신체가 나타남"
        ],
        "physics": "로봇은 안정적으로 서 있으나, 침대 위의 팔은 몸통과 연결되지 않은 채 허공이나 이불 밑에서 부자연스럽게 나와 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "가슴 앞의 두 주먹과 푸른 눈빛은 맞지만, 침대에 다른 사람의 팔을 남겨 찰리만 등장해야 한다는 조건을 위반하고 이전 인물의 후드 의상까지 답습했다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "찰리만 담은 미디엄 숏으로 가슴 앞의 두 주먹과 강한 푸른 눈빛을 충실히 구현했으나, 고릴라형 얼굴의 특징과 기존 병실의 구체적 일치는 다소 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 머리와 두 발광 눈은 카메라보다 약간 화면 왼쪽을 향한다. 특정 상대는 보이지 않는다. 양 주먹은 팔꿈치를 굽혀 가슴 앞에 올렸으며 무언가를 향해 뻗거나 타격하는 순간은 아니다.",
        "built_space": "뒤쪽에 병상 한 개, 가까운 발치 판 하나, 측면 난간, 큰 분할 창과 왼쪽 커튼 한 벌이 보인다. 크림색·연녹색 벽과 밤의 창밖 풍경은 참고 장소와 가깝다. 왼쪽 벽 수납함과 바닥의 금속 쓰레기통도 보인다. 찰리는 병상 옆에 서 있지만, 뒤쪽 이불 위에는 다른 사람의 맨팔과 손이 남아 있다. 창 위쪽의 실내등 반사는 이 시점에서 불가능하다고 볼 근거가 없다.",
        "entities": "찰리는 각진 금속 머리 외장, 흰 얼굴판, 고릴라처럼 돌출된 코 주변, 두 점 형태의 발광 눈, 선 형태의 입을 갖춘다. 두 손은 기계 손이며 두 주먹 모두 보인다. 낡은 짙은 후드 상의는 참고 사진의 남성이 입은 의상을 답습한다. 침대 위 맨팔은 찰리의 기계 팔과 별개인 추가 인체다. 읽을 수 있는 글자나 화면 표시 장치는 보이지 않는다.",
        "hard_violations": [
         "찰리 외에는 등장할 수 없는데도 침대 위에 다른 사람의 맨팔과 손이 보여, 이전 장면의 인체 일부가 남았다."
        ],
        "physics": "두 주먹은 손목과 팔뚝에 연결되어 있고 굽힌 팔로 지탱된다. 상체는 곧게 서 있으며 하체가 프레임 밖이라는 이유만으로 부유한다고 볼 수 없다. 뒤쪽 사람의 팔은 침구에 얹혀 지지된다. 침대와 쓰레기통은 바닥에 놓여 있고, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴과 두 푸른 눈은 거의 카메라 정면을 향하고 머리는 조금 숙여져 있다. 양 주먹은 서로 가까이 가슴 앞에 세워져 있어 결연하게 힘을 주는 동작으로 읽힌다. 공격 대상이나 별도의 응시 대상은 화면에 없다.",
        "built_space": "병상 한 개가 찰리 뒤에 있고, 화면 왼쪽에는 협탁 하나와 그 위 휴지 상자 하나·금속 용기 하나가 보인다. 오른쪽에는 야간 풍경이 보이는 창과 녹색 커튼, 왼쪽에는 병상 가림 커튼과 벽 부착 조명·의료 설비 띠가 보인다. 벽 색과 침구 재질은 참고와 유사하지만, 참고에서 확인되지 않았던 설비와 달라진 병상·창의 배치 때문에 동일 장소의 구체적 연속성은 덜 확실하다. 찰리는 병상 옆 전경에 있으며 다른 사람은 보이지 않는다.",
        "entities": "한 명의 찰리만 보이며, 흰 얼굴판과 금속 머리 외장, 두 점 형태의 푸른 발광 눈, 선 형태의 입이 있다. 다만 얼굴은 돌출된 고릴라 주둥이보다 평평한 인간형 로봇에 가깝고 머리 외장의 각도도 비교적 둥글다. 몸통과 양팔은 마모된 기계 장치로 표현되며, 이전 인물의 얼굴이나 후드 의상을 복제하지 않았다. 병상과 야간 치료실 요소가 있고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양 주먹은 손목·전완·팔꿈치 관절로 연결되어 가슴 앞에서 지탱된다. 어깨와 몸통의 연결도 보이며 자세에 명백한 불가능성은 없다. 발은 미디엄 숏 밖에 있으나 상체가 떠 있다는 징후는 없다. 휴지 상자와 금속 용기는 협탁 위에 놓여 있고 병상은 바닥에 지지된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "가슴 앞의 두 주먹과 푸른 눈빛은 맞지만, 침대에 다른 사람의 팔을 남겨 찰리만 등장해야 한다는 조건을 위반하고 이전 인물의 후드 의상까지 답습했다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "찰리만 담은 미디엄 숏으로 가슴 앞의 두 주먹과 강한 푸른 눈빛을 충실히 구현했으나, 고릴라형 얼굴의 특징과 기존 병실의 구체적 일치는 다소 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 머리와 두 발광 눈은 카메라보다 약간 화면 왼쪽을 향한다. 특정 상대는 보이지 않는다. 양 주먹은 팔꿈치를 굽혀 가슴 앞에 올렸으며 무언가를 향해 뻗거나 타격하는 순간은 아니다.",
        "built_space": "뒤쪽에 병상 한 개, 가까운 발치 판 하나, 측면 난간, 큰 분할 창과 왼쪽 커튼 한 벌이 보인다. 크림색·연녹색 벽과 밤의 창밖 풍경은 참고 장소와 가깝다. 왼쪽 벽 수납함과 바닥의 금속 쓰레기통도 보인다. 찰리는 병상 옆에 서 있지만, 뒤쪽 이불 위에는 다른 사람의 맨팔과 손이 남아 있다. 창 위쪽의 실내등 반사는 이 시점에서 불가능하다고 볼 근거가 없다.",
        "entities": "찰리는 각진 금속 머리 외장, 흰 얼굴판, 고릴라처럼 돌출된 코 주변, 두 점 형태의 발광 눈, 선 형태의 입을 갖춘다. 두 손은 기계 손이며 두 주먹 모두 보인다. 낡은 짙은 후드 상의는 참고 사진의 남성이 입은 의상을 답습한다. 침대 위 맨팔은 찰리의 기계 팔과 별개인 추가 인체다. 읽을 수 있는 글자나 화면 표시 장치는 보이지 않는다.",
        "hard_violations": [
         "찰리 외에는 등장할 수 없는데도 침대 위에 다른 사람의 맨팔과 손이 보여, 이전 장면의 인체 일부가 남았다."
        ],
        "physics": "두 주먹은 손목과 팔뚝에 연결되어 있고 굽힌 팔로 지탱된다. 상체는 곧게 서 있으며 하체가 프레임 밖이라는 이유만으로 부유한다고 볼 수 없다. 뒤쪽 사람의 팔은 침구에 얹혀 지지된다. 침대와 쓰레기통은 바닥에 놓여 있고, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴과 두 푸른 눈은 거의 카메라 정면을 향하고 머리는 조금 숙여져 있다. 양 주먹은 서로 가까이 가슴 앞에 세워져 있어 결연하게 힘을 주는 동작으로 읽힌다. 공격 대상이나 별도의 응시 대상은 화면에 없다.",
        "built_space": "병상 한 개가 찰리 뒤에 있고, 화면 왼쪽에는 협탁 하나와 그 위 휴지 상자 하나·금속 용기 하나가 보인다. 오른쪽에는 야간 풍경이 보이는 창과 녹색 커튼, 왼쪽에는 병상 가림 커튼과 벽 부착 조명·의료 설비 띠가 보인다. 벽 색과 침구 재질은 참고와 유사하지만, 참고에서 확인되지 않았던 설비와 달라진 병상·창의 배치 때문에 동일 장소의 구체적 연속성은 덜 확실하다. 찰리는 병상 옆 전경에 있으며 다른 사람은 보이지 않는다.",
        "entities": "한 명의 찰리만 보이며, 흰 얼굴판과 금속 머리 외장, 두 점 형태의 푸른 발광 눈, 선 형태의 입이 있다. 다만 얼굴은 돌출된 고릴라 주둥이보다 평평한 인간형 로봇에 가깝고 머리 외장의 각도도 비교적 둥글다. 몸통과 양팔은 마모된 기계 장치로 표현되며, 이전 인물의 얼굴이나 후드 의상을 복제하지 않았다. 병상과 야간 치료실 요소가 있고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양 주먹은 손목·전완·팔꿈치 관절로 연결되어 가슴 앞에서 지탱된다. 어깨와 몸통의 연결도 보이며 자세에 명백한 불가능성은 없다. 발은 미디엄 숏 밖에 있으나 상체가 떠 있다는 징후는 없다. 휴지 상자와 금속 용기는 협탁 위에 놓여 있고 병상은 바닥에 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.679
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.429
   },
   "violations": {
    "B": [
     "[gemini-pro] 레퍼런스 이미지 인물의 의상(후드티)을 그대로 복사하여 로봇에게 입힘",
     "[gemini-pro] 침대 위에 머리 없이 팔만 존재하는 물리적으로 불가능한 신체가 나타남",
     "[gpt-high] 찰리 외에는 등장할 수 없는데도 침대 위에 다른 사람의 맨팔과 손이 보여, 이전 장면의 인체 일부가 남았다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스의 의상을 배제하는 규칙을 지켰으며, 찰리의 기계적 외형과 결연한 자세를 프롬프트에 맞게 잘 구현함."
   },
   {
    "label": "B",
    "score": 429,
    "verdict_ko": "찰리의 머리 조형은 우수하나, 레퍼런스의 의상(후드티)을 그대로 복사해 온 점과 침대 위의 기형적 신체가 치명적인 규칙 위반임.  ★위반: [gemini-pro] 레퍼런스 이미지 인물의 의상(후드티)을 그대로 복사하여 로봇에게 입힘 / [gemini-pro] 침대 위에 머리 없이 팔만 존재하는 물리적으로 불가능한 신체가 나타남 / [gpt-high] 찰리 외에는 등장할 수 없는데도 침대 위에 다른 사람의 맨팔과 손이 보여, 이전 장면의 인체 일부가 남았다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S76sh4_sel.png",
    "asset_id": "c50d8763-3c7f-47b1-8b82-8728b399933e",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacea7-d4ba-701e-9fc7-6db755c5146f",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S76sh4"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S77sh18": {
  "input_fingerprint": "596bbfee08091899",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리가 두 주먹을 허리춤에 얹고 가슴을 쫙 편 채 위풍당당한 슈퍼히어로 자세를 취한 전신.\n\nLOCATION (lock): On the hangar floor beside the restarted light aircraft, under electric bulbs energized during the repair. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Restarted light aircraft (Engines starting, with black exhaust issuing from multiple aircraft) — Oblique bodies and wing sections are separated across the background; used as Provide visible evidence of the repair's success while preserving 찰리's full-body silhouette; Warehouse floor (Visible beneath 찰리 and between the aircraft) — Recedes from his planted feet into the aircraft spacing; used as Grounds the superhero stance and leaves an unobstructed area for the approaching companions.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The warehouse's established electric illumination and black engine exhaust support a weighty nighttime reveal while keeping 찰리's triumphant outline distinct.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several formerly inoperative light aircraft now have their engines running and emit dark exhaust inside the hangar. At least one airport bulb has burst during the power surge.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리가 두 주먹을 허리춤에 얹고 가슴을 쫙 편 채 위풍당당한 슈퍼히어로 자세를 취한 전신.\n\nLOCATION (lock): On the hangar floor beside the restarted light aircraft, under electric bulbs energized during the repair. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Restarted light aircraft (Engines starting, with black exhaust issuing from multiple aircraft) — Oblique bodies and wing sections are separated across the background; used as Provide visible evidence of the repair's success while preserving 찰리's full-body silhouette; Warehouse floor (Visible beneath 찰리 and between the aircraft) — Recedes from his planted feet into the aircraft spacing; used as Grounds the superhero stance and leaves an unobstructed area for the approaching companions.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The warehouse's established electric illumination and black engine exhaust support a weighty nighttime reveal while keeping 찰리's triumphant outline distinct.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several formerly inoperative light aircraft now have their engines running and emit dark exhaust inside the hangar. At least one airport bulb has burst during the power surge.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리가 두 주먹을 허리춤에 얹고 가슴을 쫙 편 채 위풍당당한 슈퍼히어로 자세를 취한 전신.\n\nLOCATION (lock): On the hangar floor beside the restarted light aircraft, under electric bulbs energized during the repair. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Restarted light aircraft (Engines starting, with black exhaust issuing from multiple aircraft) — Oblique bodies and wing sections are separated across the background; used as Provide visible evidence of the repair's success while preserving 찰리's full-body silhouette; Warehouse floor (Visible beneath 찰리 and between the aircraft) — Recedes from his planted feet into the aircraft spacing; used as Grounds the superhero stance and leaves an unobstructed area for the approaching companions.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The warehouse's established electric illumination and black engine exhaust support a weighty nighttime reveal while keeping 찰리's triumphant outline distinct.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several formerly inoperative light aircraft now have their engines running and emit dark exhaust inside the hangar. At least one airport bulb has burst during the power surge.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리는 카메라를 정면으로 응시하며 서 있고, 배기가스는 위로 향함.",
    "built_space": "참조 이미지의 격납고 구조, 비행기 위치, 켜진 전구가 정확히 일치함.",
    "entities": "찰리는 고릴라형 기계 머리, 흰 얼굴판, 점 눈, 선 형태의 입을 지니고 서 있음.",
    "hard_violations": [],
    "physics": "찰리의 두 발이 격납고 바닥에 안정적으로 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "찰리는 카메라를 향해 서 있고, 배기가스는 대각선 위로 솟아오름.",
    "built_space": "격납고 내부 구조와 비행기 배치가 참조 이미지와 동일하게 구현됨.",
    "entities": "찰리의 외형은 조건에 맞으나 눈이 빛나고 있으며, 배경 좌측에 미상의 인물 두 명이 서 있음.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 지시되지 않은 두 명의 인물이 배경에 추가됨 (invented people)",
     "[gpt-high] 찰리만 보여야 하는 장면에 격납고 내부의 배경 인물 두 명을 추가했다."
    ],
    "physics": "찰리는 바닥에 서 있으며, 왼쪽 비행기의 프로펠러가 회전 중임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리의 기계 머리 외형과 슈퍼히어로 자세를 정확히 구현했으며, 참조된 배경과 조명을 위반 없이 유지함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에 없는 인물들이 배경에 추가되어 핵심 지시사항을 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 카메라를 정면으로 응시하며 서 있고, 배기가스는 위로 향함.",
        "built_space": "참조 이미지의 격납고 구조, 비행기 위치, 켜진 전구가 정확히 일치함.",
        "entities": "찰리는 고릴라형 기계 머리, 흰 얼굴판, 점 눈, 선 형태의 입을 지니고 서 있음.",
        "hard_violations": [],
        "physics": "찰리의 두 발이 격납고 바닥에 안정적으로 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "찰리는 카메라를 향해 서 있고, 배기가스는 대각선 위로 솟아오름.",
        "built_space": "격납고 내부 구조와 비행기 배치가 참조 이미지와 동일하게 구현됨.",
        "entities": "찰리의 외형은 조건에 맞으나 눈이 빛나고 있으며, 배경 좌측에 미상의 인물 두 명이 서 있음.",
        "hard_violations": [
         "프롬프트에 지시되지 않은 두 명의 인물이 배경에 추가됨 (invented people)"
        ],
        "physics": "찰리는 바닥에 서 있으며, 왼쪽 비행기의 프로펠러가 회전 중임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리의 기계 머리 외형과 슈퍼히어로 자세를 정확히 구현했으며, 참조된 배경과 조명을 위반 없이 유지함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에 없는 인물들이 배경에 추가되어 핵심 지시사항을 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 카메라를 정면으로 응시하며 서 있고, 배기가스는 위로 향함.",
        "built_space": "참조 이미지의 격납고 구조, 비행기 위치, 켜진 전구가 정확히 일치함.",
        "entities": "찰리는 고릴라형 기계 머리, 흰 얼굴판, 점 눈, 선 형태의 입을 지니고 서 있음.",
        "hard_violations": [],
        "physics": "찰리의 두 발이 격납고 바닥에 안정적으로 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "찰리는 카메라를 향해 서 있고, 배기가스는 대각선 위로 솟아오름.",
        "built_space": "격납고 내부 구조와 비행기 배치가 참조 이미지와 동일하게 구현됨.",
        "entities": "찰리의 외형은 조건에 맞으나 눈이 빛나고 있으며, 배경 좌측에 미상의 인물 두 명이 서 있음.",
        "hard_violations": [
         "프롬프트에 지시되지 않은 두 명의 인물이 배경에 추가됨 (invented people)"
        ],
        "physics": "찰리는 바닥에 서 있으며, 왼쪽 비행기의 프로펠러가 회전 중임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "전신의 당당한 자세는 맞지만, 격납고 안에 허용되지 않은 두 사람을 추가한 것이 결정적인 실격 사유다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "찰리만 등장하는 전신 와이드 숏과 허리의 두 주먹, 복수 항공기의 검은 배기를 충실히 구현했지만, 머리 외장은 다소 둥글고 참고 사진의 입구 구도를 거의 반복했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 몸은 카메라를 향하고 머리와 시선은 화면 오른쪽을 향한다. 두 주먹은 양쪽 허리에 닿고 팔꿈치는 바깥으로 벌어져 있다. 특정 응시 대상은 지정되지 않았으므로 시선 방향 자체는 문제없다. 항공기들은 중앙 통로 양옆에서 비스듬히 앞쪽을 향한다.",
        "built_space": "참고 사진과 같은 녹슨 출입구, 철골 박공지붕, 주름진 금속 벽과 마모된 콘크리트 바닥이 보인다. 항공기는 앞쪽 좌우 두 대와 뒤쪽 두 대가 구분된다. 갓이 있는 점등 조명 다섯 곳과 중앙 상부의 불꽃 같은 발광부 한 곳이 보인다. 찰리는 오른쪽 전경의 빈 바닥에 서 있으나, 왼쪽 뒤 항공기 사이에 두 사람이 추가되어 있다. 카메라는 참고 사진의 입구 시점을 거의 그대로 따른다.",
        "entities": "찰리는 흰 얼굴판, 고릴라를 닮은 기계 머리, 각진 검은 외장과 선 모양 입을 갖췄다. 눈은 작은 검은 점보다는 밝게 빛나는 눈처럼 보인다. 작업복과 장갑을 착용했으며 참고 사진에는 인물이 없어 의상 연속성을 직접 대조할 수 없다. 허용되지 않은 배경 인물 두 명이 명확하다. 여러 경비행기와 배기 연무는 있지만 검은 배기가 각 엔진에서 나오는 관계는 B보다 불명확하다. 중앙의 불꽃은 전구 파손을 암시하나 깨진 전구 자체는 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "찰리만 보여야 하는 장면에 격납고 내부의 배경 인물 두 명을 추가했다."
        ],
        "physics": "찰리는 양발을 콘크리트 바닥에 붙이고 다리를 벌려 체중을 지탱한다. 주먹은 허리에 접촉하며 가슴을 편 자세도 실현 가능하다. 항공기들은 착륙장치와 바퀴로 지지되고 조명은 천장에 매달려 있다. 배경 인물들도 바닥에 서 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 몸과 흰 얼굴판은 카메라 정면을 향한다. 두 주먹은 양쪽 허리춤에 놓이고 팔꿈치는 좌우로 벌어져 있어 요청한 슈퍼히어로 자세로 읽힌다. 뒤쪽 항공기 두 대의 검은 배기는 엔진 부근에서 위쪽으로 퍼진다. 별도로 겨냥하거나 바라봐야 할 대상은 없다.",
        "built_space": "녹슨 출입구 양쪽 기둥, 철골 박공지붕, 금속 벽, 닳은 콘크리트 바닥과 항공기 네 대의 전후 배치가 참고 장소와 대응한다. 천장 조명은 앞쪽 좌우 두 곳, 중간 좌우 두 곳, 중앙 깊이 방향 두 곳이 보이며 오른쪽 앞 조명에는 검게 손상된 부분이 있다. 찰리는 중앙 통로의 빈 바닥에 서고 날개들은 뒤쪽으로 분리되어 전신을 가리지 않는다. 다만 출입구에서 안을 보는 구도는 참고 사진과 거의 같다.",
        "entities": "등장인물은 찰리 한 명뿐이다. 고릴라형 머리, 흰 얼굴판, 검은 점 형태의 눈과 선 형태의 입이 확인된다. 머리 외장은 각진 기계 상자보다는 둥근 금속 투구에 가까운 부분이 있다. 어두운 재킷과 바지를 착용했으며 참고 사진에 인물이 없어 기존 의상과의 충돌은 확인되지 않는다. 경비행기 여러 대, 서로 다른 두 엔진 부근의 짙은 검은 배기, 켜진 전등과 손상된 것으로 보이는 오른쪽 앞 전등이 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "찰리는 두 발을 바닥에 단단히 딛고 서 있으며 양손은 허리와 접촉한다. 팔꿈치를 벌리고 상체를 세운 자세는 물리적으로 가능하다. 항공기들은 바퀴와 착륙장치로 바닥에 지지된다. 뒤쪽 왼편 프로펠러에는 회전 흐림이 보이고 배기는 엔진 부근에서 상승한다. 천장 조명에도 매달린 지지 구조가 있으며 지지 없이 떠 있는 물체나 인물은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "전신의 당당한 자세는 맞지만, 격납고 안에 허용되지 않은 두 사람을 추가한 것이 결정적인 실격 사유다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "찰리만 등장하는 전신 와이드 숏과 허리의 두 주먹, 복수 항공기의 검은 배기를 충실히 구현했지만, 머리 외장은 다소 둥글고 참고 사진의 입구 구도를 거의 반복했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 몸은 카메라를 향하고 머리와 시선은 화면 오른쪽을 향한다. 두 주먹은 양쪽 허리에 닿고 팔꿈치는 바깥으로 벌어져 있다. 특정 응시 대상은 지정되지 않았으므로 시선 방향 자체는 문제없다. 항공기들은 중앙 통로 양옆에서 비스듬히 앞쪽을 향한다.",
        "built_space": "참고 사진과 같은 녹슨 출입구, 철골 박공지붕, 주름진 금속 벽과 마모된 콘크리트 바닥이 보인다. 항공기는 앞쪽 좌우 두 대와 뒤쪽 두 대가 구분된다. 갓이 있는 점등 조명 다섯 곳과 중앙 상부의 불꽃 같은 발광부 한 곳이 보인다. 찰리는 오른쪽 전경의 빈 바닥에 서 있으나, 왼쪽 뒤 항공기 사이에 두 사람이 추가되어 있다. 카메라는 참고 사진의 입구 시점을 거의 그대로 따른다.",
        "entities": "찰리는 흰 얼굴판, 고릴라를 닮은 기계 머리, 각진 검은 외장과 선 모양 입을 갖췄다. 눈은 작은 검은 점보다는 밝게 빛나는 눈처럼 보인다. 작업복과 장갑을 착용했으며 참고 사진에는 인물이 없어 의상 연속성을 직접 대조할 수 없다. 허용되지 않은 배경 인물 두 명이 명확하다. 여러 경비행기와 배기 연무는 있지만 검은 배기가 각 엔진에서 나오는 관계는 B보다 불명확하다. 중앙의 불꽃은 전구 파손을 암시하나 깨진 전구 자체는 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "찰리만 보여야 하는 장면에 격납고 내부의 배경 인물 두 명을 추가했다."
        ],
        "physics": "찰리는 양발을 콘크리트 바닥에 붙이고 다리를 벌려 체중을 지탱한다. 주먹은 허리에 접촉하며 가슴을 편 자세도 실현 가능하다. 항공기들은 착륙장치와 바퀴로 지지되고 조명은 천장에 매달려 있다. 배경 인물들도 바닥에 서 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 몸과 흰 얼굴판은 카메라 정면을 향한다. 두 주먹은 양쪽 허리춤에 놓이고 팔꿈치는 좌우로 벌어져 있어 요청한 슈퍼히어로 자세로 읽힌다. 뒤쪽 항공기 두 대의 검은 배기는 엔진 부근에서 위쪽으로 퍼진다. 별도로 겨냥하거나 바라봐야 할 대상은 없다.",
        "built_space": "녹슨 출입구 양쪽 기둥, 철골 박공지붕, 금속 벽, 닳은 콘크리트 바닥과 항공기 네 대의 전후 배치가 참고 장소와 대응한다. 천장 조명은 앞쪽 좌우 두 곳, 중간 좌우 두 곳, 중앙 깊이 방향 두 곳이 보이며 오른쪽 앞 조명에는 검게 손상된 부분이 있다. 찰리는 중앙 통로의 빈 바닥에 서고 날개들은 뒤쪽으로 분리되어 전신을 가리지 않는다. 다만 출입구에서 안을 보는 구도는 참고 사진과 거의 같다.",
        "entities": "등장인물은 찰리 한 명뿐이다. 고릴라형 머리, 흰 얼굴판, 검은 점 형태의 눈과 선 형태의 입이 확인된다. 머리 외장은 각진 기계 상자보다는 둥근 금속 투구에 가까운 부분이 있다. 어두운 재킷과 바지를 착용했으며 참고 사진에 인물이 없어 기존 의상과의 충돌은 확인되지 않는다. 경비행기 여러 대, 서로 다른 두 엔진 부근의 짙은 검은 배기, 켜진 전등과 손상된 것으로 보이는 오른쪽 앞 전등이 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "찰리는 두 발을 바닥에 단단히 딛고 서 있으며 양손은 허리와 접촉한다. 팔꿈치를 벌리고 상체를 세운 자세는 물리적으로 가능하다. 항공기들은 바퀴와 착륙장치로 바닥에 지지된다. 뒤쪽 왼편 프로펠러에는 회전 흐림이 보이고 배기는 엔진 부근에서 상승한다. 천장 조명에도 매달린 지지 구조가 있으며 지지 없이 떠 있는 물체나 인물은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.804
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.554
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트에 지시되지 않은 두 명의 인물이 배경에 추가됨 (invented people)",
     "[gpt-high] 찰리만 보여야 하는 장면에 격납고 내부의 배경 인물 두 명을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 554
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "찰리의 기계 머리 외형과 슈퍼히어로 자세를 정확히 구현했으며, 참조된 배경과 조명을 위반 없이 유지함."
   },
   {
    "label": "B",
    "score": 554,
    "verdict_ko": "프롬프트에 없는 인물들이 배경에 추가되어 핵심 지시사항을 위반함.  ★위반: [gemini-pro] 프롬프트에 지시되지 않은 두 명의 인물이 배경에 추가됨 (invented people) / [gpt-high] 찰리만 보여야 하는 장면에 격납고 내부의 배경 인물 두 명을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S75sh4_sel.png",
    "asset_id": "527ecb61-948a-4be6-8fb7-6fa4f1b1160a",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aaceac-b816-7eef-b0c0-0498e2947ba8",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S75sh4"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "era_fail::60a2115426d15f03": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S78sh19": {
  "input_fingerprint": "79cec4fe8ea66b13",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 어두운 목포 항구를 등지고 낡은 배가 거친 바다를 향해 흰 물살을 가르며 나아가는 순간의 후면 외관.\n\nLOCATION (lock): On the harbor's outward waterway at night, behind a small motor vessel leaving the quay in a white wake. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Departing boat in the upper-center of the frame, background, moves toward open sea beyond the harbor; Trailing motor wake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Old boat (Slowly departing the harbor under motor power) — Stern and one side are visible as the bow points away; used as Small receding subject with open water ahead; Motor-driven wake (White water disturbed behind the departing boat) — Extends from the stern toward the lower foreground; used as Connects the fixed viewpoint to the vessel's outward course; Harbor and sea (Dark harbor giving way to moving open water) — Harbor margins remain nearer the camera than the departing boat; used as Establishes departure rather than return.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low evening ambient light preserves the white wake against the dark harbor and moving sea without introducing visible light fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small old boat has left its berth and is moving away from the harbor under motor power, leaving disturbed water in its wake.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 크리스 (40대 초반 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 바닷물과 잦은 비에 노출되어 얼룩진 두꺼운 노란색 계열의 어부용 방수 점퍼.; 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 어두운 목포 항구를 등지고 낡은 배가 거친 바다를 향해 흰 물살을 가르며 나아가는 순간의 후면 외관.\n\nLOCATION (lock): On the harbor's outward waterway at night, behind a small motor vessel leaving the quay in a white wake. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Departing boat in the upper-center of the frame, background, moves toward open sea beyond the harbor; Trailing motor wake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Old boat (Slowly departing the harbor under motor power) — Stern and one side are visible as the bow points away; used as Small receding subject with open water ahead; Motor-driven wake (White water disturbed behind the departing boat) — Extends from the stern toward the lower foreground; used as Connects the fixed viewpoint to the vessel's outward course; Harbor and sea (Dark harbor giving way to moving open water) — Harbor margins remain nearer the camera than the departing boat; used as Establishes departure rather than return.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low evening ambient light preserves the white wake against the dark harbor and moving sea without introducing visible light fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small old boat has left its berth and is moving away from the harbor under motor power, leaving disturbed water in its wake.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 크리스 (40대 초반 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 바닷물과 잦은 비에 노출되어 얼룩진 두꺼운 노란색 계열의 어부용 방수 점퍼.; 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 어두운 목포 항구를 등지고 낡은 배가 거친 바다를 향해 흰 물살을 가르며 나아가는 순간의 후면 외관.\n\nLOCATION (lock): On the harbor's outward waterway at night, behind a small motor vessel leaving the quay in a white wake. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Departing boat in the upper-center of the frame, background, moves toward open sea beyond the harbor; Trailing motor wake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Old boat (Slowly departing the harbor under motor power) — Stern and one side are visible as the bow points away; used as Small receding subject with open water ahead; Motor-driven wake (White water disturbed behind the departing boat) — Extends from the stern toward the lower foreground; used as Connects the fixed viewpoint to the vessel's outward course; Harbor and sea (Dark harbor giving way to moving open water) — Harbor margins remain nearer the camera than the departing boat; used as Establishes departure rather than return.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low evening ambient light preserves the white wake against the dark harbor and moving sea without introducing visible light fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small old boat has left its berth and is moving away from the harbor under motor power, leaving disturbed water in its wake.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 크리스 (40대 초반 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 바닷물과 잦은 비에 노출되어 얼룩진 두꺼운 노란색 계열의 어부용 방수 점퍼.; 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "배가 화면 정중앙을 향해 직진하며 멀어지고, 항적은 바로 아래로 곧게 떨어짐.",
    "built_space": "원본에 없는 좌측 부두와 크레인을 창조하여 탁 트인 항구가 아닌 좁은 운하 형태로 변형됨.",
    "entities": "중앙에 위치한 낡은 배와 그 뒤로 이어지는 하얀 물살이 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 존재하지 않는 좌측 부두와 대형 구조물을 임의로 추가하여 장소를 왜곡함",
     "[gpt-high] 참조에 없는 좌우 돌출 부두 두 개와 정면을 가로막는 맞은편 안벽을 도입하여, 고정된 장소의 구조를 다른 항만 배치로 바꿨다."
    ],
    "physics": "수면 위 배의 부력과 모터 작동에 의한 항적 발생 등 물리적 요소에 무리가 없음."
   },
   {
    "label": "B",
    "direction": "배가 부두에서 벗어나 좌측 전방을 향해 나아가며, 화면 하단으로 넓은 항적이 이어짐.",
    "built_space": "우측에 원본 사진과 일치하는 거대한 주황색 크레인과 고무 범퍼가 있는 단일 부두가 배치됨.",
    "entities": "후면과 측면이 보이는 낡은 배와 모터가 일으키는 하얀 물살이 뚜렷하게 확인됨.",
    "hard_violations": [
     "[gpt-high] 오른쪽 주황색 크레인의 가로보에 식별 가능한 로마자 문자열이 남아 있어, 읽을 수 있는 문자 금지 조건을 위반한다."
    ],
    "physics": "배가 안정적으로 부력을 유지하며 전진하고, 추진력에 맞는 현실적인 물살을 일으킴."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스의 부두 및 크레인 구조를 정확히 구현하였고, 저녁 시간대의 조명과 멀어지는 배의 구도를 성공적으로 연출함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "레퍼런스에 없는 좌측 부두를 임의로 생성하여 장소의 기본 구조를 심각하게 훼손하는 결정적 위반을 범함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "배가 화면 정중앙을 향해 직진하며 멀어지고, 항적은 바로 아래로 곧게 떨어짐.",
        "built_space": "원본에 없는 좌측 부두와 크레인을 창조하여 탁 트인 항구가 아닌 좁은 운하 형태로 변형됨.",
        "entities": "중앙에 위치한 낡은 배와 그 뒤로 이어지는 하얀 물살이 묘사됨.",
        "hard_violations": [
         "레퍼런스에 존재하지 않는 좌측 부두와 대형 구조물을 임의로 추가하여 장소를 왜곡함"
        ],
        "physics": "수면 위 배의 부력과 모터 작동에 의한 항적 발생 등 물리적 요소에 무리가 없음."
       },
       {
        "label": "B",
        "direction": "배가 부두에서 벗어나 좌측 전방을 향해 나아가며, 화면 하단으로 넓은 항적이 이어짐.",
        "built_space": "우측에 원본 사진과 일치하는 거대한 주황색 크레인과 고무 범퍼가 있는 단일 부두가 배치됨.",
        "entities": "후면과 측면이 보이는 낡은 배와 모터가 일으키는 하얀 물살이 뚜렷하게 확인됨.",
        "hard_violations": [],
        "physics": "배가 안정적으로 부력을 유지하며 전진하고, 추진력에 맞는 현실적인 물살을 일으킴."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스의 부두 및 크레인 구조를 정확히 구현하였고, 저녁 시간대의 조명과 멀어지는 배의 구도를 성공적으로 연출함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "레퍼런스에 없는 좌측 부두를 임의로 생성하여 장소의 기본 구조를 심각하게 훼손하는 결정적 위반을 범함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "배가 화면 정중앙을 향해 직진하며 멀어지고, 항적은 바로 아래로 곧게 떨어짐.",
        "built_space": "원본에 없는 좌측 부두와 크레인을 창조하여 탁 트인 항구가 아닌 좁은 운하 형태로 변형됨.",
        "entities": "중앙에 위치한 낡은 배와 그 뒤로 이어지는 하얀 물살이 묘사됨.",
        "hard_violations": [
         "레퍼런스에 존재하지 않는 좌측 부두와 대형 구조물을 임의로 추가하여 장소를 왜곡함"
        ],
        "physics": "수면 위 배의 부력과 모터 작동에 의한 항적 발생 등 물리적 요소에 무리가 없음."
       },
       {
        "label": "B",
        "direction": "배가 부두에서 벗어나 좌측 전방을 향해 나아가며, 화면 하단으로 넓은 항적이 이어짐.",
        "built_space": "우측에 원본 사진과 일치하는 거대한 주황색 크레인과 고무 범퍼가 있는 단일 부두가 배치됨.",
        "entities": "후면과 측면이 보이는 낡은 배와 모터가 일으키는 하얀 물살이 뚜렷하게 확인됨.",
        "hard_violations": [],
        "physics": "배가 안정적으로 부력을 유지하며 전진하고, 추진력에 맞는 현실적인 물살을 일으킴."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "항구를 뒤로하고 열린 바다로 나가는 방향과 참조 부두는 더 충실하지만, 크레인의 식별 가능한 문자와 상단 중앙보다 낮고 큰 배가 지시를 위반한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "후류는 자연스럽지만 배가 열린 바다가 아닌 맞은편 컨테이너 부두를 향하며, 참조에 없는 양쪽 돌출 부두로 장소 구조를 바꿨다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "선미가 카메라를 향하고 선수는 화면 왼쪽 위의 열린 수역을 향한다. 선미와 한쪽 옆면이 함께 보이며, 흰 후류가 선미에서 화면 아래쪽으로 이어져 카메라에서 멀어지는 출항 방향이 성립한다. 사람의 시선이나 무기는 보이지 않는다.",
        "built_space": "오른쪽에 연속된 콘크리트 안벽 하나가 있고 검은 방충재와 계선주가 반복된다. 전경에는 화면 밖으로 잘린 대형 주황색 갠트리 크레인 하나, 그 뒤에는 가느다란 크레인 두 대와 노란 하역 구조물이 보인다. 참조의 안벽 재료와 주황색 기둥·가새·케이블 릴을 대체로 유지한다. 다만 배는 상단 중앙의 작은 배경 피사체가 아니라 화면 중앙보다 아래에 놓였고, 오른쪽 크레인 보에 문자가 보인다.",
        "entities": "낡고 얼룩진 소형 작업선 한 척, 흰 모터 후류, 어두운 항구와 물결치는 바다가 보인다. 배의 마모된 선체와 선실은 요구된 오래된 배에 부합한다. 식별 가능한 인물은 없어 장면 단위 인물 목록을 불필요하게 추가하지 않았다. 하늘은 해질녘으로 읽힌다.",
        "hard_violations": [
         "오른쪽 주황색 크레인의 가로보에 식별 가능한 로마자 문자열이 남아 있어, 읽을 수 있는 문자 금지 조건을 위반한다."
        ],
        "physics": "선체 하부가 수면에 잠겨 물의 부력으로 지지된다. 선미에 연결된 교란수와 뒤로 퍼지는 후류가 모터 추진으로 전진하는 상태에 부합한다. 부두와 크레인은 각각 안벽과 지상 지지부에 놓여 있으며, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "선미가 카메라를 정면으로 향하고 후류가 아래쪽으로 이어져 배는 화면 위쪽으로 멀어진다. 그러나 그 진행 방향에 열린 바다가 아니라 컨테이너와 크레인이 늘어선 맞은편 부두가 있다. 항구를 등지고 외해로 나가는 방향으로 읽히지 않는다. 거의 정후면이어서 요청한 한쪽 옆면도 뚜렷하지 않다.",
        "built_space": "전경 좌우에 돌출된 콘크리트 부두가 각각 하나씩 있고, 배 앞에는 화면을 가로지르는 별도의 안벽이 있다. 배는 두 돌출 부두 사이에 놓였다. 배경에는 왼쪽의 붐이 누운 주황색 크레인 하나와 중앙 오른쪽의 높은 쌍붐 구조물이 보인다. 참조의 한쪽 연속 안벽과 인접한 대형 갠트리 구조 대신, 양쪽 부두와 맞은편 터미널로 둘러싸인 수역을 만들었다. 배의 위치도 상단 중앙보다는 화면 중앙에 가깝다.",
        "entities": "녹슨 소형 동력선 한 척과 흰 후류, 콘크리트 부두, 컨테이너 야적장이 보인다. 낡은 배와 저녁의 어두운 색조는 요구에 맞지만, 배 앞의 외해는 보이지 않는다. 식별 가능한 인물이나 인물 목록의 의상·기계 머리는 나타나지 않는다. 확실히 읽히는 문자는 확인되지 않는다.",
        "hard_violations": [
         "참조에 없는 좌우 돌출 부두 두 개와 정면을 가로막는 맞은편 안벽을 도입하여, 고정된 장소의 구조를 다른 항만 배치로 바꿨다."
        ],
        "physics": "배는 수면에 잠긴 선체로 부력을 받고 있으며, 선미에서 시작해 전경으로 넓어지는 후류는 동력으로 멀어지는 움직임과 일치한다. 양쪽 부두와 배경 크레인에는 지면 지지가 보인다. 지지 없는 부유물이나 물리적으로 불가능한 선체 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "항구를 뒤로하고 열린 바다로 나가는 방향과 참조 부두는 더 충실하지만, 크레인의 식별 가능한 문자와 상단 중앙보다 낮고 큰 배가 지시를 위반한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "후류는 자연스럽지만 배가 열린 바다가 아닌 맞은편 컨테이너 부두를 향하며, 참조에 없는 양쪽 돌출 부두로 장소 구조를 바꿨다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "선미가 카메라를 향하고 선수는 화면 왼쪽 위의 열린 수역을 향한다. 선미와 한쪽 옆면이 함께 보이며, 흰 후류가 선미에서 화면 아래쪽으로 이어져 카메라에서 멀어지는 출항 방향이 성립한다. 사람의 시선이나 무기는 보이지 않는다.",
        "built_space": "오른쪽에 연속된 콘크리트 안벽 하나가 있고 검은 방충재와 계선주가 반복된다. 전경에는 화면 밖으로 잘린 대형 주황색 갠트리 크레인 하나, 그 뒤에는 가느다란 크레인 두 대와 노란 하역 구조물이 보인다. 참조의 안벽 재료와 주황색 기둥·가새·케이블 릴을 대체로 유지한다. 다만 배는 상단 중앙의 작은 배경 피사체가 아니라 화면 중앙보다 아래에 놓였고, 오른쪽 크레인 보에 문자가 보인다.",
        "entities": "낡고 얼룩진 소형 작업선 한 척, 흰 모터 후류, 어두운 항구와 물결치는 바다가 보인다. 배의 마모된 선체와 선실은 요구된 오래된 배에 부합한다. 식별 가능한 인물은 없어 장면 단위 인물 목록을 불필요하게 추가하지 않았다. 하늘은 해질녘으로 읽힌다.",
        "hard_violations": [
         "오른쪽 주황색 크레인의 가로보에 식별 가능한 로마자 문자열이 남아 있어, 읽을 수 있는 문자 금지 조건을 위반한다."
        ],
        "physics": "선체 하부가 수면에 잠겨 물의 부력으로 지지된다. 선미에 연결된 교란수와 뒤로 퍼지는 후류가 모터 추진으로 전진하는 상태에 부합한다. 부두와 크레인은 각각 안벽과 지상 지지부에 놓여 있으며, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "선미가 카메라를 정면으로 향하고 후류가 아래쪽으로 이어져 배는 화면 위쪽으로 멀어진다. 그러나 그 진행 방향에 열린 바다가 아니라 컨테이너와 크레인이 늘어선 맞은편 부두가 있다. 항구를 등지고 외해로 나가는 방향으로 읽히지 않는다. 거의 정후면이어서 요청한 한쪽 옆면도 뚜렷하지 않다.",
        "built_space": "전경 좌우에 돌출된 콘크리트 부두가 각각 하나씩 있고, 배 앞에는 화면을 가로지르는 별도의 안벽이 있다. 배는 두 돌출 부두 사이에 놓였다. 배경에는 왼쪽의 붐이 누운 주황색 크레인 하나와 중앙 오른쪽의 높은 쌍붐 구조물이 보인다. 참조의 한쪽 연속 안벽과 인접한 대형 갠트리 구조 대신, 양쪽 부두와 맞은편 터미널로 둘러싸인 수역을 만들었다. 배의 위치도 상단 중앙보다는 화면 중앙에 가깝다.",
        "entities": "녹슨 소형 동력선 한 척과 흰 후류, 콘크리트 부두, 컨테이너 야적장이 보인다. 낡은 배와 저녁의 어두운 색조는 요구에 맞지만, 배 앞의 외해는 보이지 않는다. 식별 가능한 인물이나 인물 목록의 의상·기계 머리는 나타나지 않는다. 확실히 읽히는 문자는 확인되지 않는다.",
        "hard_violations": [
         "참조에 없는 좌우 돌출 부두 두 개와 정면을 가로막는 맞은편 안벽을 도입하여, 고정된 장소의 구조를 다른 항만 배치로 바꿨다."
        ],
        "physics": "배는 수면에 잠긴 선체로 부력을 받고 있으며, 선미에서 시작해 전경으로 넓어지는 후류는 동력으로 멀어지는 움직임과 일치한다. 양쪽 부두와 배경 크레인에는 지면 지지가 보인다. 지지 없는 부유물이나 물리적으로 불가능한 선체 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.829,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.579,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스에 존재하지 않는 좌측 부두와 대형 구조물을 임의로 추가하여 장소를 왜곡함",
     "[gpt-high] 참조에 없는 좌우 돌출 부두 두 개와 정면을 가로막는 맞은편 안벽을 도입하여, 고정된 장소의 구조를 다른 항만 배치로 바꿨다."
    ],
    "B": [
     "[gpt-high] 오른쪽 주황색 크레인의 가로보에 식별 가능한 로마자 문자열이 남아 있어, 읽을 수 있는 문자 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 579
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "레퍼런스의 부두 및 크레인 구조를 정확히 구현하였고, 저녁 시간대의 조명과 멀어지는 배의 구도를 성공적으로 연출함.  ★위반: [gpt-high] 오른쪽 주황색 크레인의 가로보에 식별 가능한 로마자 문자열이 남아 있어, 읽을 수 있는 문자 금지 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 579,
    "verdict_ko": "레퍼런스에 없는 좌측 부두를 임의로 생성하여 장소의 기본 구조를 심각하게 훼손하는 결정적 위반을 범함.  ★위반: [gemini-pro] 레퍼런스에 존재하지 않는 좌측 부두와 대형 구조물을 임의로 추가하여 장소를 왜곡함 / [gpt-high] 참조에 없는 좌우 돌출 부두 두 개와 정면을 가로막는 맞은편 안벽을 도입하여, 고정된 장소의 구조를 다른 항만 배치로 바꿨다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_harbor_pier_sel.png",
    "asset_id": "9a426850-70d6-481f-8da7-a212290aacfa",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacecf-7f67-7125-a746-c81473ffd41f",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "era_fail::ffaee7876c805584": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S79sh8": {
  "input_fingerprint": "984ba23bbcc662c9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 폭발음에 두 눈을 번쩍 치켜뜬 이현우가 상체를 펄쩍 일으켜 몸이 반쯤 허공에 떠오른 찰나.\n\nLOCATION (lock): At the sofa inside the vessel's small sleeping cabin, in low nighttime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sofa backrest and seat (이현우's upper body has lifted away from the backrest) — Seen obliquely beneath and behind his rising torso; used as Shows the gap proving the sudden upward movement; Cabin wall (Visible beyond the sofa) — Remains behind the rising figure; used as Stable spatial reference against the handheld reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding cabin illumination unchanged, allowing pose and contrast rather than an invented explosion flash to convey alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small cabin retains its sofa and closed door. The boat remains underway at night. 이현우: He is waking from sleep on the sofa and lifting his upper body. His backpack, grime and accumulated injuries remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 폭발음에 두 눈을 번쩍 치켜뜬 이현우가 상체를 펄쩍 일으켜 몸이 반쯤 허공에 떠오른 찰나.\n\nLOCATION (lock): At the sofa inside the vessel's small sleeping cabin, in low nighttime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sofa backrest and seat (이현우's upper body has lifted away from the backrest) — Seen obliquely beneath and behind his rising torso; used as Shows the gap proving the sudden upward movement; Cabin wall (Visible beyond the sofa) — Remains behind the rising figure; used as Stable spatial reference against the handheld reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding cabin illumination unchanged, allowing pose and contrast rather than an invented explosion flash to convey alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small cabin retains its sofa and closed door. The boat remains underway at night. 이현우: He is waking from sleep on the sofa and lifting his upper body. His backpack, grime and accumulated injuries remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 폭발음에 두 눈을 번쩍 치켜뜬 이현우가 상체를 펄쩍 일으켜 몸이 반쯤 허공에 떠오른 찰나.\n\nLOCATION (lock): At the sofa inside the vessel's small sleeping cabin, in low nighttime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sofa backrest and seat (이현우's upper body has lifted away from the backrest) — Seen obliquely beneath and behind his rising torso; used as Shows the gap proving the sudden upward movement; Cabin wall (Visible beyond the sofa) — Remains behind the rising figure; used as Stable spatial reference against the handheld reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding cabin illumination unchanged, allowing pose and contrast rather than an invented explosion flash to convey alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small cabin retains its sofa and closed door. The boat remains underway at night. 이현우: He is waking from sleep on the sofa and lifting his upper body. His backpack, grime and accumulated injuries remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 좌측 밖을 향하며 폭발음에 놀란 표정을 명확히 보여줍니다.",
    "built_space": "선실의 수납장과 조명 등은 일치하나, 캐릭터 뒤편에 소파 등받이 쿠션이 본체에서 분리되어 떠 있습니다.",
    "entities": "이현우의 얼굴 특징과 지저분한 후드티, 찢어진 청바지가 레퍼런스와 정확히 일치합니다.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 배경 구조 (아무 지지 없이 허공에 떠 있는 분리된 소파 등받이 쿠션)",
     "[gpt-high] 기존 소파 등받이 앞에 참조에 없는 거대한 추가 등받이 쿠션을 세워, 유지되어야 할 소파의 고정 구조와 인물 뒤 공간을 바꾸었다."
    ],
    "physics": "오른손으로 좌석을 짚고 체중을 지지하며 상체를 일으키고 있으나, 뒤쪽의 대형 쿠션은 지지하는 힘 없이 공중에 떠 있습니다."
   },
   {
    "label": "B",
    "direction": "화면 좌측을 향해 두 눈을 크게 뜨고 응시하고 있습니다.",
    "built_space": "선실 내부의 구조물과 소파의 배치가 레퍼런스와 자연스럽게 이어집니다.",
    "entities": "이현우의 인물 묘사와 웨더링된 의상이 레퍼런스의 설정과 잘 맞습니다.",
    "hard_violations": [
     "[gemini-pro] 물리적 지지나 도약의 힘 없이 공중에 부유하는 신체 (아무것도 몸을 지탱하지 않음)"
    ],
    "physics": "오른손이 등받이에 닿아 있으나 체중을 지탱하기엔 역부족이며, 도약하는 근육의 움직임 없이 앉은 포즈 그대로 허공에 떠 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "상체를 일으키는 인물의 동작은 비교적 자연스러우나, 등받이 쿠션이 아무런 물리적 지지 없이 허공에 떠 있는 치명적인 구조적 오류가 발생했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "허공에 떠오른 상태를 묘사하려 했으나 도약의 힘이나 물리적 지지점 없이 앉은 자세 그대로 부유하고 있어 물리 법칙을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 좌측 밖을 향하며 폭발음에 놀란 표정을 명확히 보여줍니다.",
        "built_space": "선실의 수납장과 조명 등은 일치하나, 캐릭터 뒤편에 소파 등받이 쿠션이 본체에서 분리되어 떠 있습니다.",
        "entities": "이현우의 얼굴 특징과 지저분한 후드티, 찢어진 청바지가 레퍼런스와 정확히 일치합니다.",
        "hard_violations": [
         "물리적으로 불가능한 배경 구조 (아무 지지 없이 허공에 떠 있는 분리된 소파 등받이 쿠션)"
        ],
        "physics": "오른손으로 좌석을 짚고 체중을 지지하며 상체를 일으키고 있으나, 뒤쪽의 대형 쿠션은 지지하는 힘 없이 공중에 떠 있습니다."
       },
       {
        "label": "B",
        "direction": "화면 좌측을 향해 두 눈을 크게 뜨고 응시하고 있습니다.",
        "built_space": "선실 내부의 구조물과 소파의 배치가 레퍼런스와 자연스럽게 이어집니다.",
        "entities": "이현우의 인물 묘사와 웨더링된 의상이 레퍼런스의 설정과 잘 맞습니다.",
        "hard_violations": [
         "물리적 지지나 도약의 힘 없이 공중에 부유하는 신체 (아무것도 몸을 지탱하지 않음)"
        ],
        "physics": "오른손이 등받이에 닿아 있으나 체중을 지탱하기엔 역부족이며, 도약하는 근육의 움직임 없이 앉은 포즈 그대로 허공에 떠 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "상체를 일으키는 인물의 동작은 비교적 자연스러우나, 등받이 쿠션이 아무런 물리적 지지 없이 허공에 떠 있는 치명적인 구조적 오류가 발생했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "허공에 떠오른 상태를 묘사하려 했으나 도약의 힘이나 물리적 지지점 없이 앉은 자세 그대로 부유하고 있어 물리 법칙을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 좌측 밖을 향하며 폭발음에 놀란 표정을 명확히 보여줍니다.",
        "built_space": "선실의 수납장과 조명 등은 일치하나, 캐릭터 뒤편에 소파 등받이 쿠션이 본체에서 분리되어 떠 있습니다.",
        "entities": "이현우의 얼굴 특징과 지저분한 후드티, 찢어진 청바지가 레퍼런스와 정확히 일치합니다.",
        "hard_violations": [
         "물리적으로 불가능한 배경 구조 (아무 지지 없이 허공에 떠 있는 분리된 소파 등받이 쿠션)"
        ],
        "physics": "오른손으로 좌석을 짚고 체중을 지지하며 상체를 일으키고 있으나, 뒤쪽의 대형 쿠션은 지지하는 힘 없이 공중에 떠 있습니다."
       },
       {
        "label": "B",
        "direction": "화면 좌측을 향해 두 눈을 크게 뜨고 응시하고 있습니다.",
        "built_space": "선실 내부의 구조물과 소파의 배치가 레퍼런스와 자연스럽게 이어집니다.",
        "entities": "이현우의 인물 묘사와 웨더링된 의상이 레퍼런스의 설정과 잘 맞습니다.",
        "hard_violations": [
         "물리적 지지나 도약의 힘 없이 공중에 부유하는 신체 (아무것도 몸을 지탱하지 않음)"
        ],
        "physics": "오른손이 등받이에 닿아 있으나 체중을 지탱하기엔 역부족이며, 도약하는 근육의 움직임 없이 앉은 포즈 그대로 허공에 떠 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구한 미디엄 숏보다 다소 넓지만, 기존 선실과 소파를 유지하면서 손으로 버티고 상체를 갑자기 들어 올린 순간과 등받이 사이의 틈을 설득력 있게 보여준다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물 크기는 미디엄 숏에 더 가깝지만, 기존 등받이 앞에 거대한 추가 쿠션을 세워 장소의 고정 구조를 바꾸었고 동작도 솟구치기보다 앉아서 돌아보기에 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈을 크게 뜨고 화면 왼쪽 위의 프레임 밖을 바라본다. 소리의 원인은 보이지 않으며, 프롬프트도 그 위치를 지정하지 않는다. 몸통은 등받이에서 떨어져 위쪽과 앞으로 일어나고, 화면 왼쪽 손은 뒤의 등받이를 짚는다. 무기나 겨누는 물체는 없다.",
        "built_space": "회색 천 소파 한 개의 좌판과 연속된 등받이, 오른쪽의 닫힌 금속 문 한 개, 상부 책 선반과 목재 수납장, 벽의 배 사진들, 오른쪽 남색 쿠션 한 개가 보인다. 기존 선실의 재료와 배치는 대체로 유지된다. 인물은 소파 오른쪽 자리에서 일어나며 몸통 뒤와 아래로 등받이 및 좌판이 비스듬히 드러난다. 따뜻하고 낮은 실내조명이 유지되며 폭발 섬광이나 반사는 없다.",
        "entities": "이현우로 보이는 젊은 한국계 남성 한 명만 있다. 헝클어진 검은 머리와 얼굴은 참조에 가깝고, 얼굴의 오염과 상처, 낡은 짙은 후드티, 찢어진 청바지, 어깨에 멘 배낭이 유지된다. 참조의 다른 인물은 등장하지 않는다. 문에는 기존의 녹색 도형 표지가 있지만 판독 가능한 문장이나 자막은 보이지 않는다.",
        "hard_violations": [],
        "physics": "화면 왼쪽 손바닥과 손가락이 등받이에 닿아 몸을 지탱하고, 골반은 좌판에서 살짝 들린 듯 보이며 무릎은 굽혀져 있다. 발은 프레임 밖이라 바닥 접촉을 확인할 수 없지만, 손으로 소파를 밀며 몸을 일으키는 순간으로 해석 가능한 자세다. 지지 없이 공중에 누운 몸은 아니다. 배낭은 양어깨 끈으로 지지되고 반대 손은 몸 옆에서 균형을 잡는다."
       },
       {
        "label": "B",
        "direction": "눈과 얼굴은 화면 왼쪽의 프레임 밖으로 돌아가 있다. 놀란 표정은 보이지만 몸통은 약간 뒤로 기울고 골반은 좌판에 놓여 있어, 위로 솟구치는 방향성보다 앉은 채 소리 쪽으로 돌아보는 동작이 강하다. 무기나 겨누는 물체는 없다.",
        "built_space": "회색 소파 한 개, 오른쪽 닫힌 금속 문 한 개, 상부 책 선반과 수납장, 수납장 아래 조명 한 개, 벽 사진들과 남색 쿠션이 보인다. 그러나 기존의 낮은 등받이 앞에 별도의 거대한 직사각형 등받이 쿠션이 높이 세워져 벽 사진을 가린다. 이는 참조에 없던 추가 구조이며, 인물 뒤 공간도 그 쿠션이 차지해 상승을 증명할 등받이와의 틈이 불명확하다. 야간의 따뜻한 조명은 유지된다.",
        "entities": "젊은 한국계 남성 한 명이며 얼굴, 검은 헝클어진 머리, 낡은 후드티와 찢어진 청바지, 얼굴의 때와 상처는 참조에 가깝다. 배낭과 어깨 끈도 보인다. 다른 사람은 없고 읽을 수 있는 글자도 보이지 않는다. 다만 인물 뒤의 대형 직사각형 쿠션은 참조 소파에 없던 물체다.",
        "hard_violations": [
         "기존 소파 등받이 앞에 참조에 없는 거대한 추가 등받이 쿠션을 세워, 유지되어야 할 소파의 고정 구조와 인물 뒤 공간을 바꾸었다."
        ],
        "physics": "화면 아래쪽 손바닥이 좌판을 짚고 골반과 허벅지가 좌판에 놓여 있어 몸의 지지는 자연스럽다. 배낭도 어깨 끈으로 지지된다. 큰 추가 쿠션은 아래쪽이 소파에 걸리고 뒤로 기대어 있어 무지지 부유로 단정할 수는 없다. 다만 신체가 반쯤 들리는 추진이나 좌판과 골반 사이의 명확한 틈은 보이지 않아, 요구한 급작스러운 상승 순간은 약하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구한 미디엄 숏보다 다소 넓지만, 기존 선실과 소파를 유지하면서 손으로 버티고 상체를 갑자기 들어 올린 순간과 등받이 사이의 틈을 설득력 있게 보여준다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물 크기는 미디엄 숏에 더 가깝지만, 기존 등받이 앞에 거대한 추가 쿠션을 세워 장소의 고정 구조를 바꾸었고 동작도 솟구치기보다 앉아서 돌아보기에 가깝다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 눈을 크게 뜨고 화면 왼쪽 위의 프레임 밖을 바라본다. 소리의 원인은 보이지 않으며, 프롬프트도 그 위치를 지정하지 않는다. 몸통은 등받이에서 떨어져 위쪽과 앞으로 일어나고, 화면 왼쪽 손은 뒤의 등받이를 짚는다. 무기나 겨누는 물체는 없다.",
        "built_space": "회색 천 소파 한 개의 좌판과 연속된 등받이, 오른쪽의 닫힌 금속 문 한 개, 상부 책 선반과 목재 수납장, 벽의 배 사진들, 오른쪽 남색 쿠션 한 개가 보인다. 기존 선실의 재료와 배치는 대체로 유지된다. 인물은 소파 오른쪽 자리에서 일어나며 몸통 뒤와 아래로 등받이 및 좌판이 비스듬히 드러난다. 따뜻하고 낮은 실내조명이 유지되며 폭발 섬광이나 반사는 없다.",
        "entities": "이현우로 보이는 젊은 한국계 남성 한 명만 있다. 헝클어진 검은 머리와 얼굴은 참조에 가깝고, 얼굴의 오염과 상처, 낡은 짙은 후드티, 찢어진 청바지, 어깨에 멘 배낭이 유지된다. 참조의 다른 인물은 등장하지 않는다. 문에는 기존의 녹색 도형 표지가 있지만 판독 가능한 문장이나 자막은 보이지 않는다.",
        "hard_violations": [],
        "physics": "화면 왼쪽 손바닥과 손가락이 등받이에 닿아 몸을 지탱하고, 골반은 좌판에서 살짝 들린 듯 보이며 무릎은 굽혀져 있다. 발은 프레임 밖이라 바닥 접촉을 확인할 수 없지만, 손으로 소파를 밀며 몸을 일으키는 순간으로 해석 가능한 자세다. 지지 없이 공중에 누운 몸은 아니다. 배낭은 양어깨 끈으로 지지되고 반대 손은 몸 옆에서 균형을 잡는다."
       },
       {
        "label": "A",
        "direction": "눈과 얼굴은 화면 왼쪽의 프레임 밖으로 돌아가 있다. 놀란 표정은 보이지만 몸통은 약간 뒤로 기울고 골반은 좌판에 놓여 있어, 위로 솟구치는 방향성보다 앉은 채 소리 쪽으로 돌아보는 동작이 강하다. 무기나 겨누는 물체는 없다.",
        "built_space": "회색 소파 한 개, 오른쪽 닫힌 금속 문 한 개, 상부 책 선반과 수납장, 수납장 아래 조명 한 개, 벽 사진들과 남색 쿠션이 보인다. 그러나 기존의 낮은 등받이 앞에 별도의 거대한 직사각형 등받이 쿠션이 높이 세워져 벽 사진을 가린다. 이는 참조에 없던 추가 구조이며, 인물 뒤 공간도 그 쿠션이 차지해 상승을 증명할 등받이와의 틈이 불명확하다. 야간의 따뜻한 조명은 유지된다.",
        "entities": "젊은 한국계 남성 한 명이며 얼굴, 검은 헝클어진 머리, 낡은 후드티와 찢어진 청바지, 얼굴의 때와 상처는 참조에 가깝다. 배낭과 어깨 끈도 보인다. 다른 사람은 없고 읽을 수 있는 글자도 보이지 않는다. 다만 인물 뒤의 대형 직사각형 쿠션은 참조 소파에 없던 물체다.",
        "hard_violations": [
         "기존 소파 등받이 앞에 참조에 없는 거대한 추가 등받이 쿠션을 세워, 유지되어야 할 소파의 고정 구조와 인물 뒤 공간을 바꾸었다."
        ],
        "physics": "화면 아래쪽 손바닥이 좌판을 짚고 골반과 허벅지가 좌판에 놓여 있어 몸의 지지는 자연스럽다. 배낭도 어깨 끈으로 지지된다. 큰 추가 쿠션은 아래쪽이 소파에 걸리고 뒤로 기대어 있어 무지지 부유로 단정할 수는 없다. 다만 신체가 반쯤 들리는 추진이나 좌판과 골반 사이의 명확한 틈은 보이지 않아, 요구한 급작스러운 상승 순간은 약하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 배경 구조 (아무 지지 없이 허공에 떠 있는 분리된 소파 등받이 쿠션)",
     "[gpt-high] 기존 소파 등받이 앞에 참조에 없는 거대한 추가 등받이 쿠션을 세워, 유지되어야 할 소파의 고정 구조와 인물 뒤 공간을 바꾸었다."
    ],
    "B": [
     "[gemini-pro] 물리적 지지나 도약의 힘 없이 공중에 부유하는 신체 (아무것도 몸을 지탱하지 않음)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1179,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "상체를 일으키는 인물의 동작은 비교적 자연스러우나, 등받이 쿠션이 아무런 물리적 지지 없이 허공에 떠 있는 치명적인 구조적 오류가 발생했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 배경 구조 (아무 지지 없이 허공에 떠 있는 분리된 소파 등받이 쿠션) / [gpt-high] 기존 소파 등받이 앞에 참조에 없는 거대한 추가 등받이 쿠션을 세워, 유지되어야 할 소파의 고정 구조와 인물 뒤 공간을 바꾸었다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "허공에 떠오른 상태를 묘사하려 했으나 도약의 힘이나 물리적 지지점 없이 앉은 자세 그대로 부유하고 있어 물리 법칙을 심각하게 위반했습니다.  ★위반: [gemini-pro] 물리적 지지나 도약의 힘 없이 공중에 부유하는 신체 (아무것도 몸을 지탱하지 않음)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S79sh6_sel.png",
    "asset_id": "3c8e5f0b-b45d-46ab-927d-9fc3fc6a5b12",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacedd-5204-7d01-8029-0bda48596621",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S79sh6"
  }
 },
 "S79sh9": {
  "input_fingerprint": "224164cf8979c822",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 기울어지는 선원실 바닥 위로 찰리의 거대한 금속 몸통이 벽 쪽으로 쏠려 허공에 뜬 위태로운 순간.\n\nLOCATION (lock): Inside the vessel's small crew sleeping cabin, between the sofa and side wall as the room tilts under low cabin light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Wall in 찰리's displacement direction in the middle-right of the frame, background; Floor beneath the airborne torso in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Cabin floor (Tilting with the vessel beneath 찰리) — Visible diagonally across the lower frame; used as Establishes clearance below the lifted torso; Cabin wall (Directly ahead of the figures' displacement) — Its interior face lies at screen right beyond 찰리; used as Impact destination with visible intervening space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established cabin illumination and readable floor-to-body separation throughout the violent tilt.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The boat and its small cabin are sharply heeled to one side. The cabin door remains closed. The cabin interior is visibly tilted steeply to one side.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 기울어지는 선원실 바닥 위로 찰리의 거대한 금속 몸통이 벽 쪽으로 쏠려 허공에 뜬 위태로운 순간.\n\nLOCATION (lock): Inside the vessel's small crew sleeping cabin, between the sofa and side wall as the room tilts under low cabin light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Wall in 찰리's displacement direction in the middle-right of the frame, background; Floor beneath the airborne torso in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Cabin floor (Tilting with the vessel beneath 찰리) — Visible diagonally across the lower frame; used as Establishes clearance below the lifted torso; Cabin wall (Directly ahead of the figures' displacement) — Its interior face lies at screen right beyond 찰리; used as Impact destination with visible intervening space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established cabin illumination and readable floor-to-body separation throughout the violent tilt.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The boat and its small cabin are sharply heeled to one side. The cabin door remains closed. The cabin interior is visibly tilted steeply to one side.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 기울어지는 선원실 바닥 위로 찰리의 거대한 금속 몸통이 벽 쪽으로 쏠려 허공에 뜬 위태로운 순간.\n\nLOCATION (lock): Inside the vessel's small crew sleeping cabin, between the sofa and side wall as the room tilts under low cabin light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Wall in 찰리's displacement direction in the middle-right of the frame, background; Floor beneath the airborne torso in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Cabin floor (Tilting with the vessel beneath 찰리) — Visible diagonally across the lower frame; used as Establishes clearance below the lifted torso; Cabin wall (Directly ahead of the figures' displacement) — Its interior face lies at screen right beyond 찰리; used as Impact destination with visible intervening space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established cabin illumination and readable floor-to-body separation throughout the violent tilt.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The boat and its small cabin are sharply heeled to one side. The cabin door remains closed. The cabin interior is visibly tilted steeply to one side.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "로봇의 시선은 정면 아래를 향하며, 신체는 우측 문과 벽 쪽으로 기울어짐.",
    "built_space": "소파, 책이 있는 선반, 우측의 십자가 마크 금속 문 등 레퍼런스의 공간 요소가 정확히 배치됨. 공간 전체가 심하게 기울어짐.",
    "entities": "찰리: 고릴라형 기계 머리, 각진 외장, 흰 얼굴판, 점 눈과 선 입이 프롬프트와 완벽히 일치함.",
    "hard_violations": [],
    "physics": "오른손으로 소파를 짚어 체중을 지탱하고 있으며, 방의 기울기에 의해 나머지 몸통은 공중에 떠 있음."
   },
   {
    "label": "B",
    "direction": "로봇의 상반신과 시선이 우측 벽면을 명확히 향하고 있음.",
    "built_space": "좌측의 소파와 선반은 유지되었으나, 바닥에 레퍼런스에 존재하지 않는 대형 금속 통풍구와 흩어진 펜들이 추가됨.",
    "entities": "찰리: 지정된 얼굴 특징을 갖추고 있으나, 하반신이 완전히 생략되고 허리에 전선만 늘어져 있음.",
    "hard_violations": [
     "[gemini-pro] 물리 법칙 위반: 신체를 지지하거나 도약한 지점이 전혀 없이 금속 몸통이 허공에 떠 있음 (nothing supports it).",
     "[gpt-high] 몸통과 양손이 모두 공중에 있으며 이를 받치는 접촉점이 없다. 발의 밀어내기나 지지면에서 이탈하는 동작도 보이지 않아 아무것도 지지하지 않는 부유 상체로 보인다.",
     "[gpt-high] 참조에 없는 대형 격자형 바닥 설비를 주요 고정 시설로 추가했다."
    ],
    "physics": "지지하는 것이 없음(nothing supports it). 하반신 없이 무거운 금속 몸통이 허공에 수평으로 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 공간을 정확히 유지하면서, 기울어지는 방에서 소파를 짚고 허공에 뜬 찰리의 위태로운 자세와 외형을 물리적으로 설득력 있게 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 얼굴 묘사는 좋으나 바닥 구조를 임의로 변경했으며, 물리적 지지대 없이 상반신만 허공에 떠 있는 치명적인 물리 법칙 오류가 있음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 시선은 정면 아래를 향하며, 신체는 우측 문과 벽 쪽으로 기울어짐.",
        "built_space": "소파, 책이 있는 선반, 우측의 십자가 마크 금속 문 등 레퍼런스의 공간 요소가 정확히 배치됨. 공간 전체가 심하게 기울어짐.",
        "entities": "찰리: 고릴라형 기계 머리, 각진 외장, 흰 얼굴판, 점 눈과 선 입이 프롬프트와 완벽히 일치함.",
        "hard_violations": [],
        "physics": "오른손으로 소파를 짚어 체중을 지탱하고 있으며, 방의 기울기에 의해 나머지 몸통은 공중에 떠 있음."
       },
       {
        "label": "B",
        "direction": "로봇의 상반신과 시선이 우측 벽면을 명확히 향하고 있음.",
        "built_space": "좌측의 소파와 선반은 유지되었으나, 바닥에 레퍼런스에 존재하지 않는 대형 금속 통풍구와 흩어진 펜들이 추가됨.",
        "entities": "찰리: 지정된 얼굴 특징을 갖추고 있으나, 하반신이 완전히 생략되고 허리에 전선만 늘어져 있음.",
        "hard_violations": [
         "물리 법칙 위반: 신체를 지지하거나 도약한 지점이 전혀 없이 금속 몸통이 허공에 떠 있음 (nothing supports it)."
        ],
        "physics": "지지하는 것이 없음(nothing supports it). 하반신 없이 무거운 금속 몸통이 허공에 수평으로 떠 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 공간을 정확히 유지하면서, 기울어지는 방에서 소파를 짚고 허공에 뜬 찰리의 위태로운 자세와 외형을 물리적으로 설득력 있게 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 얼굴 묘사는 좋으나 바닥 구조를 임의로 변경했으며, 물리적 지지대 없이 상반신만 허공에 떠 있는 치명적인 물리 법칙 오류가 있음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 시선은 정면 아래를 향하며, 신체는 우측 문과 벽 쪽으로 기울어짐.",
        "built_space": "소파, 책이 있는 선반, 우측의 십자가 마크 금속 문 등 레퍼런스의 공간 요소가 정확히 배치됨. 공간 전체가 심하게 기울어짐.",
        "entities": "찰리: 고릴라형 기계 머리, 각진 외장, 흰 얼굴판, 점 눈과 선 입이 프롬프트와 완벽히 일치함.",
        "hard_violations": [],
        "physics": "오른손으로 소파를 짚어 체중을 지탱하고 있으며, 방의 기울기에 의해 나머지 몸통은 공중에 떠 있음."
       },
       {
        "label": "B",
        "direction": "로봇의 상반신과 시선이 우측 벽면을 명확히 향하고 있음.",
        "built_space": "좌측의 소파와 선반은 유지되었으나, 바닥에 레퍼런스에 존재하지 않는 대형 금속 통풍구와 흩어진 펜들이 추가됨.",
        "entities": "찰리: 지정된 얼굴 특징을 갖추고 있으나, 하반신이 완전히 생략되고 허리에 전선만 늘어져 있음.",
        "hard_violations": [
         "물리 법칙 위반: 신체를 지지하거나 도약한 지점이 전혀 없이 금속 몸통이 허공에 떠 있음 (nothing supports it)."
        ],
        "physics": "지지하는 것이 없음(nothing supports it). 하반신 없이 무거운 금속 몸통이 허공에 수평으로 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "오른쪽 벽을 향한 몸통과 바닥 사이 간격은 명확하지만, 지지점이나 이탈 동작 없이 몸통만 부유해 물리적 연출을 위반한다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "소파를 짚은 손이 몸통의 들림을 뒷받침하며 기울어진 선실과 닫힌 문도 유지하지만, 얼굴의 고릴라형 특징과 오른쪽 벽으로 쏠리는 방향성은 덜 명확하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리와 가슴이 화면 오른쪽 벽을 향하고, 두 주먹은 몸통 아래로 당겨져 있다. 얼굴 앞에 벽까지의 빈 공간이 있어 충돌 목적지는 분명하다. 무기나 별도의 겨냥 물체는 없다.",
        "built_space": "왼쪽에 소파 한 개와 상부 책 선반, 뒤쪽에 닫힌 문 한 개, 오른쪽에 얼룩진 패널 벽이 보인다. 바닥은 하단을 대각선으로 가로지른다. 다만 몸통 아래에 큰 격자형 바닥 설비가 새로 나타나며, 참조의 단순한 선실 바닥과 달라졌다.",
        "entities": "등장 개체는 찰리 한 대뿐이며 이전 장면의 인간은 없다. 거대한 금속 상체, 각진 고릴라형 머리, 흰 얼굴판, 작은 점눈과 선형 입은 요구에 가깝다. 바닥에는 필기구처럼 보이는 물건과 작은 부품들이 추가되어 있다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "몸통과 양손이 모두 공중에 있으며 이를 받치는 접촉점이 없다. 발의 밀어내기나 지지면에서 이탈하는 동작도 보이지 않아 아무것도 지지하지 않는 부유 상체로 보인다.",
         "참조에 없는 대형 격자형 바닥 설비를 주요 고정 시설로 추가했다."
        ],
        "physics": "두 손은 아무것도 잡거나 짚지 않고, 소파·바닥·벽 어느 곳도 몸에 접촉하지 않는다. 하체도 몸통 뒤에서 확인되지 않는다. 선실의 기울기만으로 이 수평 부유 자세를 설명할 출발점이나 착지 동작은 보이지 않는다. 바닥의 작은 물건들은 바닥에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "얼굴은 카메라 쪽을 향하며 오른쪽 벽을 직접 바라보지는 않는다. 소파를 짚은 왼손을 기준으로 몸통과 굽힌 다리가 오른쪽 통로로 뻗어 있다. 오른쪽 벽으로 넘어가는 동작으로 읽을 수 있지만, 머리는 왼쪽에 남아 있어 A보다 충돌 방향이 덜 선명하다.",
        "built_space": "왼쪽 소파 한 개, 그 위 책 선반과 사진들, 따뜻한 벽등 한 개, 뒤쪽의 닫힌 금속문 한 개가 보인다. 소파의 직물과 벽의 마모는 참조에 가깝다. 오른쪽에는 좁은 패널과 수납 선반이 보이며, 참조에서 확인되지 않은 전자레인지 같은 기기가 놓여 있다. 바닥은 하단 중앙에 드러나고 선실 전체가 기울어져 있다.",
        "entities": "찰리 한 대만 등장한다. 큰 금속 몸통과 각진 머리 외장, 흰 얼굴판, 점눈 두 개와 선형 입은 맞는다. 다만 얼굴은 고릴라형 주둥이보다 평평한 가면에 가깝다. 이전 장면의 인간이나 의상은 복제되지 않았다. 배경의 종이와 책에는 인쇄 흔적이 있으나 내용을 확실히 읽기는 어렵다.",
        "hard_violations": [],
        "physics": "화면 왼쪽 손이 소파 좌석을 짚어 몸을 지탱하고 회전시키는 접촉점이 된다. 다리는 무릎을 굽힌 채 바닥에서 떨어져 있으며, 손을 축으로 몸통이 통로 쪽으로 넘어가는 자세가 가능하다. 몸 아래에는 바닥과의 간격과 떨어질 공간이 보인다. 오른쪽 기기는 선반에 지지되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "오른쪽 벽을 향한 몸통과 바닥 사이 간격은 명확하지만, 지지점이나 이탈 동작 없이 몸통만 부유해 물리적 연출을 위반한다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "소파를 짚은 손이 몸통의 들림을 뒷받침하며 기울어진 선실과 닫힌 문도 유지하지만, 얼굴의 고릴라형 특징과 오른쪽 벽으로 쏠리는 방향성은 덜 명확하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "머리와 가슴이 화면 오른쪽 벽을 향하고, 두 주먹은 몸통 아래로 당겨져 있다. 얼굴 앞에 벽까지의 빈 공간이 있어 충돌 목적지는 분명하다. 무기나 별도의 겨냥 물체는 없다.",
        "built_space": "왼쪽에 소파 한 개와 상부 책 선반, 뒤쪽에 닫힌 문 한 개, 오른쪽에 얼룩진 패널 벽이 보인다. 바닥은 하단을 대각선으로 가로지른다. 다만 몸통 아래에 큰 격자형 바닥 설비가 새로 나타나며, 참조의 단순한 선실 바닥과 달라졌다.",
        "entities": "등장 개체는 찰리 한 대뿐이며 이전 장면의 인간은 없다. 거대한 금속 상체, 각진 고릴라형 머리, 흰 얼굴판, 작은 점눈과 선형 입은 요구에 가깝다. 바닥에는 필기구처럼 보이는 물건과 작은 부품들이 추가되어 있다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "몸통과 양손이 모두 공중에 있으며 이를 받치는 접촉점이 없다. 발의 밀어내기나 지지면에서 이탈하는 동작도 보이지 않아 아무것도 지지하지 않는 부유 상체로 보인다.",
         "참조에 없는 대형 격자형 바닥 설비를 주요 고정 시설로 추가했다."
        ],
        "physics": "두 손은 아무것도 잡거나 짚지 않고, 소파·바닥·벽 어느 곳도 몸에 접촉하지 않는다. 하체도 몸통 뒤에서 확인되지 않는다. 선실의 기울기만으로 이 수평 부유 자세를 설명할 출발점이나 착지 동작은 보이지 않는다. 바닥의 작은 물건들은 바닥에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "얼굴은 카메라 쪽을 향하며 오른쪽 벽을 직접 바라보지는 않는다. 소파를 짚은 왼손을 기준으로 몸통과 굽힌 다리가 오른쪽 통로로 뻗어 있다. 오른쪽 벽으로 넘어가는 동작으로 읽을 수 있지만, 머리는 왼쪽에 남아 있어 A보다 충돌 방향이 덜 선명하다.",
        "built_space": "왼쪽 소파 한 개, 그 위 책 선반과 사진들, 따뜻한 벽등 한 개, 뒤쪽의 닫힌 금속문 한 개가 보인다. 소파의 직물과 벽의 마모는 참조에 가깝다. 오른쪽에는 좁은 패널과 수납 선반이 보이며, 참조에서 확인되지 않은 전자레인지 같은 기기가 놓여 있다. 바닥은 하단 중앙에 드러나고 선실 전체가 기울어져 있다.",
        "entities": "찰리 한 대만 등장한다. 큰 금속 몸통과 각진 머리 외장, 흰 얼굴판, 점눈 두 개와 선형 입은 맞는다. 다만 얼굴은 고릴라형 주둥이보다 평평한 가면에 가깝다. 이전 장면의 인간이나 의상은 복제되지 않았다. 배경의 종이와 책에는 인쇄 흔적이 있으나 내용을 확실히 읽기는 어렵다.",
        "hard_violations": [],
        "physics": "화면 왼쪽 손이 소파 좌석을 짚어 몸을 지탱하고 회전시키는 접촉점이 된다. 다리는 무릎을 굽힌 채 바닥에서 떨어져 있으며, 손을 축으로 몸통이 통로 쪽으로 넘어가는 자세가 가능하다. 몸 아래에는 바닥과의 간격과 떨어질 공간이 보인다. 오른쪽 기기는 선반에 지지되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.714
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.464
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리 법칙 위반: 신체를 지지하거나 도약한 지점이 전혀 없이 금속 몸통이 허공에 떠 있음 (nothing supports it).",
     "[gpt-high] 몸통과 양손이 모두 공중에 있으며 이를 받치는 접촉점이 없다. 발의 밀어내기나 지지면에서 이탈하는 동작도 보이지 않아 아무것도 지지하지 않는 부유 상체로 보인다.",
     "[gpt-high] 참조에 없는 대형 격자형 바닥 설비를 주요 고정 시설로 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 464
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스의 공간을 정확히 유지하면서, 기울어지는 방에서 소파를 짚고 허공에 뜬 찰리의 위태로운 자세와 외형을 물리적으로 설득력 있게 구현함."
   },
   {
    "label": "B",
    "score": 464,
    "verdict_ko": "찰리의 얼굴 묘사는 좋으나 바닥 구조를 임의로 변경했으며, 물리적 지지대 없이 상반신만 허공에 떠 있는 치명적인 물리 법칙 오류가 있음.  ★위반: [gemini-pro] 물리 법칙 위반: 신체를 지지하거나 도약한 지점이 전혀 없이 금속 몸통이 허공에 떠 있음 (nothing supports it). / [gpt-high] 몸통과 양손이 모두 공중에 있으며 이를 받치는 접촉점이 없다. 발의 밀어내기나 지지면에서 이탈하는 동작도 보이지 않아 아무것도 지지하지 않는 부유 상체로 보인다. / [gpt-high] 참조에 없는 대형 격자형 바닥 설비를 주요 고정 시설로 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S79sh8_sel.png",
    "asset_id": "5eea6e2f-08e1-45c4-a266-3d573a70d01c",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacee2-c029-79e9-b6df-5672ba20c6ed",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S79sh8"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "era_fail::6c3851ac6a9e965d": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S80sh12": {
  "input_fingerprint": "8948737480beb56b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, typhoon and torrential rain.\n\nSHOT TEXT (authoritative, Korean): 물기둥 속에서 찰리의 금속 손아귀를 빠져나간 이현우의 손끝 클로즈업.\n\nLOCATION (lock): At the wave-swept edge of the vessel's exposed deck during the nighttime typhoon. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ship's edge (Being struck by the wave) — A narrow oblique portion preserves the earlier support-side orientation; used as Continuity anchor behind the separating hands; Striking wave (Surging through the hand connection as contact fails) — Water passes across the foreground and around the separating fingers; used as Partial occlusion that motivates the subsequent downward transition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the grip shot's low nighttime tonal structure as rain and the striking wave intermittently obscure the wet hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A massive wave crashes across the boat amid torrential nighttime rain. Water engulfs the deck and obscures visibility. 이현우: His outstretched hand has lost its grip, and he is being swept away into the sea. He is soaked, with his backpack and accumulated injuries still present. 찰리: He remains at the deck post with one hand braced and the other extended into the breaking water. His soaked covering, backpack, unrepaired body damage and salvaged chest component remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, typhoon and torrential rain.\n\nSHOT TEXT (authoritative, Korean): 물기둥 속에서 찰리의 금속 손아귀를 빠져나간 이현우의 손끝 클로즈업.\n\nLOCATION (lock): At the wave-swept edge of the vessel's exposed deck during the nighttime typhoon. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ship's edge (Being struck by the wave) — A narrow oblique portion preserves the earlier support-side orientation; used as Continuity anchor behind the separating hands; Striking wave (Surging through the hand connection as contact fails) — Water passes across the foreground and around the separating fingers; used as Partial occlusion that motivates the subsequent downward transition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the grip shot's low nighttime tonal structure as rain and the striking wave intermittently obscure the wet hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A massive wave crashes across the boat amid torrential nighttime rain. Water engulfs the deck and obscures visibility. 이현우: His outstretched hand has lost its grip, and he is being swept away into the sea. He is soaked, with his backpack and accumulated injuries still present. 찰리: He remains at the deck post with one hand braced and the other extended into the breaking water. His soaked covering, backpack, unrepaired body damage and salvaged chest component remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, typhoon and torrential rain.\n\nSHOT TEXT (authoritative, Korean): 물기둥 속에서 찰리의 금속 손아귀를 빠져나간 이현우의 손끝 클로즈업.\n\nLOCATION (lock): At the wave-swept edge of the vessel's exposed deck during the nighttime typhoon. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ship's edge (Being struck by the wave) — A narrow oblique portion preserves the earlier support-side orientation; used as Continuity anchor behind the separating hands; Striking wave (Surging through the hand connection as contact fails) — Water passes across the foreground and around the separating fingers; used as Partial occlusion that motivates the subsequent downward transition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the grip shot's low nighttime tonal structure as rain and the striking wave intermittently obscure the wet hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A massive wave crashes across the boat amid torrential nighttime rain. Water engulfs the deck and obscures visibility. 이현우: His outstretched hand has lost its grip, and he is being swept away into the sea. He is soaked, with his backpack and accumulated injuries still present. 찰리: He remains at the deck post with one hand braced and the other extended into the breaking water. His soaked covering, backpack, unrepaired body damage and salvaged chest component remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 금속 손은 화면 우측에서 좌측 아래를 향해 뻗어 있고, 이현우의 손은 좌측 하단에서 우측 위를 향해 뻗은 채 빠져나가고 있음. 물줄기가 전경을 가로지르며 손 사이로 쏟아짐.",
    "built_space": "프레임 하단과 우측에 걸쳐 비스듬한 선박의 난간(모서리)이 위치하며, 금속 팔이 그 위를 가로지름. 왼쪽 배경에 흐릿한 선박 구조물이 보여 역방향 앵글(Reverse angle)의 공간 배치를 이룸.",
    "entities": "찰리의 금속 팔(우측, 질감 및 마모 상태 일치), 이현우의 젖은 맨손과 어두운 소매(좌측). 두 손 모두 정상적인 5개의 손가락을 갖추고 있음.",
    "hard_violations": [],
    "physics": "금속 팔은 선박 난간에 의해 안정적으로 지탱되고 있으며, 인간의 손은 허공에서 지지력을 잃고 미끄러지는 상태임. 쏟아지는 물의 방향이 중력과 일치하게 연출됨."
   },
   {
    "label": "B",
    "direction": "찰리의 금속 손은 화면 좌측에서 우측을 향해 뻗고, 이현우의 손은 우측 하단에서 좌측 위를 향해 뻗어 있으나 거리가 멂.",
    "built_space": "화면 왼쪽에 선박의 갑판과 구조물이 있고, 오른쪽에는 파도치는 바다가 위치하여 이전 샷의 좌우 공간 배치를 그대로 유지함.",
    "entities": "찰리의 상반신, 망토, 금속 팔(이전 샷과 달리 왼팔로 렌더링됨). 이현우의 팔과 머리 일부, 거대한 파도 배경.",
    "hard_violations": [],
    "physics": "찰리는 선박 갑판 위에 서서 자세를 지탱하고 있으며, 이현우는 바다를 향해 추락하며 손을 뻗고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 최우선으로 요구한 '클로즈업' 프레이밍을 완벽하게 구현하여 두 손이 분리되는 순간에 집중했으며, 지정된 배경 요소(물줄기와 선박의 모서리)를 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물의 상반신과 거대한 파도를 넓게 보여주는 미디엄 샷으로 렌더링되어, 지정된 '클로즈업' 프레이밍 및 피사체 크기 제약을 명백히 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 금속 손은 화면 우측에서 좌측 아래를 향해 뻗어 있고, 이현우의 손은 좌측 하단에서 우측 위를 향해 뻗은 채 빠져나가고 있음. 물줄기가 전경을 가로지르며 손 사이로 쏟아짐.",
        "built_space": "프레임 하단과 우측에 걸쳐 비스듬한 선박의 난간(모서리)이 위치하며, 금속 팔이 그 위를 가로지름. 왼쪽 배경에 흐릿한 선박 구조물이 보여 역방향 앵글(Reverse angle)의 공간 배치를 이룸.",
        "entities": "찰리의 금속 팔(우측, 질감 및 마모 상태 일치), 이현우의 젖은 맨손과 어두운 소매(좌측). 두 손 모두 정상적인 5개의 손가락을 갖추고 있음.",
        "hard_violations": [],
        "physics": "금속 팔은 선박 난간에 의해 안정적으로 지탱되고 있으며, 인간의 손은 허공에서 지지력을 잃고 미끄러지는 상태임. 쏟아지는 물의 방향이 중력과 일치하게 연출됨."
       },
       {
        "label": "B",
        "direction": "찰리의 금속 손은 화면 좌측에서 우측을 향해 뻗고, 이현우의 손은 우측 하단에서 좌측 위를 향해 뻗어 있으나 거리가 멂.",
        "built_space": "화면 왼쪽에 선박의 갑판과 구조물이 있고, 오른쪽에는 파도치는 바다가 위치하여 이전 샷의 좌우 공간 배치를 그대로 유지함.",
        "entities": "찰리의 상반신, 망토, 금속 팔(이전 샷과 달리 왼팔로 렌더링됨). 이현우의 팔과 머리 일부, 거대한 파도 배경.",
        "hard_violations": [],
        "physics": "찰리는 선박 갑판 위에 서서 자세를 지탱하고 있으며, 이현우는 바다를 향해 추락하며 손을 뻗고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 최우선으로 요구한 '클로즈업' 프레이밍을 완벽하게 구현하여 두 손이 분리되는 순간에 집중했으며, 지정된 배경 요소(물줄기와 선박의 모서리)를 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물의 상반신과 거대한 파도를 넓게 보여주는 미디엄 샷으로 렌더링되어, 지정된 '클로즈업' 프레이밍 및 피사체 크기 제약을 명백히 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 금속 손은 화면 우측에서 좌측 아래를 향해 뻗어 있고, 이현우의 손은 좌측 하단에서 우측 위를 향해 뻗은 채 빠져나가고 있음. 물줄기가 전경을 가로지르며 손 사이로 쏟아짐.",
        "built_space": "프레임 하단과 우측에 걸쳐 비스듬한 선박의 난간(모서리)이 위치하며, 금속 팔이 그 위를 가로지름. 왼쪽 배경에 흐릿한 선박 구조물이 보여 역방향 앵글(Reverse angle)의 공간 배치를 이룸.",
        "entities": "찰리의 금속 팔(우측, 질감 및 마모 상태 일치), 이현우의 젖은 맨손과 어두운 소매(좌측). 두 손 모두 정상적인 5개의 손가락을 갖추고 있음.",
        "hard_violations": [],
        "physics": "금속 팔은 선박 난간에 의해 안정적으로 지탱되고 있으며, 인간의 손은 허공에서 지지력을 잃고 미끄러지는 상태임. 쏟아지는 물의 방향이 중력과 일치하게 연출됨."
       },
       {
        "label": "B",
        "direction": "찰리의 금속 손은 화면 좌측에서 우측을 향해 뻗고, 이현우의 손은 우측 하단에서 좌측 위를 향해 뻗어 있으나 거리가 멂.",
        "built_space": "화면 왼쪽에 선박의 갑판과 구조물이 있고, 오른쪽에는 파도치는 바다가 위치하여 이전 샷의 좌우 공간 배치를 그대로 유지함.",
        "entities": "찰리의 상반신, 망토, 금속 팔(이전 샷과 달리 왼팔로 렌더링됨). 이현우의 팔과 머리 일부, 거대한 파도 배경.",
        "hard_violations": [],
        "physics": "찰리는 선박 갑판 위에 서서 자세를 지탱하고 있으며, 이현우는 바다를 향해 추락하며 손을 뻗고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 손 사이의 명확한 틈과 이전 장면의 갑판·바다 방향을 살렸지만, 손끝 클로즈업치고 넓고 전경을 삼키는 물기둥이 약하다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "손의 크기는 클로즈업에 더 가깝지만 손끝이 금속 손과 여전히 겹쳐 이탈 순간이 불명확하고, 이전 장면의 지지측 화면 방향도 뒤집혔다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 금속 손은 왼쪽 갑판 쪽에서 오른쪽 아래의 이현우 손을 향한다. 이현우의 손끝은 오른쪽 아래에서 금속 손 쪽으로 뻗지만 이미 틈이 벌어져 있다. 팔이 화면 아래로 이어져 바다 쪽으로 떨어지는 관계와 맞는다. 시선은 프레임 밖이다.",
        "built_space": "녹슨 흰색 현측벽 하나가 왼쪽 아래에서 중앙 뒤로 이어지고, 갑판은 왼쪽, 바다는 오른쪽에 있다. 이전 장면의 지지측 방향과 일치한다. 손 뒤에 밝은 수직 기둥 하나, 왼쪽 갑판에 검은 계선주 하나가 보인다. 계선주는 이전 사진에서 인물에 가려졌을 수 있어 새로 추가된 구조물이라고 단정할 수 없다. 다만 현측벽과 배경이 넓게 보여 '좁은 사선 부분'이라는 요구보다 비중이 크다.",
        "entities": "찰리의 마모된 금속 손·전완과 어두운 젖은 덮개, 이현우의 젊어 보이는 맨손과 검은 젖은 소매가 보인다. 이현우의 얼굴은 사실상 제외되어 한국계 정체성과 정확한 나이는 확인할 수 없지만 보이는 손은 참조와 충돌하지 않는다. 배낭과 가슴 부품은 크롭 밖이다. 밤의 비와 거친 바다, 손 사이에서 흘러내리는 물은 있으나 손끝을 에워싸고 가리는 거대한 파도의 효과는 약하다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "금속 손은 손목과 전완, 화면 왼쪽의 상완에 연결되어 있고 사람 손도 소매 속 팔에 연결되어 있다. 허공에 분리된 신체 부위는 없다. 손가락을 펼치고 접촉을 잃은 자세는 미끄러져 놓친 직후로 가능하다. 물은 손 주변에서 아래로 떨어지고 배경 파도는 현측에 부딪친다. 전신과 찰리의 버팀손은 크롭 밖이므로 지지 상태를 추가로 판단할 수 없다."
       },
       {
        "label": "B",
        "direction": "찰리의 금속 손은 오른쪽에서 왼쪽 아래의 이현우 손을 향하고, 이현우는 왼쪽 아래에서 위로 손을 뻗는다. 인간 손끝과 금속 손가락이 상당히 겹쳐 있어 완전히 손아귀를 빠져나간 순간인지 명료하지 않다. 시선은 보이지 않는다.",
        "built_space": "녹슨 흰색 현측벽 하나가 오른쪽 아래에서 왼쪽 뒤로 이어진다. 갑판과 찰리는 오른쪽, 바다는 왼쪽으로 배치되어 이전 사진의 지지측 화면 방향이 반대다. 배경에는 밝은 기둥 하나와 검은 계선주 하나가 보이며, 명백한 중복 구조물은 없다. 손은 더 크게 잡혔지만 현측벽 역시 화면 아래와 배경의 넓은 면적을 차지한다. 카메라 각도 자체는 자유지만 연속성 기준이 되는 이전 방향은 유지하지 못했다.",
        "entities": "찰리의 젖고 긁힌 금속 손·팔과 어두운 덮개, 이현우의 맨손과 젖은 검은 소매가 보인다. 얼굴·머리·배낭·가슴 부품은 프레임 밖이므로 확인 대상이 아니다. 사람 손의 피부와 가는 형태는 젊은 이현우와 모순되지 않는다. 밤비와 현측을 타고 흐르는 물은 있으나 연결부를 덮치는 물기둥보다는 벽에서 쏟아지는 물막에 가깝다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "두 손 모두 각자의 손목과 팔에 연결되어 있어 지지 없는 부유물은 없다. 손가락이 서로 겹친 자세는 접촉이 풀리는 도중으로는 가능하지만 이미 떨어져 나갔다는 행동을 확실히 보여주지 않는다. 물은 현측 상단에서 바깥쪽 아래로 흐르므로 중력 방향은 자연스럽다. 프레임 밖 전신의 지지나 낙하 자세는 판단할 수 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "두 손 사이의 명확한 틈과 이전 장면의 갑판·바다 방향을 살렸지만, 손끝 클로즈업치고 넓고 전경을 삼키는 물기둥이 약하다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "손의 크기는 클로즈업에 더 가깝지만 손끝이 금속 손과 여전히 겹쳐 이탈 순간이 불명확하고, 이전 장면의 지지측 화면 방향도 뒤집혔다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 금속 손은 왼쪽 갑판 쪽에서 오른쪽 아래의 이현우 손을 향한다. 이현우의 손끝은 오른쪽 아래에서 금속 손 쪽으로 뻗지만 이미 틈이 벌어져 있다. 팔이 화면 아래로 이어져 바다 쪽으로 떨어지는 관계와 맞는다. 시선은 프레임 밖이다.",
        "built_space": "녹슨 흰색 현측벽 하나가 왼쪽 아래에서 중앙 뒤로 이어지고, 갑판은 왼쪽, 바다는 오른쪽에 있다. 이전 장면의 지지측 방향과 일치한다. 손 뒤에 밝은 수직 기둥 하나, 왼쪽 갑판에 검은 계선주 하나가 보인다. 계선주는 이전 사진에서 인물에 가려졌을 수 있어 새로 추가된 구조물이라고 단정할 수 없다. 다만 현측벽과 배경이 넓게 보여 '좁은 사선 부분'이라는 요구보다 비중이 크다.",
        "entities": "찰리의 마모된 금속 손·전완과 어두운 젖은 덮개, 이현우의 젊어 보이는 맨손과 검은 젖은 소매가 보인다. 이현우의 얼굴은 사실상 제외되어 한국계 정체성과 정확한 나이는 확인할 수 없지만 보이는 손은 참조와 충돌하지 않는다. 배낭과 가슴 부품은 크롭 밖이다. 밤의 비와 거친 바다, 손 사이에서 흘러내리는 물은 있으나 손끝을 에워싸고 가리는 거대한 파도의 효과는 약하다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "금속 손은 손목과 전완, 화면 왼쪽의 상완에 연결되어 있고 사람 손도 소매 속 팔에 연결되어 있다. 허공에 분리된 신체 부위는 없다. 손가락을 펼치고 접촉을 잃은 자세는 미끄러져 놓친 직후로 가능하다. 물은 손 주변에서 아래로 떨어지고 배경 파도는 현측에 부딪친다. 전신과 찰리의 버팀손은 크롭 밖이므로 지지 상태를 추가로 판단할 수 없다."
       },
       {
        "label": "A",
        "direction": "찰리의 금속 손은 오른쪽에서 왼쪽 아래의 이현우 손을 향하고, 이현우는 왼쪽 아래에서 위로 손을 뻗는다. 인간 손끝과 금속 손가락이 상당히 겹쳐 있어 완전히 손아귀를 빠져나간 순간인지 명료하지 않다. 시선은 보이지 않는다.",
        "built_space": "녹슨 흰색 현측벽 하나가 오른쪽 아래에서 왼쪽 뒤로 이어진다. 갑판과 찰리는 오른쪽, 바다는 왼쪽으로 배치되어 이전 사진의 지지측 화면 방향이 반대다. 배경에는 밝은 기둥 하나와 검은 계선주 하나가 보이며, 명백한 중복 구조물은 없다. 손은 더 크게 잡혔지만 현측벽 역시 화면 아래와 배경의 넓은 면적을 차지한다. 카메라 각도 자체는 자유지만 연속성 기준이 되는 이전 방향은 유지하지 못했다.",
        "entities": "찰리의 젖고 긁힌 금속 손·팔과 어두운 덮개, 이현우의 맨손과 젖은 검은 소매가 보인다. 얼굴·머리·배낭·가슴 부품은 프레임 밖이므로 확인 대상이 아니다. 사람 손의 피부와 가는 형태는 젊은 이현우와 모순되지 않는다. 밤비와 현측을 타고 흐르는 물은 있으나 연결부를 덮치는 물기둥보다는 벽에서 쏟아지는 물막에 가깝다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "두 손 모두 각자의 손목과 팔에 연결되어 있어 지지 없는 부유물은 없다. 손가락이 서로 겹친 자세는 접촉이 풀리는 도중으로는 가능하지만 이미 떨어져 나갔다는 행동을 확실히 보여주지 않는다. 물은 현측 상단에서 바깥쪽 아래로 흐르므로 중력 방향은 자연스럽다. 프레임 밖 전신의 지지나 낙하 자세는 판단할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.444
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.444
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1714,
   "B": 1444
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "프롬프트가 최우선으로 요구한 '클로즈업' 프레이밍을 완벽하게 구현하여 두 손이 분리되는 순간에 집중했으며, 지정된 배경 요소(물줄기와 선박의 모서리)를 충실히 반영했습니다."
   },
   {
    "label": "B",
    "score": 1444,
    "verdict_ko": "인물의 상반신과 거대한 파도를 넓게 보여주는 미디엄 샷으로 렌더링되어, 지정된 '클로즈업' 프레이밍 및 피사체 크기 제약을 명백히 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S80sh6_sel.png",
    "asset_id": "bf09bed8-2429-4e00-9b07-45bce50db38a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:948699>",
    "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aaceee-b78e-72f3-8986-1fee034c0a25",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S80sh6"
  }
 },
 "era_fail::ff9db9d9f1a953f4": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S81sh5": {
  "input_fingerprint": "7057eb8d590d848c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흰색 연구복을 입은 서지민이 이현우를 향해 옅은 미소를 머금고 있는 찰나.\n\nLOCATION (lock): Beside the white bed in the research institute's shelter room, lit by bright fluorescent fixtures and coastal daylight through the window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: White bed (Supporting the now-seated 이현우) — A partial upper surface remains beneath the foreground shoulder; used as Retain the bedside location beneath the conversational framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established fluorescent illumination and readable white clothing, keeping her reassuring expression clear without adding a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The temporary shelter retains its white bed and strong fluorescent lighting, with the emerald-blue sea visible outside the window. 서지민: She wears a research coat and stands in the temporary shelter.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 서지민 (20세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흰색 연구복을 입은 서지민이 이현우를 향해 옅은 미소를 머금고 있는 찰나.\n\nLOCATION (lock): Beside the white bed in the research institute's shelter room, lit by bright fluorescent fixtures and coastal daylight through the window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: White bed (Supporting the now-seated 이현우) — A partial upper surface remains beneath the foreground shoulder; used as Retain the bedside location beneath the conversational framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established fluorescent illumination and readable white clothing, keeping her reassuring expression clear without adding a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The temporary shelter retains its white bed and strong fluorescent lighting, with the emerald-blue sea visible outside the window. 서지민: She wears a research coat and stands in the temporary shelter.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 서지민 (20세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흰색 연구복을 입은 서지민이 이현우를 향해 옅은 미소를 머금고 있는 찰나.\n\nLOCATION (lock): Beside the white bed in the research institute's shelter room, lit by bright fluorescent fixtures and coastal daylight through the window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: White bed (Supporting the now-seated 이현우) — A partial upper surface remains beneath the foreground shoulder; used as Retain the bedside location beneath the conversational framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established fluorescent illumination and readable white clothing, keeping her reassuring expression clear without adding a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The temporary shelter retains its white bed and strong fluorescent lighting, with the emerald-blue sea visible outside the window. 서지민: She wears a research coat and stands in the temporary shelter.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 서지민 (20세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "서지민이 우측 전경의 이현우를 향해 시선을 두고 옅은 미소를 지음.",
    "built_space": "우측에 창문이 있고 배경 중심에 빈 침대가 놓임. 기준 샷의 '침대 바로 뒤 창문' 구조와 일치하지 않음.",
    "entities": "서지민의 인상과 가운, 넥타이가 참고 이미지와 일치함. 전경 인물은 기준에 맞는 흰색 상의를 입음.",
    "hard_violations": [
     "[gemini-pro] 인물 배치 오류: 이현우가 지지받아야 할 침대가 배경에 완전히 비어 있음",
     "[gemini-pro] 공간 구조 위반: 이전 샷의 침대-창문 밀착 구조를 무시함"
    ],
    "physics": "서지민은 바닥에 안정적으로 서 있고, 전경 인물은 프레임 밖 하단 구조물에 몸을 지지함."
   },
   {
    "label": "B",
    "direction": "서지민이 우측 하단의 이현우를 내려다보며 옅은 미소를 지음.",
    "built_space": "정면에 바다가 보이는 창문이 있고 좌측에 빈 침대가 있음. 기준 샷의 공간 구조와 크게 다름.",
    "entities": "서지민의 외모는 비슷하나 넥타이가 누락됨. 전경 인물은 기준(흰색)과 다른 회색 상의를 입음.",
    "hard_violations": [
     "[gemini-pro] 인물 배치 오류: 이현우가 지지받아야 할 침대가 배경에 완전히 비어 있음",
     "[gemini-pro] 공간 구조 위반: 이전 샷의 침대-창문 밀착 구조를 무시함"
    ],
    "physics": "서지민은 두 발로 서 있고, 전경 인물은 화면 밖 보이지 않는 곳에 앉거나 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이현우가 앉아야 할 침대가 배경에 비어 있는 치명적인 배치 오류가 있으나, 서지민의 외형과 의상 디테일은 비교적 잘 구현됨."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "침대가 배경에 비어 있는 심각한 배치 오류가 발생했으며, 서지민의 넥타이 누락 및 전경 인물의 의상 색상 오류까지 겹침."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "서지민이 우측 전경의 이현우를 향해 시선을 두고 옅은 미소를 지음.",
        "built_space": "우측에 창문이 있고 배경 중심에 빈 침대가 놓임. 기준 샷의 '침대 바로 뒤 창문' 구조와 일치하지 않음.",
        "entities": "서지민의 인상과 가운, 넥타이가 참고 이미지와 일치함. 전경 인물은 기준에 맞는 흰색 상의를 입음.",
        "hard_violations": [
         "인물 배치 오류: 이현우가 지지받아야 할 침대가 배경에 완전히 비어 있음",
         "공간 구조 위반: 이전 샷의 침대-창문 밀착 구조를 무시함"
        ],
        "physics": "서지민은 바닥에 안정적으로 서 있고, 전경 인물은 프레임 밖 하단 구조물에 몸을 지지함."
       },
       {
        "label": "B",
        "direction": "서지민이 우측 하단의 이현우를 내려다보며 옅은 미소를 지음.",
        "built_space": "정면에 바다가 보이는 창문이 있고 좌측에 빈 침대가 있음. 기준 샷의 공간 구조와 크게 다름.",
        "entities": "서지민의 외모는 비슷하나 넥타이가 누락됨. 전경 인물은 기준(흰색)과 다른 회색 상의를 입음.",
        "hard_violations": [
         "인물 배치 오류: 이현우가 지지받아야 할 침대가 배경에 완전히 비어 있음",
         "공간 구조 위반: 이전 샷의 침대-창문 밀착 구조를 무시함"
        ],
        "physics": "서지민은 두 발로 서 있고, 전경 인물은 화면 밖 보이지 않는 곳에 앉거나 서 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이현우가 앉아야 할 침대가 배경에 비어 있는 치명적인 배치 오류가 있으나, 서지민의 외형과 의상 디테일은 비교적 잘 구현됨."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "침대가 배경에 비어 있는 심각한 배치 오류가 발생했으며, 서지민의 넥타이 누락 및 전경 인물의 의상 색상 오류까지 겹침."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "서지민이 우측 전경의 이현우를 향해 시선을 두고 옅은 미소를 지음.",
        "built_space": "우측에 창문이 있고 배경 중심에 빈 침대가 놓임. 기준 샷의 '침대 바로 뒤 창문' 구조와 일치하지 않음.",
        "entities": "서지민의 인상과 가운, 넥타이가 참고 이미지와 일치함. 전경 인물은 기준에 맞는 흰색 상의를 입음.",
        "hard_violations": [
         "인물 배치 오류: 이현우가 지지받아야 할 침대가 배경에 완전히 비어 있음",
         "공간 구조 위반: 이전 샷의 침대-창문 밀착 구조를 무시함"
        ],
        "physics": "서지민은 바닥에 안정적으로 서 있고, 전경 인물은 프레임 밖 하단 구조물에 몸을 지지함."
       },
       {
        "label": "B",
        "direction": "서지민이 우측 하단의 이현우를 내려다보며 옅은 미소를 지음.",
        "built_space": "정면에 바다가 보이는 창문이 있고 좌측에 빈 침대가 있음. 기준 샷의 공간 구조와 크게 다름.",
        "entities": "서지민의 외모는 비슷하나 넥타이가 누락됨. 전경 인물은 기준(흰색)과 다른 회색 상의를 입음.",
        "hard_violations": [
         "인물 배치 오류: 이현우가 지지받아야 할 침대가 배경에 완전히 비어 있음",
         "공간 구조 위반: 이전 샷의 침대-창문 밀착 구조를 무시함"
        ],
        "physics": "서지민은 두 발로 서 있고, 전경 인물은 화면 밖 보이지 않는 곳에 앉거나 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "미디엄 숏과 옅은 미소는 구현했지만 시선이 이현우의 얼굴을 향하지 않고, 침대와 창문의 배치 및 서지민의 머리·복장이 참조와 다릅니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "서지민이 전경의 이현우를 바라보며 옅게 웃는 미디엄 숏으로, 참조의 인물 외형과 창문 아래 침대 배치를 더 충실히 유지합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "서지민의 눈은 화면 왼쪽 아래, 침대나 전경 남성의 등 쪽을 향합니다. 오른쪽 끝에 있는 이현우의 얼굴과 눈을 마주치는 방향은 아닙니다. 남성은 뒷머리와 목·등만 보여 시선의 도착점을 확인할 수 없습니다.",
        "built_space": "흰 침대 한 개의 베개와 머리 쪽이 왼쪽 벽에 놓이고, 바다가 보이는 창문 한 개는 뒤쪽 벽 중앙에 있습니다. 참조에서는 베개 바로 뒤에 창문과 창턱이 있었으므로 침대와 창문의 관계가 달라졌습니다. 천장 형광등 두 조와 환기구 하나, 왼쪽 벽 모니터 하나와 의료용 벽 패널, 오른쪽 수납장 하나와 출입구가 보입니다. 서지민은 침대 옆에 있고 남성의 넓은 등·어깨가 하단 전경을 차지합니다.",
        "entities": "젊은 성인 여성 한 명과 전경 남성의 일부가 보입니다. 여성의 피부색과 전반적인 얼굴 인상은 인물 참조에 가깝지만, 머리는 참조보다 길고 풍성하며 남색 넥타이가 빠졌습니다. 흰 연구복, 흰 침구, 낮의 청록색 바다와 형광등은 있습니다. 남성은 얼굴이 거의 가려져 신원을 검증하기 어렵습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "여성의 상체는 곧게 서 있고 팔은 아래로 내려가 전경에 가려집니다. 발은 프레임 밖이지만 공중에 떠 있다는 증거는 없습니다. 남성은 상체를 크게 숙인 듯 보이며 골반과 침대의 접촉은 가려져 있어, 요청된 착석 상태를 명확히 확인하기 어렵습니다. 베개와 이불은 매트리스에 지지되고 가구와 장비에도 지지면이 있습니다."
       },
       {
        "label": "B",
        "direction": "서지민은 고개와 눈을 화면 오른쪽 전경의 이현우 얼굴 쪽으로 돌리고 옅게 웃습니다. 이현우의 보이는 머리 방향도 서지민 쪽을 향해 대화 관계가 성립합니다.",
        "built_space": "흰 침대 한 개가 화면 하단을 차지하며, 베개와 머리판이 오른쪽 창문 바로 아래에 놓여 참조의 공간 관계를 유지합니다. 남성의 어깨 아래에도 침대 윗면이 남습니다. 서지민은 침대 옆에 서 있습니다. 천장 형광등 두 조, 오른쪽 창문 구획, 뒤쪽 모니터 세 대, 의자 한 개와 의료용 카트·작업대가 보입니다. 뒤쪽 설비의 정확한 구성은 좁게 촬영된 장소 참조만으로 확인할 수 없습니다.",
        "entities": "젊은 성인 여성 한 명과 오른쪽 전경 남성의 머리·어깨 일부가 보입니다. 여성의 얼굴형, 피부색, 귀 뒤로 넘긴 짧은 검은 머리와 체형이 인물 참조에 가깝습니다. 흰 연구복과 셔츠, 남색 넥타이도 유지됩니다. 흰 침대와 베개, 창밖 청록색 바다, 형광등이 보이며 문서와 표지는 판독되지 않습니다. 남성의 얼굴은 부분적으로만 보여 정확한 신원 확인은 제한됩니다.",
        "hard_violations": [],
        "physics": "여성은 침대 옆에서 상체를 자연스럽게 세우고 두 손을 맞잡고 있습니다. 손과 팔의 연결 및 접촉이 자연스럽고, 발이 프레임 밖인 것을 제외하면 비정상적인 지지 상태는 보이지 않습니다. 남성의 골반은 프레임 밖이지만 창가 침대에 앉아 여성을 마주보는 상체 배치와 양립합니다. 베개와 침구는 매트리스 위에 놓이고 장비는 작업대와 카트에 지지됩니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "미디엄 숏과 옅은 미소는 구현했지만 시선이 이현우의 얼굴을 향하지 않고, 침대와 창문의 배치 및 서지민의 머리·복장이 참조와 다릅니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "서지민이 전경의 이현우를 바라보며 옅게 웃는 미디엄 숏으로, 참조의 인물 외형과 창문 아래 침대 배치를 더 충실히 유지합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "서지민의 눈은 화면 왼쪽 아래, 침대나 전경 남성의 등 쪽을 향합니다. 오른쪽 끝에 있는 이현우의 얼굴과 눈을 마주치는 방향은 아닙니다. 남성은 뒷머리와 목·등만 보여 시선의 도착점을 확인할 수 없습니다.",
        "built_space": "흰 침대 한 개의 베개와 머리 쪽이 왼쪽 벽에 놓이고, 바다가 보이는 창문 한 개는 뒤쪽 벽 중앙에 있습니다. 참조에서는 베개 바로 뒤에 창문과 창턱이 있었으므로 침대와 창문의 관계가 달라졌습니다. 천장 형광등 두 조와 환기구 하나, 왼쪽 벽 모니터 하나와 의료용 벽 패널, 오른쪽 수납장 하나와 출입구가 보입니다. 서지민은 침대 옆에 있고 남성의 넓은 등·어깨가 하단 전경을 차지합니다.",
        "entities": "젊은 성인 여성 한 명과 전경 남성의 일부가 보입니다. 여성의 피부색과 전반적인 얼굴 인상은 인물 참조에 가깝지만, 머리는 참조보다 길고 풍성하며 남색 넥타이가 빠졌습니다. 흰 연구복, 흰 침구, 낮의 청록색 바다와 형광등은 있습니다. 남성은 얼굴이 거의 가려져 신원을 검증하기 어렵습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "여성의 상체는 곧게 서 있고 팔은 아래로 내려가 전경에 가려집니다. 발은 프레임 밖이지만 공중에 떠 있다는 증거는 없습니다. 남성은 상체를 크게 숙인 듯 보이며 골반과 침대의 접촉은 가려져 있어, 요청된 착석 상태를 명확히 확인하기 어렵습니다. 베개와 이불은 매트리스에 지지되고 가구와 장비에도 지지면이 있습니다."
       },
       {
        "label": "A",
        "direction": "서지민은 고개와 눈을 화면 오른쪽 전경의 이현우 얼굴 쪽으로 돌리고 옅게 웃습니다. 이현우의 보이는 머리 방향도 서지민 쪽을 향해 대화 관계가 성립합니다.",
        "built_space": "흰 침대 한 개가 화면 하단을 차지하며, 베개와 머리판이 오른쪽 창문 바로 아래에 놓여 참조의 공간 관계를 유지합니다. 남성의 어깨 아래에도 침대 윗면이 남습니다. 서지민은 침대 옆에 서 있습니다. 천장 형광등 두 조, 오른쪽 창문 구획, 뒤쪽 모니터 세 대, 의자 한 개와 의료용 카트·작업대가 보입니다. 뒤쪽 설비의 정확한 구성은 좁게 촬영된 장소 참조만으로 확인할 수 없습니다.",
        "entities": "젊은 성인 여성 한 명과 오른쪽 전경 남성의 머리·어깨 일부가 보입니다. 여성의 얼굴형, 피부색, 귀 뒤로 넘긴 짧은 검은 머리와 체형이 인물 참조에 가깝습니다. 흰 연구복과 셔츠, 남색 넥타이도 유지됩니다. 흰 침대와 베개, 창밖 청록색 바다, 형광등이 보이며 문서와 표지는 판독되지 않습니다. 남성의 얼굴은 부분적으로만 보여 정확한 신원 확인은 제한됩니다.",
        "hard_violations": [],
        "physics": "여성은 침대 옆에서 상체를 자연스럽게 세우고 두 손을 맞잡고 있습니다. 손과 팔의 연결 및 접촉이 자연스럽고, 발이 프레임 밖인 것을 제외하면 비정상적인 지지 상태는 보이지 않습니다. 남성의 골반은 프레임 밖이지만 창가 침대에 앉아 여성을 마주보는 상체 배치와 양립합니다. 베개와 침구는 매트리스 위에 놓이고 장비는 작업대와 카트에 지지됩니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.167
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.917
   },
   "violations": {
    "A": [
     "[gemini-pro] 인물 배치 오류: 이현우가 지지받아야 할 침대가 배경에 완전히 비어 있음",
     "[gemini-pro] 공간 구조 위반: 이전 샷의 침대-창문 밀착 구조를 무시함"
    ],
    "B": [
     "[gemini-pro] 인물 배치 오류: 이현우가 지지받아야 할 침대가 배경에 완전히 비어 있음",
     "[gemini-pro] 공간 구조 위반: 이전 샷의 침대-창문 밀착 구조를 무시함"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 917
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "이현우가 앉아야 할 침대가 배경에 비어 있는 치명적인 배치 오류가 있으나, 서지민의 외형과 의상 디테일은 비교적 잘 구현됨.  ★위반: [gemini-pro] 인물 배치 오류: 이현우가 지지받아야 할 침대가 배경에 완전히 비어 있음 / [gemini-pro] 공간 구조 위반: 이전 샷의 침대-창문 밀착 구조를 무시함"
   },
   {
    "label": "B",
    "score": 917,
    "verdict_ko": "침대가 배경에 비어 있는 심각한 배치 오류가 발생했으며, 서지민의 넥타이 누락 및 전경 인물의 의상 색상 오류까지 겹침.  ★위반: [gemini-pro] 인물 배치 오류: 이현우가 지지받아야 할 침대가 배경에 완전히 비어 있음 / [gemini-pro] 공간 구조 위반: 이전 샷의 침대-창문 밀착 구조를 무시함"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S81sh1_sel.png",
    "asset_id": "27f9c6b1-89aa-471c-9ff2-d1d0692dddcb",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 서지민: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:746082>",
    "asset_id": "a9a13f64-38f2-4a18-91db-eab511b31c43",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf05-e8cb-705b-9c07-a11a248b4d4c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S81sh1"
  }
 },
 "era_fail::917fdaa6dcf743be": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S82sh11": {
  "input_fingerprint": "76797cd7469555a8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 침대 위 찰리의 가슴팍을 향해 얇은 인공지능 기계 팔들이 붉은 레이저 빛을 쏘아 내리는 찰나.\n\nLOCATION (lock): At the stainless-steel bed inside the laboratory's circular glass enclosure, where robotic arms cast red scanning light across the robot. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Circular glass wall (Separating the observing camera from the scanning area) — The bed and scanning arms are seen through the glass, not as reflected images; used as Preserve the outside-the-enclosure viewpoint without introducing prominent reflections; Artificial-intelligence scanning arms (Moving over 찰리 with their tips directed toward his chest) — Thin tips angle downward and inward from opposite upper edges; used as Bracket the chest and make the scanning action legible without enlarging any arm beyond believable scale; Stainless-steel bed (Supporting 찰리 during scanning) — Partial upper surface and a diagonal side edge remain visible around his torso; used as Provide material context and a stable scale reference beneath the mechanical action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The red scanning laser specified in the shot provides the selective color accent against otherwise neutral laboratory illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie lies inactive on a stainless-steel bed inside a circular glass enclosure, with hardware damaged by prolonged immersion and multiple wires attached to his body. Robotic arms scan him around the bed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 침대 위 찰리의 가슴팍을 향해 얇은 인공지능 기계 팔들이 붉은 레이저 빛을 쏘아 내리는 찰나.\n\nLOCATION (lock): At the stainless-steel bed inside the laboratory's circular glass enclosure, where robotic arms cast red scanning light across the robot. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Circular glass wall (Separating the observing camera from the scanning area) — The bed and scanning arms are seen through the glass, not as reflected images; used as Preserve the outside-the-enclosure viewpoint without introducing prominent reflections; Artificial-intelligence scanning arms (Moving over 찰리 with their tips directed toward his chest) — Thin tips angle downward and inward from opposite upper edges; used as Bracket the chest and make the scanning action legible without enlarging any arm beyond believable scale; Stainless-steel bed (Supporting 찰리 during scanning) — Partial upper surface and a diagonal side edge remain visible around his torso; used as Provide material context and a stable scale reference beneath the mechanical action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The red scanning laser specified in the shot provides the selective color accent against otherwise neutral laboratory illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie lies inactive on a stainless-steel bed inside a circular glass enclosure, with hardware damaged by prolonged immersion and multiple wires attached to his body. Robotic arms scan him around the bed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 침대 위 찰리의 가슴팍을 향해 얇은 인공지능 기계 팔들이 붉은 레이저 빛을 쏘아 내리는 찰나.\n\nLOCATION (lock): At the stainless-steel bed inside the laboratory's circular glass enclosure, where robotic arms cast red scanning light across the robot. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Circular glass wall (Separating the observing camera from the scanning area) — The bed and scanning arms are seen through the glass, not as reflected images; used as Preserve the outside-the-enclosure viewpoint without introducing prominent reflections; Artificial-intelligence scanning arms (Moving over 찰리 with their tips directed toward his chest) — Thin tips angle downward and inward from opposite upper edges; used as Bracket the chest and make the scanning action legible without enlarging any arm beyond believable scale; Stainless-steel bed (Supporting 찰리 during scanning) — Partial upper surface and a diagonal side edge remain visible around his torso; used as Provide material context and a stable scale reference beneath the mechanical action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The red scanning laser specified in the shot provides the selective color accent against otherwise neutral laboratory illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie lies inactive on a stainless-steel bed inside a circular glass enclosure, with hardware damaged by prolonged immersion and multiple wires attached to his body. Robotic arms scan him around the bed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "두 개의 얇은 기계 팔이 로봇의 가슴을 향해 붉은 레이저를 쏘고 있으며, 배경의 기계들에서도 레이저가 뿜어져 나옵니다.",
    "built_space": "레퍼런스의 원형 유리실 대신 전경에 두꺼운 금속 링이 떠 있고, 그 너머로 수술대와 기계들이 배치되어 있습니다.",
    "entities": "스테인리스 스틸 침대와 기계 팔은 존재하나, 찰리의 머리가 각진 흰 얼굴판이 아닌 사람 형태의 안면으로 잘못 묘사되었습니다.",
    "hard_violations": [
     "[gemini-pro] 지정된 인물 외양 위반 (사람 형태의 로봇 얼굴)",
     "[gemini-pro] 공간 구조 위반 (원형 유리벽이 금속 링으로 변형됨)",
     "[gemini-pro] 프레이밍 위반 (클로즈업이 아닌 와이드 샷에 가까움)",
     "[gpt-high] 기준 장소의 좌우 로봇 팔 두 개를 배경에 그대로 두고 전경에 별도 스캔 팔 두 개를 추가해, 고정 스캔 설비가 두 개에서 네 개로 중복되었다."
    ],
    "physics": "로봇은 침대 위에 누워 지탱되고 있으며, 기계 팔들은 화면 밖 구조물에 연결되어 있습니다."
   },
   {
    "label": "B",
    "direction": "화면 양쪽 상단에서 내려온 두 기계 팔이 로봇의 가슴팍을 향해 정확히 붉은 레이저를 조준하고 있습니다.",
    "built_space": "카메라가 원형 유리벽 바깥에서 안을 들여다보는 구도이며, 레퍼런스와 일치하는 유리 공간 내부에 스테인리스 침대가 놓여 있습니다.",
    "entities": "찰리는 프롬프트대로 각진 흰색 얼굴판, 점 형태의 눈, 선 형태의 입을 가졌으며 가슴에 전선이 연결되어 있습니다.",
    "hard_violations": [],
    "physics": "로봇은 금속 침대 위에 안정적으로 누워 있으며, 기계 팔은 화면 상단에서 뻗어 내려옵니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요청된 찰리의 얼굴 디자인(각진 흰 얼굴판, 점 눈, 선 입)과 원형 유리벽 너머로 보이는 구도를 정확하게 구현했으며, 레이저 방향도 적절합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 얼굴이 사람과 유사한 형태로 잘못 묘사되었으며, 원형 유리벽 대신 금속 링이 뜬금없이 전경에 배치되어 공간적 오류를 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 개의 얇은 기계 팔이 로봇의 가슴을 향해 붉은 레이저를 쏘고 있으며, 배경의 기계들에서도 레이저가 뿜어져 나옵니다.",
        "built_space": "레퍼런스의 원형 유리실 대신 전경에 두꺼운 금속 링이 떠 있고, 그 너머로 수술대와 기계들이 배치되어 있습니다.",
        "entities": "스테인리스 스틸 침대와 기계 팔은 존재하나, 찰리의 머리가 각진 흰 얼굴판이 아닌 사람 형태의 안면으로 잘못 묘사되었습니다.",
        "hard_violations": [
         "지정된 인물 외양 위반 (사람 형태의 로봇 얼굴)",
         "공간 구조 위반 (원형 유리벽이 금속 링으로 변형됨)",
         "프레이밍 위반 (클로즈업이 아닌 와이드 샷에 가까움)"
        ],
        "physics": "로봇은 침대 위에 누워 지탱되고 있으며, 기계 팔들은 화면 밖 구조물에 연결되어 있습니다."
       },
       {
        "label": "B",
        "direction": "화면 양쪽 상단에서 내려온 두 기계 팔이 로봇의 가슴팍을 향해 정확히 붉은 레이저를 조준하고 있습니다.",
        "built_space": "카메라가 원형 유리벽 바깥에서 안을 들여다보는 구도이며, 레퍼런스와 일치하는 유리 공간 내부에 스테인리스 침대가 놓여 있습니다.",
        "entities": "찰리는 프롬프트대로 각진 흰색 얼굴판, 점 형태의 눈, 선 형태의 입을 가졌으며 가슴에 전선이 연결되어 있습니다.",
        "hard_violations": [],
        "physics": "로봇은 금속 침대 위에 안정적으로 누워 있으며, 기계 팔은 화면 상단에서 뻗어 내려옵니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요청된 찰리의 얼굴 디자인(각진 흰 얼굴판, 점 눈, 선 입)과 원형 유리벽 너머로 보이는 구도를 정확하게 구현했으며, 레이저 방향도 적절합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 얼굴이 사람과 유사한 형태로 잘못 묘사되었으며, 원형 유리벽 대신 금속 링이 뜬금없이 전경에 배치되어 공간적 오류를 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 개의 얇은 기계 팔이 로봇의 가슴을 향해 붉은 레이저를 쏘고 있으며, 배경의 기계들에서도 레이저가 뿜어져 나옵니다.",
        "built_space": "레퍼런스의 원형 유리실 대신 전경에 두꺼운 금속 링이 떠 있고, 그 너머로 수술대와 기계들이 배치되어 있습니다.",
        "entities": "스테인리스 스틸 침대와 기계 팔은 존재하나, 찰리의 머리가 각진 흰 얼굴판이 아닌 사람 형태의 안면으로 잘못 묘사되었습니다.",
        "hard_violations": [
         "지정된 인물 외양 위반 (사람 형태의 로봇 얼굴)",
         "공간 구조 위반 (원형 유리벽이 금속 링으로 변형됨)",
         "프레이밍 위반 (클로즈업이 아닌 와이드 샷에 가까움)"
        ],
        "physics": "로봇은 침대 위에 누워 지탱되고 있으며, 기계 팔들은 화면 밖 구조물에 연결되어 있습니다."
       },
       {
        "label": "B",
        "direction": "화면 양쪽 상단에서 내려온 두 기계 팔이 로봇의 가슴팍을 향해 정확히 붉은 레이저를 조준하고 있습니다.",
        "built_space": "카메라가 원형 유리벽 바깥에서 안을 들여다보는 구도이며, 레퍼런스와 일치하는 유리 공간 내부에 스테인리스 침대가 놓여 있습니다.",
        "entities": "찰리는 프롬프트대로 각진 흰색 얼굴판, 점 형태의 눈, 선 형태의 입을 가졌으며 가슴에 전선이 연결되어 있습니다.",
        "hard_violations": [],
        "physics": "로봇은 금속 침대 위에 안정적으로 누워 있으며, 기계 팔은 화면 상단에서 뻗어 내려옵니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "두 가느다란 팔의 레이저가 찰리의 가슴에 정확히 닿지만, 허벅지까지 보이는 넓은 구도와 불명확한 유리벽 밖 시점은 지정된 클로즈업에 미달한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "전경 레이저는 가슴을 겨냥하지만, 기준 장소의 두 팔에 별도 스캔 팔 두 개를 추가했고 전신과 침대 하부까지 보여 클로즈업 지시를 크게 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "좌우 상단에서 들어오는 두 스캔 팔의 끝이 아래쪽 안으로 향하며, 두 붉은 광선은 찰리의 가슴판에 모인다. 찰리의 얼굴은 위쪽과 카메라 쪽을 향하고 몸은 움직이지 않는다.",
        "built_space": "침대 한 개, 스캔 팔 두 개, 천장의 원형 조명 한 개가 보인다. 뒤와 양옆에는 유리벽과 수직 프레임, 실험 장비와 창문이 있다. 기준 장소의 주요 재료와 설비는 이어지지만, 카메라와 침대 사이를 가르는 유리면은 뚜렷하지 않아 enclosure 밖 관찰 시점이 명확하지 않다. 침대 상판과 양쪽 대각선 가장자리가 보이나, 가슴뿐 아니라 머리부터 허벅지까지 포함해 클로즈업보다 넓다.",
        "entities": "찰리로 제시된 기계 몸체 하나만 있으며 다른 사람은 없다. 각진 흰 얼굴판, 점 모양 눈, 선 모양 입은 맞지만 머리는 고릴라형보다 좁은 인간형에 가깝다. 몸체에는 긁힘과 오염이 있고 여러 전선이 연결되어 있다. 스테인리스 침대와 가느다란 금속 스캔 끝, 붉은 레이저가 보인다. 중립적인 주간 실험실 조명이며 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 등과 몸통은 기울어진 금속 침대 상판에 받쳐져 있고 팔도 침대 위에 놓여 있다. 스캔 끝은 화면 밖으로 이어지는 기계 팔에 연결되어 있으며, 전선은 몸체와 침대 위에 걸쳐 있다. 지지 없이 떠 있는 몸체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "전경 좌우 상단의 가느다란 스캔 끝 두 개는 아래쪽 안으로 기울어 가슴판의 밝은 지점을 겨냥한다. 반면 배경의 굵은 로봇 팔 두 개가 내보내는 붉은 광선은 가슴보다 위쪽의 빈 공간으로 향한다. 찰리의 얼굴은 천장을 향하고 몸은 수평으로 누워 있다.",
        "built_space": "침대 한 개의 상판과 두꺼운 측면, 하부 지지대까지 보인다. 배경에는 기준 사진과 비슷한 좌우 로봇 팔 두 개가 있고, 그 앞에 가느다란 스캔 팔 두 개가 별도로 추가되어 총 네 개다. 전경의 굵은 수직 원형 테두리는 기준 장소의 원통형 유리벽과 천장 수평 원형 조명과 다른 구조로 읽힌다. 유리를 통한 관찰 느낌은 있으나, 전신과 침대 대부분을 포함한 측면 구도여서 가슴 클로즈업이 아니다.",
        "entities": "기계 몸체 하나만 침대에 누워 있으며 추가 인물은 없다. 흰 얼굴판은 옆으로 보이지만 둥근 두개부와 큰 원형 귀 관절이 두드러져 지정된 각진 고릴라형 머리와 차이가 있다. 점 눈과 선 입은 이 각도에서 명확히 확인하기 어렵다. 몸체의 손상과 노출 배선, 연결된 여러 전선, 금속 침대, 붉은 스캔 빛은 보인다. 배경 표시는 흐려 읽기 어렵다.",
        "hard_violations": [
         "기준 장소의 좌우 로봇 팔 두 개를 배경에 그대로 두고 전경에 별도 스캔 팔 두 개를 추가해, 고정 스캔 설비가 두 개에서 네 개로 중복되었다."
        ],
        "physics": "찰리의 머리 뒤, 등과 골반, 다리는 침대 상판에 지지되어 있다. 손은 몸 옆의 상판에 놓이며, 전선은 몸과 침대에 걸쳐 아래로 늘어진다. 전경 스캔 끝은 위쪽으로 이어지는 기계 축에 연결되고 배경 팔은 각자의 수직 기반에 붙어 있다. 공중에 지지 없이 떠 있는 신체나 고체 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "두 가느다란 팔의 레이저가 찰리의 가슴에 정확히 닿지만, 허벅지까지 보이는 넓은 구도와 불명확한 유리벽 밖 시점은 지정된 클로즈업에 미달한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "전경 레이저는 가슴을 겨냥하지만, 기준 장소의 두 팔에 별도 스캔 팔 두 개를 추가했고 전신과 침대 하부까지 보여 클로즈업 지시를 크게 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "좌우 상단에서 들어오는 두 스캔 팔의 끝이 아래쪽 안으로 향하며, 두 붉은 광선은 찰리의 가슴판에 모인다. 찰리의 얼굴은 위쪽과 카메라 쪽을 향하고 몸은 움직이지 않는다.",
        "built_space": "침대 한 개, 스캔 팔 두 개, 천장의 원형 조명 한 개가 보인다. 뒤와 양옆에는 유리벽과 수직 프레임, 실험 장비와 창문이 있다. 기준 장소의 주요 재료와 설비는 이어지지만, 카메라와 침대 사이를 가르는 유리면은 뚜렷하지 않아 enclosure 밖 관찰 시점이 명확하지 않다. 침대 상판과 양쪽 대각선 가장자리가 보이나, 가슴뿐 아니라 머리부터 허벅지까지 포함해 클로즈업보다 넓다.",
        "entities": "찰리로 제시된 기계 몸체 하나만 있으며 다른 사람은 없다. 각진 흰 얼굴판, 점 모양 눈, 선 모양 입은 맞지만 머리는 고릴라형보다 좁은 인간형에 가깝다. 몸체에는 긁힘과 오염이 있고 여러 전선이 연결되어 있다. 스테인리스 침대와 가느다란 금속 스캔 끝, 붉은 레이저가 보인다. 중립적인 주간 실험실 조명이며 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 등과 몸통은 기울어진 금속 침대 상판에 받쳐져 있고 팔도 침대 위에 놓여 있다. 스캔 끝은 화면 밖으로 이어지는 기계 팔에 연결되어 있으며, 전선은 몸체와 침대 위에 걸쳐 있다. 지지 없이 떠 있는 몸체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "전경 좌우 상단의 가느다란 스캔 끝 두 개는 아래쪽 안으로 기울어 가슴판의 밝은 지점을 겨냥한다. 반면 배경의 굵은 로봇 팔 두 개가 내보내는 붉은 광선은 가슴보다 위쪽의 빈 공간으로 향한다. 찰리의 얼굴은 천장을 향하고 몸은 수평으로 누워 있다.",
        "built_space": "침대 한 개의 상판과 두꺼운 측면, 하부 지지대까지 보인다. 배경에는 기준 사진과 비슷한 좌우 로봇 팔 두 개가 있고, 그 앞에 가느다란 스캔 팔 두 개가 별도로 추가되어 총 네 개다. 전경의 굵은 수직 원형 테두리는 기준 장소의 원통형 유리벽과 천장 수평 원형 조명과 다른 구조로 읽힌다. 유리를 통한 관찰 느낌은 있으나, 전신과 침대 대부분을 포함한 측면 구도여서 가슴 클로즈업이 아니다.",
        "entities": "기계 몸체 하나만 침대에 누워 있으며 추가 인물은 없다. 흰 얼굴판은 옆으로 보이지만 둥근 두개부와 큰 원형 귀 관절이 두드러져 지정된 각진 고릴라형 머리와 차이가 있다. 점 눈과 선 입은 이 각도에서 명확히 확인하기 어렵다. 몸체의 손상과 노출 배선, 연결된 여러 전선, 금속 침대, 붉은 스캔 빛은 보인다. 배경 표시는 흐려 읽기 어렵다.",
        "hard_violations": [
         "기준 장소의 좌우 로봇 팔 두 개를 배경에 그대로 두고 전경에 별도 스캔 팔 두 개를 추가해, 고정 스캔 설비가 두 개에서 네 개로 중복되었다."
        ],
        "physics": "찰리의 머리 뒤, 등과 골반, 다리는 침대 상판에 지지되어 있다. 손은 몸 옆의 상판에 놓이며, 전선은 몸과 침대에 걸쳐 아래로 늘어진다. 전경 스캔 끝은 위쪽으로 이어지는 기계 축에 연결되고 배경 팔은 각자의 수직 기반에 붙어 있다. 공중에 지지 없이 떠 있는 신체나 고체 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.829,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.579,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지정된 인물 외양 위반 (사람 형태의 로봇 얼굴)",
     "[gemini-pro] 공간 구조 위반 (원형 유리벽이 금속 링으로 변형됨)",
     "[gemini-pro] 프레이밍 위반 (클로즈업이 아닌 와이드 샷에 가까움)",
     "[gpt-high] 기준 장소의 좌우 로봇 팔 두 개를 배경에 그대로 두고 전경에 별도 스캔 팔 두 개를 추가해, 고정 스캔 설비가 두 개에서 네 개로 중복되었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 579
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요청된 찰리의 얼굴 디자인(각진 흰 얼굴판, 점 눈, 선 입)과 원형 유리벽 너머로 보이는 구도를 정확하게 구현했으며, 레이저 방향도 적절합니다."
   },
   {
    "label": "A",
    "score": 579,
    "verdict_ko": "찰리의 얼굴이 사람과 유사한 형태로 잘못 묘사되었으며, 원형 유리벽 대신 금속 링이 뜬금없이 전경에 배치되어 공간적 오류를 범했습니다.  ★위반: [gemini-pro] 지정된 인물 외양 위반 (사람 형태의 로봇 얼굴) / [gemini-pro] 공간 구조 위반 (원형 유리벽이 금속 링으로 변형됨) / [gemini-pro] 프레이밍 위반 (클로즈업이 아닌 와이드 샷에 가까움) / [gpt-high] 기준 장소의 좌우 로봇 팔 두 개를 배경에 그대로 두고 전경에 별도 스캔 팔 두 개를 추가해, 고정 스캔 설비가 두 개에서 네 개로 중복되었다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L129B01.png",
    "asset_id": "4a6aa631-eaf4-4be4-a449-2ca4a933e26a",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf17-75f3-759f-a29f-250990b07787",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S82sh16": {
  "input_fingerprint": "667dd515819b5e38",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 지소영의 건조한 표정을 마주하며 턱을 빳빳하게 굳힌 위화감 섞인 구도.\n\nLOCATION (lock): On the observation side of the circular glass enclosure in the main center's adjoining laboratory, with workstation screens and scanning equipment lit nearby. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Circular glass enclosure (Remaining behind the conversation in the laboratory) — A peripheral section is seen obliquely, with interior detail left indistinct; used as Retain a minimal trace of the laboratory location without drawing attention away from the faces.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral laboratory illumination with enough tonal separation to read his tightened jaw and her restrained foreground profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains inactive on the stainless-steel bed inside the circular glass enclosure, with multiple wires attached during treatment. The surrounding robotic arms and laboratory computers remain in place. 이현우: He stands outside the glass enclosure in his white clothing, visibly worried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 지소영의 건조한 표정을 마주하며 턱을 빳빳하게 굳힌 위화감 섞인 구도.\n\nLOCATION (lock): On the observation side of the circular glass enclosure in the main center's adjoining laboratory, with workstation screens and scanning equipment lit nearby. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Circular glass enclosure (Remaining behind the conversation in the laboratory) — A peripheral section is seen obliquely, with interior detail left indistinct; used as Retain a minimal trace of the laboratory location without drawing attention away from the faces.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral laboratory illumination with enough tonal separation to read his tightened jaw and her restrained foreground profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains inactive on the stainless-steel bed inside the circular glass enclosure, with multiple wires attached during treatment. The surrounding robotic arms and laboratory computers remain in place. 이현우: He stands outside the glass enclosure in his white clothing, visibly worried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 지소영의 건조한 표정을 마주하며 턱을 빳빳하게 굳힌 위화감 섞인 구도.\n\nLOCATION (lock): On the observation side of the circular glass enclosure in the main center's adjoining laboratory, with workstation screens and scanning equipment lit nearby. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Circular glass enclosure (Remaining behind the conversation in the laboratory) — A peripheral section is seen obliquely, with interior detail left indistinct; used as Retain a minimal trace of the laboratory location without drawing attention away from the faces.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral laboratory illumination with enough tonal separation to read his tightened jaw and her restrained foreground profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains inactive on the stainless-steel bed inside the circular glass enclosure, with multiple wires attached during treatment. The surrounding robotic arms and laboratory computers remain in place. 이현우: He stands outside the glass enclosure in his white clothing, visibly worried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "이현우와 지소영이 서로 정면으로 마주보고 있음.",
    "built_space": "원형 유리 부스 안 로봇 침대에 바퀴가 생겼으며, 로봇 팔 2개가 로봇을 향하지 않고 위로 들려 있음.",
    "entities": "이현우의 외형이 일치함. 부스 안 로봇에 연결된 와이어와 레이저가 누락됨.",
    "hard_violations": [
     "[gemini-pro] 부스 내 고정형 침대를 바퀴 달린 이동식 병상으로 변형 (고정 요소 위반)",
     "[gemini-pro] 로봇 팔 위치 변경 및 레이저 누락 ('제자리에 유지' 지시 위반)"
    ],
    "physics": "인물들은 지면에 서 있고, 로봇은 침대 위에 놓여 있음."
   },
   {
    "label": "B",
    "direction": "이현우가 전경에 위치한 지소영을 향해 시선을 두고 있음.",
    "built_space": "원형 유리 부스 내 로봇 침대가 평평한 형태로 변경되었으며, 대칭을 이루어야 할 오른쪽 로봇 팔이 보이지 않음.",
    "entities": "이현우의 외형이 레퍼런스와 일치하며 지소영이 전경에 배치됨. 로봇 하단에 검은 신발이 추가됨.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 존재하는 오른쪽 로봇 팔 누락 (고정 요소 훼손)",
     "[gemini-pro] 로봇을 눕혀둔 침대의 형태 변경"
    ],
    "physics": "인물들은 바닥에 안정적으로 서 있고, 로봇은 테이블 위에 안착됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시된 클로즈업 샷과 전경의 측면 구도를 정확히 구현했으나, 고정되어야 할 오른쪽 로봇 팔이 누락되어 실격입니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "요구된 전경 측면 구도를 따르지 않았으며, 락킹된 배경의 침대에 바퀴를 추가하고 로봇 팔 위치를 변경해 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "이현우가 전경에 위치한 지소영을 향해 시선을 두고 있음.",
        "built_space": "원형 유리 부스 내 로봇 침대가 평평한 형태로 변경되었으며, 대칭을 이루어야 할 오른쪽 로봇 팔이 보이지 않음.",
        "entities": "이현우의 외형이 레퍼런스와 일치하며 지소영이 전경에 배치됨. 로봇 하단에 검은 신발이 추가됨.",
        "hard_violations": [
         "레퍼런스에 존재하는 오른쪽 로봇 팔 누락 (고정 요소 훼손)",
         "로봇을 눕혀둔 침대의 형태 변경"
        ],
        "physics": "인물들은 바닥에 안정적으로 서 있고, 로봇은 테이블 위에 안착됨."
       },
       {
        "label": "A",
        "direction": "이현우와 지소영이 서로 정면으로 마주보고 있음.",
        "built_space": "원형 유리 부스 안 로봇 침대에 바퀴가 생겼으며, 로봇 팔 2개가 로봇을 향하지 않고 위로 들려 있음.",
        "entities": "이현우의 외형이 일치함. 부스 안 로봇에 연결된 와이어와 레이저가 누락됨.",
        "hard_violations": [
         "부스 내 고정형 침대를 바퀴 달린 이동식 병상으로 변형 (고정 요소 위반)",
         "로봇 팔 위치 변경 및 레이저 누락 ('제자리에 유지' 지시 위반)"
        ],
        "physics": "인물들은 지면에 서 있고, 로봇은 침대 위에 놓여 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시된 클로즈업 샷과 전경의 측면 구도를 정확히 구현했으나, 고정되어야 할 오른쪽 로봇 팔이 누락되어 실격입니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "요구된 전경 측면 구도를 따르지 않았으며, 락킹된 배경의 침대에 바퀴를 추가하고 로봇 팔 위치를 변경해 실격입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우가 전경에 위치한 지소영을 향해 시선을 두고 있음.",
        "built_space": "원형 유리 부스 내 로봇 침대가 평평한 형태로 변경되었으며, 대칭을 이루어야 할 오른쪽 로봇 팔이 보이지 않음.",
        "entities": "이현우의 외형이 레퍼런스와 일치하며 지소영이 전경에 배치됨. 로봇 하단에 검은 신발이 추가됨.",
        "hard_violations": [
         "레퍼런스에 존재하는 오른쪽 로봇 팔 누락 (고정 요소 훼손)",
         "로봇을 눕혀둔 침대의 형태 변경"
        ],
        "physics": "인물들은 바닥에 안정적으로 서 있고, 로봇은 테이블 위에 안착됨."
       },
       {
        "label": "A",
        "direction": "이현우와 지소영이 서로 정면으로 마주보고 있음.",
        "built_space": "원형 유리 부스 안 로봇 침대에 바퀴가 생겼으며, 로봇 팔 2개가 로봇을 향하지 않고 위로 들려 있음.",
        "entities": "이현우의 외형이 일치함. 부스 안 로봇에 연결된 와이어와 레이저가 누락됨.",
        "hard_violations": [
         "부스 내 고정형 침대를 바퀴 달린 이동식 병상으로 변형 (고정 요소 위반)",
         "로봇 팔 위치 변경 및 레이저 누락 ('제자리에 유지' 지시 위반)"
        ],
        "physics": "인물들은 지면에 서 있고, 로봇은 침대 위에 놓여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지소영의 전경 옆얼굴과 이를 바라보는 이현우의 비대칭 배치는 더 충실하지만, 턱의 긴장이 약하고 내부 장비까지 드러나 요구된 얼굴 중심 클로즈업보다 넓습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 사람의 맞시선은 명확하지만, 대칭적인 상반신 투샷과 중앙에 노출된 치료대가 전경 옆얼굴 중심의 클로즈업 및 최소 배경 지시에서 멀어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 오른쪽 전경의 지소영 얼굴을 바라봅니다. 지소영은 화면 왼쪽을 향하지만 이현우와 눈을 맞추기보다는 그의 앞쪽 공간을 보는 듯합니다. 이현우가 그녀의 건조한 표정을 살피는 관계는 읽힙니다. 배경 왼쪽 로봇 팔 끝은 치료대 쪽으로 내려가며, 오른쪽 팔 끝은 인물에 가려 정확한 목표를 확인하기 어렵습니다.",
        "built_space": "곡면 유리 인클로저 하나와 여러 수직 금속 지지대, 안쪽 치료대 하나, 양쪽 로봇 팔 두 개가 보입니다. 관찰 측에는 좌우 화면 두 개와 작은 스캐너 하나가 보이고, 두 사람은 유리 밖에 있습니다. 유리·금속 재질과 밝은 주간 조명은 참조 장소에 부합하지만, 인클로저 주변부만 비스듬히 남기라는 지시와 달리 침상과 내부가 상당히 식별됩니다. 불가능한 반사나 명백한 중복 설비는 보이지 않습니다.",
        "entities": "검은 헝클어진 머리와 흰 상의를 입은 젊은 동아시아계 남성은 이현우 참조의 외형에 대체로 부합합니다. 얼굴에는 걱정이 있으나 입술이 조금 벌어져 턱을 빳빳하게 굳힌 연기는 약합니다. 전경의 짧은 검은 머리 성인 여성은 샷에 명시된 지소영으로 읽히며 표정은 절제되어 있습니다. 지소영의 별도 외모 참조는 없습니다. 배경에는 금속성 찰리, 치료대, 로봇 팔, 컴퓨터와 스캐너가 보이며 전선의 세부 연결은 흐립니다. 뚜렷하게 읽히는 문구는 없습니다.",
        "hard_violations": [],
        "physics": "두 사람의 머리와 어깨, 몸통은 자연스럽게 연결되어 있고 서 있는 자세로 읽힙니다. 발은 프레임 밖이므로 접지는 확인할 수 없지만 공중에 떠 있다는 징후는 없습니다. 찰리는 기울어진 금속 침상에 몸을 기대고 있으며 침상 하부 지지대가 보입니다. 화면과 스캐너는 책상 또는 받침에 지지되고 로봇 팔은 기계 관절과 연결됩니다."
       },
       {
        "label": "B",
        "direction": "왼쪽 이현우는 오른쪽 지소영의 눈을 보고, 지소영도 이현우를 올려다봐 맞시선이 명확합니다. 두 로봇 팔의 끝은 중앙 치료대 쪽을 향하지만 가시적인 조사 광선이나 접촉은 없습니다.",
        "built_space": "곡면 유리 인클로저 하나가 배경 중앙에 정면에 가깝게 펼쳐지고, 그 안에 치료대 하나와 양쪽 로봇 팔 두 개가 보입니다. 관찰 측에는 왼쪽 화면 하나와 오른쪽 화면 두 개, 스캐너 하나가 보입니다. 두 사람은 유리 밖에서 서로 마주 섭니다. 참조의 금속 프레임, 유리, 창을 통한 주간광은 유지되지만, 중앙 침상과 인클로저 내부가 넓고 선명해 최소한의 비스듬한 주변부 배경이라는 지시에 어긋납니다. 화면 수는 참조에 다수의 작업대가 있어 중복 설비라고 단정할 수 없습니다.",
        "entities": "이현우는 젊은 동아시아계 남성으로, 헝클어진 검은 머리와 흰 상의가 참조에 대체로 맞습니다. 입을 다문 옆얼굴에서 턱의 긴장은 읽히지만 걱정은 비교적 약합니다. 짧은 검은 머리의 성인 여성은 지소영으로 읽히며 건조한 표정을 보입니다. 찰리는 중앙 치료대 위에 있고 전선, 로봇 팔, 컴퓨터와 스캐너도 보입니다. 뚜렷하게 읽히는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 사람은 자연스럽게 선 상반신 자세이며 발은 화면 밖입니다. 찰리의 등과 다리는 경사진 침상에 받쳐져 있고 침상은 하부 기둥과 바퀴 달린 받침으로 지지됩니다. 늘어진 전선은 침상에서 아래로 이어집니다. 모니터와 스캐너는 작업대에 놓여 있으며, 지지 없이 떠 있는 물체나 불가능한 신체 자세는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지소영의 전경 옆얼굴과 이를 바라보는 이현우의 비대칭 배치는 더 충실하지만, 턱의 긴장이 약하고 내부 장비까지 드러나 요구된 얼굴 중심 클로즈업보다 넓습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 사람의 맞시선은 명확하지만, 대칭적인 상반신 투샷과 중앙에 노출된 치료대가 전경 옆얼굴 중심의 클로즈업 및 최소 배경 지시에서 멀어집니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 오른쪽 전경의 지소영 얼굴을 바라봅니다. 지소영은 화면 왼쪽을 향하지만 이현우와 눈을 맞추기보다는 그의 앞쪽 공간을 보는 듯합니다. 이현우가 그녀의 건조한 표정을 살피는 관계는 읽힙니다. 배경 왼쪽 로봇 팔 끝은 치료대 쪽으로 내려가며, 오른쪽 팔 끝은 인물에 가려 정확한 목표를 확인하기 어렵습니다.",
        "built_space": "곡면 유리 인클로저 하나와 여러 수직 금속 지지대, 안쪽 치료대 하나, 양쪽 로봇 팔 두 개가 보입니다. 관찰 측에는 좌우 화면 두 개와 작은 스캐너 하나가 보이고, 두 사람은 유리 밖에 있습니다. 유리·금속 재질과 밝은 주간 조명은 참조 장소에 부합하지만, 인클로저 주변부만 비스듬히 남기라는 지시와 달리 침상과 내부가 상당히 식별됩니다. 불가능한 반사나 명백한 중복 설비는 보이지 않습니다.",
        "entities": "검은 헝클어진 머리와 흰 상의를 입은 젊은 동아시아계 남성은 이현우 참조의 외형에 대체로 부합합니다. 얼굴에는 걱정이 있으나 입술이 조금 벌어져 턱을 빳빳하게 굳힌 연기는 약합니다. 전경의 짧은 검은 머리 성인 여성은 샷에 명시된 지소영으로 읽히며 표정은 절제되어 있습니다. 지소영의 별도 외모 참조는 없습니다. 배경에는 금속성 찰리, 치료대, 로봇 팔, 컴퓨터와 스캐너가 보이며 전선의 세부 연결은 흐립니다. 뚜렷하게 읽히는 문구는 없습니다.",
        "hard_violations": [],
        "physics": "두 사람의 머리와 어깨, 몸통은 자연스럽게 연결되어 있고 서 있는 자세로 읽힙니다. 발은 프레임 밖이므로 접지는 확인할 수 없지만 공중에 떠 있다는 징후는 없습니다. 찰리는 기울어진 금속 침상에 몸을 기대고 있으며 침상 하부 지지대가 보입니다. 화면과 스캐너는 책상 또는 받침에 지지되고 로봇 팔은 기계 관절과 연결됩니다."
       },
       {
        "label": "A",
        "direction": "왼쪽 이현우는 오른쪽 지소영의 눈을 보고, 지소영도 이현우를 올려다봐 맞시선이 명확합니다. 두 로봇 팔의 끝은 중앙 치료대 쪽을 향하지만 가시적인 조사 광선이나 접촉은 없습니다.",
        "built_space": "곡면 유리 인클로저 하나가 배경 중앙에 정면에 가깝게 펼쳐지고, 그 안에 치료대 하나와 양쪽 로봇 팔 두 개가 보입니다. 관찰 측에는 왼쪽 화면 하나와 오른쪽 화면 두 개, 스캐너 하나가 보입니다. 두 사람은 유리 밖에서 서로 마주 섭니다. 참조의 금속 프레임, 유리, 창을 통한 주간광은 유지되지만, 중앙 침상과 인클로저 내부가 넓고 선명해 최소한의 비스듬한 주변부 배경이라는 지시에 어긋납니다. 화면 수는 참조에 다수의 작업대가 있어 중복 설비라고 단정할 수 없습니다.",
        "entities": "이현우는 젊은 동아시아계 남성으로, 헝클어진 검은 머리와 흰 상의가 참조에 대체로 맞습니다. 입을 다문 옆얼굴에서 턱의 긴장은 읽히지만 걱정은 비교적 약합니다. 짧은 검은 머리의 성인 여성은 지소영으로 읽히며 건조한 표정을 보입니다. 찰리는 중앙 치료대 위에 있고 전선, 로봇 팔, 컴퓨터와 스캐너도 보입니다. 뚜렷하게 읽히는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "두 사람은 자연스럽게 선 상반신 자세이며 발은 화면 밖입니다. 찰리의 등과 다리는 경사진 침상에 받쳐져 있고 침상은 하부 기둥과 바퀴 달린 받침으로 지지됩니다. 늘어진 전선은 침상에서 아래로 이어집니다. 모니터와 스캐너는 작업대에 놓여 있으며, 지지 없이 떠 있는 물체나 불가능한 신체 자세는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.417,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.167,
    "B": 1.75
   },
   "violations": {
    "B": [
     "[gemini-pro] 레퍼런스에 존재하는 오른쪽 로봇 팔 누락 (고정 요소 훼손)",
     "[gemini-pro] 로봇을 눕혀둔 침대의 형태 변경"
    ],
    "A": [
     "[gemini-pro] 부스 내 고정형 침대를 바퀴 달린 이동식 병상으로 변형 (고정 요소 위반)",
     "[gemini-pro] 로봇 팔 위치 변경 및 레이저 누락 ('제자리에 유지' 지시 위반)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 1167
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지시된 클로즈업 샷과 전경의 측면 구도를 정확히 구현했으나, 고정되어야 할 오른쪽 로봇 팔이 누락되어 실격입니다.  ★위반: [gemini-pro] 레퍼런스에 존재하는 오른쪽 로봇 팔 누락 (고정 요소 훼손) / [gemini-pro] 로봇을 눕혀둔 침대의 형태 변경"
   },
   {
    "label": "A",
    "score": 1167,
    "verdict_ko": "요구된 전경 측면 구도를 따르지 않았으며, 락킹된 배경의 침대에 바퀴를 추가하고 로봇 팔 위치를 변경해 실격입니다.  ★위반: [gemini-pro] 부스 내 고정형 침대를 바퀴 달린 이동식 병상으로 변형 (고정 요소 위반) / [gemini-pro] 로봇 팔 위치 변경 및 레이저 누락 ('제자리에 유지' 지시 위반)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S82sh11_sel.png",
    "asset_id": "5c9a06dc-1703-4fa5-8a75-90552c758d97",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf1b-1d60-72ea-9f26-4c94dcbbadba",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S82sh11"
  }
 },
 "S83sh9": {
  "input_fingerprint": "d808db12f4936a05",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면 안, 환자복을 입은 앰버가 환하게 웃고 있는 클로즈업.\n\nLOCATION (lock): On the video-call monitor inside a room at the island temporary shelter, where the illuminated screen shows a remote recovering patient. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Obliquely viewed active video-call monitor with visible screen edge in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Displaying 앰버 smiling in patient clothing) — The active content face is visible at an oblique angle, with a narrow physical edge confirming that this is a screen image; used as Contain the remote face while retaining a small room-side margin; do not reinterpret the image as a hologram.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render 앰버's image with the mediated brightness and tonal compression of an active video-call screen, without inventing illumination in her remote location.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The call is now in video mode, showing Amber's face on the screen.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면 안, 환자복을 입은 앰버가 환하게 웃고 있는 클로즈업.\n\nLOCATION (lock): On the video-call monitor inside a room at the island temporary shelter, where the illuminated screen shows a remote recovering patient. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Obliquely viewed active video-call monitor with visible screen edge in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Displaying 앰버 smiling in patient clothing) — The active content face is visible at an oblique angle, with a narrow physical edge confirming that this is a screen image; used as Contain the remote face while retaining a small room-side margin; do not reinterpret the image as a hologram.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render 앰버's image with the mediated brightness and tonal compression of an active video-call screen, without inventing illumination in her remote location.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The call is now in video mode, showing Amber's face on the screen.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면 안, 환자복을 입은 앰버가 환하게 웃고 있는 클로즈업.\n\nLOCATION (lock): On the video-call monitor inside a room at the island temporary shelter, where the illuminated screen shows a remote recovering patient. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Obliquely viewed active video-call monitor with visible screen edge in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Displaying 앰버 smiling in patient clothing) — The active content face is visible at an oblique angle, with a narrow physical edge confirming that this is a screen image; used as Contain the remote face while retaining a small room-side margin; do not reinterpret the image as a hologram.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render 앰버's image with the mediated brightness and tonal compression of an active video-call screen, without inventing illumination in her remote location.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The call is now in video mode, showing Amber's face on the screen.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "화면 속 소녀가 정면을 바라보며 미소 짓고 있으며, 모니터 화면은 비스듬한 각도로 배치됨.",
    "built_space": "흰색 벽면에 모니터가 부착되어 공간을 형성함.",
    "entities": "금발의 10세 혼혈 소녀가 묘사되었으나, 환자복이 아닌 무늬가 있는 옷을 입고 있음.",
    "hard_violations": [],
    "physics": "모니터가 벽면에 정상적으로 고정되어 있음."
   },
   {
    "label": "B",
    "direction": "화면 속 소녀가 정면을 응시하며 밝게 웃고 있고, 모니터는 비스듬히 벽에 걸려 있음.",
    "built_space": "흰색 벽면에 모니터가 부착되어 있으며, 우측 끝에 레퍼런스와 일치하는 문틀이 보임.",
    "entities": "금발의 10세 혼혈 소녀가 프롬프트에 명시된 하늘색 환자복을 정확히 입고 있음.",
    "hard_violations": [],
    "physics": "모니터가 벽면에 단단히 고정되어 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "명시된 환자복을 정확히 묘사하였으며, 우측 문틀을 통해 레퍼런스의 공간적 특징도 훌륭하게 반영했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "전반적인 구도는 좋으나, 요구된 환자복 대신 패턴이 있는 평상복을 입고 있어 프롬프트 지시를 완벽히 따르지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 속 소녀가 정면을 바라보며 미소 짓고 있으며, 모니터 화면은 비스듬한 각도로 배치됨.",
        "built_space": "흰색 벽면에 모니터가 부착되어 공간을 형성함.",
        "entities": "금발의 10세 혼혈 소녀가 묘사되었으나, 환자복이 아닌 무늬가 있는 옷을 입고 있음.",
        "hard_violations": [],
        "physics": "모니터가 벽면에 정상적으로 고정되어 있음."
       },
       {
        "label": "B",
        "direction": "화면 속 소녀가 정면을 응시하며 밝게 웃고 있고, 모니터는 비스듬히 벽에 걸려 있음.",
        "built_space": "흰색 벽면에 모니터가 부착되어 있으며, 우측 끝에 레퍼런스와 일치하는 문틀이 보임.",
        "entities": "금발의 10세 혼혈 소녀가 프롬프트에 명시된 하늘색 환자복을 정확히 입고 있음.",
        "hard_violations": [],
        "physics": "모니터가 벽면에 단단히 고정되어 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "명시된 환자복을 정확히 묘사하였으며, 우측 문틀을 통해 레퍼런스의 공간적 특징도 훌륭하게 반영했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "전반적인 구도는 좋으나, 요구된 환자복 대신 패턴이 있는 평상복을 입고 있어 프롬프트 지시를 완벽히 따르지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "화면 속 소녀가 정면을 바라보며 미소 짓고 있으며, 모니터 화면은 비스듬한 각도로 배치됨.",
        "built_space": "흰색 벽면에 모니터가 부착되어 공간을 형성함.",
        "entities": "금발의 10세 혼혈 소녀가 묘사되었으나, 환자복이 아닌 무늬가 있는 옷을 입고 있음.",
        "hard_violations": [],
        "physics": "모니터가 벽면에 정상적으로 고정되어 있음."
       },
       {
        "label": "B",
        "direction": "화면 속 소녀가 정면을 응시하며 밝게 웃고 있고, 모니터는 비스듬히 벽에 걸려 있음.",
        "built_space": "흰색 벽면에 모니터가 부착되어 있으며, 우측 끝에 레퍼런스와 일치하는 문틀이 보임.",
        "entities": "금발의 10세 혼혈 소녀가 프롬프트에 명시된 하늘색 환자복을 정확히 입고 있음.",
        "hard_violations": [],
        "physics": "모니터가 벽면에 단단히 고정되어 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "비스듬한 실물 모니터와 환자복 차림의 밝은 미소는 맞지만, 가슴까지 담겨 화면 속 얼굴 클로즈업이 B보다 느슨하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "모니터가 약간 오른쪽에 치우쳤지만, 물리적 테두리 안을 크게 채운 앰버의 얼굴과 환한 웃음이 핵심 클로즈업 지시를 가장 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "아이는 영상통화 카메라를 정면으로 바라보며 이를 드러내고 웃는다. 모니터 표시 면은 촬영 카메라를 향해 비스듬히 열려 있어 영상과 오른쪽 측면 두께가 함께 보인다.",
        "built_space": "밝은 벽에 모니터 한 대가 설치되어 있고 오른쪽에 문틀과 경첩 일부가 보인다. 검은 테두리와 은색 하단부는 참고 사진의 모니터와 유사하다. 참고 사진의 전화기와 창문은 좁은 구도 밖이므로 누락 결함이 아니다. 화면 속에는 흐린 침대 머리판 하나와 침구 일부가 보이며, 중복 설비나 불가능한 반사는 없다.",
        "entities": "화면 안에 금발의 초등학생 연령대 여자아이 한 명만 보인다. 큰 눈, 밝은 미소, 연한 청색 환자복은 지시와 맞는다. 한국계·백인 혼혈이라는 세부 정체성은 외모만으로 확정하기 어렵다. 실물 모니터의 영상으로 읽히며 홀로그램이나 추가 인물, 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "모니터는 벽면에 밀착된 벽걸이 상태다. 아이는 화면 안에서 상체를 곧게 세우고 있으며 하체와 좌석은 크롭 밖이다. 머리와 어깨의 연결 및 환자복 주름은 자연스럽고, 공중에 떠 있거나 지지 없이 놓인 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "아이는 영상통화 카메라 쪽으로 눈을 맞추며 입을 벌려 환하게 웃는다. 촬영 카메라는 모니터의 표시 면을 비스듬히 보고 있으며 오른쪽 물리적 가장자리도 확인된다.",
        "built_space": "밝은 벽에 모니터 한 대가 걸려 있고 오른쪽에 수직 벽 경계가 보인다. 화면은 프레임 중앙보다 다소 오른쪽에 있으며 왼쪽 벽 여백이 상대적으로 넓다. 참고 사진과 벽 재질 및 벽걸이 배치는 부합하지만 모니터 하단은 참고의 은색과 달리 검다. 화면 내부의 침구와 배경 설비는 흐리고 작게 유지되며 중복 설비나 불가능한 반사는 없다.",
        "entities": "금발의 초등학생 연령대 여자아이 한 명이 화면을 크게 채운다. 자연스러운 큰 눈과 환한 웃음이 보이고, 양어깨에는 청색 무늬의 환자복 일부가 드러난다. 한국계·백인 혼혈이라는 정체성을 외모만으로 확정할 수는 없다. 화면의 미세한 표시 질감과 압축된 명암이 영상통화 이미지로 읽히며, 추가 인물이나 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "모니터는 벽에 설치된 상태로 보이며 떠 있는 장치로 표현되지 않는다. 아이의 머리는 목과 어깨에 자연스럽게 연결되고 몸의 나머지는 클로즈업 밖이다. 화면 속 인물에게 도약이나 부유 동작은 없고, 지지 없는 물체도 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "비스듬한 실물 모니터와 환자복 차림의 밝은 미소는 맞지만, 가슴까지 담겨 화면 속 얼굴 클로즈업이 B보다 느슨하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "모니터가 약간 오른쪽에 치우쳤지만, 물리적 테두리 안을 크게 채운 앰버의 얼굴과 환한 웃음이 핵심 클로즈업 지시를 가장 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "아이는 영상통화 카메라를 정면으로 바라보며 이를 드러내고 웃는다. 모니터 표시 면은 촬영 카메라를 향해 비스듬히 열려 있어 영상과 오른쪽 측면 두께가 함께 보인다.",
        "built_space": "밝은 벽에 모니터 한 대가 설치되어 있고 오른쪽에 문틀과 경첩 일부가 보인다. 검은 테두리와 은색 하단부는 참고 사진의 모니터와 유사하다. 참고 사진의 전화기와 창문은 좁은 구도 밖이므로 누락 결함이 아니다. 화면 속에는 흐린 침대 머리판 하나와 침구 일부가 보이며, 중복 설비나 불가능한 반사는 없다.",
        "entities": "화면 안에 금발의 초등학생 연령대 여자아이 한 명만 보인다. 큰 눈, 밝은 미소, 연한 청색 환자복은 지시와 맞는다. 한국계·백인 혼혈이라는 세부 정체성은 외모만으로 확정하기 어렵다. 실물 모니터의 영상으로 읽히며 홀로그램이나 추가 인물, 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "모니터는 벽면에 밀착된 벽걸이 상태다. 아이는 화면 안에서 상체를 곧게 세우고 있으며 하체와 좌석은 크롭 밖이다. 머리와 어깨의 연결 및 환자복 주름은 자연스럽고, 공중에 떠 있거나 지지 없이 놓인 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "아이는 영상통화 카메라 쪽으로 눈을 맞추며 입을 벌려 환하게 웃는다. 촬영 카메라는 모니터의 표시 면을 비스듬히 보고 있으며 오른쪽 물리적 가장자리도 확인된다.",
        "built_space": "밝은 벽에 모니터 한 대가 걸려 있고 오른쪽에 수직 벽 경계가 보인다. 화면은 프레임 중앙보다 다소 오른쪽에 있으며 왼쪽 벽 여백이 상대적으로 넓다. 참고 사진과 벽 재질 및 벽걸이 배치는 부합하지만 모니터 하단은 참고의 은색과 달리 검다. 화면 내부의 침구와 배경 설비는 흐리고 작게 유지되며 중복 설비나 불가능한 반사는 없다.",
        "entities": "금발의 초등학생 연령대 여자아이 한 명이 화면을 크게 채운다. 자연스러운 큰 눈과 환한 웃음이 보이고, 양어깨에는 청색 무늬의 환자복 일부가 드러난다. 한국계·백인 혼혈이라는 정체성을 외모만으로 확정할 수는 없다. 화면의 미세한 표시 질감과 압축된 명암이 영상통화 이미지로 읽히며, 추가 인물이나 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "모니터는 벽에 설치된 상태로 보이며 떠 있는 장치로 표현되지 않는다. 아이의 머리는 목과 어깨에 자연스럽게 연결되고 몸의 나머지는 클로즈업 밖이다. 화면 속 인물에게 도약이나 부유 동작은 없고, 지지 없는 물체도 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.889
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.889
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1889,
   "A": 1714
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1889,
    "verdict_ko": "명시된 환자복을 정확히 묘사하였으며, 우측 문틀을 통해 레퍼런스의 공간적 특징도 훌륭하게 반영했습니다."
   },
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "전반적인 구도는 좋으나, 요구된 환자복 대신 패턴이 있는 평상복을 입고 있어 프롬프트 지시를 완벽히 따르지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L127B01.png",
    "asset_id": "7921e778-417f-49f1-9784-6734e0e6fcd3",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf20-3fec-7732-a297-e2b3a466702f",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S83sh11": {
  "input_fingerprint": "24b7de8b0ea0e941",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 모니터를 향해 애써 밝은 척 활짝 웃음 짓는 얼굴.\n\nLOCATION (lock): At the video-call station inside the temporary shelter room, lit by the active monitor and the room's interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Monitor edge (The video call remains active) — Only a narrow side edge is visible; the content face is directed toward 이현우 and outside the camera's readable view; used as Provide a small foreground eyeline anchor without introducing another visible caller.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained room-appropriate ambient illumination, preserving the contrast between his broad smile and the strain around his eyes without introducing a new source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The video call remains connected; Raul appears on the screen during his turn in the conversation. 이현우: He remains at the video call in his shelter room, dressed in white.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 모니터를 향해 애써 밝은 척 활짝 웃음 짓는 얼굴.\n\nLOCATION (lock): At the video-call station inside the temporary shelter room, lit by the active monitor and the room's interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Monitor edge (The video call remains active) — Only a narrow side edge is visible; the content face is directed toward 이현우 and outside the camera's readable view; used as Provide a small foreground eyeline anchor without introducing another visible caller.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained room-appropriate ambient illumination, preserving the contrast between his broad smile and the strain around his eyes without introducing a new source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The video call remains connected; Raul appears on the screen during his turn in the conversation. 이현우: He remains at the video call in his shelter room, dressed in white.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 모니터를 향해 애써 밝은 척 활짝 웃음 짓는 얼굴.\n\nLOCATION (lock): At the video-call station inside the temporary shelter room, lit by the active monitor and the room's interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Monitor edge (The video call remains active) — Only a narrow side edge is visible; the content face is directed toward 이현우 and outside the camera's readable view; used as Provide a small foreground eyeline anchor without introducing another visible caller.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained room-appropriate ambient illumination, preserving the contrast between his broad smile and the strain around his eyes without introducing a new source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The video call remains connected; Raul appears on the screen during his turn in the conversation. 이현우: He remains at the video call in his shelter room, dressed in white.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 프레임 왼쪽 전경에 배치된 모니터 화면 쪽을 향하고 있습니다.",
    "built_space": "레퍼런스 이미지 좌측 벽의 얼룩과 부착된 종이, 스테인리스 의료용 카트, 병상 테두리 등 고정 요소들이 배경에 알맞게 배치되어 있습니다.",
    "entities": "이현우의 얼굴과 헤어스타일이 캐릭터 레퍼런스와 일치하며 지정된 흰색 긴팔 의상을 입고 있습니다.",
    "hard_violations": [],
    "physics": "자연스럽게 앉아있는 자세를 유지하고 있으며, 전경의 모니터는 프레임 밖 바닥면(책상)에 안정적으로 놓여 있습니다."
   },
   {
    "label": "B",
    "direction": "시선은 프레임 왼쪽 전경의 모니터를 향하고 있습니다.",
    "built_space": "천장의 형광등과 병상은 보이나, 레퍼런스에 없는 야전침대와 옷가지 등 잡동사니가 임의로 추가되었습니다.",
    "entities": "이현우의 외모는 일치하지만, 프롬프트에 명시되지 않은 정체불명의 인물이 배경에 포함되었습니다.",
    "hard_violations": [
     "[gemini-pro] invented people: 배경 우측에 초점이 흐려진 다른 인물이 서 있음",
     "[gpt-high] 오른쪽 배경에 이현우가 아닌 추가 인물이 보이므로, 이 장면에는 이현우만 보여야 한다는 인물 제한을 위반합니다."
    ],
    "physics": "앉아 있는 자세이며 모니터 뒷면과 스탠드가 묘사되어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 샷과 '애써 밝게 웃는 얼굴'의 표정 연기를 잘 담아냈으며, 레퍼런스 이미지의 벽면 얼룩과 의료용 카트 등 공간 디테일을 충실히 재현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에 없는 인물이 배경에 등장하는 치명적인 오류가 있으며, 공간 구성과 의상 디테일(셔츠 하단 글자)이 지시사항과 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 프레임 왼쪽 전경에 배치된 모니터 화면 쪽을 향하고 있습니다.",
        "built_space": "레퍼런스 이미지 좌측 벽의 얼룩과 부착된 종이, 스테인리스 의료용 카트, 병상 테두리 등 고정 요소들이 배경에 알맞게 배치되어 있습니다.",
        "entities": "이현우의 얼굴과 헤어스타일이 캐릭터 레퍼런스와 일치하며 지정된 흰색 긴팔 의상을 입고 있습니다.",
        "hard_violations": [],
        "physics": "자연스럽게 앉아있는 자세를 유지하고 있으며, 전경의 모니터는 프레임 밖 바닥면(책상)에 안정적으로 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "시선은 프레임 왼쪽 전경의 모니터를 향하고 있습니다.",
        "built_space": "천장의 형광등과 병상은 보이나, 레퍼런스에 없는 야전침대와 옷가지 등 잡동사니가 임의로 추가되었습니다.",
        "entities": "이현우의 외모는 일치하지만, 프롬프트에 명시되지 않은 정체불명의 인물이 배경에 포함되었습니다.",
        "hard_violations": [
         "invented people: 배경 우측에 초점이 흐려진 다른 인물이 서 있음"
        ],
        "physics": "앉아 있는 자세이며 모니터 뒷면과 스탠드가 묘사되어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 샷과 '애써 밝게 웃는 얼굴'의 표정 연기를 잘 담아냈으며, 레퍼런스 이미지의 벽면 얼룩과 의료용 카트 등 공간 디테일을 충실히 재현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에 없는 인물이 배경에 등장하는 치명적인 오류가 있으며, 공간 구성과 의상 디테일(셔츠 하단 글자)이 지시사항과 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 프레임 왼쪽 전경에 배치된 모니터 화면 쪽을 향하고 있습니다.",
        "built_space": "레퍼런스 이미지 좌측 벽의 얼룩과 부착된 종이, 스테인리스 의료용 카트, 병상 테두리 등 고정 요소들이 배경에 알맞게 배치되어 있습니다.",
        "entities": "이현우의 얼굴과 헤어스타일이 캐릭터 레퍼런스와 일치하며 지정된 흰색 긴팔 의상을 입고 있습니다.",
        "hard_violations": [],
        "physics": "자연스럽게 앉아있는 자세를 유지하고 있으며, 전경의 모니터는 프레임 밖 바닥면(책상)에 안정적으로 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "시선은 프레임 왼쪽 전경의 모니터를 향하고 있습니다.",
        "built_space": "천장의 형광등과 병상은 보이나, 레퍼런스에 없는 야전침대와 옷가지 등 잡동사니가 임의로 추가되었습니다.",
        "entities": "이현우의 외모는 일치하지만, 프롬프트에 명시되지 않은 정체불명의 인물이 배경에 포함되었습니다.",
        "hard_violations": [
         "invented people: 배경 우측에 초점이 흐려진 다른 인물이 서 있음"
        ],
        "physics": "앉아 있는 자세이며 모니터 뒷면과 스탠드가 묘사되어 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "모니터를 보며 웃는 동작은 맞지만, 오른쪽 배경에 금지된 추가 인물이 있고 구도도 얼굴 클로즈업보다 넓으며 모니터 뒷면이 지나치게 크게 보입니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이현우의 얼굴과 모니터를 향한 활짝 웃음을 더 가깝게 담고 장소도 잘 이어지지만, 모니터를 좁은 옆면만 보이라는 구도 지시는 충족하지 못합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 얼굴과 두 눈은 왼쪽 전경 모니터를 향합니다. 카메라에는 통풍구가 있는 모니터 뒷면이 보여 화면이 이현우 쪽을 향하는 사용 방향은 맞습니다. 통화 상대나 화면 내용은 보이지 않습니다.",
        "built_space": "왼쪽 전경 모니터 1대, 인물 뒤 침대 1개, 왼쪽 아래 금속 카트 1개, 뒤쪽 선반형 카트 1개와 여러 천장 조명이 보입니다. 흰 벽과 의료용 가구는 참고 장소의 재질을 따르지만, 뒤쪽에는 야전침대처럼 보이는 녹색 구조물도 있습니다. 이현우는 오른쪽 의자에 앉아 모니터 앞에 위치합니다. 모니터의 넓은 뒷면과 상체·팔까지 포함해 지정된 얼굴 클로즈업보다 넓습니다.",
        "entities": "주인공은 검은 헝클어진 머리와 흰 긴팔 상의를 가진 한국계 청년 남성으로 보이며 참고 인물의 특징을 대체로 따릅니다. 다만 얼굴은 18세보다 다소 성숙하게 읽힙니다. 치아를 드러낸 웃음과 좁아진 눈은 있으나 애써 밝게 행동하는 긴장감은 약합니다. 오른쪽 끝 배경에 별도 인물의 머리와 몸 일부가 보입니다. 라울은 노출되지 않았고, 모니터의 통화 연결 상태는 확인할 수 없습니다.",
        "hard_violations": [
         "오른쪽 배경에 이현우가 아닌 추가 인물이 보이므로, 이 장면에는 이현우만 보여야 한다는 인물 제한을 위반합니다."
        ],
        "physics": "이현우의 등 뒤에 의자 등받이가 있고 상체가 그 앞에서 자연스럽게 앞으로 기울어 있습니다. 팔은 아래 작업면 쪽으로 내려갑니다. 모니터 하부에는 지지 구조와 케이블이 보이며 침구와 용기는 각각 침대와 카트 위에 놓여 있습니다. 공중에 떠 있거나 해부학적으로 불가능한 자세는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "이현우의 얼굴과 시선은 왼쪽 앞 모니터 쪽을 향하고 렌즈를 직접 보지 않습니다. 모니터에는 넓고 흐린 외장면처럼 보이는 면이 노출되며 통화 내용은 보이지 않습니다. 화면 방향이 뒤집혔다고 단정할 근거는 없지만, 요구한 좁은 측면 노출은 아닙니다.",
        "built_space": "왼쪽 전경 통화용 모니터 1대, 뒤쪽 금속 카트 1개와 그 위 용기들, 오른쪽 침대 난간 1개, 오른쪽 끝 배경 모니터 1대, 창틀과 흰 벽이 보입니다. 참고 장소의 침대·의료 장비·마모된 벽과 일관성이 있습니다. 이현우 뒤에는 의자 등받이가 보이며 모니터 앞에 앉은 배치도 자연스럽습니다. 얼굴 비중은 A보다 크지만 어깨와 가슴도 상당 부분 포함되고, 모니터가 왼쪽 전경을 넓게 차지합니다.",
        "entities": "추가 인물 없이 이현우 한 명만 보입니다. 한국계 청년 남성의 얼굴, 헝클어진 검은 머리, 흰 둥근 목 긴팔 상의가 참고 인물과 대체로 맞습니다. 활짝 드러난 치아와 눈가의 조임은 요구된 웃음에 가깝지만, 힘겨움을 감춘 표정이라기보다 자연스러운 기쁨으로도 읽힙니다. 라울이나 다른 통화 상대는 보이지 않으며 통화 연결 여부 자체는 확인되지 않습니다. 벽의 종이와 용기 표시는 흐려 읽을 수 없습니다.",
        "hard_violations": [],
        "physics": "오른쪽에 보이는 의자 등받이가 앉은 자세를 뒷받침하고, 목과 어깨 및 앞으로 뻗은 팔의 연결은 자연스럽습니다. 좌판과 손, 모니터 받침은 프레임 밖이므로 접촉 상태를 직접 확인할 수 없지만 떠 있다는 증거도 없습니다. 뒤쪽 용기들은 카트 상판에 놓여 있고 침대 난간도 침대에 연결되어 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "모니터를 보며 웃는 동작은 맞지만, 오른쪽 배경에 금지된 추가 인물이 있고 구도도 얼굴 클로즈업보다 넓으며 모니터 뒷면이 지나치게 크게 보입니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이현우의 얼굴과 모니터를 향한 활짝 웃음을 더 가깝게 담고 장소도 잘 이어지지만, 모니터를 좁은 옆면만 보이라는 구도 지시는 충족하지 못합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 얼굴과 두 눈은 왼쪽 전경 모니터를 향합니다. 카메라에는 통풍구가 있는 모니터 뒷면이 보여 화면이 이현우 쪽을 향하는 사용 방향은 맞습니다. 통화 상대나 화면 내용은 보이지 않습니다.",
        "built_space": "왼쪽 전경 모니터 1대, 인물 뒤 침대 1개, 왼쪽 아래 금속 카트 1개, 뒤쪽 선반형 카트 1개와 여러 천장 조명이 보입니다. 흰 벽과 의료용 가구는 참고 장소의 재질을 따르지만, 뒤쪽에는 야전침대처럼 보이는 녹색 구조물도 있습니다. 이현우는 오른쪽 의자에 앉아 모니터 앞에 위치합니다. 모니터의 넓은 뒷면과 상체·팔까지 포함해 지정된 얼굴 클로즈업보다 넓습니다.",
        "entities": "주인공은 검은 헝클어진 머리와 흰 긴팔 상의를 가진 한국계 청년 남성으로 보이며 참고 인물의 특징을 대체로 따릅니다. 다만 얼굴은 18세보다 다소 성숙하게 읽힙니다. 치아를 드러낸 웃음과 좁아진 눈은 있으나 애써 밝게 행동하는 긴장감은 약합니다. 오른쪽 끝 배경에 별도 인물의 머리와 몸 일부가 보입니다. 라울은 노출되지 않았고, 모니터의 통화 연결 상태는 확인할 수 없습니다.",
        "hard_violations": [
         "오른쪽 배경에 이현우가 아닌 추가 인물이 보이므로, 이 장면에는 이현우만 보여야 한다는 인물 제한을 위반합니다."
        ],
        "physics": "이현우의 등 뒤에 의자 등받이가 있고 상체가 그 앞에서 자연스럽게 앞으로 기울어 있습니다. 팔은 아래 작업면 쪽으로 내려갑니다. 모니터 하부에는 지지 구조와 케이블이 보이며 침구와 용기는 각각 침대와 카트 위에 놓여 있습니다. 공중에 떠 있거나 해부학적으로 불가능한 자세는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "이현우의 얼굴과 시선은 왼쪽 앞 모니터 쪽을 향하고 렌즈를 직접 보지 않습니다. 모니터에는 넓고 흐린 외장면처럼 보이는 면이 노출되며 통화 내용은 보이지 않습니다. 화면 방향이 뒤집혔다고 단정할 근거는 없지만, 요구한 좁은 측면 노출은 아닙니다.",
        "built_space": "왼쪽 전경 통화용 모니터 1대, 뒤쪽 금속 카트 1개와 그 위 용기들, 오른쪽 침대 난간 1개, 오른쪽 끝 배경 모니터 1대, 창틀과 흰 벽이 보입니다. 참고 장소의 침대·의료 장비·마모된 벽과 일관성이 있습니다. 이현우 뒤에는 의자 등받이가 보이며 모니터 앞에 앉은 배치도 자연스럽습니다. 얼굴 비중은 A보다 크지만 어깨와 가슴도 상당 부분 포함되고, 모니터가 왼쪽 전경을 넓게 차지합니다.",
        "entities": "추가 인물 없이 이현우 한 명만 보입니다. 한국계 청년 남성의 얼굴, 헝클어진 검은 머리, 흰 둥근 목 긴팔 상의가 참고 인물과 대체로 맞습니다. 활짝 드러난 치아와 눈가의 조임은 요구된 웃음에 가깝지만, 힘겨움을 감춘 표정이라기보다 자연스러운 기쁨으로도 읽힙니다. 라울이나 다른 통화 상대는 보이지 않으며 통화 연결 여부 자체는 확인되지 않습니다. 벽의 종이와 용기 표시는 흐려 읽을 수 없습니다.",
        "hard_violations": [],
        "physics": "오른쪽에 보이는 의자 등받이가 앉은 자세를 뒷받침하고, 목과 어깨 및 앞으로 뻗은 팔의 연결은 자연스럽습니다. 좌판과 손, 모니터 받침은 프레임 밖이므로 접촉 상태를 직접 확인할 수 없지만 떠 있다는 증거도 없습니다. 뒤쪽 용기들은 카트 상판에 놓여 있고 침대 난간도 침대에 연결되어 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.714
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.464
   },
   "violations": {
    "B": [
     "[gemini-pro] invented people: 배경 우측에 초점이 흐려진 다른 인물이 서 있음",
     "[gpt-high] 오른쪽 배경에 이현우가 아닌 추가 인물이 보이므로, 이 장면에는 이현우만 보여야 한다는 인물 제한을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 464
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 클로즈업 샷과 '애써 밝게 웃는 얼굴'의 표정 연기를 잘 담아냈으며, 레퍼런스 이미지의 벽면 얼룩과 의료용 카트 등 공간 디테일을 충실히 재현했습니다."
   },
   {
    "label": "B",
    "score": 464,
    "verdict_ko": "프롬프트에 없는 인물이 배경에 등장하는 치명적인 오류가 있으며, 공간 구성과 의상 디테일(셔츠 하단 글자)이 지시사항과 어긋납니다.  ★위반: [gemini-pro] invented people: 배경 우측에 초점이 흐려진 다른 인물이 서 있음 / [gpt-high] 오른쪽 배경에 이현우가 아닌 추가 인물이 보이므로, 이 장면에는 이현우만 보여야 한다는 인물 제한을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S81sh5_sel.png",
    "asset_id": "6dcf2374-bce6-4988-86e2-821e5a912464",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf23-ea7c-759b-bbb3-c9b88766ed24",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S81sh5"
  }
 },
 "S83sh17": {
  "input_fingerprint": "ef631c1dc9a6ecb7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 침대에 누운 이현우가 천장을 멍하니 바라보며 미간에 깊은 주름을 잡은 심란한 얼굴 클로즈업.\n\nLOCATION (lock): On the bed inside the temporary shelter room, beneath the room's fluorescent lighting after the call. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bed in 이현우's room (Occupied by 이현우 after the call) — A limited area of the upper surface borders his head and shoulder; used as Ground the isolated face in his resting position without revealing unrelated room features.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued, room-appropriate ambient illumination that reveals the furrowed brow without adding a motivated color shift or a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The call has ended and the video image has disappeared. The shelter-room bed remains available. 이현우: He is alone, lying on the bed in his white clothing with a worried expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 침대에 누운 이현우가 천장을 멍하니 바라보며 미간에 깊은 주름을 잡은 심란한 얼굴 클로즈업.\n\nLOCATION (lock): On the bed inside the temporary shelter room, beneath the room's fluorescent lighting after the call. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bed in 이현우's room (Occupied by 이현우 after the call) — A limited area of the upper surface borders his head and shoulder; used as Ground the isolated face in his resting position without revealing unrelated room features.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued, room-appropriate ambient illumination that reveals the furrowed brow without adding a motivated color shift or a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The call has ended and the video image has disappeared. The shelter-room bed remains available. 이현우: He is alone, lying on the bed in his white clothing with a worried expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 침대에 누운 이현우가 천장을 멍하니 바라보며 미간에 깊은 주름을 잡은 심란한 얼굴 클로즈업.\n\nLOCATION (lock): On the bed inside the temporary shelter room, beneath the room's fluorescent lighting after the call. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bed in 이현우's room (Occupied by 이현우 after the call) — A limited area of the upper surface borders his head and shoulder; used as Ground the isolated face in his resting position without revealing unrelated room features.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued, room-appropriate ambient illumination that reveals the furrowed brow without adding a motivated color shift or a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The call has ended and the video image has disappeared. The shelter-room bed remains available. 이현우: He is alone, lying on the bed in his white clothing with a worried expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선이 카메라 렌즈 아래쪽을 향하고 있어 천장을 바라보지 않음.",
    "built_space": "침대, 하얀 베개, 머리 위쪽의 하얀 구조물 일부가 보임.",
    "entities": "하얀 옷을 입고 미간에 주름을 지은 이현우와 일치하는 남성.",
    "hard_violations": [],
    "physics": "머리와 등 상부가 베개와 침대 표면에 안정적으로 지탱됨."
   },
   {
    "label": "B",
    "direction": "시선이 위쪽 허공을 향해 천장을 바라보는 묘사가 명확함.",
    "built_space": "침대와 하얀 베개, 머리 뒤로 하얀 구조물과 파란색 배경이 이전 샷과 일치함.",
    "entities": "하얀 옷을 입고 미간에 깊은 주름을 지은 이현우와 일치하는 남성.",
    "hard_violations": [],
    "physics": "머리와 상체가 베개와 침대 표면에 안정적으로 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "자유로운 카메라 각도를 활용하여 천장을 바라보는 시선과 심란한 표정을 완벽하게 포착했으며, 이전 샷의 배경과도 잘 일치합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "미간의 주름과 인물 외형은 일치하나, 시선이 천장이 아닌 아래쪽을 향하고 있어 행동 묘사가 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선이 카메라 렌즈 아래쪽을 향하고 있어 천장을 바라보지 않음.",
        "built_space": "침대, 하얀 베개, 머리 위쪽의 하얀 구조물 일부가 보임.",
        "entities": "하얀 옷을 입고 미간에 주름을 지은 이현우와 일치하는 남성.",
        "hard_violations": [],
        "physics": "머리와 등 상부가 베개와 침대 표면에 안정적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "시선이 위쪽 허공을 향해 천장을 바라보는 묘사가 명확함.",
        "built_space": "침대와 하얀 베개, 머리 뒤로 하얀 구조물과 파란색 배경이 이전 샷과 일치함.",
        "entities": "하얀 옷을 입고 미간에 깊은 주름을 지은 이현우와 일치하는 남성.",
        "hard_violations": [],
        "physics": "머리와 상체가 베개와 침대 표면에 안정적으로 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "자유로운 카메라 각도를 활용하여 천장을 바라보는 시선과 심란한 표정을 완벽하게 포착했으며, 이전 샷의 배경과도 잘 일치합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "미간의 주름과 인물 외형은 일치하나, 시선이 천장이 아닌 아래쪽을 향하고 있어 행동 묘사가 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선이 카메라 렌즈 아래쪽을 향하고 있어 천장을 바라보지 않음.",
        "built_space": "침대, 하얀 베개, 머리 위쪽의 하얀 구조물 일부가 보임.",
        "entities": "하얀 옷을 입고 미간에 주름을 지은 이현우와 일치하는 남성.",
        "hard_violations": [],
        "physics": "머리와 등 상부가 베개와 침대 표면에 안정적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "시선이 위쪽 허공을 향해 천장을 바라보는 묘사가 명확함.",
        "built_space": "침대와 하얀 베개, 머리 뒤로 하얀 구조물과 파란색 배경이 이전 샷과 일치함.",
        "entities": "하얀 옷을 입고 미간에 깊은 주름을 지은 이현우와 일치하는 남성.",
        "hard_violations": [],
        "physics": "머리와 상체가 베개와 침대 표면에 안정적으로 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "얼굴 중심의 밀착 클로즈업과 카메라를 비껴 천장으로 향하는 시선이 정확하며, 깊게 접힌 미간과 침구에 실린 머리 무게가 심란한 휴식 순간을 구현한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "인물과 침구의 연속성, 깊은 미간 주름은 충실하지만, 이전 사진과 거의 같은 정면 구도여서 천장을 멍하니 보는 시선이 렌즈 응시처럼 읽히고 어깨·가슴의 비중도 더 크다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "누운 얼굴이 화면 오른쪽 위를 향하고 두 눈도 그쪽 상방을 바라본다. 비스듬히 내려다보는 카메라가 아니라 화면 밖 천장 쪽이 시선의 대상으로 읽힌다. 무기나 방향을 확인할 휴대 소품은 없다.",
        "built_space": "침대 한 개의 상면, 머리를 받친 흰 베개 한 개, 흰 덮개 일부가 보인다. 상단 가장자리에는 이전 사진의 회색 머리맡 구조와 밝은 창턱에 해당하는 부분이 비스듬히 남아 있다. 머리와 어깨 주변으로 공간이 제한되어 있으며 중복 설비나 반사는 없다.",
        "entities": "젊은 한국계 남성으로 설정된 이현우 한 명만 보이며, 참고 사진의 얼굴 윤곽과 굵은 눈썹, 헝클어진 검은 머리, 흰 둥근목 상의가 잘 이어진다. 눈은 정상적인 홍채와 동공을 유지한다. 미간이 깊게 접히고 입이 다물려 걱정스러운 표정이다. 흰 천 침구는 참고와 부합하며 통화 영상, 다른 사람, 읽을 수 있는 글자는 없다. 전체 색은 참고보다 다소 차갑다.",
        "hard_violations": [],
        "physics": "뒤통수와 목 주변을 베개가 받치고, 어깨와 몸통은 침대 위에 놓여 있다. 베개 눌림과 천 주름이 접촉에 맞으며 덮개도 몸 위에 자연스럽게 얹혀 있다. 떠 있는 신체나 지지 없는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "얼굴은 거의 수직 위를 향하며 두 눈은 얼굴 바로 위 카메라 근처를 바라본다. 누운 자세상 천장 방향과 양립하지만, 천장에 풀린 시선보다는 렌즈 부근을 집중해서 보는 인상이 강하다. 방향성 있는 소품은 없다.",
        "built_space": "침대 한 개와 흰 베개 한 개가 보이고, 상단에 회색 머리맡 구조와 밝은 창턱 부분이 수평으로 이어진다. 이전 사진의 배치와 일치하며 머리는 베개 중앙에 놓여 있다. 얼굴 외에 양어깨와 윗가슴, 베개 양쪽 면적이 비교적 넓게 포함된다. 중복 설비나 광학적으로 문제 되는 반사는 없다.",
        "entities": "이현우 한 명만 있으며 참고의 젊은 남성 얼굴, 검은 헝클어진 머리, 흰 둥근목 상의가 유지된다. 정상적인 눈과 깊은 세로 미간 주름, 굳은 입매로 심란함을 표현한다. 흰 베개와 시트의 소재도 참고에 맞는다. 통화 화면이나 다른 인물, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리 뒤가 베개에 닿고 어깨와 등은 침대에 지지되어 있다. 목과 몸통의 정렬은 바로 누운 자세로 가능하며 베개와 상의의 주름도 자연스럽다. 지지 없이 뜬 물체나 불가능한 신체 배치는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "얼굴 중심의 밀착 클로즈업과 카메라를 비껴 천장으로 향하는 시선이 정확하며, 깊게 접힌 미간과 침구에 실린 머리 무게가 심란한 휴식 순간을 구현한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "인물과 침구의 연속성, 깊은 미간 주름은 충실하지만, 이전 사진과 거의 같은 정면 구도여서 천장을 멍하니 보는 시선이 렌즈 응시처럼 읽히고 어깨·가슴의 비중도 더 크다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "누운 얼굴이 화면 오른쪽 위를 향하고 두 눈도 그쪽 상방을 바라본다. 비스듬히 내려다보는 카메라가 아니라 화면 밖 천장 쪽이 시선의 대상으로 읽힌다. 무기나 방향을 확인할 휴대 소품은 없다.",
        "built_space": "침대 한 개의 상면, 머리를 받친 흰 베개 한 개, 흰 덮개 일부가 보인다. 상단 가장자리에는 이전 사진의 회색 머리맡 구조와 밝은 창턱에 해당하는 부분이 비스듬히 남아 있다. 머리와 어깨 주변으로 공간이 제한되어 있으며 중복 설비나 반사는 없다.",
        "entities": "젊은 한국계 남성으로 설정된 이현우 한 명만 보이며, 참고 사진의 얼굴 윤곽과 굵은 눈썹, 헝클어진 검은 머리, 흰 둥근목 상의가 잘 이어진다. 눈은 정상적인 홍채와 동공을 유지한다. 미간이 깊게 접히고 입이 다물려 걱정스러운 표정이다. 흰 천 침구는 참고와 부합하며 통화 영상, 다른 사람, 읽을 수 있는 글자는 없다. 전체 색은 참고보다 다소 차갑다.",
        "hard_violations": [],
        "physics": "뒤통수와 목 주변을 베개가 받치고, 어깨와 몸통은 침대 위에 놓여 있다. 베개 눌림과 천 주름이 접촉에 맞으며 덮개도 몸 위에 자연스럽게 얹혀 있다. 떠 있는 신체나 지지 없는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "얼굴은 거의 수직 위를 향하며 두 눈은 얼굴 바로 위 카메라 근처를 바라본다. 누운 자세상 천장 방향과 양립하지만, 천장에 풀린 시선보다는 렌즈 부근을 집중해서 보는 인상이 강하다. 방향성 있는 소품은 없다.",
        "built_space": "침대 한 개와 흰 베개 한 개가 보이고, 상단에 회색 머리맡 구조와 밝은 창턱 부분이 수평으로 이어진다. 이전 사진의 배치와 일치하며 머리는 베개 중앙에 놓여 있다. 얼굴 외에 양어깨와 윗가슴, 베개 양쪽 면적이 비교적 넓게 포함된다. 중복 설비나 광학적으로 문제 되는 반사는 없다.",
        "entities": "이현우 한 명만 있으며 참고의 젊은 남성 얼굴, 검은 헝클어진 머리, 흰 둥근목 상의가 유지된다. 정상적인 눈과 깊은 세로 미간 주름, 굳은 입매로 심란함을 표현한다. 흰 베개와 시트의 소재도 참고에 맞는다. 통화 화면이나 다른 인물, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리 뒤가 베개에 닿고 어깨와 등은 침대에 지지되어 있다. 목과 몸통의 정렬은 바로 누운 자세로 가능하며 베개와 상의의 주름도 자연스럽다. 지지 없이 뜬 물체나 불가능한 신체 배치는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.603,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.603,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1603
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "자유로운 카메라 각도를 활용하여 천장을 바라보는 시선과 심란한 표정을 완벽하게 포착했으며, 이전 샷의 배경과도 잘 일치합니다."
   },
   {
    "label": "A",
    "score": 1603,
    "verdict_ko": "미간의 주름과 인물 외형은 일치하나, 시선이 천장이 아닌 아래쪽을 향하고 있어 행동 묘사가 아쉽습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S81sh1_sel.png",
    "asset_id": "27f9c6b1-89aa-471c-9ff2-d1d0692dddcb",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf27-7c5d-76ad-8d40-31601a57ee64",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S81sh1"
  }
 },
 "S84sh1": {
  "input_fingerprint": "bfa64d3178ddf83d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차가운 실험실 침대 위, 수많은 전선이 연결된 채 누워 있던 찰리의 두 눈에 푸른빛이 번쩍 켜지는 찰나.\n\nLOCATION (lock): On the wired examination bed inside the main laboratory's glass enclosure at night, with the robot's blue eyes newly illuminated. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Laboratory bed (Supporting 찰리 as he awakens) — A narrow portion of the upper surface is visible beside his head and shoulder; used as Provides a stable reclining plane beneath the oblique facial composition; Attached wires (Still connected to 찰리's body); used as Enter along the lower frame edge as evidence of the procedure without obscuring his eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination lets the brief blue activation in 찰리's eyes register clearly without extending that color throughout the laboratory.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie lies on the stainless-steel laboratory bed inside the glass enclosure, with multiple wires still attached across his body. His power has just returned; the wires have not yet been pulled off.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차가운 실험실 침대 위, 수많은 전선이 연결된 채 누워 있던 찰리의 두 눈에 푸른빛이 번쩍 켜지는 찰나.\n\nLOCATION (lock): On the wired examination bed inside the main laboratory's glass enclosure at night, with the robot's blue eyes newly illuminated. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Laboratory bed (Supporting 찰리 as he awakens) — A narrow portion of the upper surface is visible beside his head and shoulder; used as Provides a stable reclining plane beneath the oblique facial composition; Attached wires (Still connected to 찰리's body); used as Enter along the lower frame edge as evidence of the procedure without obscuring his eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination lets the brief blue activation in 찰리's eyes register clearly without extending that color throughout the laboratory.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie lies on the stainless-steel laboratory bed inside the glass enclosure, with multiple wires still attached across his body. His power has just returned; the wires have not yet been pulled off.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차가운 실험실 침대 위, 수많은 전선이 연결된 채 누워 있던 찰리의 두 눈에 푸른빛이 번쩍 켜지는 찰나.\n\nLOCATION (lock): On the wired examination bed inside the main laboratory's glass enclosure at night, with the robot's blue eyes newly illuminated. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Laboratory bed (Supporting 찰리 as he awakens) — A narrow portion of the upper surface is visible beside his head and shoulder; used as Provides a stable reclining plane beneath the oblique facial composition; Attached wires (Still connected to 찰리's body); used as Enter along the lower frame edge as evidence of the procedure without obscuring his eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination lets the brief blue activation in 찰리's eyes register clearly without extending that color throughout the laboratory.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie lies on the stainless-steel laboratory bed inside the glass enclosure, with multiple wires still attached across his body. His power has just returned; the wires have not yet been pulled off.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 얼굴과 시선은 위쪽 허공을 향해 있음.",
    "built_space": "유리벽으로 둘러싸인 어두운 실험실 내부. 천장의 둥근 조명은 꺼져 있거나 어두우며, 찰리는 중앙의 금속 침대에 누워 있음.",
    "entities": "찰리(흰색 각진 얼굴, 파란색으로 빛나는 점 형태의 눈). 레퍼런스와 달리 선 형태의 입에서도 파란 불빛이 남.",
    "hard_violations": [],
    "physics": "찰리의 몸은 실험실 금속 침대 위에 안정적으로 지탱되어 있음."
   },
   {
    "label": "B",
    "direction": "찰리의 얼굴과 시선은 위쪽 허공을 향해 있음.",
    "built_space": "유리벽으로 둘러싸인 밝은 실험실 내부. 천장의 둥근 조명과 배경 조명들이 켜져 있으며, 찰리는 금속 침대에 누워 있음.",
    "entities": "찰리(흰색 각진 얼굴, 파란색으로 빛나는 점 형태의 눈, 검은 선 형태의 입).",
    "hard_violations": [],
    "physics": "찰리의 몸은 금속 침대 위에 안정적으로 지탱되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "클로즈업 지시에는 다소 못 미치나 B보다 화각이 좁으며, '억제된 조명' 지시를 잘 따름(입이 파랗게 빛나는 사소한 오류 존재)."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "클로즈업 화각 지시를 무시하고 하반신까지 노출했으며, 주변 조명이 너무 밝아 프롬프트의 무드와 어긋남."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴과 시선은 위쪽 허공을 향해 있음.",
        "built_space": "유리벽으로 둘러싸인 어두운 실험실 내부. 천장의 둥근 조명은 꺼져 있거나 어두우며, 찰리는 중앙의 금속 침대에 누워 있음.",
        "entities": "찰리(흰색 각진 얼굴, 파란색으로 빛나는 점 형태의 눈). 레퍼런스와 달리 선 형태의 입에서도 파란 불빛이 남.",
        "hard_violations": [],
        "physics": "찰리의 몸은 실험실 금속 침대 위에 안정적으로 지탱되어 있음."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴과 시선은 위쪽 허공을 향해 있음.",
        "built_space": "유리벽으로 둘러싸인 밝은 실험실 내부. 천장의 둥근 조명과 배경 조명들이 켜져 있으며, 찰리는 금속 침대에 누워 있음.",
        "entities": "찰리(흰색 각진 얼굴, 파란색으로 빛나는 점 형태의 눈, 검은 선 형태의 입).",
        "hard_violations": [],
        "physics": "찰리의 몸은 금속 침대 위에 안정적으로 지탱되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "클로즈업 지시에는 다소 못 미치나 B보다 화각이 좁으며, '억제된 조명' 지시를 잘 따름(입이 파랗게 빛나는 사소한 오류 존재)."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "클로즈업 화각 지시를 무시하고 하반신까지 노출했으며, 주변 조명이 너무 밝아 프롬프트의 무드와 어긋남."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴과 시선은 위쪽 허공을 향해 있음.",
        "built_space": "유리벽으로 둘러싸인 어두운 실험실 내부. 천장의 둥근 조명은 꺼져 있거나 어두우며, 찰리는 중앙의 금속 침대에 누워 있음.",
        "entities": "찰리(흰색 각진 얼굴, 파란색으로 빛나는 점 형태의 눈). 레퍼런스와 달리 선 형태의 입에서도 파란 불빛이 남.",
        "hard_violations": [],
        "physics": "찰리의 몸은 실험실 금속 침대 위에 안정적으로 지탱되어 있음."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴과 시선은 위쪽 허공을 향해 있음.",
        "built_space": "유리벽으로 둘러싸인 밝은 실험실 내부. 천장의 둥근 조명과 배경 조명들이 켜져 있으며, 찰리는 금속 침대에 누워 있음.",
        "entities": "찰리(흰색 각진 얼굴, 파란색으로 빛나는 점 형태의 눈, 검은 선 형태의 입).",
        "hard_violations": [],
        "physics": "찰리의 몸은 금속 침대 위에 안정적으로 지탱되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "A보다 얼굴에 가까이 접근했지만 여전히 상반신 중심의 정면 구도로, 요구한 비스듬한 얼굴 클로즈업과 좁은 침대 노출을 충족하지 못하며 입에도 불필요한 푸른빛이 생겼다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 외형·자세와 두 눈의 점등은 충실하지만, 이전 숏의 넓은 구도를 거의 그대로 반복해 핵심 지시인 얼굴 클로즈업을 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴과 두 점 형태의 눈은 침대 발치 쪽 카메라를 정면으로 향한다. 특정 응시 대상은 제시되지 않았으며 눈을 가리는 전선은 없다. 얼굴은 비스듬한 배치가 아니라 화면 중앙에 수직으로 놓였다.",
        "built_space": "경사진 금속 검사 침대 한 개와 이를 둘러싼 유리 enclosure, 여러 수직 프레임, 상부의 원형 조명 구조가 보인다. 왼쪽 장비군과 오른쪽 작업대·의자·대형 장비는 참조의 배치를 대체로 유지한다. 다만 머리와 어깨 옆의 좁은 침대 표면만 보여야 하는데 침대 대부분과 실험실 전경이 드러난다. 참조보다 천장 점 조명이 늘어났으며 불가능한 반사는 보이지 않는다.",
        "entities": "찰리 한 개체만 보이며 각진 흰 얼굴판, 작은 원형 눈 두 개, 검은 선 형태의 입, 마모된 회색 금속 몸체가 참조와 가깝다. 비인간 기계이므로 인간의 나이·성별·민족성 판단은 해당하지 않는다. 양쪽 눈에 푸른빛이 켜졌고 가슴과 몸 주변의 여러 전선은 연결된 상태다. 어두운 창과 절제된 주변광은 밤 설정에 맞으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몸통과 팔다리는 경사진 침대 면에 놓이고 머리는 기계 목으로 지지된다. 전선은 가슴 접속부와 몸 주변에서 이어져 침대에 걸쳐 늘어지므로 부유하지 않는다. 참조의 누운 자세를 유지하며 점등만 일어난 장면으로 물리적으로 성립한다."
       },
       {
        "label": "B",
        "direction": "두 눈과 얼굴은 침대 발치 쪽 카메라를 정면으로 향하고 별도의 응시 대상은 없다. 전선은 가슴에서 화면 아래와 오른쪽으로 이어져 눈을 가리지 않는다. 얼굴이 더 크게 보이지만 요구한 비스듬한 얼굴 구성은 아니다.",
        "built_space": "경사진 금속 침대 한 개, 주변 유리벽과 수직 프레임, 상부 원형 조명 구조가 보인다. 왼쪽 실험 장비와 오른쪽 작업대·의자·대형 장비의 관계는 참조와 대체로 일치한다. 침대 위 상반신을 가까이 보지만 어깨 양옆과 몸통 주변의 침대 면, 넓은 실험실 배경이 여전히 포함되어 지정된 좁은 배경 범위를 초과한다. 중복 침대나 불가능한 반사는 보이지 않는다.",
        "entities": "찰리 한 개체의 각진 흰 얼굴판, 두 점 형태의 기계 눈, 선 형태의 입과 낡은 금속 몸체가 보인다. 인간의 나이·성별·민족성 조건은 해당하지 않는다. 눈 두 개에 푸른빛이 있으나 A보다 점등의 번쩍임이 약하고, 참조에서 검던 입선까지 파랗게 발광한다. 가슴 접속부에는 여러 전선이 남아 있다. 어두운 실험실은 밤에 부합하며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "상반신은 경사진 침대에 기대어 있고 머리는 노출된 기계 목이 받친다. 전선은 가슴 접속부에 연결되어 어깨와 침대 가장자리를 따라 처진다. 지지 없이 뜬 몸이나 물체는 없으며, 보이는 범위의 자세는 참조의 누운 상태와 양립한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "A보다 얼굴에 가까이 접근했지만 여전히 상반신 중심의 정면 구도로, 요구한 비스듬한 얼굴 클로즈업과 좁은 침대 노출을 충족하지 못하며 입에도 불필요한 푸른빛이 생겼다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 외형·자세와 두 눈의 점등은 충실하지만, 이전 숏의 넓은 구도를 거의 그대로 반복해 핵심 지시인 얼굴 클로즈업을 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴과 두 점 형태의 눈은 침대 발치 쪽 카메라를 정면으로 향한다. 특정 응시 대상은 제시되지 않았으며 눈을 가리는 전선은 없다. 얼굴은 비스듬한 배치가 아니라 화면 중앙에 수직으로 놓였다.",
        "built_space": "경사진 금속 검사 침대 한 개와 이를 둘러싼 유리 enclosure, 여러 수직 프레임, 상부의 원형 조명 구조가 보인다. 왼쪽 장비군과 오른쪽 작업대·의자·대형 장비는 참조의 배치를 대체로 유지한다. 다만 머리와 어깨 옆의 좁은 침대 표면만 보여야 하는데 침대 대부분과 실험실 전경이 드러난다. 참조보다 천장 점 조명이 늘어났으며 불가능한 반사는 보이지 않는다.",
        "entities": "찰리 한 개체만 보이며 각진 흰 얼굴판, 작은 원형 눈 두 개, 검은 선 형태의 입, 마모된 회색 금속 몸체가 참조와 가깝다. 비인간 기계이므로 인간의 나이·성별·민족성 판단은 해당하지 않는다. 양쪽 눈에 푸른빛이 켜졌고 가슴과 몸 주변의 여러 전선은 연결된 상태다. 어두운 창과 절제된 주변광은 밤 설정에 맞으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몸통과 팔다리는 경사진 침대 면에 놓이고 머리는 기계 목으로 지지된다. 전선은 가슴 접속부와 몸 주변에서 이어져 침대에 걸쳐 늘어지므로 부유하지 않는다. 참조의 누운 자세를 유지하며 점등만 일어난 장면으로 물리적으로 성립한다."
       },
       {
        "label": "A",
        "direction": "두 눈과 얼굴은 침대 발치 쪽 카메라를 정면으로 향하고 별도의 응시 대상은 없다. 전선은 가슴에서 화면 아래와 오른쪽으로 이어져 눈을 가리지 않는다. 얼굴이 더 크게 보이지만 요구한 비스듬한 얼굴 구성은 아니다.",
        "built_space": "경사진 금속 침대 한 개, 주변 유리벽과 수직 프레임, 상부 원형 조명 구조가 보인다. 왼쪽 실험 장비와 오른쪽 작업대·의자·대형 장비의 관계는 참조와 대체로 일치한다. 침대 위 상반신을 가까이 보지만 어깨 양옆과 몸통 주변의 침대 면, 넓은 실험실 배경이 여전히 포함되어 지정된 좁은 배경 범위를 초과한다. 중복 침대나 불가능한 반사는 보이지 않는다.",
        "entities": "찰리 한 개체의 각진 흰 얼굴판, 두 점 형태의 기계 눈, 선 형태의 입과 낡은 금속 몸체가 보인다. 인간의 나이·성별·민족성 조건은 해당하지 않는다. 눈 두 개에 푸른빛이 있으나 A보다 점등의 번쩍임이 약하고, 참조에서 검던 입선까지 파랗게 발광한다. 가슴 접속부에는 여러 전선이 남아 있다. 어두운 실험실은 밤에 부합하며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "상반신은 경사진 침대에 기대어 있고 머리는 노출된 기계 목이 받친다. 전선은 가슴 접속부에 연결되어 어깨와 침대 가장자리를 따라 처진다. 지지 없이 뜬 몸이나 물체는 없으며, 보이는 범위의 자세는 참조의 누운 상태와 양립한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.267
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.267
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1267
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "클로즈업 지시에는 다소 못 미치나 B보다 화각이 좁으며, '억제된 조명' 지시를 잘 따름(입이 파랗게 빛나는 사소한 오류 존재)."
   },
   {
    "label": "B",
    "score": 1267,
    "verdict_ko": "클로즈업 화각 지시를 무시하고 하반신까지 노출했으며, 주변 조명이 너무 밝아 프롬프트의 무드와 어긋남."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S82sh11_sel.png",
    "asset_id": "5c9a06dc-1703-4fa5-8a75-90552c758d97",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf2a-f63d-7bd2-a68d-3519824f47ae",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S82sh11"
  },
  "lane_policy": "share_plan_prev_bgonly",
  "locked_char_refs_excluded": [
   "찰리(C01)"
  ]
 },
 "S84sh18": {
  "input_fingerprint": "25c6463a1f0f998f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 시야 인터페이스 속, 피아노를 치는 지소영의 얼굴 위로 어린 지소영의 얼굴 데이터가 정확히 겹쳐진 주관 시점 찰나.\n\nLOCATION (lock): Beside the old piano in the central hall of the director's office adjoining the research center, in subdued nighttime interior light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Old piano (Being played by 지소영) — The keyboard is viewed obliquely from its side, beneath her hands; used as Occupies less than the lower third and anchors the recognition image in the ongoing performance; Director's office interior (Neatly arranged and modern); used as Remains softly readable behind her to establish that the interface has not replaced the present location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral ambient illumination on the present-day face and room, keeping the recognition overlay legible without introducing a new colored light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An old piano stands in the central hall of the otherwise clean, modern director's office. Charlie is upright beside the piano, with the laboratory wires already detached.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 시야 인터페이스 속, 피아노를 치는 지소영의 얼굴 위로 어린 지소영의 얼굴 데이터가 정확히 겹쳐진 주관 시점 찰나.\n\nLOCATION (lock): Beside the old piano in the central hall of the director's office adjoining the research center, in subdued nighttime interior light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Old piano (Being played by 지소영) — The keyboard is viewed obliquely from its side, beneath her hands; used as Occupies less than the lower third and anchors the recognition image in the ongoing performance; Director's office interior (Neatly arranged and modern); used as Remains softly readable behind her to establish that the interface has not replaced the present location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral ambient illumination on the present-day face and room, keeping the recognition overlay legible without introducing a new colored light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An old piano stands in the central hall of the otherwise clean, modern director's office. Charlie is upright beside the piano, with the laboratory wires already detached.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 시야 인터페이스 속, 피아노를 치는 지소영의 얼굴 위로 어린 지소영의 얼굴 데이터가 정확히 겹쳐진 주관 시점 찰나.\n\nLOCATION (lock): Beside the old piano in the central hall of the director's office adjoining the research center, in subdued nighttime interior light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Old piano (Being played by 지소영) — The keyboard is viewed obliquely from its side, beneath her hands; used as Occupies less than the lower third and anchors the recognition image in the ongoing performance; Director's office interior (Neatly arranged and modern); used as Remains softly readable behind her to establish that the interface has not replaced the present location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral ambient illumination on the present-day face and room, keeping the recognition overlay legible without introducing a new colored light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An old piano stands in the central hall of the otherwise clean, modern director's office. Charlie is upright beside the piano, with the laboratory wires already detached.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "여성의 시선과 몸의 방향은 아래쪽 피아노 건반을 향하고 있음.",
    "built_space": "배경에 원통형 유리 구조물과 연구실 장비들이 배치되어 레퍼런스 사진의 구조와 일치함.",
    "entities": "50대 여성 피사체가 존재하며, 프롬프트가 지시한 젊은 시절의 얼굴이 여성의 본래 얼굴 좌측으로 어긋나게 떠 있음.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 인체 구조 (피아노 건반 위에 놓인 왼손 손가락들이 융합되고 형태가 무너짐)"
    ],
    "physics": "오른손은 건반 위에 얹혀 있으나, 왼손은 형태가 뭉개져 정상적인 지지나 연주 동작을 취하지 못함."
   },
   {
    "label": "B",
    "direction": "여성의 시선이 피아노 건반 쪽을 바라보고 있음.",
    "built_space": "중앙의 유리로 된 둥근 연구실 공간과 내부 수술대, 주변 데스크가 레퍼런스 이미지와 매우 흡사하게 구현됨.",
    "entities": "50대 여성이 피아노를 치고 있으며, 인터페이스 창과 함께 젊은 여성의 얼굴이 우측으로 치우쳐 나타남.",
    "hard_violations": [
     "[gpt-high] 얼굴 주변 인터페이스에 판독 가능한 숫자 표기가 노출되어 무문자 조건을 위반한다.",
     "[gpt-high] 요청과 장소 참조에 없는 황동 피아노 조명과 펼친 악보를 추가했다."
    ],
    "physics": "양손이 정상적인 해부학적 구조를 유지한 채 피아노 건반을 누르고 있어 물리적 지지가 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "레퍼런스의 공간 구조를 잘 반영하고 신체 구조가 안정적이나, 프롬프트가 요구한 '정확히 겹쳐진' 얼굴 데이터 대신 우측으로 치우친 홀로그램 인터페이스로 잘못 표현되었습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴 데이터가 어긋나게 겹쳐져 지시사항을 어겼으며, 무엇보다 피아노를 치는 왼손 손가락이 융합되어 물리적으로 불가능한 해부학적 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "여성의 시선과 몸의 방향은 아래쪽 피아노 건반을 향하고 있음.",
        "built_space": "배경에 원통형 유리 구조물과 연구실 장비들이 배치되어 레퍼런스 사진의 구조와 일치함.",
        "entities": "50대 여성 피사체가 존재하며, 프롬프트가 지시한 젊은 시절의 얼굴이 여성의 본래 얼굴 좌측으로 어긋나게 떠 있음.",
        "hard_violations": [
         "물리적으로 불가능한 인체 구조 (피아노 건반 위에 놓인 왼손 손가락들이 융합되고 형태가 무너짐)"
        ],
        "physics": "오른손은 건반 위에 얹혀 있으나, 왼손은 형태가 뭉개져 정상적인 지지나 연주 동작을 취하지 못함."
       },
       {
        "label": "B",
        "direction": "여성의 시선이 피아노 건반 쪽을 바라보고 있음.",
        "built_space": "중앙의 유리로 된 둥근 연구실 공간과 내부 수술대, 주변 데스크가 레퍼런스 이미지와 매우 흡사하게 구현됨.",
        "entities": "50대 여성이 피아노를 치고 있으며, 인터페이스 창과 함께 젊은 여성의 얼굴이 우측으로 치우쳐 나타남.",
        "hard_violations": [],
        "physics": "양손이 정상적인 해부학적 구조를 유지한 채 피아노 건반을 누르고 있어 물리적 지지가 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "레퍼런스의 공간 구조를 잘 반영하고 신체 구조가 안정적이나, 프롬프트가 요구한 '정확히 겹쳐진' 얼굴 데이터 대신 우측으로 치우친 홀로그램 인터페이스로 잘못 표현되었습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴 데이터가 어긋나게 겹쳐져 지시사항을 어겼으며, 무엇보다 피아노를 치는 왼손 손가락이 융합되어 물리적으로 불가능한 해부학적 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "여성의 시선과 몸의 방향은 아래쪽 피아노 건반을 향하고 있음.",
        "built_space": "배경에 원통형 유리 구조물과 연구실 장비들이 배치되어 레퍼런스 사진의 구조와 일치함.",
        "entities": "50대 여성 피사체가 존재하며, 프롬프트가 지시한 젊은 시절의 얼굴이 여성의 본래 얼굴 좌측으로 어긋나게 떠 있음.",
        "hard_violations": [
         "물리적으로 불가능한 인체 구조 (피아노 건반 위에 놓인 왼손 손가락들이 융합되고 형태가 무너짐)"
        ],
        "physics": "오른손은 건반 위에 얹혀 있으나, 왼손은 형태가 뭉개져 정상적인 지지나 연주 동작을 취하지 못함."
       },
       {
        "label": "B",
        "direction": "여성의 시선이 피아노 건반 쪽을 바라보고 있음.",
        "built_space": "중앙의 유리로 된 둥근 연구실 공간과 내부 수술대, 주변 데스크가 레퍼런스 이미지와 매우 흡사하게 구현됨.",
        "entities": "50대 여성이 피아노를 치고 있으며, 인터페이스 창과 함께 젊은 여성의 얼굴이 우측으로 치우쳐 나타남.",
        "hard_violations": [],
        "physics": "양손이 정상적인 해부학적 구조를 유지한 채 피아노 건반을 누르고 있어 물리적 지지가 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "어린 얼굴이 현재 얼굴 오른쪽으로 크게 어긋나 핵심 중첩 동작을 놓쳤고, 추가된 조명과 판독 가능한 인터페이스 숫자가 중립 조명·무문자 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "중년 여성의 연주를 담은 미디엄 숏과 중립 조명은 더 충실하지만, 어린 얼굴 데이터가 왼쪽으로 밀려 정확한 얼굴 중첩을 구현하지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현재 여성은 왼쪽 아래 건반 쪽을 바라보고 양손으로 건반을 누른다. 어린 얼굴은 화면 왼쪽을 향하지만 현재 얼굴보다 오른쪽에 있어 눈·코·입의 위치가 서로 맞지 않는다. 건반은 손 아래에서 비스듬히 보인다.",
        "built_space": "왼쪽에 업라이트 피아노 한 대, 그 위에 황동 조명 한 개와 펼친 악보가 있다. 뒤에는 책상과 사무용 의자, 오른쪽에는 곡면 유리 연구실과 천장 관절식 조명 한 개가 보인다. 유리·금속 연구실의 재료감은 참조와 유사하지만 참조의 그랜드 피아노가 업라이트로 바뀌었다. 건반은 대체로 하단에 있으나 피아노 몸체는 왼쪽 화면을 크게 차지한다.",
        "entities": "실제 연주자는 동아시아계로 보이는 중년 여성 한 명이며, 어두운 니트 상의를 입었다. 한국인이라는 국적은 외형만으로 확정할 수 없다. 얼굴 데이터에는 소녀 또는 청소년으로 보이는 여성 얼굴이 있지만, 별도 인물 사진처럼 선명하게 분리된다. 찰리는 주관 시점 밖에 있으며 노출된 실험실 전선은 없다. 인터페이스에는 판독 가능한 숫자들이 있고, 악보와 황동 조명은 요청에 없는 추가 소품이다.",
        "hard_violations": [
         "얼굴 주변 인터페이스에 판독 가능한 숫자 표기가 노출되어 무문자 조건을 위반한다.",
         "요청과 장소 참조에 없는 황동 피아노 조명과 펼친 악보를 추가했다."
        ],
        "physics": "양손은 팔에 자연스럽게 연결되어 손가락으로 건반을 누르고 있다. 조명은 피아노 상판에 받쳐져 있고 악보는 보면대에 기대어 있다. 하체와 좌석은 프레임 밖이므로 지지 상태를 단정할 수 없지만 몸이 공중에 떠 있다는 증거는 없다. 반투명 얼굴은 지시된 시야 데이터 표현이므로 실제로 떠 있는 신체로 보지는 않는다."
       },
       {
        "label": "B",
        "direction": "여성은 왼쪽 아래 건반을 내려다보며 양손으로 연주한다. 반투명 얼굴 역시 왼쪽 아래를 향하지만 현재 얼굴의 왼쪽에 치우쳐 눈·코·입이 이중으로 보인다. 얼굴 데이터의 정확한 정합에는 실패했다.",
        "built_space": "왼쪽에 업라이트 피아노 한 대, 뒤쪽에 책상과 의자, 오른쪽 곡면 유리실에 수술대 한 대와 천장 관절식 장치 두 개가 보이며 그중 하나는 가장자리에서 일부 잘린다. 연구실의 유리 벽과 금속 프레임은 참조 장소를 비교적 잘 이어간다. 다만 참조의 그랜드 피아노와 종류가 다르다. 허리와 손을 포함한 미디엄 숏이며 건반은 하단 약 삼분의 일에 비스듬히 배치된다.",
        "entities": "동아시아계로 보이는 50대 전후 여성 한 명이 검은 재킷과 밝은 블라우스를 입고 연주한다. 나이 인상은 요청한 50대 중반에 비교적 가깝다. 반투명 얼굴은 더 젊어 보이지만 어린아이인지 명확하지 않다. 찰리는 주관 시점 밖에 있고 전선은 보이지 않는다. 인터페이스의 작은 표식은 보이지만 명확히 읽히는 문구는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "손목과 손가락이 건반 연주에 맞게 연결되고 손끝이 건반에 닿아 있다. 하단에 앉은 자세의 허벅지가 일부 보이며 좌석 자체는 가려져 있다. 피아노와 배경 설비는 정상적인 설치물로 보이고 지지 없이 떠 있는 실제 물체는 없다. 겹쳐진 얼굴은 시야 인터페이스 데이터로 읽힌다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "어린 얼굴이 현재 얼굴 오른쪽으로 크게 어긋나 핵심 중첩 동작을 놓쳤고, 추가된 조명과 판독 가능한 인터페이스 숫자가 중립 조명·무문자 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "중년 여성의 연주를 담은 미디엄 숏과 중립 조명은 더 충실하지만, 어린 얼굴 데이터가 왼쪽으로 밀려 정확한 얼굴 중첩을 구현하지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현재 여성은 왼쪽 아래 건반 쪽을 바라보고 양손으로 건반을 누른다. 어린 얼굴은 화면 왼쪽을 향하지만 현재 얼굴보다 오른쪽에 있어 눈·코·입의 위치가 서로 맞지 않는다. 건반은 손 아래에서 비스듬히 보인다.",
        "built_space": "왼쪽에 업라이트 피아노 한 대, 그 위에 황동 조명 한 개와 펼친 악보가 있다. 뒤에는 책상과 사무용 의자, 오른쪽에는 곡면 유리 연구실과 천장 관절식 조명 한 개가 보인다. 유리·금속 연구실의 재료감은 참조와 유사하지만 참조의 그랜드 피아노가 업라이트로 바뀌었다. 건반은 대체로 하단에 있으나 피아노 몸체는 왼쪽 화면을 크게 차지한다.",
        "entities": "실제 연주자는 동아시아계로 보이는 중년 여성 한 명이며, 어두운 니트 상의를 입었다. 한국인이라는 국적은 외형만으로 확정할 수 없다. 얼굴 데이터에는 소녀 또는 청소년으로 보이는 여성 얼굴이 있지만, 별도 인물 사진처럼 선명하게 분리된다. 찰리는 주관 시점 밖에 있으며 노출된 실험실 전선은 없다. 인터페이스에는 판독 가능한 숫자들이 있고, 악보와 황동 조명은 요청에 없는 추가 소품이다.",
        "hard_violations": [
         "얼굴 주변 인터페이스에 판독 가능한 숫자 표기가 노출되어 무문자 조건을 위반한다.",
         "요청과 장소 참조에 없는 황동 피아노 조명과 펼친 악보를 추가했다."
        ],
        "physics": "양손은 팔에 자연스럽게 연결되어 손가락으로 건반을 누르고 있다. 조명은 피아노 상판에 받쳐져 있고 악보는 보면대에 기대어 있다. 하체와 좌석은 프레임 밖이므로 지지 상태를 단정할 수 없지만 몸이 공중에 떠 있다는 증거는 없다. 반투명 얼굴은 지시된 시야 데이터 표현이므로 실제로 떠 있는 신체로 보지는 않는다."
       },
       {
        "label": "A",
        "direction": "여성은 왼쪽 아래 건반을 내려다보며 양손으로 연주한다. 반투명 얼굴 역시 왼쪽 아래를 향하지만 현재 얼굴의 왼쪽에 치우쳐 눈·코·입이 이중으로 보인다. 얼굴 데이터의 정확한 정합에는 실패했다.",
        "built_space": "왼쪽에 업라이트 피아노 한 대, 뒤쪽에 책상과 의자, 오른쪽 곡면 유리실에 수술대 한 대와 천장 관절식 장치 두 개가 보이며 그중 하나는 가장자리에서 일부 잘린다. 연구실의 유리 벽과 금속 프레임은 참조 장소를 비교적 잘 이어간다. 다만 참조의 그랜드 피아노와 종류가 다르다. 허리와 손을 포함한 미디엄 숏이며 건반은 하단 약 삼분의 일에 비스듬히 배치된다.",
        "entities": "동아시아계로 보이는 50대 전후 여성 한 명이 검은 재킷과 밝은 블라우스를 입고 연주한다. 나이 인상은 요청한 50대 중반에 비교적 가깝다. 반투명 얼굴은 더 젊어 보이지만 어린아이인지 명확하지 않다. 찰리는 주관 시점 밖에 있고 전선은 보이지 않는다. 인터페이스의 작은 표식은 보이지만 명확히 읽히는 문구는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "손목과 손가락이 건반 연주에 맞게 연결되고 손끝이 건반에 닿아 있다. 하단에 앉은 자세의 허벅지가 일부 보이며 좌석 자체는 가려져 있다. 피아노와 배경 설비는 정상적인 설치물로 보이고 지지 없이 떠 있는 실제 물체는 없다. 겹쳐진 얼굴은 시야 인터페이스 데이터로 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.4
   },
   "adjusted": {
    "A": 1.35,
    "B": 1.15
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 인체 구조 (피아노 건반 위에 놓인 왼손 손가락들이 융합되고 형태가 무너짐)"
    ],
    "B": [
     "[gpt-high] 얼굴 주변 인터페이스에 판독 가능한 숫자 표기가 노출되어 무문자 조건을 위반한다.",
     "[gpt-high] 요청과 장소 참조에 없는 황동 피아노 조명과 펼친 악보를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1150,
   "A": 1350
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1150,
    "verdict_ko": "레퍼런스의 공간 구조를 잘 반영하고 신체 구조가 안정적이나, 프롬프트가 요구한 '정확히 겹쳐진' 얼굴 데이터 대신 우측으로 치우친 홀로그램 인터페이스로 잘못 표현되었습니다.  ★위반: [gpt-high] 얼굴 주변 인터페이스에 판독 가능한 숫자 표기가 노출되어 무문자 조건을 위반한다. / [gpt-high] 요청과 장소 참조에 없는 황동 피아노 조명과 펼친 악보를 추가했다."
   },
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "얼굴 데이터가 어긋나게 겹쳐져 지시사항을 어겼으며, 무엇보다 피아노를 치는 왼손 손가락이 융합되어 물리적으로 불가능한 해부학적 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 인체 구조 (피아노 건반 위에 놓인 왼손 손가락들이 융합되고 형태가 무너짐)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L130B04.png",
    "asset_id": "f02daf94-b300-45a0-9788-8839a25b7ab6",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf2e-fdaf-77b1-9a44-f3632d13032f",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S84sh21": {
  "input_fingerprint": "5e8df5bfe6af42ee",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 품에 안긴 지소영이 두 눈을 지그시 감은 채 평온한 미소를 짓는 감동적인 구도.\n\nLOCATION (lock): At the piano inside the director's spacious modern office hall, under quiet nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Director's office interior (Remaining behind the embracing figures); used as A subdued background sliver preserves location without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly controlled ambient contrast preserves her expression and the detail of his metal embrace before the image fades to black.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old piano remains in the central hall of the modern director's office. Charlie is awake and free of the laboratory wires. 지소영: She wears her clean research coat and has closed her eyes, relaxing into the embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 품에 안긴 지소영이 두 눈을 지그시 감은 채 평온한 미소를 짓는 감동적인 구도.\n\nLOCATION (lock): At the piano inside the director's spacious modern office hall, under quiet nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Director's office interior (Remaining behind the embracing figures); used as A subdued background sliver preserves location without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly controlled ambient contrast preserves her expression and the detail of his metal embrace before the image fades to black.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old piano remains in the central hall of the modern director's office. Charlie is awake and free of the laboratory wires. 지소영: She wears her clean research coat and has closed her eyes, relaxing into the embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 품에 안긴 지소영이 두 눈을 지그시 감은 채 평온한 미소를 짓는 감동적인 구도.\n\nLOCATION (lock): At the piano inside the director's spacious modern office hall, under quiet nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Director's office interior (Remaining behind the embracing figures); used as A subdued background sliver preserves location without competing with her face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly controlled ambient contrast preserves her expression and the detail of his metal embrace before the image fades to black.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old piano remains in the central hall of the modern director's office. Charlie is awake and free of the laboratory wires. 지소영: She wears her clean research coat and has closed her eyes, relaxing into the embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "지소영은 눈을 감고 오른쪽의 찰리를 향해 미소 지으며 안겨 있음.",
    "built_space": "왼쪽에 피아노가 위치하며, 유리벽을 포함한 이전 샷의 사무실 배경이 정확히 나타남.",
    "entities": "지소영은 흰색 연구복을 착용함. 찰리는 금속 몸체로 보이나 요구된 흰 얼굴판은 각도상 명확히 보이지 않음.",
    "hard_violations": [
     "[gemini-pro] 화면 왼쪽에 찰리의 신체와 연결될 수 없는 추가 기계 팔이 등장함 (duplicated or extra bodies / physically impossible anatomy)"
    ],
    "physics": "두 인물이 서서 포옹 중이며, 지소영의 손은 찰리 가슴에 닿아 있음. 하지만 왼쪽의 기계 팔은 해부학적으로 불가능한 위치에 떠 있음."
   },
   {
    "label": "B",
    "direction": "지소영은 눈을 감고 왼쪽의 찰리를 향해 미소 지으며 안겨 있음.",
    "built_space": "피아노가 보이지 않으며, 배경은 매우 어둡게 처리되어 지정된 공간의 세부 묘사가 부족함.",
    "entities": "지소영은 흰색 연구복을 착용함. 찰리는 요구된 '흰 얼굴판'이 아닌 금속 얼굴과 파란 눈의 다른 디자인으로 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 화면 오른쪽 아래에 찰리의 신체와 연결 불가능한 여분의 기계 팔이 등장함 (duplicated or extra bodies / physically impossible anatomy)"
    ],
    "physics": "두 인물이 포옹하는 자세이나, 오른쪽 아래에서 지소영의 어깨를 잡는 기계 팔은 화면 밖에서 뻗어나와 해부학적으로 성립하지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "피아노와 사무실 배경을 잘 살렸으나, 화면 왼쪽에 찰리의 몸과 무관한 추가 기계 팔이 생성되는 치명적 구조 오류가 발생함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 찰리의 흰 얼굴판 디자인과 피아노가 누락되었으며, 화면 오른쪽에 출처 불명의 여분 기계 팔이 나타남."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지소영은 눈을 감고 오른쪽의 찰리를 향해 미소 지으며 안겨 있음.",
        "built_space": "왼쪽에 피아노가 위치하며, 유리벽을 포함한 이전 샷의 사무실 배경이 정확히 나타남.",
        "entities": "지소영은 흰색 연구복을 착용함. 찰리는 금속 몸체로 보이나 요구된 흰 얼굴판은 각도상 명확히 보이지 않음.",
        "hard_violations": [
         "화면 왼쪽에 찰리의 신체와 연결될 수 없는 추가 기계 팔이 등장함 (duplicated or extra bodies / physically impossible anatomy)"
        ],
        "physics": "두 인물이 서서 포옹 중이며, 지소영의 손은 찰리 가슴에 닿아 있음. 하지만 왼쪽의 기계 팔은 해부학적으로 불가능한 위치에 떠 있음."
       },
       {
        "label": "B",
        "direction": "지소영은 눈을 감고 왼쪽의 찰리를 향해 미소 지으며 안겨 있음.",
        "built_space": "피아노가 보이지 않으며, 배경은 매우 어둡게 처리되어 지정된 공간의 세부 묘사가 부족함.",
        "entities": "지소영은 흰색 연구복을 착용함. 찰리는 요구된 '흰 얼굴판'이 아닌 금속 얼굴과 파란 눈의 다른 디자인으로 묘사됨.",
        "hard_violations": [
         "화면 오른쪽 아래에 찰리의 신체와 연결 불가능한 여분의 기계 팔이 등장함 (duplicated or extra bodies / physically impossible anatomy)"
        ],
        "physics": "두 인물이 포옹하는 자세이나, 오른쪽 아래에서 지소영의 어깨를 잡는 기계 팔은 화면 밖에서 뻗어나와 해부학적으로 성립하지 않음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "피아노와 사무실 배경을 잘 살렸으나, 화면 왼쪽에 찰리의 몸과 무관한 추가 기계 팔이 생성되는 치명적 구조 오류가 발생함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 찰리의 흰 얼굴판 디자인과 피아노가 누락되었으며, 화면 오른쪽에 출처 불명의 여분 기계 팔이 나타남."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "지소영은 눈을 감고 오른쪽의 찰리를 향해 미소 지으며 안겨 있음.",
        "built_space": "왼쪽에 피아노가 위치하며, 유리벽을 포함한 이전 샷의 사무실 배경이 정확히 나타남.",
        "entities": "지소영은 흰색 연구복을 착용함. 찰리는 금속 몸체로 보이나 요구된 흰 얼굴판은 각도상 명확히 보이지 않음.",
        "hard_violations": [
         "화면 왼쪽에 찰리의 신체와 연결될 수 없는 추가 기계 팔이 등장함 (duplicated or extra bodies / physically impossible anatomy)"
        ],
        "physics": "두 인물이 서서 포옹 중이며, 지소영의 손은 찰리 가슴에 닿아 있음. 하지만 왼쪽의 기계 팔은 해부학적으로 불가능한 위치에 떠 있음."
       },
       {
        "label": "B",
        "direction": "지소영은 눈을 감고 왼쪽의 찰리를 향해 미소 지으며 안겨 있음.",
        "built_space": "피아노가 보이지 않으며, 배경은 매우 어둡게 처리되어 지정된 공간의 세부 묘사가 부족함.",
        "entities": "지소영은 흰색 연구복을 착용함. 찰리는 요구된 '흰 얼굴판'이 아닌 금속 얼굴과 파란 눈의 다른 디자인으로 묘사됨.",
        "hard_violations": [
         "화면 오른쪽 아래에 찰리의 신체와 연결 불가능한 여분의 기계 팔이 등장함 (duplicated or extra bodies / physically impossible anatomy)"
        ],
        "physics": "두 인물이 포옹하는 자세이나, 오른쪽 아래에서 지소영의 어깨를 잡는 기계 팔은 화면 밖에서 뻗어나와 해부학적으로 성립하지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "눈을 감고 미소 짓는 얼굴과 금속 포옹에 집중한 클로즈업, 어둡게 억제된 배경이 우세하나 찰리의 흰 얼굴판은 누락됐다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이전 장소와 피아노는 더 정확하지만, 넓고 밝게 드러난 배경이 얼굴 중심 클로즈업과 조용한 야간 분위기를 약화한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지소영은 두 눈을 완전히 감고 찰리의 가슴 쪽으로 머리를 기댄 채 미소 짓는다. 찰리의 머리와 눈 부위는 지소영 쪽 아래로 기울어 있다. 두 금속 손은 각각 지소영의 먼쪽 어깨와 가까운 위팔을 감싸며, 포옹의 대상이 명확하다.",
        "built_space": "인물 뒤 좁은 영역에 어두운 유리 칸막이와 수직 프레임, 흐릿한 가로 조명 하나가 보인다. 이전 장소의 유리·금속 계열 실내와 양립하지만, 피아노나 사무 설비는 이 크롭에서 확인되지 않아 동일 장소의 식별력은 낮다. 중복된 고정 설비나 불가능한 반사는 보이지 않는다. 배경은 요구대로 인물보다 억제되어 있다.",
        "entities": "찰리 한 명과 지소영 한 명만 보인다. 지소영은 이전 스틸과 유사한 중년 동아시아계 여성의 얼굴, 회색이 섞인 뒤로 넘긴 머리이며 깨끗한 흰 연구복을 입었다. 연구복은 현재 상태 지시에는 맞지만 이전 스틸의 검은 재킷과는 다르며, 이는 프롬프트 내부의 상충된 복장 지시와 관련된다. 찰리는 거대한 낡은 금속 몸체와 원형 가슴 부품을 갖췄으나, 얼굴은 입체적인 금속 고릴라형으로 명시된 흰 얼굴판·점 눈·선 입과 다르다. 실험실 연결선이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "지소영의 머리는 찰리의 가슴에 닿고, 찰리의 두 팔과 손은 어깨와 위팔에 접촉해 포옹을 지지한다. 금속 팔은 몸체에서 팔꿈치와 손목으로 이어진다. 하체와 바닥은 화면 밖이므로 발의 지지는 확인할 수 없지만, 공중에 들리거나 지지 없이 떠 있는 자세는 아니다."
       },
       {
        "label": "B",
        "direction": "지소영은 두 눈을 감고 찰리의 가슴에 관자놀이를 기댄 채 평온하게 미소 짓는다. 보이는 사람 손은 찰리의 가슴에 펼쳐져 있다. 찰리의 눈은 화면 밖이라 시선 대상은 확인할 수 없으며, 두 금속 손은 지소영의 양쪽 위팔과 어깨를 향해 감싸고 있다.",
        "built_space": "왼쪽에 낡은 업라이트 피아노 한 대의 몸통과 건반이 보이고, 뒤에는 유리 칸막이, 수직 기둥, 여러 직사각형 천장등과 오른쪽의 흐릿한 장비대 일부가 보인다. 두 인물은 피아노 옆에 있어 장소 관계가 자연스럽고 이전 스틸의 구조와 재료도 잘 이어진다. 설비 중복이나 불가능한 반사는 보이지 않는다. 다만 배경과 피아노가 넓고 밝게 노출되어 요구된 작은 배경 조각보다 존재감이 크다.",
        "entities": "찰리와 지소영만 등장한다. 지소영의 중년 여성 얼굴과 회색이 섞인 머리는 이전 스틸에 가깝고, 깨끗한 흰 연구복은 현재 상태 지시에 맞는다. 이전 스틸의 검은 재킷과는 다르다는 복장 지시의 상충은 여기에도 적용된다. 찰리의 큰 금속 체격과 마모된 장갑은 참고에 부합한다. 머리 대부분은 잘렸으며, 보이는 하안면 일부는 흰 판이 아닌 어두운 입체 금속이다. 피아노는 이전 스틸처럼 낡은 목재 업라이트형이다. 연결선과 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "지소영의 머리와 손바닥이 찰리의 가슴에 닿고, 찰리의 양팔이 그녀의 상체를 감싼다. 앞쪽 금속 팔은 굽힌 팔꿈치에서 손목과 손으로 연결되며 손은 위팔에 접촉한다. 기대는 자세와 접촉점은 자연스럽다. 하체는 잘려 있지만 떠 있거나 별도 지지가 필요한 동작은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "눈을 감고 미소 짓는 얼굴과 금속 포옹에 집중한 클로즈업, 어둡게 억제된 배경이 우세하나 찰리의 흰 얼굴판은 누락됐다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이전 장소와 피아노는 더 정확하지만, 넓고 밝게 드러난 배경이 얼굴 중심 클로즈업과 조용한 야간 분위기를 약화한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "지소영은 두 눈을 완전히 감고 찰리의 가슴 쪽으로 머리를 기댄 채 미소 짓는다. 찰리의 머리와 눈 부위는 지소영 쪽 아래로 기울어 있다. 두 금속 손은 각각 지소영의 먼쪽 어깨와 가까운 위팔을 감싸며, 포옹의 대상이 명확하다.",
        "built_space": "인물 뒤 좁은 영역에 어두운 유리 칸막이와 수직 프레임, 흐릿한 가로 조명 하나가 보인다. 이전 장소의 유리·금속 계열 실내와 양립하지만, 피아노나 사무 설비는 이 크롭에서 확인되지 않아 동일 장소의 식별력은 낮다. 중복된 고정 설비나 불가능한 반사는 보이지 않는다. 배경은 요구대로 인물보다 억제되어 있다.",
        "entities": "찰리 한 명과 지소영 한 명만 보인다. 지소영은 이전 스틸과 유사한 중년 동아시아계 여성의 얼굴, 회색이 섞인 뒤로 넘긴 머리이며 깨끗한 흰 연구복을 입었다. 연구복은 현재 상태 지시에는 맞지만 이전 스틸의 검은 재킷과는 다르며, 이는 프롬프트 내부의 상충된 복장 지시와 관련된다. 찰리는 거대한 낡은 금속 몸체와 원형 가슴 부품을 갖췄으나, 얼굴은 입체적인 금속 고릴라형으로 명시된 흰 얼굴판·점 눈·선 입과 다르다. 실험실 연결선이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "지소영의 머리는 찰리의 가슴에 닿고, 찰리의 두 팔과 손은 어깨와 위팔에 접촉해 포옹을 지지한다. 금속 팔은 몸체에서 팔꿈치와 손목으로 이어진다. 하체와 바닥은 화면 밖이므로 발의 지지는 확인할 수 없지만, 공중에 들리거나 지지 없이 떠 있는 자세는 아니다."
       },
       {
        "label": "A",
        "direction": "지소영은 두 눈을 감고 찰리의 가슴에 관자놀이를 기댄 채 평온하게 미소 짓는다. 보이는 사람 손은 찰리의 가슴에 펼쳐져 있다. 찰리의 눈은 화면 밖이라 시선 대상은 확인할 수 없으며, 두 금속 손은 지소영의 양쪽 위팔과 어깨를 향해 감싸고 있다.",
        "built_space": "왼쪽에 낡은 업라이트 피아노 한 대의 몸통과 건반이 보이고, 뒤에는 유리 칸막이, 수직 기둥, 여러 직사각형 천장등과 오른쪽의 흐릿한 장비대 일부가 보인다. 두 인물은 피아노 옆에 있어 장소 관계가 자연스럽고 이전 스틸의 구조와 재료도 잘 이어진다. 설비 중복이나 불가능한 반사는 보이지 않는다. 다만 배경과 피아노가 넓고 밝게 노출되어 요구된 작은 배경 조각보다 존재감이 크다.",
        "entities": "찰리와 지소영만 등장한다. 지소영의 중년 여성 얼굴과 회색이 섞인 머리는 이전 스틸에 가깝고, 깨끗한 흰 연구복은 현재 상태 지시에 맞는다. 이전 스틸의 검은 재킷과는 다르다는 복장 지시의 상충은 여기에도 적용된다. 찰리의 큰 금속 체격과 마모된 장갑은 참고에 부합한다. 머리 대부분은 잘렸으며, 보이는 하안면 일부는 흰 판이 아닌 어두운 입체 금속이다. 피아노는 이전 스틸처럼 낡은 목재 업라이트형이다. 연결선과 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "지소영의 머리와 손바닥이 찰리의 가슴에 닿고, 찰리의 양팔이 그녀의 상체를 감싼다. 앞쪽 금속 팔은 굽힌 팔꿈치에서 손목과 손으로 연결되며 손은 위팔에 접촉한다. 기대는 자세와 접촉점은 자연스럽다. 하체는 잘려 있지만 떠 있거나 별도 지지가 필요한 동작은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.625,
    "B": 1.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 화면 왼쪽에 찰리의 신체와 연결될 수 없는 추가 기계 팔이 등장함 (duplicated or extra bodies / physically impossible anatomy)"
    ],
    "B": [
     "[gemini-pro] 화면 오른쪽 아래에 찰리의 신체와 연결 불가능한 여분의 기계 팔이 등장함 (duplicated or extra bodies / physically impossible anatomy)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1625,
   "B": 1500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1625,
    "verdict_ko": "피아노와 사무실 배경을 잘 살렸으나, 화면 왼쪽에 찰리의 몸과 무관한 추가 기계 팔이 생성되는 치명적 구조 오류가 발생함.  ★위반: [gemini-pro] 화면 왼쪽에 찰리의 신체와 연결될 수 없는 추가 기계 팔이 등장함 (duplicated or extra bodies / physically impossible anatomy)"
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "요구된 찰리의 흰 얼굴판 디자인과 피아노가 누락되었으며, 화면 오른쪽에 출처 불명의 여분 기계 팔이 나타남.  ★위반: [gemini-pro] 화면 오른쪽 아래에 찰리의 신체와 연결 불가능한 여분의 기계 팔이 등장함 (duplicated or extra bodies / physically impossible anatomy)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S84sh18_sel.png",
    "asset_id": "70823952-f0ae-482b-86b8-12df963c2674",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1597480>",
    "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:808347>",
    "asset_id": "f15ee474-495b-45e2-8cb1-804018e9010e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf33-9f16-7334-86e5-d53f361d33a5",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S84sh18"
  }
 },
 "S85sh1": {
  "input_fingerprint": "018f41b0256e9211",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 유리 벽 너머, 가슴과 팔에 여러 가닥의 굵은 전선을 붙인 찰리가 차가운 금속 의자에 꼿꼿이 앉아 있는 구도.\n\nLOCATION (lock): At the wired test chair inside the laboratory's glass enclosure, facing researchers and illuminated computer workstations beyond the glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Glass partition separating the preparation area from Charlie in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Glass partition (Separating 찰리 from the preparation area) — 찰리 and his chair are visible through the glass from the researchers' side; used as Establishes two distinct depth zones without using a reflected image of 찰리; Metal chair (Occupied by 찰리) — Its seat and side are visible obliquely beneath and beside him; used as Supports the full-body scale and the controlled test posture; Body-connected wires (Attached across 찰리's chest and arms); used as Make the experimental restraint of the situation legible across the partition; Researchers' computers (Present in the active preparation area) — Seen primarily from their backs and sides, without readable screen content; used as Anchor the near-side preparation activity while leaving the view of 찰리 clear.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination keeps both sides of the glass readable, with restrained contrast emphasizing the separation between preparation and confinement.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is seated inside the glass enclosure with several wires reattached to his body for the experiment. Computers are set up outside the glass for monitoring and recording.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 유리 벽 너머, 가슴과 팔에 여러 가닥의 굵은 전선을 붙인 찰리가 차가운 금속 의자에 꼿꼿이 앉아 있는 구도.\n\nLOCATION (lock): At the wired test chair inside the laboratory's glass enclosure, facing researchers and illuminated computer workstations beyond the glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Glass partition separating the preparation area from Charlie in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Glass partition (Separating 찰리 from the preparation area) — 찰리 and his chair are visible through the glass from the researchers' side; used as Establishes two distinct depth zones without using a reflected image of 찰리; Metal chair (Occupied by 찰리) — Its seat and side are visible obliquely beneath and beside him; used as Supports the full-body scale and the controlled test posture; Body-connected wires (Attached across 찰리's chest and arms); used as Make the experimental restraint of the situation legible across the partition; Researchers' computers (Present in the active preparation area) — Seen primarily from their backs and sides, without readable screen content; used as Anchor the near-side preparation activity while leaving the view of 찰리 clear.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination keeps both sides of the glass readable, with restrained contrast emphasizing the separation between preparation and confinement.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is seated inside the glass enclosure with several wires reattached to his body for the experiment. Computers are set up outside the glass for monitoring and recording.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 유리 벽 너머, 가슴과 팔에 여러 가닥의 굵은 전선을 붙인 찰리가 차가운 금속 의자에 꼿꼿이 앉아 있는 구도.\n\nLOCATION (lock): At the wired test chair inside the laboratory's glass enclosure, facing researchers and illuminated computer workstations beyond the glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Glass partition separating the preparation area from Charlie in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Glass partition (Separating 찰리 from the preparation area) — 찰리 and his chair are visible through the glass from the researchers' side; used as Establishes two distinct depth zones without using a reflected image of 찰리; Metal chair (Occupied by 찰리) — Its seat and side are visible obliquely beneath and beside him; used as Supports the full-body scale and the controlled test posture; Body-connected wires (Attached across 찰리's chest and arms); used as Make the experimental restraint of the situation legible across the partition; Researchers' computers (Present in the active preparation area) — Seen primarily from their backs and sides, without readable screen content; used as Anchor the near-side preparation activity while leaving the view of 찰리 clear.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination keeps both sides of the glass readable, with restrained contrast emphasizing the separation between preparation and confinement.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is seated inside the glass enclosure with several wires reattached to his body for the experiment. Computers are set up outside the glass for monitoring and recording.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라와 찰리는 정면을 향하며, 연구원들은 각자의 모니터와 유리 너머를 주시함.",
    "built_space": "레퍼런스와 동일한 형태의 실험실. 유리 가벽을 기준으로 내부에 찰리가 앉은 금속 의자가, 외부에 연구원들의 책상과 모니터가 정확히 배치됨.",
    "entities": "찰리의 얼굴과 흉부 구조가 레퍼런스와 완벽히 일치함. 굵은 전선이 몸에 부착되었고, 모니터 텍스트는 지시대로 식별 불가능하게 블러 처리됨.",
    "hard_violations": [
     "[gpt-high] 찰리만 등장해야 하는 장면에 좌우 작업대의 연구자들을 추가했다."
    ],
    "physics": "찰리가 의자에 안정적으로 앉아 있으며, 몸에 연결된 전선들은 중력에 의해 바닥으로 자연스럽게 늘어져 지탱됨."
   },
   {
    "label": "B",
    "direction": "카메라와 찰리는 정면을 향함. 서 있는 연구원은 찰리를 향하고 다른 연구원들은 모니터를 봄.",
    "built_space": "레퍼런스의 구조를 잘 반영한 실험실. 유리 가벽 내부에 찰리와 장비 카트가 존재하고 외부에 연구 공간이 있음.",
    "entities": "찰리의 외형은 레퍼런스와 일치하나, 좌측 하단 모니터에 식별 가능한 선명한 텍스트(코드)가 노출되어 제약 사항과 충돌함.",
    "hard_violations": [
     "[gemini-pro] 화면 내 식별 가능한 텍스트 금지(No readable writing anywhere) 지시를 위반하여 좌측 모니터에 선명한 코드가 나타남.",
     "[gpt-high] 찰리 외에는 아무도 등장시키지 말라는 명시적 제한을 어기고 연구자 네 명을 추가했다."
    ],
    "physics": "찰리가 의자에 바르게 앉아 있고, 전선들은 양옆의 카트 장비로 팽팽하게 연결되어 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 로봇 외형과 실험실 배경을 완벽히 유지했으며, 모니터 화면을 흐리게 처리하여 텍스트 비노출 제약을 훌륭하게 준수했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전반적인 구도와 형태는 좋으나, 좌측 모니터에 선명한 코드가 노출되어 텍스트 금지 제약을 위반했으며 중앙의 연구원이 시야를 다소 가립니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라와 찰리는 정면을 향하며, 연구원들은 각자의 모니터와 유리 너머를 주시함.",
        "built_space": "레퍼런스와 동일한 형태의 실험실. 유리 가벽을 기준으로 내부에 찰리가 앉은 금속 의자가, 외부에 연구원들의 책상과 모니터가 정확히 배치됨.",
        "entities": "찰리의 얼굴과 흉부 구조가 레퍼런스와 완벽히 일치함. 굵은 전선이 몸에 부착되었고, 모니터 텍스트는 지시대로 식별 불가능하게 블러 처리됨.",
        "hard_violations": [],
        "physics": "찰리가 의자에 안정적으로 앉아 있으며, 몸에 연결된 전선들은 중력에 의해 바닥으로 자연스럽게 늘어져 지탱됨."
       },
       {
        "label": "B",
        "direction": "카메라와 찰리는 정면을 향함. 서 있는 연구원은 찰리를 향하고 다른 연구원들은 모니터를 봄.",
        "built_space": "레퍼런스의 구조를 잘 반영한 실험실. 유리 가벽 내부에 찰리와 장비 카트가 존재하고 외부에 연구 공간이 있음.",
        "entities": "찰리의 외형은 레퍼런스와 일치하나, 좌측 하단 모니터에 식별 가능한 선명한 텍스트(코드)가 노출되어 제약 사항과 충돌함.",
        "hard_violations": [
         "화면 내 식별 가능한 텍스트 금지(No readable writing anywhere) 지시를 위반하여 좌측 모니터에 선명한 코드가 나타남."
        ],
        "physics": "찰리가 의자에 바르게 앉아 있고, 전선들은 양옆의 카트 장비로 팽팽하게 연결되어 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 로봇 외형과 실험실 배경을 완벽히 유지했으며, 모니터 화면을 흐리게 처리하여 텍스트 비노출 제약을 훌륭하게 준수했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전반적인 구도와 형태는 좋으나, 좌측 모니터에 선명한 코드가 노출되어 텍스트 금지 제약을 위반했으며 중앙의 연구원이 시야를 다소 가립니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라와 찰리는 정면을 향하며, 연구원들은 각자의 모니터와 유리 너머를 주시함.",
        "built_space": "레퍼런스와 동일한 형태의 실험실. 유리 가벽을 기준으로 내부에 찰리가 앉은 금속 의자가, 외부에 연구원들의 책상과 모니터가 정확히 배치됨.",
        "entities": "찰리의 얼굴과 흉부 구조가 레퍼런스와 완벽히 일치함. 굵은 전선이 몸에 부착되었고, 모니터 텍스트는 지시대로 식별 불가능하게 블러 처리됨.",
        "hard_violations": [],
        "physics": "찰리가 의자에 안정적으로 앉아 있으며, 몸에 연결된 전선들은 중력에 의해 바닥으로 자연스럽게 늘어져 지탱됨."
       },
       {
        "label": "B",
        "direction": "카메라와 찰리는 정면을 향함. 서 있는 연구원은 찰리를 향하고 다른 연구원들은 모니터를 봄.",
        "built_space": "레퍼런스의 구조를 잘 반영한 실험실. 유리 가벽 내부에 찰리와 장비 카트가 존재하고 외부에 연구 공간이 있음.",
        "entities": "찰리의 외형은 레퍼런스와 일치하나, 좌측 하단 모니터에 식별 가능한 선명한 텍스트(코드)가 노출되어 제약 사항과 충돌함.",
        "hard_violations": [
         "화면 내 식별 가능한 텍스트 금지(No readable writing anywhere) 지시를 위반하여 좌측 모니터에 선명한 코드가 나타남."
        ],
        "physics": "찰리가 의자에 바르게 앉아 있고, 전선들은 양옆의 카트 장비로 팽팽하게 연결되어 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "중앙 중경의 직립한 찰리와 굵은 배선은 구현했지만, 금지된 연구자 네 명을 추가했고 컴퓨터 화면을 정면으로 노출해 지시를 크게 위반한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "참조의 넓은 금속 등받이와 찰리의 외형을 더 잘 유지하지만, 금지된 연구자들을 추가했으며 컴퓨터의 뒷면·측면 중심 배치도 지키지 못해 부적합하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴과 몸은 유리 밖 준비 구역 및 카메라 쪽을 향한다. 왼쪽에 선 연구자는 찰리 쪽을 보고, 앉은 연구자들은 앞의 모니터 쪽을 향한다. 모니터는 사용자 쪽으로 향해 사용 관계는 가능하지만, 카메라에도 화면 앞면이 크게 드러나므로 뒷면과 측면 위주라는 지시에는 맞지 않는다.",
        "built_space": "중앙 유리 enclosure 안에 시험 의자 한 개, 그 좌우에 배선 장치 카트 두 개가 있다. 전경 양쪽에는 책상과 여러 모니터가 있으며 연구자 네 명이 서거나 앉아 있다. 원형 천장 조명, 유리 수직 프레임, 뒤쪽 실험 장비와 식재는 참조와 유사하다. 다만 참조의 넓은 금속 등받이가 좁은 의자로 달라졌고, 거의 정면 구도여서 의자의 사선 측면은 약하다. 찰리는 반사상이 아니라 유리 너머 실체로 보이며, 참조보다 공간이 훨씬 밝다.",
        "entities": "찰리 한 명의 각진 흰 얼굴판, 작은 푸른 눈, 선형 입과 마모된 금속 몸체는 참조의 기계형 외형을 대체로 따른다. 굵은 다색 전선은 양팔 주변에서 좌우 장치로 이어지지만 가슴에 여러 가닥이 직접 부착된 상태는 뚜렷하지 않다. 금속 의자, 유리벽과 컴퓨터는 존재한다. 흰 가운을 입은 성인 연구자 네 명은 찰리만 허용한 인물 제한을 위반한다. 화면에 글자 같은 행은 많지만 개별 문구의 판독 가능성은 확실하지 않다.",
        "hard_violations": [
         "찰리 외에는 아무도 등장시키지 말라는 명시적 제한을 어기고 연구자 네 명을 추가했다."
        ],
        "physics": "찰리의 골반은 의자 좌판, 팔은 팔걸이, 양발은 바닥에 지지되어 꼿꼿한 착석이 가능하다. 전선은 팔 주변과 장치 연결부 사이에서 아래로 처지고 일부는 바닥에 놓인다. 서 있는 연구자들은 바닥에, 앉은 연구자들은 작업 의자에 지지된다. 근거 없이 공중에 뜬 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴과 몸은 준비 구역 및 카메라 쪽을 향한다. 좌우 연구자들은 각자의 모니터 방향을 보고 있다. 모니터 화면은 사용자에게 향하지만 카메라도 그 앞면을 넓게 보므로, 요구된 컴퓨터 뒷면·측면 중심 구도는 아니다.",
        "built_space": "유리 구획 중앙에 넓은 금속 등받이를 가진 시험 의자 한 개가 있고, 전경 좌우에는 모니터 여러 대가 놓인 작업대가 있다. 유리 프레임, 원형 상부 조명, 뒤쪽 식재와 오른쪽 대형 실험 장비는 참조 장소와 연결된다. 찰리 전신이 중앙 중경에서 유리 너머 직접 보이며 전경 작업대와 깊이 구역이 나뉜다. 의자 좌판과 측면 구조는 일부 보이지만 거의 정면 구도다. 참조보다 주광이 강하고 밝다.",
        "entities": "찰리의 각진 흰 얼굴판, 작은 푸른 눈, 선형 입, 금속 흉갑과 관절 외형은 참조와 유사하다. 넓은 금속 등받이는 A보다 참조에 가깝다. 가슴에는 연결구가 있고 팔 주변 배선은 보이지만, 가슴과 팔에 여러 굵은 전선이 직접 붙어 있다는 표현은 충분히 선명하지 않다. 유리벽과 준비 구역 컴퓨터는 존재한다. 좌우에 보이는 연구자 최소 두 명은 허용되지 않은 추가 인물이다. 화면 내용은 대체로 흐려 명확히 읽히는 문구는 확인되지 않는다.",
        "hard_violations": [
         "찰리만 등장해야 하는 장면에 좌우 작업대의 연구자들을 추가했다."
        ],
        "physics": "찰리의 몸은 좌판과 등받이에 지지되고 양손은 팔걸이에, 양발은 의자 하부 발판에 놓여 있다. 자세는 안정적이지만 등받이에 기대어 A보다 덜 꼿꼿해 보인다. 배선은 팔과 몸 주변에서 오른쪽으로 내려와 바닥에 놓여 있으며, 컴퓨터와 작업 도구는 책상에 지지된다. 지지 없이 떠 있는 물체나 신체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "중앙 중경의 직립한 찰리와 굵은 배선은 구현했지만, 금지된 연구자 네 명을 추가했고 컴퓨터 화면을 정면으로 노출해 지시를 크게 위반한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "참조의 넓은 금속 등받이와 찰리의 외형을 더 잘 유지하지만, 금지된 연구자들을 추가했으며 컴퓨터의 뒷면·측면 중심 배치도 지키지 못해 부적합하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴과 몸은 유리 밖 준비 구역 및 카메라 쪽을 향한다. 왼쪽에 선 연구자는 찰리 쪽을 보고, 앉은 연구자들은 앞의 모니터 쪽을 향한다. 모니터는 사용자 쪽으로 향해 사용 관계는 가능하지만, 카메라에도 화면 앞면이 크게 드러나므로 뒷면과 측면 위주라는 지시에는 맞지 않는다.",
        "built_space": "중앙 유리 enclosure 안에 시험 의자 한 개, 그 좌우에 배선 장치 카트 두 개가 있다. 전경 양쪽에는 책상과 여러 모니터가 있으며 연구자 네 명이 서거나 앉아 있다. 원형 천장 조명, 유리 수직 프레임, 뒤쪽 실험 장비와 식재는 참조와 유사하다. 다만 참조의 넓은 금속 등받이가 좁은 의자로 달라졌고, 거의 정면 구도여서 의자의 사선 측면은 약하다. 찰리는 반사상이 아니라 유리 너머 실체로 보이며, 참조보다 공간이 훨씬 밝다.",
        "entities": "찰리 한 명의 각진 흰 얼굴판, 작은 푸른 눈, 선형 입과 마모된 금속 몸체는 참조의 기계형 외형을 대체로 따른다. 굵은 다색 전선은 양팔 주변에서 좌우 장치로 이어지지만 가슴에 여러 가닥이 직접 부착된 상태는 뚜렷하지 않다. 금속 의자, 유리벽과 컴퓨터는 존재한다. 흰 가운을 입은 성인 연구자 네 명은 찰리만 허용한 인물 제한을 위반한다. 화면에 글자 같은 행은 많지만 개별 문구의 판독 가능성은 확실하지 않다.",
        "hard_violations": [
         "찰리 외에는 아무도 등장시키지 말라는 명시적 제한을 어기고 연구자 네 명을 추가했다."
        ],
        "physics": "찰리의 골반은 의자 좌판, 팔은 팔걸이, 양발은 바닥에 지지되어 꼿꼿한 착석이 가능하다. 전선은 팔 주변과 장치 연결부 사이에서 아래로 처지고 일부는 바닥에 놓인다. 서 있는 연구자들은 바닥에, 앉은 연구자들은 작업 의자에 지지된다. 근거 없이 공중에 뜬 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴과 몸은 준비 구역 및 카메라 쪽을 향한다. 좌우 연구자들은 각자의 모니터 방향을 보고 있다. 모니터 화면은 사용자에게 향하지만 카메라도 그 앞면을 넓게 보므로, 요구된 컴퓨터 뒷면·측면 중심 구도는 아니다.",
        "built_space": "유리 구획 중앙에 넓은 금속 등받이를 가진 시험 의자 한 개가 있고, 전경 좌우에는 모니터 여러 대가 놓인 작업대가 있다. 유리 프레임, 원형 상부 조명, 뒤쪽 식재와 오른쪽 대형 실험 장비는 참조 장소와 연결된다. 찰리 전신이 중앙 중경에서 유리 너머 직접 보이며 전경 작업대와 깊이 구역이 나뉜다. 의자 좌판과 측면 구조는 일부 보이지만 거의 정면 구도다. 참조보다 주광이 강하고 밝다.",
        "entities": "찰리의 각진 흰 얼굴판, 작은 푸른 눈, 선형 입, 금속 흉갑과 관절 외형은 참조와 유사하다. 넓은 금속 등받이는 A보다 참조에 가깝다. 가슴에는 연결구가 있고 팔 주변 배선은 보이지만, 가슴과 팔에 여러 굵은 전선이 직접 붙어 있다는 표현은 충분히 선명하지 않다. 유리벽과 준비 구역 컴퓨터는 존재한다. 좌우에 보이는 연구자 최소 두 명은 허용되지 않은 추가 인물이다. 화면 내용은 대체로 흐려 명확히 읽히는 문구는 확인되지 않는다.",
        "hard_violations": [
         "찰리만 등장해야 하는 장면에 좌우 작업대의 연구자들을 추가했다."
        ],
        "physics": "찰리의 몸은 좌판과 등받이에 지지되고 양손은 팔걸이에, 양발은 의자 하부 발판에 놓여 있다. 자세는 안정적이지만 등받이에 기대어 A보다 덜 꼿꼿해 보인다. 배선은 팔과 몸 주변에서 오른쪽으로 내려와 바닥에 놓여 있으며, 컴퓨터와 작업 도구는 책상에 지지된다. 지지 없이 떠 있는 물체나 신체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.381
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.131
   },
   "violations": {
    "B": [
     "[gemini-pro] 화면 내 식별 가능한 텍스트 금지(No readable writing anywhere) 지시를 위반하여 좌측 모니터에 선명한 코드가 나타남.",
     "[gpt-high] 찰리 외에는 아무도 등장시키지 말라는 명시적 제한을 어기고 연구자 네 명을 추가했다."
    ],
    "A": [
     "[gpt-high] 찰리만 등장해야 하는 장면에 좌우 작업대의 연구자들을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1131
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "레퍼런스의 로봇 외형과 실험실 배경을 완벽히 유지했으며, 모니터 화면을 흐리게 처리하여 텍스트 비노출 제약을 훌륭하게 준수했습니다.  ★위반: [gpt-high] 찰리만 등장해야 하는 장면에 좌우 작업대의 연구자들을 추가했다."
   },
   {
    "label": "B",
    "score": 1131,
    "verdict_ko": "전반적인 구도와 형태는 좋으나, 좌측 모니터에 선명한 코드가 노출되어 텍스트 금지 제약을 위반했으며 중앙의 연구원이 시야를 다소 가립니다.  ★위반: [gemini-pro] 화면 내 식별 가능한 텍스트 금지(No readable writing anywhere) 지시를 위반하여 좌측 모니터에 선명한 코드가 나타남. / [gpt-high] 찰리 외에는 아무도 등장시키지 말라는 명시적 제한을 어기고 연구자 네 명을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S84sh1_sel.png",
    "asset_id": "aab7da1e-cd8c-4a16-a151-803530ebbe82",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf39-f02a-77de-b23f-f9cec75ca424",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S84sh1"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S85sh12": {
  "input_fingerprint": "5f9261fb6f34e343",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링에서 시커먼 연기가 훅 피어오르던 도중의 한 형태로 포착된 위태로운 순간.\n\nLOCATION (lock): Inside the glass-walled energy test area of the main laboratory, opposite glowing data screens as smoke rises from the robot's chest. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Glass partition (Remaining between camera and 찰리) — The wired upper body is seen through the glass; reflections are not the subject; used as Maintains the researchers' external observation position throughout the close view; Chest ring (Emitting black smoke as its rotation slows) — Its outward face is seen obliquely within 찰리's chest; used as Forms the focal detail below center while occupying only a limited portion of the body-centered frame; Connected wires (Still attached during the malfunction); used as Retain procedural context along the sides of the torso; Black smoke (Rising from the slowing chest ring); used as Creates an irregular upward shape beside the face without concealing the ring's origin.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the laboratory's neutral ambient exposure so the black smoke reads distinctly against the body without adding an unsupported glow or change of light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains seated and wired inside the glass enclosure; black smoke rises from his chest ring as it loses speed. The monitoring screen displays “ERROR” after the recorded energy graphs have climbed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링에서 시커먼 연기가 훅 피어오르던 도중의 한 형태로 포착된 위태로운 순간.\n\nLOCATION (lock): Inside the glass-walled energy test area of the main laboratory, opposite glowing data screens as smoke rises from the robot's chest. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Glass partition (Remaining between camera and 찰리) — The wired upper body is seen through the glass; reflections are not the subject; used as Maintains the researchers' external observation position throughout the close view; Chest ring (Emitting black smoke as its rotation slows) — Its outward face is seen obliquely within 찰리's chest; used as Forms the focal detail below center while occupying only a limited portion of the body-centered frame; Connected wires (Still attached during the malfunction); used as Retain procedural context along the sides of the torso; Black smoke (Rising from the slowing chest ring); used as Creates an irregular upward shape beside the face without concealing the ring's origin.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the laboratory's neutral ambient exposure so the black smoke reads distinctly against the body without adding an unsupported glow or change of light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains seated and wired inside the glass enclosure; black smoke rises from his chest ring as it loses speed. The monitoring screen displays “ERROR” after the recorded energy graphs have climbed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링에서 시커먼 연기가 훅 피어오르던 도중의 한 형태로 포착된 위태로운 순간.\n\nLOCATION (lock): Inside the glass-walled energy test area of the main laboratory, opposite glowing data screens as smoke rises from the robot's chest. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Glass partition (Remaining between camera and 찰리) — The wired upper body is seen through the glass; reflections are not the subject; used as Maintains the researchers' external observation position throughout the close view; Chest ring (Emitting black smoke as its rotation slows) — Its outward face is seen obliquely within 찰리's chest; used as Forms the focal detail below center while occupying only a limited portion of the body-centered frame; Connected wires (Still attached during the malfunction); used as Retain procedural context along the sides of the torso; Black smoke (Rising from the slowing chest ring); used as Creates an irregular upward shape beside the face without concealing the ring's origin.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the laboratory's neutral ambient exposure so the black smoke reads distinctly against the body without adding an unsupported glow or change of light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains seated and wired inside the glass enclosure; black smoke rises from his chest ring as it loses speed. The monitoring screen displays “ERROR” after the recorded energy graphs have climbed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리는 정면을 향해 앉아있고 가슴의 파이프 형태 구멍에서 검은 연기가 위로 솟아오름. 우측 배경의 모니터에 'ERROR' 텍스트가 표시됨.",
    "built_space": "실험실 내부. 찰리 뒤쪽으로 유리벽들이 보이나, 카메라와 찰리 사이에는 유리 파티션이나 반사가 없어 카메라가 격리실 내부에 위치한 것으로 보임.",
    "entities": "찰리의 머리와 외형은 레퍼런스와 매우 유사하게 유지됨. 프롬프트 지시대로 다른 인물은 전혀 등장하지 않음.",
    "hard_violations": [
     "[gpt-high] 배경 모니터의 ‘ERROR’가 선명하게 읽혀, 어디에도 읽을 수 있는 글자를 두지 말라는 지시를 위반합니다."
    ],
    "physics": "찰리는 금속 테이블 위에 안정적으로 앉아 지탱되고 있으며 연기는 자연스럽게 위로 퍼짐."
   },
   {
    "label": "B",
    "direction": "찰리는 약간 우측을 향하고 있으며, 가슴의 거대한 원형 장치에서 짙은 검은 연기가 뿜어져 나옴.",
    "built_space": "실험실 내부. 카메라와 찰리 사이에 세로 형태의 유리 파티션 구조물과 반사가 명확하게 보여 관찰자 시점을 유지함.",
    "entities": "찰리의 뺨에 레퍼런스에 없는 점 형태의 문양이 추가되어 외형이 일부 변형됨. 배경에 흰 가운을 입은 두 명의 연구원이 배치됨.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에서 엄격히 금지한 추가 인물(배경의 연구원 2명) 생성",
     "[gpt-high] 찰리 외에는 누구도 등장시키지 말라는 지시와 달리 배경에 연구원 두 명이 보입니다."
    ],
    "physics": "찰리는 금속 테이블 위에 앉아있으며 연기는 공기 중으로 솟아오름."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "요구된 클로즈업 대신 미디엄 샷으로 연출되고 카메라 앞의 유리벽이 생략되었으나, 프롬프트가 금지한 추가 인물을 철저히 배제하여 더 나은 결과를 보여줍니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "카메라 앞의 유리벽 구도는 잘 살렸으나, 프롬프트에서 등장시키지 말라고 엄격히 지시한 추가 인물(배경의 연구원들)을 포함시킨 치명적인 규정 위반이 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 정면을 향해 앉아있고 가슴의 파이프 형태 구멍에서 검은 연기가 위로 솟아오름. 우측 배경의 모니터에 'ERROR' 텍스트가 표시됨.",
        "built_space": "실험실 내부. 찰리 뒤쪽으로 유리벽들이 보이나, 카메라와 찰리 사이에는 유리 파티션이나 반사가 없어 카메라가 격리실 내부에 위치한 것으로 보임.",
        "entities": "찰리의 머리와 외형은 레퍼런스와 매우 유사하게 유지됨. 프롬프트 지시대로 다른 인물은 전혀 등장하지 않음.",
        "hard_violations": [],
        "physics": "찰리는 금속 테이블 위에 안정적으로 앉아 지탱되고 있으며 연기는 자연스럽게 위로 퍼짐."
       },
       {
        "label": "B",
        "direction": "찰리는 약간 우측을 향하고 있으며, 가슴의 거대한 원형 장치에서 짙은 검은 연기가 뿜어져 나옴.",
        "built_space": "실험실 내부. 카메라와 찰리 사이에 세로 형태의 유리 파티션 구조물과 반사가 명확하게 보여 관찰자 시점을 유지함.",
        "entities": "찰리의 뺨에 레퍼런스에 없는 점 형태의 문양이 추가되어 외형이 일부 변형됨. 배경에 흰 가운을 입은 두 명의 연구원이 배치됨.",
        "hard_violations": [
         "프롬프트에서 엄격히 금지한 추가 인물(배경의 연구원 2명) 생성"
        ],
        "physics": "찰리는 금속 테이블 위에 앉아있으며 연기는 공기 중으로 솟아오름."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "요구된 클로즈업 대신 미디엄 샷으로 연출되고 카메라 앞의 유리벽이 생략되었으나, 프롬프트가 금지한 추가 인물을 철저히 배제하여 더 나은 결과를 보여줍니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "카메라 앞의 유리벽 구도는 잘 살렸으나, 프롬프트에서 등장시키지 말라고 엄격히 지시한 추가 인물(배경의 연구원들)을 포함시킨 치명적인 규정 위반이 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 정면을 향해 앉아있고 가슴의 파이프 형태 구멍에서 검은 연기가 위로 솟아오름. 우측 배경의 모니터에 'ERROR' 텍스트가 표시됨.",
        "built_space": "실험실 내부. 찰리 뒤쪽으로 유리벽들이 보이나, 카메라와 찰리 사이에는 유리 파티션이나 반사가 없어 카메라가 격리실 내부에 위치한 것으로 보임.",
        "entities": "찰리의 머리와 외형은 레퍼런스와 매우 유사하게 유지됨. 프롬프트 지시대로 다른 인물은 전혀 등장하지 않음.",
        "hard_violations": [],
        "physics": "찰리는 금속 테이블 위에 안정적으로 앉아 지탱되고 있으며 연기는 자연스럽게 위로 퍼짐."
       },
       {
        "label": "B",
        "direction": "찰리는 약간 우측을 향하고 있으며, 가슴의 거대한 원형 장치에서 짙은 검은 연기가 뿜어져 나옴.",
        "built_space": "실험실 내부. 카메라와 찰리 사이에 세로 형태의 유리 파티션 구조물과 반사가 명확하게 보여 관찰자 시점을 유지함.",
        "entities": "찰리의 뺨에 레퍼런스에 없는 점 형태의 문양이 추가되어 외형이 일부 변형됨. 배경에 흰 가운을 입은 두 명의 연구원이 배치됨.",
        "hard_violations": [
         "프롬프트에서 엄격히 금지한 추가 인물(배경의 연구원 2명) 생성"
        ],
        "physics": "찰리는 금속 테이블 위에 앉아있으며 연기는 공기 중으로 솟아오름."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "비스듬히 보이는 링과 짙은 연기는 맞지만, 금지된 연구원 두 명이 등장하고 가슴 링이 과도하게 커져 인물 중심의 제한된 초점이라는 지시에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "단독으로 앉은 찰리, 작은 가슴 링과 연결선, 유리 너머의 관찰 위치를 더 충실히 유지하지만, 선명한 ‘ERROR’ 표기로 읽을 수 있는 글자 금지를 위반합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴은 화면 오른쪽 앞을 향하고, 가슴 링의 바깥 면도 카메라에 비스듬히 드러납니다. 검은 연기는 링 내부에서 시작해 얼굴 오른쪽을 따라 위로 치솟으며 발생 지점이 보입니다. 배경 연구원 두 명은 각각 앞의 모니터를 향합니다.",
        "built_space": "찰리 뒤에 금속 등받이 하나가 있고, 카메라와 찰리 사이에는 굵은 수직 유리 프레임들이 보입니다. 배경에는 왼쪽 화면 하나와 오른쪽 연구원 자리의 화면 두 개가 보입니다. 유리와 금속, 낮의 차가운 실험실 조명은 이어지지만, 등장하면 안 되는 연구원 두 명을 배경 작업석에 배치했습니다. 광학적으로 불가능한 반사는 보이지 않습니다.",
        "entities": "찰리 한 대와 성인으로 보이는 연구원 두 명이 있습니다. 찰리의 각진 흰 얼굴판, 작은 눈, 선형 입과 마모된 금속 몸체는 참조의 특징을 대체로 유지합니다. 몸통과 팔 옆에는 여러 연결선이 남아 있습니다. 가슴 링은 하나이나 참조의 작은 상부 가슴 부품에 비해 매우 크게 돌출되어 몸통에서 차지하는 비중이 큽니다. 짙은 검은 연기는 명확하며, 읽을 수 있는 글자는 뚜렷하지 않습니다.",
        "hard_violations": [
         "찰리 외에는 누구도 등장시키지 말라는 지시와 달리 배경에 연구원 두 명이 보입니다."
        ],
        "physics": "찰리의 몸통은 기울어진 금속 등받이에 기대고 있으며, 하체는 화면 아래로 이어져 착석 상태로 읽힙니다. 전선들은 팔과 몸통에 걸쳐 지지되고 아래로 늘어집니다. 연기는 가슴 링에서 연속적으로 솟아오르므로 공중에 고립된 물체가 아닙니다. 지지 없이 떠 있는 신체나 물건은 보이지 않으며, 링의 감속 여부는 정지 화면만으로 확인하기 어렵습니다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴과 몸통은 거의 정면으로 카메라를 향합니다. 작은 가슴 링의 면은 약간 위로 기울어져 타원형으로 보이지만, 몸통 전체의 사선 관찰 각도는 약합니다. 연기는 링에서 나와 화면 왼쪽으로 휘며 얼굴 옆으로 상승하고 발생 지점을 남겨 둡니다.",
        "built_space": "금속 의자 등받이 하나, 유리 칸막이의 수직 프레임과 중앙 이음선, 천장의 원형 조명 하나가 보입니다. 뒤쪽 식물과 오른쪽 창문·실험 장비는 참조 장소의 배치 및 재질과 잘 이어집니다. 카메라는 유리 바깥에서 찰리를 관찰하는 위치로 읽힙니다. 오른쪽 배경 모니터 하나가 보이며 다른 사람은 없습니다. 다만 머리부터 허벅지까지 포함해 요구한 클로즈업보다 다소 넓습니다.",
        "entities": "찰리 한 대만 보이며, 각진 흰 머리 외장과 점 형태의 눈, 선 형태의 입, 회색 금속 장갑이 유지됩니다. 작은 가슴 링 하나가 화면 중심 아래에 놓이고, 연결선들이 가슴과 양팔 옆으로 이어집니다. 검은 연기는 있지만 A보다 가늘고 반투명해 거세게 피어오르는 느낌은 약합니다. 오른쪽 모니터에는 ‘ERROR’가 명확히 읽혀 최종 무문자 지시와 충돌합니다.",
        "hard_violations": [
         "배경 모니터의 ‘ERROR’가 선명하게 읽혀, 어디에도 읽을 수 있는 글자를 두지 말라는 지시를 위반합니다."
        ],
        "physics": "찰리는 금속 의자에 앉아 등받이에 몸을 기대고 있습니다. 팔은 몸 옆으로 내려가고 전선은 가슴 접속부와 팔, 허벅지에 걸쳐 지지됩니다. 연기는 링에서 끊김 없이 위로 흐르므로 자연스러운 상승으로 읽힙니다. 지지 없이 떠 있는 신체나 물건은 없으며, 링이 회전하다 느려지는 상태인지는 이 정지 화면에서 확정할 수 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "비스듬히 보이는 링과 짙은 연기는 맞지만, 금지된 연구원 두 명이 등장하고 가슴 링이 과도하게 커져 인물 중심의 제한된 초점이라는 지시에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "단독으로 앉은 찰리, 작은 가슴 링과 연결선, 유리 너머의 관찰 위치를 더 충실히 유지하지만, 선명한 ‘ERROR’ 표기로 읽을 수 있는 글자 금지를 위반합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴은 화면 오른쪽 앞을 향하고, 가슴 링의 바깥 면도 카메라에 비스듬히 드러납니다. 검은 연기는 링 내부에서 시작해 얼굴 오른쪽을 따라 위로 치솟으며 발생 지점이 보입니다. 배경 연구원 두 명은 각각 앞의 모니터를 향합니다.",
        "built_space": "찰리 뒤에 금속 등받이 하나가 있고, 카메라와 찰리 사이에는 굵은 수직 유리 프레임들이 보입니다. 배경에는 왼쪽 화면 하나와 오른쪽 연구원 자리의 화면 두 개가 보입니다. 유리와 금속, 낮의 차가운 실험실 조명은 이어지지만, 등장하면 안 되는 연구원 두 명을 배경 작업석에 배치했습니다. 광학적으로 불가능한 반사는 보이지 않습니다.",
        "entities": "찰리 한 대와 성인으로 보이는 연구원 두 명이 있습니다. 찰리의 각진 흰 얼굴판, 작은 눈, 선형 입과 마모된 금속 몸체는 참조의 특징을 대체로 유지합니다. 몸통과 팔 옆에는 여러 연결선이 남아 있습니다. 가슴 링은 하나이나 참조의 작은 상부 가슴 부품에 비해 매우 크게 돌출되어 몸통에서 차지하는 비중이 큽니다. 짙은 검은 연기는 명확하며, 읽을 수 있는 글자는 뚜렷하지 않습니다.",
        "hard_violations": [
         "찰리 외에는 누구도 등장시키지 말라는 지시와 달리 배경에 연구원 두 명이 보입니다."
        ],
        "physics": "찰리의 몸통은 기울어진 금속 등받이에 기대고 있으며, 하체는 화면 아래로 이어져 착석 상태로 읽힙니다. 전선들은 팔과 몸통에 걸쳐 지지되고 아래로 늘어집니다. 연기는 가슴 링에서 연속적으로 솟아오르므로 공중에 고립된 물체가 아닙니다. 지지 없이 떠 있는 신체나 물건은 보이지 않으며, 링의 감속 여부는 정지 화면만으로 확인하기 어렵습니다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴과 몸통은 거의 정면으로 카메라를 향합니다. 작은 가슴 링의 면은 약간 위로 기울어져 타원형으로 보이지만, 몸통 전체의 사선 관찰 각도는 약합니다. 연기는 링에서 나와 화면 왼쪽으로 휘며 얼굴 옆으로 상승하고 발생 지점을 남겨 둡니다.",
        "built_space": "금속 의자 등받이 하나, 유리 칸막이의 수직 프레임과 중앙 이음선, 천장의 원형 조명 하나가 보입니다. 뒤쪽 식물과 오른쪽 창문·실험 장비는 참조 장소의 배치 및 재질과 잘 이어집니다. 카메라는 유리 바깥에서 찰리를 관찰하는 위치로 읽힙니다. 오른쪽 배경 모니터 하나가 보이며 다른 사람은 없습니다. 다만 머리부터 허벅지까지 포함해 요구한 클로즈업보다 다소 넓습니다.",
        "entities": "찰리 한 대만 보이며, 각진 흰 머리 외장과 점 형태의 눈, 선 형태의 입, 회색 금속 장갑이 유지됩니다. 작은 가슴 링 하나가 화면 중심 아래에 놓이고, 연결선들이 가슴과 양팔 옆으로 이어집니다. 검은 연기는 있지만 A보다 가늘고 반투명해 거세게 피어오르는 느낌은 약합니다. 오른쪽 모니터에는 ‘ERROR’가 명확히 읽혀 최종 무문자 지시와 충돌합니다.",
        "hard_violations": [
         "배경 모니터의 ‘ERROR’가 선명하게 읽혀, 어디에도 읽을 수 있는 글자를 두지 말라는 지시를 위반합니다."
        ],
        "physics": "찰리는 금속 의자에 앉아 등받이에 몸을 기대고 있습니다. 팔은 몸 옆으로 내려가고 전선은 가슴 접속부와 팔, 허벅지에 걸쳐 지지됩니다. 연기는 링에서 끊김 없이 위로 흐르므로 자연스러운 상승으로 읽힙니다. 지지 없이 떠 있는 신체나 물건은 없으며, 링이 회전하다 느려지는 상태인지는 이 정지 화면에서 확정할 수 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.0
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.75
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트에서 엄격히 금지한 추가 인물(배경의 연구원 2명) 생성",
     "[gpt-high] 찰리 외에는 누구도 등장시키지 말라는 지시와 달리 배경에 연구원 두 명이 보입니다."
    ],
    "A": [
     "[gpt-high] 배경 모니터의 ‘ERROR’가 선명하게 읽혀, 어디에도 읽을 수 있는 글자를 두지 말라는 지시를 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "요구된 클로즈업 대신 미디엄 샷으로 연출되고 카메라 앞의 유리벽이 생략되었으나, 프롬프트가 금지한 추가 인물을 철저히 배제하여 더 나은 결과를 보여줍니다.  ★위반: [gpt-high] 배경 모니터의 ‘ERROR’가 선명하게 읽혀, 어디에도 읽을 수 있는 글자를 두지 말라는 지시를 위반합니다."
   },
   {
    "label": "B",
    "score": 750,
    "verdict_ko": "카메라 앞의 유리벽 구도는 잘 살렸으나, 프롬프트에서 등장시키지 말라고 엄격히 지시한 추가 인물(배경의 연구원들)을 포함시킨 치명적인 규정 위반이 있습니다.  ★위반: [gemini-pro] 프롬프트에서 엄격히 금지한 추가 인물(배경의 연구원 2명) 생성 / [gpt-high] 찰리 외에는 누구도 등장시키지 말라는 지시와 달리 배경에 연구원 두 명이 보입니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S85sh1_sel.png",
    "asset_id": "e583f190-96c0-41d6-b177-91d4a62c3d7b",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf3f-e41d-7f11-bcdf-640b466458ff",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S85sh1"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "era_fail::7be52edca62f73f8": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "era_fail::d16bd2a8299abc18": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S86sh6": {
  "input_fingerprint": "a8edc1ebbfa81117",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 지소영을 향해 삿대질을 하며 눈에 핏대를 세운 격앙된 얼굴 클로즈업.\n\nLOCATION (lock): Across the desk-side conversation area inside the director's office, in nighttime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Desk edge (Remaining beside the confrontation) — A short oblique edge is visible beneath the pointing forearm; used as Provides a minimal continuity anchor without competing with his face or gesture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's existing ambient treatment and facial contrast so the closer view intensifies his agitation without an additional lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the director's desk, modern office finishes, and nighttime lighting. EXCLUDE laboratory beds, experimental cables, and equipment from the separate testing room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office retains its clean, modern furnishings and the old piano in the central hall. 이현우: He remains standing in the office in his white clothing, visibly furious.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 지소영을 향해 삿대질을 하며 눈에 핏대를 세운 격앙된 얼굴 클로즈업.\n\nLOCATION (lock): Across the desk-side conversation area inside the director's office, in nighttime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Desk edge (Remaining beside the confrontation) — A short oblique edge is visible beneath the pointing forearm; used as Provides a minimal continuity anchor without competing with his face or gesture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's existing ambient treatment and facial contrast so the closer view intensifies his agitation without an additional lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the director's desk, modern office finishes, and nighttime lighting. EXCLUDE laboratory beds, experimental cables, and equipment from the separate testing room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office retains its clean, modern furnishings and the old piano in the central hall. 이현우: He remains standing in the office in his white clothing, visibly furious.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 지소영을 향해 삿대질을 하며 눈에 핏대를 세운 격앙된 얼굴 클로즈업.\n\nLOCATION (lock): Across the desk-side conversation area inside the director's office, in nighttime interior light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Desk edge (Remaining beside the confrontation) — A short oblique edge is visible beneath the pointing forearm; used as Provides a minimal continuity anchor without competing with his face or gesture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's existing ambient treatment and facial contrast so the closer view intensifies his agitation without an additional lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the director's desk, modern office finishes, and nighttime lighting. EXCLUDE laboratory beds, experimental cables, and equipment from the separate testing room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office retains its clean, modern furnishings and the old piano in the central hall. 이현우: He remains standing in the office in his white clothing, visibly furious.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "남자는 화면 왼쪽 앞부분에 등지고 앉아 있는 여자를 향해 삿대질을 하고 있으며 시선도 그녀를 향하고 있다.",
    "built_space": "배경에 밤의 도시가 내다보이는 통유리창이 배치되어 있어, 레퍼런스 이미지의 막힌 회색 벽면 및 피아노가 있는 공간과 완전히 다른 구조를 띠고 있다.",
    "entities": "이현우의 전체적인 인상착의는 비슷하나 눈 흰자가 인위적으로 붉게 변형되어 있다. 프레임 앞쪽에 뒷모습의 여성(지소영)이 포함되어 있다.",
    "hard_violations": [
     "[gemini-pro] 지정된 장소와 완전히 다른 배경(유리창과 도시 야경) 생성",
     "[gemini-pro] 해부학적으로 불가능하게 변형된 눈(비정상적으로 붉게 칠해진 흰자)",
     "[gpt-high] 이현우만 보여야 하는 장면에 이전 장면의 여성으로 보이는 인물의 얼굴과 상체를 추가했다."
    ],
    "physics": "남자는 두 발로 서서 상체를 앞으로 숙인 채 오른팔을 들어 가리키고 있으며, 자세를 유지하는 데 물리적인 어색함은 없다."
   },
   {
    "label": "B",
    "direction": "남자는 화면 오른쪽 프레임 밖의 대상을 향해 팔을 뻗어 삿대질을 하고 있으며, 시선 역시 강렬하게 같은 곳을 주시하고 있다.",
    "built_space": "레퍼런스와 완벽히 동일한 공간으로, 왼쪽 뒤의 피아노, 중앙의 책장, 오른쪽의 회색 패널 벽면이 정확히 묘사되었다. 뻗은 팔 아래로 책상의 모서리가 올바르게 위치해 있다.",
    "entities": "이현우의 얼굴, 머리 스타일, 입고 있는 흰색 옷이 레퍼런스와 정확히 일치하며 지시된 격앙된 표정을 자연스럽게 짓고 있다.",
    "hard_violations": [],
    "physics": "남자는 서 있는 상태에서 상체를 약간 기울여 오른팔을 힘껏 뻗고 있으며, 몸통이 이 동작을 안정적으로 지탱하고 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "레퍼런스와 전혀 다른 유리창 배경을 생성하여 공간 고정 지시를 어겼고, 눈 흰자를 인위적으로 붉게 변형하여 해부학적 제약을 위반했습니다."
       },
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "클로즈업 앵글과 뻗은 팔 아래의 책상 모서리 배치를 정확히 구현했으며, 지정된 실내 공간과 인물의 격앙된 표정까지 완벽하게 재현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남자는 화면 왼쪽 앞부분에 등지고 앉아 있는 여자를 향해 삿대질을 하고 있으며 시선도 그녀를 향하고 있다.",
        "built_space": "배경에 밤의 도시가 내다보이는 통유리창이 배치되어 있어, 레퍼런스 이미지의 막힌 회색 벽면 및 피아노가 있는 공간과 완전히 다른 구조를 띠고 있다.",
        "entities": "이현우의 전체적인 인상착의는 비슷하나 눈 흰자가 인위적으로 붉게 변형되어 있다. 프레임 앞쪽에 뒷모습의 여성(지소영)이 포함되어 있다.",
        "hard_violations": [
         "지정된 장소와 완전히 다른 배경(유리창과 도시 야경) 생성",
         "해부학적으로 불가능하게 변형된 눈(비정상적으로 붉게 칠해진 흰자)"
        ],
        "physics": "남자는 두 발로 서서 상체를 앞으로 숙인 채 오른팔을 들어 가리키고 있으며, 자세를 유지하는 데 물리적인 어색함은 없다."
       },
       {
        "label": "B",
        "direction": "남자는 화면 오른쪽 프레임 밖의 대상을 향해 팔을 뻗어 삿대질을 하고 있으며, 시선 역시 강렬하게 같은 곳을 주시하고 있다.",
        "built_space": "레퍼런스와 완벽히 동일한 공간으로, 왼쪽 뒤의 피아노, 중앙의 책장, 오른쪽의 회색 패널 벽면이 정확히 묘사되었다. 뻗은 팔 아래로 책상의 모서리가 올바르게 위치해 있다.",
        "entities": "이현우의 얼굴, 머리 스타일, 입고 있는 흰색 옷이 레퍼런스와 정확히 일치하며 지시된 격앙된 표정을 자연스럽게 짓고 있다.",
        "hard_violations": [],
        "physics": "남자는 서 있는 상태에서 상체를 약간 기울여 오른팔을 힘껏 뻗고 있으며, 몸통이 이 동작을 안정적으로 지탱하고 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "레퍼런스와 전혀 다른 유리창 배경을 생성하여 공간 고정 지시를 어겼고, 눈 흰자를 인위적으로 붉게 변형하여 해부학적 제약을 위반했습니다."
       },
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "클로즈업 앵글과 뻗은 팔 아래의 책상 모서리 배치를 정확히 구현했으며, 지정된 실내 공간과 인물의 격앙된 표정까지 완벽하게 재현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "남자는 화면 왼쪽 앞부분에 등지고 앉아 있는 여자를 향해 삿대질을 하고 있으며 시선도 그녀를 향하고 있다.",
        "built_space": "배경에 밤의 도시가 내다보이는 통유리창이 배치되어 있어, 레퍼런스 이미지의 막힌 회색 벽면 및 피아노가 있는 공간과 완전히 다른 구조를 띠고 있다.",
        "entities": "이현우의 전체적인 인상착의는 비슷하나 눈 흰자가 인위적으로 붉게 변형되어 있다. 프레임 앞쪽에 뒷모습의 여성(지소영)이 포함되어 있다.",
        "hard_violations": [
         "지정된 장소와 완전히 다른 배경(유리창과 도시 야경) 생성",
         "해부학적으로 불가능하게 변형된 눈(비정상적으로 붉게 칠해진 흰자)"
        ],
        "physics": "남자는 두 발로 서서 상체를 앞으로 숙인 채 오른팔을 들어 가리키고 있으며, 자세를 유지하는 데 물리적인 어색함은 없다."
       },
       {
        "label": "B",
        "direction": "남자는 화면 오른쪽 프레임 밖의 대상을 향해 팔을 뻗어 삿대질을 하고 있으며, 시선 역시 강렬하게 같은 곳을 주시하고 있다.",
        "built_space": "레퍼런스와 완벽히 동일한 공간으로, 왼쪽 뒤의 피아노, 중앙의 책장, 오른쪽의 회색 패널 벽면이 정확히 묘사되었다. 뻗은 팔 아래로 책상의 모서리가 올바르게 위치해 있다.",
        "entities": "이현우의 얼굴, 머리 스타일, 입고 있는 흰색 옷이 레퍼런스와 정확히 일치하며 지시된 격앙된 표정을 자연스럽게 짓고 있다.",
        "hard_violations": [],
        "physics": "남자는 서 있는 상태에서 상체를 약간 기울여 오른팔을 힘껏 뻗고 있으며, 몸통이 이 동작을 안정적으로 지탱하고 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "화면 밖 상대를 향한 시선과 삿대질, 격앙된 얼굴 및 기존 사무실의 연속성이 맞지만, 얼굴 클로즈업에 비해 상체와 책상 노출이 다소 넓다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "상대를 향한 삿대질은 명확하지만, 제외해야 할 여성까지 등장시키고 허리까지 보이는 대화 구도로 넓혀 단독 얼굴 클로즈업 지시를 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "검지는 화면 오른쪽 바깥을 향하고 두 눈도 같은 방향의 상대를 노려본다. 이전 장면에서 지소영이 있던 책상 건너편 방향과 부합한다. 대상은 화면 밖이므로 손끝이 상대의 어느 부위를 가리키는지는 확인할 수 없다.",
        "built_space": "왼쪽에 검은 피아노 한 대와 그 위 액자 하나, 뒤쪽 틈에 수납장 하나와 화분 하나, 오른쪽에 회색 벽 패널이 보인다. 천장에는 선형 조명 하나와 작은 매입등들이 있다. 오른쪽 아래 책상 하나에는 검은 탁상 기기 하나, 책 더미와 작은 소품들이 놓여 있어 이전 사무실과 잘 이어진다. 다만 팔 아래 책상이 짧은 모서리 이상으로 넓게 드러난다.",
        "entities": "검은 헝클어진 머리와 무지 흰색 긴팔 상의를 입은 젊은 동아시아계 남성 한 명만 보인다. 얼굴과 머리 모양은 이현우의 참고 이미지에 가깝다. 찡그린 미간, 드러낸 치아와 긴장된 눈매로 분노를 표현하며 눈의 구조는 정상적이다. 다른 인물이나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "내민 손과 팔은 손목·팔꿈치·어깨로 자연스럽게 연결되어 있으며 팔 근육으로 들어 올린 삿대질 자세가 가능하다. 하체와 발은 프레임 밖이어서 바닥 접촉은 확인할 수 없지만 공중에 뜬 몸으로 보이지 않는다. 책상 소품은 상판에, 검은 탁상 기기는 받침에 지지되어 있다."
       },
       {
        "label": "B",
        "direction": "남성의 검지는 왼쪽 전경 여성의 얼굴 쪽을 가리키며 시선도 그 여성을 향한다. 여성 역시 남성 쪽으로 얼굴을 돌리고 있어 대치 방향 자체는 명확하다.",
        "built_space": "전경의 책상 하나 위에 서류 묶음, 문서 받침, 연필꽂이 하나가 있고 남성 뒤에는 빈 사무용 의자 하나가 보인다. 배경은 유리 칸막이와 어두운 창, 목재 프레임이 크게 차지하며 오른쪽 끝에 피아노 일부가 보인다. 이전 장면의 회색 패널 중심 공간보다 연속성이 약하다. 창에는 남성의 등처럼 보이는 반사상이 있으나 이 시점에서 불가능한 반사라고 단정할 근거는 없다.",
        "entities": "흰 긴팔 상의와 검은 머리의 동아시아계 남성이 보이지만, 얼굴은 참고의 18세 인상보다 나이 들어 보인다. 눈 주변이 붉고 눈을 크게 뜬 격앙된 표정이다. 왼쪽 전경에는 검은 옷을 입은 단발 여성의 얼굴과 상체가 추가되어, 이전 장면의 다른 인물을 이 장면에 포함하지 말라는 지시와 충돌한다. 서류의 글자 형태는 보이지만 내용을 판독하기는 어렵다.",
        "hard_violations": [
         "이현우만 보여야 하는 장면에 이전 장면의 여성으로 보이는 인물의 얼굴과 상체를 추가했다."
        ],
        "physics": "남성은 상체를 앞으로 숙이고 한쪽 팔을 뻗으며 반대 손은 주먹을 쥐고 있다. 손과 팔의 연결 및 몸의 기울기는 실제로 가능한 동작이고, 하체가 잘려 발 지지는 확인할 수 없다. 여성의 앉거나 선 상태도 크롭 때문에 확정할 수 없다. 서류와 연필꽂이는 책상에 놓여 있으며 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "화면 밖 상대를 향한 시선과 삿대질, 격앙된 얼굴 및 기존 사무실의 연속성이 맞지만, 얼굴 클로즈업에 비해 상체와 책상 노출이 다소 넓다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "상대를 향한 삿대질은 명확하지만, 제외해야 할 여성까지 등장시키고 허리까지 보이는 대화 구도로 넓혀 단독 얼굴 클로즈업 지시를 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "검지는 화면 오른쪽 바깥을 향하고 두 눈도 같은 방향의 상대를 노려본다. 이전 장면에서 지소영이 있던 책상 건너편 방향과 부합한다. 대상은 화면 밖이므로 손끝이 상대의 어느 부위를 가리키는지는 확인할 수 없다.",
        "built_space": "왼쪽에 검은 피아노 한 대와 그 위 액자 하나, 뒤쪽 틈에 수납장 하나와 화분 하나, 오른쪽에 회색 벽 패널이 보인다. 천장에는 선형 조명 하나와 작은 매입등들이 있다. 오른쪽 아래 책상 하나에는 검은 탁상 기기 하나, 책 더미와 작은 소품들이 놓여 있어 이전 사무실과 잘 이어진다. 다만 팔 아래 책상이 짧은 모서리 이상으로 넓게 드러난다.",
        "entities": "검은 헝클어진 머리와 무지 흰색 긴팔 상의를 입은 젊은 동아시아계 남성 한 명만 보인다. 얼굴과 머리 모양은 이현우의 참고 이미지에 가깝다. 찡그린 미간, 드러낸 치아와 긴장된 눈매로 분노를 표현하며 눈의 구조는 정상적이다. 다른 인물이나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "내민 손과 팔은 손목·팔꿈치·어깨로 자연스럽게 연결되어 있으며 팔 근육으로 들어 올린 삿대질 자세가 가능하다. 하체와 발은 프레임 밖이어서 바닥 접촉은 확인할 수 없지만 공중에 뜬 몸으로 보이지 않는다. 책상 소품은 상판에, 검은 탁상 기기는 받침에 지지되어 있다."
       },
       {
        "label": "A",
        "direction": "남성의 검지는 왼쪽 전경 여성의 얼굴 쪽을 가리키며 시선도 그 여성을 향한다. 여성 역시 남성 쪽으로 얼굴을 돌리고 있어 대치 방향 자체는 명확하다.",
        "built_space": "전경의 책상 하나 위에 서류 묶음, 문서 받침, 연필꽂이 하나가 있고 남성 뒤에는 빈 사무용 의자 하나가 보인다. 배경은 유리 칸막이와 어두운 창, 목재 프레임이 크게 차지하며 오른쪽 끝에 피아노 일부가 보인다. 이전 장면의 회색 패널 중심 공간보다 연속성이 약하다. 창에는 남성의 등처럼 보이는 반사상이 있으나 이 시점에서 불가능한 반사라고 단정할 근거는 없다.",
        "entities": "흰 긴팔 상의와 검은 머리의 동아시아계 남성이 보이지만, 얼굴은 참고의 18세 인상보다 나이 들어 보인다. 눈 주변이 붉고 눈을 크게 뜬 격앙된 표정이다. 왼쪽 전경에는 검은 옷을 입은 단발 여성의 얼굴과 상체가 추가되어, 이전 장면의 다른 인물을 이 장면에 포함하지 말라는 지시와 충돌한다. 서류의 글자 형태는 보이지만 내용을 판독하기는 어렵다.",
        "hard_violations": [
         "이현우만 보여야 하는 장면에 이전 장면의 여성으로 보이는 인물의 얼굴과 상체를 추가했다."
        ],
        "physics": "남성은 상체를 앞으로 숙이고 한쪽 팔을 뻗으며 반대 손은 주먹을 쥐고 있다. 손과 팔의 연결 및 몸의 기울기는 실제로 가능한 동작이고, 하체가 잘려 발 지지는 확인할 수 없다. 여성의 앉거나 선 상태도 크롭 때문에 확정할 수 없다. 서류와 연필꽂이는 책상에 놓여 있으며 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.45,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.2,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지정된 장소와 완전히 다른 배경(유리창과 도시 야경) 생성",
     "[gemini-pro] 해부학적으로 불가능하게 변형된 눈(비정상적으로 붉게 칠해진 흰자)",
     "[gpt-high] 이현우만 보여야 하는 장면에 이전 장면의 여성으로 보이는 인물의 얼굴과 상체를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 200,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 200,
    "verdict_ko": "레퍼런스와 전혀 다른 유리창 배경을 생성하여 공간 고정 지시를 어겼고, 눈 흰자를 인위적으로 붉게 변형하여 해부학적 제약을 위반했습니다.  ★위반: [gemini-pro] 지정된 장소와 완전히 다른 배경(유리창과 도시 야경) 생성 / [gemini-pro] 해부학적으로 불가능하게 변형된 눈(비정상적으로 붉게 칠해진 흰자) / [gpt-high] 이현우만 보여야 하는 장면에 이전 장면의 여성으로 보이는 인물의 얼굴과 상체를 추가했다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "클로즈업 앵글과 뻗은 팔 아래의 책상 모서리 배치를 정확히 구현했으며, 지정된 실내 공간과 인물의 격앙된 표정까지 완벽하게 재현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S86sh2_sel.png",
    "asset_id": "8a3860d5-2ed5-4320-b641-58484ecfacf3",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf53-2d12-7bd3-a900-eb6bfca51f42",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S86sh2"
  }
 },
 "era_fail::faa38306b054fd1a": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S88sh7": {
  "input_fingerprint": "d14659eec1dd4ee1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 차가운 스테인리스 침대 위에 거대한 금속 몸통이 단단히 결박된 찰리의 전신.\n\nLOCATION (lock): On the restraint-equipped stainless-steel bed inside the main laboratory's glass enclosure, under clinical work lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed (Supporting 찰리's restrained body) — The top and one long side are visible along a diagonal away from the camera; used as Establishes bodily scale and the physical geometry of confinement; Restraints (Fastened around 찰리) — Their secured spans remain legible across the visible body; used as Makes his enforced immobility readable within the full-body composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination separates the metal body, restraints, and stainless-steel bed without introducing a specific light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is restrained on the stainless-steel laboratory bed, with experiment sensors attached; his chest-ring malfunction has not been shown repaired. A stun gun is concealed in a laboratory drawer.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 차가운 스테인리스 침대 위에 거대한 금속 몸통이 단단히 결박된 찰리의 전신.\n\nLOCATION (lock): On the restraint-equipped stainless-steel bed inside the main laboratory's glass enclosure, under clinical work lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed (Supporting 찰리's restrained body) — The top and one long side are visible along a diagonal away from the camera; used as Establishes bodily scale and the physical geometry of confinement; Restraints (Fastened around 찰리) — Their secured spans remain legible across the visible body; used as Makes his enforced immobility readable within the full-body composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination separates the metal body, restraints, and stainless-steel bed without introducing a specific light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is restrained on the stainless-steel laboratory bed, with experiment sensors attached; his chest-ring malfunction has not been shown repaired. A stun gun is concealed in a laboratory drawer.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 차가운 스테인리스 침대 위에 거대한 금속 몸통이 단단히 결박된 찰리의 전신.\n\nLOCATION (lock): On the restraint-equipped stainless-steel bed inside the main laboratory's glass enclosure, under clinical work lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed (Supporting 찰리's restrained body) — The top and one long side are visible along a diagonal away from the camera; used as Establishes bodily scale and the physical geometry of confinement; Restraints (Fastened around 찰리) — Their secured spans remain legible across the visible body; used as Makes his enforced immobility readable within the full-body composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination separates the metal body, restraints, and stainless-steel bed without introducing a specific light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is restrained on the stainless-steel laboratory bed, with experiment sensors attached; his chest-ring malfunction has not been shown repaired. A stun gun is concealed in a laboratory drawer.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 침대의 상단과 긴 측면을 대각선 구도로 비추며, 찰리의 고개는 우측으로 돌아가 있음.",
    "built_space": "유리로 둘러싸인 실험실 내부로 둥근 천장 조명이 있으며, 중앙에 평평한 스테인리스 침대가 놓여 있음. 레퍼런스의 우측 돔형 기계들은 보이지 않음.",
    "entities": "흰색 각진 머리, 점 형태의 눈, 선 형태의 입을 가진 기계 인간(찰리)의 전신이 정확히 묘사되었으며, 여러 개의 끈으로 단단히 결박되어 있음.",
    "hard_violations": [],
    "physics": "찰리의 몸은 평평한 침대 표면에 안정적으로 누워 있으며, 결박 끈이 몸을 단단히 지지하고 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 침대의 대각선 측면을 비추며, 찰리의 고개는 우측으로 약간 돌아가 있음.",
    "built_space": "유리 벽, 천장의 원형 조명, 우측의 대형 흰색 기계 등 레퍼런스의 배경 요소가 매우 정확하게 재현되었으나, 중앙의 침대가 상단으로 기울어진 형태로 변형됨.",
    "entities": "찰리의 형태적 특징(머리, 금속 몸통)과 전신 결박 상태는 잘 나타나 있으나, 왼쪽 손목 부근의 케이블과 손의 형태가 다소 뭉개져 있음.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에서 수평으로 평평하게 누워 움직일 수 없는 캐릭터의 자세(orientation)가 고정(LOCKED)되어야 함에도, 침대와 몸 전체가 약 30도 각도로 기울어진 채 렌더링됨."
    ],
    "physics": "기울어진 침대 위에 몸이 놓여 있으나, 레퍼런스의 물리적 수평 상태가 강제로 변경됨. 지지 상태는 끈에 의해 유지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 와이드 샷과 대각선 구도를 잘 구현했으며, 레퍼런스에 설정된 평평한 스테인리스 침대와 수평으로 누운 캐릭터의 자세를 안정적으로 유지했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "배경 구조물은 레퍼런스와 정확히 일치하지만, 수평으로 고정되어야 할 캐릭터의 몸과 침대 전체를 임의로 기울어지게 렌더링하여 고정된 자세 지침을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 침대의 상단과 긴 측면을 대각선 구도로 비추며, 찰리의 고개는 우측으로 돌아가 있음.",
        "built_space": "유리로 둘러싸인 실험실 내부로 둥근 천장 조명이 있으며, 중앙에 평평한 스테인리스 침대가 놓여 있음. 레퍼런스의 우측 돔형 기계들은 보이지 않음.",
        "entities": "흰색 각진 머리, 점 형태의 눈, 선 형태의 입을 가진 기계 인간(찰리)의 전신이 정확히 묘사되었으며, 여러 개의 끈으로 단단히 결박되어 있음.",
        "hard_violations": [],
        "physics": "찰리의 몸은 평평한 침대 표면에 안정적으로 누워 있으며, 결박 끈이 몸을 단단히 지지하고 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 침대의 대각선 측면을 비추며, 찰리의 고개는 우측으로 약간 돌아가 있음.",
        "built_space": "유리 벽, 천장의 원형 조명, 우측의 대형 흰색 기계 등 레퍼런스의 배경 요소가 매우 정확하게 재현되었으나, 중앙의 침대가 상단으로 기울어진 형태로 변형됨.",
        "entities": "찰리의 형태적 특징(머리, 금속 몸통)과 전신 결박 상태는 잘 나타나 있으나, 왼쪽 손목 부근의 케이블과 손의 형태가 다소 뭉개져 있음.",
        "hard_violations": [
         "레퍼런스에서 수평으로 평평하게 누워 움직일 수 없는 캐릭터의 자세(orientation)가 고정(LOCKED)되어야 함에도, 침대와 몸 전체가 약 30도 각도로 기울어진 채 렌더링됨."
        ],
        "physics": "기울어진 침대 위에 몸이 놓여 있으나, 레퍼런스의 물리적 수평 상태가 강제로 변경됨. 지지 상태는 끈에 의해 유지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 와이드 샷과 대각선 구도를 잘 구현했으며, 레퍼런스에 설정된 평평한 스테인리스 침대와 수평으로 누운 캐릭터의 자세를 안정적으로 유지했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "배경 구조물은 레퍼런스와 정확히 일치하지만, 수평으로 고정되어야 할 캐릭터의 몸과 침대 전체를 임의로 기울어지게 렌더링하여 고정된 자세 지침을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 침대의 상단과 긴 측면을 대각선 구도로 비추며, 찰리의 고개는 우측으로 돌아가 있음.",
        "built_space": "유리로 둘러싸인 실험실 내부로 둥근 천장 조명이 있으며, 중앙에 평평한 스테인리스 침대가 놓여 있음. 레퍼런스의 우측 돔형 기계들은 보이지 않음.",
        "entities": "흰색 각진 머리, 점 형태의 눈, 선 형태의 입을 가진 기계 인간(찰리)의 전신이 정확히 묘사되었으며, 여러 개의 끈으로 단단히 결박되어 있음.",
        "hard_violations": [],
        "physics": "찰리의 몸은 평평한 침대 표면에 안정적으로 누워 있으며, 결박 끈이 몸을 단단히 지지하고 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 침대의 대각선 측면을 비추며, 찰리의 고개는 우측으로 약간 돌아가 있음.",
        "built_space": "유리 벽, 천장의 원형 조명, 우측의 대형 흰색 기계 등 레퍼런스의 배경 요소가 매우 정확하게 재현되었으나, 중앙의 침대가 상단으로 기울어진 형태로 변형됨.",
        "entities": "찰리의 형태적 특징(머리, 금속 몸통)과 전신 결박 상태는 잘 나타나 있으나, 왼쪽 손목 부근의 케이블과 손의 형태가 다소 뭉개져 있음.",
        "hard_violations": [
         "레퍼런스에서 수평으로 평평하게 누워 움직일 수 없는 캐릭터의 자세(orientation)가 고정(LOCKED)되어야 함에도, 침대와 몸 전체가 약 30도 각도로 기울어진 채 렌더링됨."
        ],
        "physics": "기울어진 침대 위에 몸이 놓여 있으나, 레퍼런스의 물리적 수평 상태가 강제로 변경됨. 지지 상태는 끈에 의해 유지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "전신과 침대의 대각선 상판·긴 측면을 보여주며, 가슴을 가로지르는 결박까지 선명해 거대한 금속 몸통의 강제적 부동 상태를 더 충실히 구현한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "전신 와이드 구도와 침대의 지지 관계는 맞지만, 결박이 팔과 하체에 집중되고 가슴을 고정하는 띠가 없어 몸통이 단단히 결박되었다는 핵심 표현이 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리는 화면 왼쪽 뒤쪽, 발은 오른쪽 앞쪽에 놓여 있다. 얼굴과 점 형태의 눈은 비스듬히 위쪽을 향하며 특정 사람이나 물체를 바라보지는 않는다. 침대와 몸의 장축이 카메라에서 멀어지는 대각선을 이룬다. 조준하거나 이동하는 대상은 없다.",
        "built_space": "유리 격리실 안에 스테인리스 침대 한 대가 있고, 상판과 카메라 쪽 긴 측면이 모두 보인다. 머리 쪽 판은 기울어져 있으며 아래에는 중앙 받침대가 보인다. 천장의 고리형 조명과 바닥의 원호형 조명, 유리의 수직 프레임, 바깥쪽 실험 장비들이 참조의 공간 구성을 이어간다. 오른쪽 뒤에는 서랍장 한 개가 있다. 몸통·팔·허벅지·하퇴를 가로지르는 여러 결박띠와 침대 쪽 고정점이 보이며, 인물이나 침대가 중복되지는 않는다. 유리 반사에 명백한 광학적 모순은 없다.",
        "entities": "찰리 한 개체의 머리부터 양발까지 보인다. 각진 흰 얼굴판, 작은 푸른 점 형태의 눈, 짧은 선 형태의 입, 노출된 목 기구와 크고 마모된 금속 흉갑이 참조와 부합한다. 인간의 나이·성별·민족적 외양을 판별할 피부는 없으며, 명시된 기계형 찰리로 표현되어 있다. 흉부 연결부와 몸 옆으로 내려오는 실험 케이블이 보인다. 흉부 고리가 수리되었다고 확인할 표현은 없다. 숨겨진 전기충격기는 드러나지 않고 읽을 수 있는 글자도 없다.",
        "hard_violations": [],
        "physics": "등과 골반, 팔, 다리는 침대 상판에 받쳐져 있고 머리는 기계 목에 연결되어 있다. 가슴을 지나는 띠와 팔다리의 띠는 몸을 감싸 침대의 고정 부위로 이어진다. 케이블은 몸과 침대 가장자리에서 아래로 늘어져 중력에 맞게 처진다. 침대는 하부 받침대에 지지되며, 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "머리는 화면 왼쪽 뒤쪽, 발은 오른쪽 앞쪽이며 얼굴은 위쪽과 발 방향 사이를 향한다. 눈이 겨냥하는 별도 대상은 없다. 몸과 침대는 대각선으로 놓여 있고 움직임이나 무기 조준은 없다.",
        "built_space": "유리 격리실 안의 침대 한 대를 비스듬히 내려다보며 상판과 긴 측면을 볼 수 있다. 머리 쪽 경사판, 가까운 쪽 금속 난간 한 줄, 중앙 받침대와 하부 바퀴가 드러난다. 뒤쪽에는 대형 실험 장비들이 있고 오른쪽 작업대에는 작업등 두 개가 보인다. 유리 수직 프레임과 천장·바닥의 원호형 조명은 참조 공간과 통한다. 팔, 골반·허벅지, 발목 부근의 결박은 보이지만 상부 흉갑을 가로지르는 결박은 없다. 명백한 중복 침대나 불가능한 반사는 없다.",
        "entities": "찰리 한 개체의 전신이 보이며 흰 각진 얼굴판, 작은 푸른 눈, 선 형태의 입과 낡은 금속 몸통은 참조의 정체성을 유지한다. 인간의 민족·나이·성별을 판별할 외양은 노출되지 않는다. 흉부에는 작은 열린 연결부가 있고 몸 옆에는 케이블이 보이나, 센서의 개별 부착점은 뚜렷하지 않다. 가슴 고리의 수리가 완료되었다는 증거는 없다. 전기충격기는 노출되지 않으며 판독 가능한 문구도 없다.",
        "hard_violations": [],
        "physics": "상체는 경사진 침대 부분에 기대고 골반과 다리, 팔은 상판에 놓여 있다. 머리는 목 관절로 지지된다. 팔과 하체의 띠는 침대 가장자리 아래로 이어지며 남는 띠 끝은 아래로 처져 있다. 침대는 중앙 받침과 바닥 쪽 하부 구조에 지지된다. 부유하는 신체나 물체는 없지만 흉부 자체를 눌러 고정하는 결박은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "전신과 침대의 대각선 상판·긴 측면을 보여주며, 가슴을 가로지르는 결박까지 선명해 거대한 금속 몸통의 강제적 부동 상태를 더 충실히 구현한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "전신 와이드 구도와 침대의 지지 관계는 맞지만, 결박이 팔과 하체에 집중되고 가슴을 고정하는 띠가 없어 몸통이 단단히 결박되었다는 핵심 표현이 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "머리는 화면 왼쪽 뒤쪽, 발은 오른쪽 앞쪽에 놓여 있다. 얼굴과 점 형태의 눈은 비스듬히 위쪽을 향하며 특정 사람이나 물체를 바라보지는 않는다. 침대와 몸의 장축이 카메라에서 멀어지는 대각선을 이룬다. 조준하거나 이동하는 대상은 없다.",
        "built_space": "유리 격리실 안에 스테인리스 침대 한 대가 있고, 상판과 카메라 쪽 긴 측면이 모두 보인다. 머리 쪽 판은 기울어져 있으며 아래에는 중앙 받침대가 보인다. 천장의 고리형 조명과 바닥의 원호형 조명, 유리의 수직 프레임, 바깥쪽 실험 장비들이 참조의 공간 구성을 이어간다. 오른쪽 뒤에는 서랍장 한 개가 있다. 몸통·팔·허벅지·하퇴를 가로지르는 여러 결박띠와 침대 쪽 고정점이 보이며, 인물이나 침대가 중복되지는 않는다. 유리 반사에 명백한 광학적 모순은 없다.",
        "entities": "찰리 한 개체의 머리부터 양발까지 보인다. 각진 흰 얼굴판, 작은 푸른 점 형태의 눈, 짧은 선 형태의 입, 노출된 목 기구와 크고 마모된 금속 흉갑이 참조와 부합한다. 인간의 나이·성별·민족적 외양을 판별할 피부는 없으며, 명시된 기계형 찰리로 표현되어 있다. 흉부 연결부와 몸 옆으로 내려오는 실험 케이블이 보인다. 흉부 고리가 수리되었다고 확인할 표현은 없다. 숨겨진 전기충격기는 드러나지 않고 읽을 수 있는 글자도 없다.",
        "hard_violations": [],
        "physics": "등과 골반, 팔, 다리는 침대 상판에 받쳐져 있고 머리는 기계 목에 연결되어 있다. 가슴을 지나는 띠와 팔다리의 띠는 몸을 감싸 침대의 고정 부위로 이어진다. 케이블은 몸과 침대 가장자리에서 아래로 늘어져 중력에 맞게 처진다. 침대는 하부 받침대에 지지되며, 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "머리는 화면 왼쪽 뒤쪽, 발은 오른쪽 앞쪽이며 얼굴은 위쪽과 발 방향 사이를 향한다. 눈이 겨냥하는 별도 대상은 없다. 몸과 침대는 대각선으로 놓여 있고 움직임이나 무기 조준은 없다.",
        "built_space": "유리 격리실 안의 침대 한 대를 비스듬히 내려다보며 상판과 긴 측면을 볼 수 있다. 머리 쪽 경사판, 가까운 쪽 금속 난간 한 줄, 중앙 받침대와 하부 바퀴가 드러난다. 뒤쪽에는 대형 실험 장비들이 있고 오른쪽 작업대에는 작업등 두 개가 보인다. 유리 수직 프레임과 천장·바닥의 원호형 조명은 참조 공간과 통한다. 팔, 골반·허벅지, 발목 부근의 결박은 보이지만 상부 흉갑을 가로지르는 결박은 없다. 명백한 중복 침대나 불가능한 반사는 없다.",
        "entities": "찰리 한 개체의 전신이 보이며 흰 각진 얼굴판, 작은 푸른 눈, 선 형태의 입과 낡은 금속 몸통은 참조의 정체성을 유지한다. 인간의 민족·나이·성별을 판별할 외양은 노출되지 않는다. 흉부에는 작은 열린 연결부가 있고 몸 옆에는 케이블이 보이나, 센서의 개별 부착점은 뚜렷하지 않다. 가슴 고리의 수리가 완료되었다는 증거는 없다. 전기충격기는 노출되지 않으며 판독 가능한 문구도 없다.",
        "hard_violations": [],
        "physics": "상체는 경사진 침대 부분에 기대고 골반과 다리, 팔은 상판에 놓여 있다. 머리는 목 관절로 지지된다. 팔과 하체의 띠는 침대 가장자리 아래로 이어지며 남는 띠 끝은 아래로 처져 있다. 침대는 중앙 받침과 바닥 쪽 하부 구조에 지지된다. 부유하는 신체나 물체는 없지만 흉부 자체를 눌러 고정하는 결박은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.889,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.889,
    "B": 1.321
   },
   "violations": {
    "B": [
     "[gemini-pro] 레퍼런스에서 수평으로 평평하게 누워 움직일 수 없는 캐릭터의 자세(orientation)가 고정(LOCKED)되어야 함에도, 침대와 몸 전체가 약 30도 각도로 기울어진 채 렌더링됨."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1889,
   "B": 1321
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1889,
    "verdict_ko": "지시된 와이드 샷과 대각선 구도를 잘 구현했으며, 레퍼런스에 설정된 평평한 스테인리스 침대와 수평으로 누운 캐릭터의 자세를 안정적으로 유지했습니다."
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "배경 구조물은 레퍼런스와 정확히 일치하지만, 수평으로 고정되어야 할 캐릭터의 몸과 침대 전체를 임의로 기울어지게 렌더링하여 고정된 자세 지침을 위반했습니다.  ★위반: [gemini-pro] 레퍼런스에서 수평으로 평평하게 누워 움직일 수 없는 캐릭터의 자세(orientation)가 고정(LOCKED)되어야 함에도, 침대와 몸 전체가 약 30도 각도로 기울어진 채 렌더링됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S84sh1_sel.png",
    "asset_id": "aab7da1e-cd8c-4a16-a151-803530ebbe82",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf66-9a84-70bf-a473-69fa54f2df95",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S84sh1"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S88sh37": {
  "input_fingerprint": "c5a0bcb8fc52dc83",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 굵은 금속 손가락 끝이 이현우의 눈물 젖은 뺨에 부드럽게 닿아 있는 손끝 클로즈업.\n\nLOCATION (lock): At the bedside inside the main laboratory's glass enclosure, beneath clinical lighting as the robot reaches toward its visitor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed edge (Partially visible beneath the touching hand) — A narrow portion of the same long side recedes behind the contact; used as Provides a subdued spatial reference without competing with the cheek.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the bedside ambient illumination, using restrained highlight contrast to distinguish tears, skin, and the gently touching metal fingertip.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the stainless-steel bed, glass enclosure, laboratory surfaces, and surrounding equipment. EXCLUDE restraints locking the hand used for contact, since that hand has been released.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie remains reclining on the stainless-steel bed after his hand restraints are released, with one freed hand raised so that a fingertip touches Hyunwoo's tear-wet cheek. His exact head angle, other hand position, leg positions, and any remaining body restraints are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains on the stainless-steel bed with sensors attached; his hands have been released, but the rest of the restraints have not yet been removed. 이현우: He stands beside the bed in white clothing, crying heavily with tears covering his cheeks; the stand he picked up has not been shown put down.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 굵은 금속 손가락 끝이 이현우의 눈물 젖은 뺨에 부드럽게 닿아 있는 손끝 클로즈업.\n\nLOCATION (lock): At the bedside inside the main laboratory's glass enclosure, beneath clinical lighting as the robot reaches toward its visitor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed edge (Partially visible beneath the touching hand) — A narrow portion of the same long side recedes behind the contact; used as Provides a subdued spatial reference without competing with the cheek.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the bedside ambient illumination, using restrained highlight contrast to distinguish tears, skin, and the gently touching metal fingertip.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the stainless-steel bed, glass enclosure, laboratory surfaces, and surrounding equipment. EXCLUDE restraints locking the hand used for contact, since that hand has been released.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie remains reclining on the stainless-steel bed after his hand restraints are released, with one freed hand raised so that a fingertip touches Hyunwoo's tear-wet cheek. His exact head angle, other hand position, leg positions, and any remaining body restraints are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains on the stainless-steel bed with sensors attached; his hands have been released, but the rest of the restraints have not yet been removed. 이현우: He stands beside the bed in white clothing, crying heavily with tears covering his cheeks; the stand he picked up has not been shown put down.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 굵은 금속 손가락 끝이 이현우의 눈물 젖은 뺨에 부드럽게 닿아 있는 손끝 클로즈업.\n\nLOCATION (lock): At the bedside inside the main laboratory's glass enclosure, beneath clinical lighting as the robot reaches toward its visitor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed edge (Partially visible beneath the touching hand) — A narrow portion of the same long side recedes behind the contact; used as Provides a subdued spatial reference without competing with the cheek.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the bedside ambient illumination, using restrained highlight contrast to distinguish tears, skin, and the gently touching metal fingertip.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the stainless-steel bed, glass enclosure, laboratory surfaces, and surrounding equipment. EXCLUDE restraints locking the hand used for contact, since that hand has been released.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie remains reclining on the stainless-steel bed after his hand restraints are released, with one freed hand raised so that a fingertip touches Hyunwoo's tear-wet cheek. His exact head angle, other hand position, leg positions, and any remaining body restraints are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains on the stainless-steel bed with sensors attached; his hands have been released, but the rest of the restraints have not yet been removed. 이현우: He stands beside the bed in white clothing, crying heavily with tears covering his cheeks; the stand he picked up has not been shown put down.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "로봇의 손가락이 텍스트대로 이현우의 뺨을 정확히 짚고 있음.",
    "built_space": "연구실 배경과 금속 침대 모서리가 보이나, 로봇의 몸통이 침대 밖에 수직으로 배치됨.",
    "entities": "찰리의 금속 외장과 이현우의 외형(눈물, 흰옷)이 참조 이미지와 잘 일치함.",
    "hard_violations": [
     "[gemini-pro] 찰리가 침대에 누워 있어야 하는 고정 설정(Canonical pose)을 위반하고 프레임 앞쪽에 수직으로 서 있음."
    ],
    "physics": "로봇의 팔은 서 있는 몸통에 연결되어 공중에 떠 있고, 이현우는 서서 얼굴을 맞대고 있음."
   },
   {
    "label": "B",
    "direction": "로봇의 손가락이 이현우의 뺨이 아닌 인중과 윗입술 부위를 찌르고 있음.",
    "built_space": "금속 침대 가장자리가 팔 아래에 놓여 뒤로 멀어지는 지정된 공간 구도를 정확하게 따름.",
    "entities": "찰리의 팔 질감과 이현우의 특징(눈물, 흰옷, 18세 외모)이 모두 참조와 일치함.",
    "hard_violations": [
     "[gpt-high] 침대 옆에 서 있어야 하는 현우가 무릎을 굽혀 침대 높이로 몸을 낮춘 자세로 배치되어 있다."
    ],
    "physics": "로봇 팔이 침대 모서리에 자연스럽게 얹혀 있어 누운 자세를 뒷받침하며, 이현우는 웅크려 지지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "지정된 침대 구도와 누운 자세의 배치는 훌륭히 구현했으나, 손가락이 뺨이 아닌 인중을 향해 아쉽습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "뺨을 짚는 동작 자체는 정확하나, 침대에 누워있어야 할 로봇이 침대 밖에서 수직으로 서 있는 치명적 자세 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 손가락이 텍스트대로 이현우의 뺨을 정확히 짚고 있음.",
        "built_space": "연구실 배경과 금속 침대 모서리가 보이나, 로봇의 몸통이 침대 밖에 수직으로 배치됨.",
        "entities": "찰리의 금속 외장과 이현우의 외형(눈물, 흰옷)이 참조 이미지와 잘 일치함.",
        "hard_violations": [
         "찰리가 침대에 누워 있어야 하는 고정 설정(Canonical pose)을 위반하고 프레임 앞쪽에 수직으로 서 있음."
        ],
        "physics": "로봇의 팔은 서 있는 몸통에 연결되어 공중에 떠 있고, 이현우는 서서 얼굴을 맞대고 있음."
       },
       {
        "label": "B",
        "direction": "로봇의 손가락이 이현우의 뺨이 아닌 인중과 윗입술 부위를 찌르고 있음.",
        "built_space": "금속 침대 가장자리가 팔 아래에 놓여 뒤로 멀어지는 지정된 공간 구도를 정확하게 따름.",
        "entities": "찰리의 팔 질감과 이현우의 특징(눈물, 흰옷, 18세 외모)이 모두 참조와 일치함.",
        "hard_violations": [],
        "physics": "로봇 팔이 침대 모서리에 자연스럽게 얹혀 있어 누운 자세를 뒷받침하며, 이현우는 웅크려 지지하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "지정된 침대 구도와 누운 자세의 배치는 훌륭히 구현했으나, 손가락이 뺨이 아닌 인중을 향해 아쉽습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "뺨을 짚는 동작 자체는 정확하나, 침대에 누워있어야 할 로봇이 침대 밖에서 수직으로 서 있는 치명적 자세 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 손가락이 텍스트대로 이현우의 뺨을 정확히 짚고 있음.",
        "built_space": "연구실 배경과 금속 침대 모서리가 보이나, 로봇의 몸통이 침대 밖에 수직으로 배치됨.",
        "entities": "찰리의 금속 외장과 이현우의 외형(눈물, 흰옷)이 참조 이미지와 잘 일치함.",
        "hard_violations": [
         "찰리가 침대에 누워 있어야 하는 고정 설정(Canonical pose)을 위반하고 프레임 앞쪽에 수직으로 서 있음."
        ],
        "physics": "로봇의 팔은 서 있는 몸통에 연결되어 공중에 떠 있고, 이현우는 서서 얼굴을 맞대고 있음."
       },
       {
        "label": "B",
        "direction": "로봇의 손가락이 이현우의 뺨이 아닌 인중과 윗입술 부위를 찌르고 있음.",
        "built_space": "금속 침대 가장자리가 팔 아래에 놓여 뒤로 멀어지는 지정된 공간 구도를 정확하게 따름.",
        "entities": "찰리의 팔 질감과 이현우의 특징(눈물, 흰옷, 18세 외모)이 모두 참조와 일치함.",
        "hard_violations": [],
        "physics": "로봇 팔이 침대 모서리에 자연스럽게 얹혀 있어 누운 자세를 뒷받침하며, 이현우는 웅크려 지지하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "접촉과 눈물은 보이지만 현우가 침대 옆에 서 있지 않고 몸을 낮춘 자세이며, 넓은 침대 상판과 상반신이 손끝 중심 구도를 방해한다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "들어 올린 금속 손끝의 뺨 접촉, 현우의 자세, 유리 실험실의 연속성이 더 정확하지만 손끝 클로즈업보다 얼굴·상반신 중심으로 넓게 잡혔다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽에서 수평으로 뻗은 금속 검지 끝이 현우의 코 아래, 입가에 가까운 뺨에 닿는다. 접촉은 보이지만 눈물 흐르는 뺨 중앙보다 입술 가까이에 치우쳤다. 현우는 카메라가 아니라 왼쪽 아래의 로봇 쪽을 바라본다.",
        "built_space": "스테인리스 침대 한 개의 넓은 상판과 긴 측면 테두리, 그 아래 평행한 측면 레일이 보인다. 뒤에는 흐릿한 실험 장비와 유리 칸막이가 있다. 요구된 좁은 침대 가장자리보다 상판이 훨씬 크게 드러난다. 현우는 침대 높이까지 몸을 낮추었고 화면 아래에 굽힌 무릎이 보여, 서 있는 방문객의 배치와 다르다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "젊은 한국계 남성으로 읽히는 현우의 검은 헝클어진 머리, 흰 상의, 붉어진 정상적인 눈과 젖은 뺨이 보인다. 찰리는 굵고 마모된 금속 손과 팔로만 나타나며 재질과 체격은 참조에 부합한다. 얼굴, 센서, 나머지 구속구와 현우가 들던 스탠드는 프레임 밖이라 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "침대 옆에 서 있어야 하는 현우가 무릎을 굽혀 침대 높이로 몸을 낮춘 자세로 배치되어 있다."
        ],
        "physics": "금속 손은 손목과 왼쪽 프레임 밖으로 이어지는 전완에 연결되어 있고 검지 끝이 피부에 실제로 맞닿는다. 나머지 손가락은 아래로 굽혀져 있으며 지지 없이 떠 있는 별도 물체는 없다. 현우의 자세는 물리적으로 가능하지만 서 있다는 지정과 맞지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 아래에서 올라온 금속 검지가 현우의 코 옆 위쪽 뺨을 향하며 끝부분이 피부에 닿아 있다. 현우의 시선은 왼쪽의 찰리 쪽으로 향한다. 손끝과 뺨의 접촉 관계가 명확하다.",
        "built_space": "침대 한 개의 상판 일부와 테두리가 손 아래로 보이고, 뒤에는 유리 enclosure의 세로 프레임, 창, 선형 임상 조명과 흐릿한 장비가 이어진다. 이전 장면의 차가운 실험실 공간과 잘 연결된다. 현우의 상체는 침대 옆에서 곧게 올라와 서 있는 배치와 양립한다. 다만 침대는 요구된 좁은 측면 조각보다 넓게 노출된다. 설비 중복이나 광학적으로 불가능한 반사는 보이지 않는다.",
        "entities": "현우는 참조와 유사한 젊은 한국계 남성 얼굴, 검은 헝클어진 머리와 흰 상의를 갖추고 있으며 양쪽 뺨에 눈물 자국과 물방울이 보인다. 찰리의 두꺼운 금속 팔과 관절 손에는 마모와 실제 금속 반사가 있다. 찰리의 머리, 센서, 남은 구속구 및 현우의 스탠드는 이 구도에서 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "왼쪽 아래의 로봇 팔 관절에서 전완, 손목, 손가락까지 연결이 보이며, 관절을 굽혀 손끝을 들어 올린 동작으로 읽힌다. 이는 해방된 한 손을 올린다는 지정 자세와 맞는다. 손끝은 뺨 표면에 닿고 다른 손가락은 접혀 있어 부드러운 접촉이 가능하다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "접촉과 눈물은 보이지만 현우가 침대 옆에 서 있지 않고 몸을 낮춘 자세이며, 넓은 침대 상판과 상반신이 손끝 중심 구도를 방해한다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "들어 올린 금속 손끝의 뺨 접촉, 현우의 자세, 유리 실험실의 연속성이 더 정확하지만 손끝 클로즈업보다 얼굴·상반신 중심으로 넓게 잡혔다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽에서 수평으로 뻗은 금속 검지 끝이 현우의 코 아래, 입가에 가까운 뺨에 닿는다. 접촉은 보이지만 눈물 흐르는 뺨 중앙보다 입술 가까이에 치우쳤다. 현우는 카메라가 아니라 왼쪽 아래의 로봇 쪽을 바라본다.",
        "built_space": "스테인리스 침대 한 개의 넓은 상판과 긴 측면 테두리, 그 아래 평행한 측면 레일이 보인다. 뒤에는 흐릿한 실험 장비와 유리 칸막이가 있다. 요구된 좁은 침대 가장자리보다 상판이 훨씬 크게 드러난다. 현우는 침대 높이까지 몸을 낮추었고 화면 아래에 굽힌 무릎이 보여, 서 있는 방문객의 배치와 다르다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "젊은 한국계 남성으로 읽히는 현우의 검은 헝클어진 머리, 흰 상의, 붉어진 정상적인 눈과 젖은 뺨이 보인다. 찰리는 굵고 마모된 금속 손과 팔로만 나타나며 재질과 체격은 참조에 부합한다. 얼굴, 센서, 나머지 구속구와 현우가 들던 스탠드는 프레임 밖이라 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "침대 옆에 서 있어야 하는 현우가 무릎을 굽혀 침대 높이로 몸을 낮춘 자세로 배치되어 있다."
        ],
        "physics": "금속 손은 손목과 왼쪽 프레임 밖으로 이어지는 전완에 연결되어 있고 검지 끝이 피부에 실제로 맞닿는다. 나머지 손가락은 아래로 굽혀져 있으며 지지 없이 떠 있는 별도 물체는 없다. 현우의 자세는 물리적으로 가능하지만 서 있다는 지정과 맞지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 아래에서 올라온 금속 검지가 현우의 코 옆 위쪽 뺨을 향하며 끝부분이 피부에 닿아 있다. 현우의 시선은 왼쪽의 찰리 쪽으로 향한다. 손끝과 뺨의 접촉 관계가 명확하다.",
        "built_space": "침대 한 개의 상판 일부와 테두리가 손 아래로 보이고, 뒤에는 유리 enclosure의 세로 프레임, 창, 선형 임상 조명과 흐릿한 장비가 이어진다. 이전 장면의 차가운 실험실 공간과 잘 연결된다. 현우의 상체는 침대 옆에서 곧게 올라와 서 있는 배치와 양립한다. 다만 침대는 요구된 좁은 측면 조각보다 넓게 노출된다. 설비 중복이나 광학적으로 불가능한 반사는 보이지 않는다.",
        "entities": "현우는 참조와 유사한 젊은 한국계 남성 얼굴, 검은 헝클어진 머리와 흰 상의를 갖추고 있으며 양쪽 뺨에 눈물 자국과 물방울이 보인다. 찰리의 두꺼운 금속 팔과 관절 손에는 마모와 실제 금속 반사가 있다. 찰리의 머리, 센서, 남은 구속구 및 현우의 스탠드는 이 구도에서 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "왼쪽 아래의 로봇 팔 관절에서 전완, 손목, 손가락까지 연결이 보이며, 관절을 굽혀 손끝을 들어 올린 동작으로 읽힌다. 이는 해방된 한 손을 올린다는 지정 자세와 맞는다. 손끝은 뺨 표면에 닿고 다른 손가락은 접혀 있어 부드러운 접촉이 가능하다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.35,
    "B": 1.321
   },
   "violations": {
    "A": [
     "[gemini-pro] 찰리가 침대에 누워 있어야 하는 고정 설정(Canonical pose)을 위반하고 프레임 앞쪽에 수직으로 서 있음."
    ],
    "B": [
     "[gpt-high] 침대 옆에 서 있어야 하는 현우가 무릎을 굽혀 침대 높이로 몸을 낮춘 자세로 배치되어 있다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1321,
   "A": 1350
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "지정된 침대 구도와 누운 자세의 배치는 훌륭히 구현했으나, 손가락이 뺨이 아닌 인중을 향해 아쉽습니다.  ★위반: [gpt-high] 침대 옆에 서 있어야 하는 현우가 무릎을 굽혀 침대 높이로 몸을 낮춘 자세로 배치되어 있다."
   },
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "뺨을 짚는 동작 자체는 정확하나, 침대에 누워있어야 할 로봇이 침대 밖에서 수직으로 서 있는 치명적 자세 위반이 발생했습니다.  ★위반: [gemini-pro] 찰리가 침대에 누워 있어야 하는 고정 설정(Canonical pose)을 위반하고 프레임 앞쪽에 수직으로 서 있음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S88sh7_sel.png",
    "asset_id": "96e1d3ba-947f-4da0-9d38-283029728d14",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1597480>",
    "asset_id": "26fbe654-f04e-4351-bcfb-03323163514e",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf6b-942d-7117-ac66-d9444c4dc2aa",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S88sh7"
  }
 },
 "S88sh41": {
  "input_fingerprint": "03ddb9d18e031061",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 실험실 천장의 붉은 경광등이 눈부신 섬광을 내뿜는 최고조의 밝기에서 멈춘 찰나.\n\nLOCATION (lock): At the ceiling-mounted warning beacon inside the main laboratory, where an intense red alarm flash lights the room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Ceiling warning-light fixture (Mounted on the laboratory ceiling) — Its underside and oblique side are visible from the bedside; used as Terminates the upward tilt and anchors the alarm reveal; Laboratory ceiling (Visible around the warning-light fixture) — Seen obliquely from below; used as Maintains environmental context and prevents an isolated fixture insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The warning light reaches a dazzling red peak, with controlled highlight bloom preserving the housing's outline against the ceiling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laboratory's red alarm warning is active. Charlie remains on the stainless-steel bed with sensors attached, his hands released but his remaining restraints still in place.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 실험실 천장의 붉은 경광등이 눈부신 섬광을 내뿜는 최고조의 밝기에서 멈춘 찰나.\n\nLOCATION (lock): At the ceiling-mounted warning beacon inside the main laboratory, where an intense red alarm flash lights the room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Ceiling warning-light fixture (Mounted on the laboratory ceiling) — Its underside and oblique side are visible from the bedside; used as Terminates the upward tilt and anchors the alarm reveal; Laboratory ceiling (Visible around the warning-light fixture) — Seen obliquely from below; used as Maintains environmental context and prevents an isolated fixture insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The warning light reaches a dazzling red peak, with controlled highlight bloom preserving the housing's outline against the ceiling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laboratory's red alarm warning is active. Charlie remains on the stainless-steel bed with sensors attached, his hands released but his remaining restraints still in place.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 실험실 천장의 붉은 경광등이 눈부신 섬광을 내뿜는 최고조의 밝기에서 멈춘 찰나.\n\nLOCATION (lock): At the ceiling-mounted warning beacon inside the main laboratory, where an intense red alarm flash lights the room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Ceiling warning-light fixture (Mounted on the laboratory ceiling) — Its underside and oblique side are visible from the bedside; used as Terminates the upward tilt and anchors the alarm reveal; Laboratory ceiling (Visible around the warning-light fixture) — Seen obliquely from below; used as Maintains environmental context and prevents an isolated fixture insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The warning light reaches a dazzling red peak, with controlled highlight bloom preserving the housing's outline against the ceiling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laboratory's red alarm warning is active. Charlie remains on the stainless-steel bed with sensors attached, his hands released but his remaining restraints still in place.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 천장 위쪽을 향해 붉은 경광등을 똑바로 바라보고 있습니다.",
    "built_space": "콘크리트 십자 보 위에 경광등이 설치되어 있으며, 배관이 지나갑니다. 배경으로 유리 벽면이 보입니다.",
    "entities": "강렬한 붉은빛을 발산하는 천장 부착형 경광등이 명확히 묘사되었습니다.",
    "hard_violations": [],
    "physics": "경광등은 천장 보에 안정적으로 고정되어 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 로봇의 뒷모습과 천장의 8각형 조명을 향하고 있습니다.",
    "built_space": "일반적인 사무실용 패널 천장에 붉은 조명이 있으며, 창밖으로 도시 건물이 보입니다.",
    "entities": "붉은 조명과 찰리(로봇)의 뒷모습이 보이나, 창밖에 한글과 영문 간판 텍스트가 있습니다.",
    "hard_violations": [
     "[gemini-pro] 배경 창밖 건물에 읽을 수 있는 텍스트(SAMSUNG, 한글 간판) 노출",
     "[gemini-pro] 레퍼런스에서 누워있어야 할 캐릭터가 똑바로 앉아있거나 서 있는 형태의 물리적/자세 위반",
     "[gpt-high] 인물을 보여 주지 않는 경광등 숏에 찰리의 머리와 상체를 추가하여, 숏 텍스트에 없는 인물을 넣지 말라는 조건을 위반했다.",
     "[gpt-high] 왼쪽 창밖 간판에 판독 가능한 숫자가 보여, 어디에도 읽을 수 있는 문자를 두지 말라는 조건을 위반했다."
    ],
    "physics": "로봇의 머리와 몸통이 똑바로 서 있으나, 레퍼런스의 누워있는 상태와 물리적으로 모순됩니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프레임 지시(천장 경광등 미디엄 샷)를 정확히 따르며 텍스트 제한을 준수했으나, 천장 재질이 레퍼런스의 분위기와 다소 차이가 있습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배경에 읽을 수 있는 텍스트가 노출되었고, 레퍼런스에 고정된 캐릭터의 누워있는 자세가 변경되는 치명적인 규정 위반이 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 천장 위쪽을 향해 붉은 경광등을 똑바로 바라보고 있습니다.",
        "built_space": "콘크리트 십자 보 위에 경광등이 설치되어 있으며, 배관이 지나갑니다. 배경으로 유리 벽면이 보입니다.",
        "entities": "강렬한 붉은빛을 발산하는 천장 부착형 경광등이 명확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "경광등은 천장 보에 안정적으로 고정되어 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 로봇의 뒷모습과 천장의 8각형 조명을 향하고 있습니다.",
        "built_space": "일반적인 사무실용 패널 천장에 붉은 조명이 있으며, 창밖으로 도시 건물이 보입니다.",
        "entities": "붉은 조명과 찰리(로봇)의 뒷모습이 보이나, 창밖에 한글과 영문 간판 텍스트가 있습니다.",
        "hard_violations": [
         "배경 창밖 건물에 읽을 수 있는 텍스트(SAMSUNG, 한글 간판) 노출",
         "레퍼런스에서 누워있어야 할 캐릭터가 똑바로 앉아있거나 서 있는 형태의 물리적/자세 위반"
        ],
        "physics": "로봇의 머리와 몸통이 똑바로 서 있으나, 레퍼런스의 누워있는 상태와 물리적으로 모순됩니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프레임 지시(천장 경광등 미디엄 샷)를 정확히 따르며 텍스트 제한을 준수했으나, 천장 재질이 레퍼런스의 분위기와 다소 차이가 있습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배경에 읽을 수 있는 텍스트가 노출되었고, 레퍼런스에 고정된 캐릭터의 누워있는 자세가 변경되는 치명적인 규정 위반이 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 천장 위쪽을 향해 붉은 경광등을 똑바로 바라보고 있습니다.",
        "built_space": "콘크리트 십자 보 위에 경광등이 설치되어 있으며, 배관이 지나갑니다. 배경으로 유리 벽면이 보입니다.",
        "entities": "강렬한 붉은빛을 발산하는 천장 부착형 경광등이 명확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "경광등은 천장 보에 안정적으로 고정되어 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 로봇의 뒷모습과 천장의 8각형 조명을 향하고 있습니다.",
        "built_space": "일반적인 사무실용 패널 천장에 붉은 조명이 있으며, 창밖으로 도시 건물이 보입니다.",
        "entities": "붉은 조명과 찰리(로봇)의 뒷모습이 보이나, 창밖에 한글과 영문 간판 텍스트가 있습니다.",
        "hard_violations": [
         "배경 창밖 건물에 읽을 수 있는 텍스트(SAMSUNG, 한글 간판) 노출",
         "레퍼런스에서 누워있어야 할 캐릭터가 똑바로 앉아있거나 서 있는 형태의 물리적/자세 위반"
        ],
        "physics": "로봇의 머리와 몸통이 똑바로 서 있으나, 레퍼런스의 누워있는 상태와 물리적으로 모순됩니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "경광등 하나의 밑면과 측면을 천장 맥락 속에서 올려다보며, 인물 없이 붉은 섬광의 정점을 보여 주어 지정된 구도와 순간에 충실하다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "천장 경광등의 강렬한 발광은 구현했지만, 요구하지 않은 찰리를 화면에 넣고 외부 간판의 읽히는 숫자까지 노출하여 명시적 금지 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 아래에서 천장을 비스듬히 올려다본다. 중앙 경광등의 원통형 측면과 둥근 밑면이 함께 보이며 붉은 빛이 주변 보와 천장으로 퍼진다. 시선이나 무기, 이동하는 인물은 없다.",
        "built_space": "경광등은 하나이며 천장 보 교차부의 사각 장착함 아래에 설치되어 있다. 장착함으로 네 방향의 가는 배관이 연결되고, 주변에는 콘크리트 보와 굵은 배관이 보인다. 하단의 유리 칸막이, 검은 창틀, 곡선형 백색 조명은 참조 장소의 주요 특징과 이어진다. 노출 콘크리트 천장 구조 자체는 참조 사진에서 확인할 수 없어 동일성을 확정할 수 없다. 경광등만 고립시키지 않고 주변 천장과 실험실 상부를 충분히 포함한다.",
        "entities": "붉게 발광하는 물리적 경광등 하나가 있으며 금속 장착부, 골이 있는 반투명 덮개와 하부 테두리가 식별된다. 밝은 중심부에도 외곽선이 유지된다. 인물과 침대는 프레임 밖으로, 이번 경광등 중심 숏에 적절하다. 읽을 수 있는 글자는 없고 창밖의 차가운 밝기는 이른 아침과 양립한다.",
        "hard_violations": [],
        "physics": "경광등은 천장에 밀착한 장착함과 금속 연결부로 지지된다. 가는 배관에는 고정 클램프가 보이고 굵은 배관은 천장 연결 지지대로 매달려 있다. 유리 칸막이는 수직 프레임에 끼워져 있으며, 지지 없이 떠 있는 물체나 인물은 없다."
       },
       {
        "label": "A",
        "direction": "카메라는 아래에서 천장을 올려다보며 큰 다각형 경광등의 밑면과 발광 측면을 보여 준다. 하단에는 찰리로 보이는 기계 머리의 뒤쪽이 보이고 얼굴은 카메라 반대편을 향한다. 실제 눈이나 시선의 표적은 보이지 않는다. 경광등의 붉은 빛은 천장과 실내로 퍼진다.",
        "built_space": "천장 중앙에 다각형 경광등 하나가 있고, 양옆에는 길쭉한 매입 설비 열 두 줄, 원형 천장 부속 두 개, 중앙에는 직사각형 패널들이 보인다. 뒤쪽에는 격자 창과 실험실 장비가 있으며 하단 중앙에는 침대 가장자리로 보이는 금속 구조와 기계 인물의 상체가 들어온다. 참조의 유리 칸막이와 곡선 조명 대신 패널 천장이 지배적이며 장소 연속성이 약하다. 경광등도 화면 폭의 약 사분의 일을 차지해 작은 천장 설비를 과도하게 전경화한 인상이다.",
        "entities": "붉은 경광등과 실험실 설비는 보이지만, 숏 텍스트가 보여 주도록 지정하지 않은 찰리의 머리와 어깨가 추가되었다. 흰색 각진 머리 외장과 금속 어깨는 참조와 대체로 연결되나 얼굴판과 센서·구속 상태는 이 구도에서 확인되지 않는다. 왼쪽 창밖 간판에는 읽을 수 있는 숫자가 노출되어 있다.",
        "hard_violations": [
         "인물을 보여 주지 않는 경광등 숏에 찰리의 머리와 상체를 추가하여, 숏 텍스트에 없는 인물을 넣지 말라는 조건을 위반했다.",
         "왼쪽 창밖 간판에 판독 가능한 숫자가 보여, 어디에도 읽을 수 있는 문자를 두지 말라는 조건을 위반했다."
        ],
        "physics": "경광등의 상부 하우징은 천장에 직접 부착되어 있어 지지가 명확하다. 찰리의 하부는 화면 아래 금속 침대 구조에 가려져 접촉점을 확인할 수 없지만, 공중에 떠 있다고 볼 근거는 없다. 머리와 어깨에 중력상 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "경광등 하나의 밑면과 측면을 천장 맥락 속에서 올려다보며, 인물 없이 붉은 섬광의 정점을 보여 주어 지정된 구도와 순간에 충실하다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "천장 경광등의 강렬한 발광은 구현했지만, 요구하지 않은 찰리를 화면에 넣고 외부 간판의 읽히는 숫자까지 노출하여 명시적 금지 조건을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 아래에서 천장을 비스듬히 올려다본다. 중앙 경광등의 원통형 측면과 둥근 밑면이 함께 보이며 붉은 빛이 주변 보와 천장으로 퍼진다. 시선이나 무기, 이동하는 인물은 없다.",
        "built_space": "경광등은 하나이며 천장 보 교차부의 사각 장착함 아래에 설치되어 있다. 장착함으로 네 방향의 가는 배관이 연결되고, 주변에는 콘크리트 보와 굵은 배관이 보인다. 하단의 유리 칸막이, 검은 창틀, 곡선형 백색 조명은 참조 장소의 주요 특징과 이어진다. 노출 콘크리트 천장 구조 자체는 참조 사진에서 확인할 수 없어 동일성을 확정할 수 없다. 경광등만 고립시키지 않고 주변 천장과 실험실 상부를 충분히 포함한다.",
        "entities": "붉게 발광하는 물리적 경광등 하나가 있으며 금속 장착부, 골이 있는 반투명 덮개와 하부 테두리가 식별된다. 밝은 중심부에도 외곽선이 유지된다. 인물과 침대는 프레임 밖으로, 이번 경광등 중심 숏에 적절하다. 읽을 수 있는 글자는 없고 창밖의 차가운 밝기는 이른 아침과 양립한다.",
        "hard_violations": [],
        "physics": "경광등은 천장에 밀착한 장착함과 금속 연결부로 지지된다. 가는 배관에는 고정 클램프가 보이고 굵은 배관은 천장 연결 지지대로 매달려 있다. 유리 칸막이는 수직 프레임에 끼워져 있으며, 지지 없이 떠 있는 물체나 인물은 없다."
       },
       {
        "label": "B",
        "direction": "카메라는 아래에서 천장을 올려다보며 큰 다각형 경광등의 밑면과 발광 측면을 보여 준다. 하단에는 찰리로 보이는 기계 머리의 뒤쪽이 보이고 얼굴은 카메라 반대편을 향한다. 실제 눈이나 시선의 표적은 보이지 않는다. 경광등의 붉은 빛은 천장과 실내로 퍼진다.",
        "built_space": "천장 중앙에 다각형 경광등 하나가 있고, 양옆에는 길쭉한 매입 설비 열 두 줄, 원형 천장 부속 두 개, 중앙에는 직사각형 패널들이 보인다. 뒤쪽에는 격자 창과 실험실 장비가 있으며 하단 중앙에는 침대 가장자리로 보이는 금속 구조와 기계 인물의 상체가 들어온다. 참조의 유리 칸막이와 곡선 조명 대신 패널 천장이 지배적이며 장소 연속성이 약하다. 경광등도 화면 폭의 약 사분의 일을 차지해 작은 천장 설비를 과도하게 전경화한 인상이다.",
        "entities": "붉은 경광등과 실험실 설비는 보이지만, 숏 텍스트가 보여 주도록 지정하지 않은 찰리의 머리와 어깨가 추가되었다. 흰색 각진 머리 외장과 금속 어깨는 참조와 대체로 연결되나 얼굴판과 센서·구속 상태는 이 구도에서 확인되지 않는다. 왼쪽 창밖 간판에는 읽을 수 있는 숫자가 노출되어 있다.",
        "hard_violations": [
         "인물을 보여 주지 않는 경광등 숏에 찰리의 머리와 상체를 추가하여, 숏 텍스트에 없는 인물을 넣지 말라는 조건을 위반했다.",
         "왼쪽 창밖 간판에 판독 가능한 숫자가 보여, 어디에도 읽을 수 있는 문자를 두지 말라는 조건을 위반했다."
        ],
        "physics": "경광등의 상부 하우징은 천장에 직접 부착되어 있어 지지가 명확하다. 찰리의 하부는 화면 아래 금속 침대 구조에 가려져 접촉점을 확인할 수 없지만, 공중에 떠 있다고 볼 근거는 없다. 머리와 어깨에 중력상 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.651
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.401
   },
   "violations": {
    "B": [
     "[gemini-pro] 배경 창밖 건물에 읽을 수 있는 텍스트(SAMSUNG, 한글 간판) 노출",
     "[gemini-pro] 레퍼런스에서 누워있어야 할 캐릭터가 똑바로 앉아있거나 서 있는 형태의 물리적/자세 위반",
     "[gpt-high] 인물을 보여 주지 않는 경광등 숏에 찰리의 머리와 상체를 추가하여, 숏 텍스트에 없는 인물을 넣지 말라는 조건을 위반했다.",
     "[gpt-high] 왼쪽 창밖 간판에 판독 가능한 숫자가 보여, 어디에도 읽을 수 있는 문자를 두지 말라는 조건을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 401
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프레임 지시(천장 경광등 미디엄 샷)를 정확히 따르며 텍스트 제한을 준수했으나, 천장 재질이 레퍼런스의 분위기와 다소 차이가 있습니다."
   },
   {
    "label": "B",
    "score": 401,
    "verdict_ko": "배경에 읽을 수 있는 텍스트가 노출되었고, 레퍼런스에 고정된 캐릭터의 누워있는 자세가 변경되는 치명적인 규정 위반이 있습니다.  ★위반: [gemini-pro] 배경 창밖 건물에 읽을 수 있는 텍스트(SAMSUNG, 한글 간판) 노출 / [gemini-pro] 레퍼런스에서 누워있어야 할 캐릭터가 똑바로 앉아있거나 서 있는 형태의 물리적/자세 위반 / [gpt-high] 인물을 보여 주지 않는 경광등 숏에 찰리의 머리와 상체를 추가하여, 숏 텍스트에 없는 인물을 넣지 말라는 조건을 위반했다. / [gpt-high] 왼쪽 창밖 간판에 판독 가능한 숫자가 보여, 어디에도 읽을 수 있는 문자를 두지 말라는 조건을 위반했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S88sh7_sel.png",
    "asset_id": "96e1d3ba-947f-4da0-9d38-283029728d14",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf71-86b9-79d4-a812-cca65088e9dc",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S88sh7"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S89sh38": {
  "input_fingerprint": "220862524d8329bf",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 유빅사 로고가 찍힌 검은색 무장 헬기 편대가 기수를 숙인 채 먹구름 낀 하늘을 가르며 맹렬히 돌진 중인 웅장한 전경.\n\nLOCATION (lock): In the open sky approaching the island research institute's rooftop, where armed helicopters advance toward the building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approaching armed helicopter formation in the upper-right of the frame, background, moves toward lower-left rooftop approach.\n- KEY BACKGROUND ELEMENTS: Armed helicopter formation (Approaching rapidly with noses lowered) — Oblique front-quarter and side views expose the 유빅사 markings without placing the camera on the flight path; used as Creates a layered diagonal advance across the sky; Clouded sky (Dark clouds visible behind the formation); used as Provides separation around individual aircraft and space for their approach; Institute rooftop edge (Visible as a narrow lower-frame reference) — Seen across the foreground from the rooftop camera position; used as Retains the observer's rooftop location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight beneath the clouded sky gives the black aircraft firm silhouettes while retaining the legibility of their markings.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Armed helicopters bearing Ubik logos are approaching the research center; the rooftop retains its lawn, large antenna and all-direction speakers. Disabled artificial soldiers lie where the amplified digital sound dropped them, while Charlie is free of restraints with his experiment sensors still attached.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 유빅사 로고가 찍힌 검은색 무장 헬기 편대가 기수를 숙인 채 먹구름 낀 하늘을 가르며 맹렬히 돌진 중인 웅장한 전경.\n\nLOCATION (lock): In the open sky approaching the island research institute's rooftop, where armed helicopters advance toward the building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approaching armed helicopter formation in the upper-right of the frame, background, moves toward lower-left rooftop approach.\n- KEY BACKGROUND ELEMENTS: Armed helicopter formation (Approaching rapidly with noses lowered) — Oblique front-quarter and side views expose the 유빅사 markings without placing the camera on the flight path; used as Creates a layered diagonal advance across the sky; Clouded sky (Dark clouds visible behind the formation); used as Provides separation around individual aircraft and space for their approach; Institute rooftop edge (Visible as a narrow lower-frame reference) — Seen across the foreground from the rooftop camera position; used as Retains the observer's rooftop location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight beneath the clouded sky gives the black aircraft firm silhouettes while retaining the legibility of their markings.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Armed helicopters bearing Ubik logos are approaching the research center; the rooftop retains its lawn, large antenna and all-direction speakers. Disabled artificial soldiers lie where the amplified digital sound dropped them, while Charlie is free of restraints with his experiment sensors still attached.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 유빅사 로고가 찍힌 검은색 무장 헬기 편대가 기수를 숙인 채 먹구름 낀 하늘을 가르며 맹렬히 돌진 중인 웅장한 전경.\n\nLOCATION (lock): In the open sky approaching the island research institute's rooftop, where armed helicopters advance toward the building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approaching armed helicopter formation in the upper-right of the frame, background, moves toward lower-left rooftop approach.\n- KEY BACKGROUND ELEMENTS: Armed helicopter formation (Approaching rapidly with noses lowered) — Oblique front-quarter and side views expose the 유빅사 markings without placing the camera on the flight path; used as Creates a layered diagonal advance across the sky; Clouded sky (Dark clouds visible behind the formation); used as Provides separation around individual aircraft and space for their approach; Institute rooftop edge (Visible as a narrow lower-frame reference) — Seen across the foreground from the rooftop camera position; used as Retains the observer's rooftop location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight beneath the clouded sky gives the black aircraft firm silhouettes while retaining the legibility of their markings.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Armed helicopters bearing Ubik logos are approaching the research center; the rooftop retains its lawn, large antenna and all-direction speakers. Disabled artificial soldiers lie where the amplified digital sound dropped them, while Charlie is free of restraints with his experiment sensors still attached.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "헬기 편대가 좌측 수평 방향으로 비행 중임.",
    "built_space": "옥상 잔디밭 공간 좌측에 스피커탑과 레이돔 철탑이 위치함.",
    "entities": "유빅사 표기가 있는 검은색 무장 헬기들, 바닥에 엎드리거나 앉아있는 사람들.",
    "hard_violations": [
     "[gemini-pro] 지시문을 위반하여 살아 움직이고 앉아있는 사람들을 추가함 (invented people)",
     "[gpt-high] 옥상 중앙에 상체를 세우고 앉거나 무릎을 꿇은 인물들이 등장하여, 앉은 인물도 금지한 명시적 무인 숏 조건을 위반한다.",
     "[gpt-high] 헬기에 선명하게 읽히는 '유빅사' 표기와 로고가 노출되어, 최종 지시의 문자·로고 금지를 위반한다."
    ],
    "physics": "헬기는 공중 비행으로, 사람들은 지면에 앉거나 엎드린 상태로 지탱됨."
   },
   {
    "label": "B",
    "direction": "헬기 기수가 좌측 하단을 향해 일제히 하강하듯 정렬됨.",
    "built_space": "옥상 잔디밭 우측에 스피커탑과 레이돔 철탑, 난간이 배치됨.",
    "entities": "유빅사 표기가 있는 검은색 무장 헬기들, 잔디밭에 완전히 쓰러진 인조 병사들.",
    "hard_violations": [
     "[gpt-high] 헬기 측면에 판독 가능한 한글 표기와 식별 가능한 로고가 노출되어, 최종 지시의 문자·로고 금지를 위반한다."
    ],
    "physics": "헬기는 로터 회전과 함께 공중 비행 중이며, 쓰러진 병사들은 지면에 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "우측 상단에서 좌측 하단으로 하강하는 헬기 편대의 배치와 비활성화되어 쓰러진 병사들의 모습 등 프레이밍과 피사체 상태를 지시문과 가장 정확하게 일치시켰습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화면에 살아있는 사람이 없어야 한다는 명시적 지시를 어기고 잔디밭에 앉아 활동하는 인물들을 포함해 치명적인 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "헬기 편대가 좌측 수평 방향으로 비행 중임.",
        "built_space": "옥상 잔디밭 공간 좌측에 스피커탑과 레이돔 철탑이 위치함.",
        "entities": "유빅사 표기가 있는 검은색 무장 헬기들, 바닥에 엎드리거나 앉아있는 사람들.",
        "hard_violations": [
         "지시문을 위반하여 살아 움직이고 앉아있는 사람들을 추가함 (invented people)"
        ],
        "physics": "헬기는 공중 비행으로, 사람들은 지면에 앉거나 엎드린 상태로 지탱됨."
       },
       {
        "label": "B",
        "direction": "헬기 기수가 좌측 하단을 향해 일제히 하강하듯 정렬됨.",
        "built_space": "옥상 잔디밭 우측에 스피커탑과 레이돔 철탑, 난간이 배치됨.",
        "entities": "유빅사 표기가 있는 검은색 무장 헬기들, 잔디밭에 완전히 쓰러진 인조 병사들.",
        "hard_violations": [],
        "physics": "헬기는 로터 회전과 함께 공중 비행 중이며, 쓰러진 병사들은 지면에 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "우측 상단에서 좌측 하단으로 하강하는 헬기 편대의 배치와 비활성화되어 쓰러진 병사들의 모습 등 프레이밍과 피사체 상태를 지시문과 가장 정확하게 일치시켰습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화면에 살아있는 사람이 없어야 한다는 명시적 지시를 어기고 잔디밭에 앉아 활동하는 인물들을 포함해 치명적인 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "헬기 편대가 좌측 수평 방향으로 비행 중임.",
        "built_space": "옥상 잔디밭 공간 좌측에 스피커탑과 레이돔 철탑이 위치함.",
        "entities": "유빅사 표기가 있는 검은색 무장 헬기들, 바닥에 엎드리거나 앉아있는 사람들.",
        "hard_violations": [
         "지시문을 위반하여 살아 움직이고 앉아있는 사람들을 추가함 (invented people)"
        ],
        "physics": "헬기는 공중 비행으로, 사람들은 지면에 앉거나 엎드린 상태로 지탱됨."
       },
       {
        "label": "B",
        "direction": "헬기 기수가 좌측 하단을 향해 일제히 하강하듯 정렬됨.",
        "built_space": "옥상 잔디밭 우측에 스피커탑과 레이돔 철탑, 난간이 배치됨.",
        "entities": "유빅사 표기가 있는 검은색 무장 헬기들, 잔디밭에 완전히 쓰러진 인조 병사들.",
        "hard_violations": [],
        "physics": "헬기는 로터 회전과 함께 공중 비행 중이며, 쓰러진 병사들은 지면에 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "기수를 낮춘 좌하향 편대와 옥상 설비는 비교적 충실하지만, 헬기와 잔디 전경을 지나치게 크게 담았고 판독 가능한 글자·로고가 금지 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "큰 헬기와 넓은 옥상으로 지정 구도를 벗어난 데 더해, 앉아 있는 인물들과 선명한 글자·로고가 인물 및 문자 금지 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "다섯 헬기의 기수는 모두 화면 왼쪽을 향하고 꼬리보다 낮다. 기수와 전방 무장의 방향은 대체로 좌하단 옥상 접근 방향에 맞는다. 편대도 우상단에서 좌하단으로 이어지지만, 우상단 선두기가 매우 커서 먼 배경의 접근 편대보다는 바로 머리 위를 지나는 헬기로 보인다.",
        "built_space": "옥상 잔디와 금속 난간이 하단 약 4분의 1 이상을 차지한다. 오른쪽에 구형 덮개가 있는 안테나탑 하나, 다방향 스피커 기둥 하나, 낮은 설비실 하나가 있으며 왼쪽 뒤에 사각 개구부 구조가 보인다. 옥상에서 바라보는 시점과 주요 설비 종류는 참조에 부합하지만, 옥상 가장자리만 좁게 남기라는 구도와 달리 잔디와 쓰러진 병사들까지 넓게 노출했다.",
        "entities": "검은 무장 헬기 다섯 대, 먹구름, 바다, 잔디 옥상, 안테나와 스피커가 보인다. 전경에는 장갑을 입고 누운 인공병사형 몸체 네 개와 오른쪽 탑 아래 부분적으로 보이는 몸체 하나가 있다. 서거나 앉아 있는 인물은 없다. 헬기 측면의 한글과 원형 로고는 식별 가능하여 최종 문자·로고 금지 조건에 어긋난다. 찰리와 센서는 보이지 않으며, 이 숏에서 이를 추가로 보여줄 필요는 없다.",
        "hard_violations": [
         "헬기 측면에 판독 가능한 한글 표기와 식별 가능한 로고가 노출되어, 최종 지시의 문자·로고 금지를 위반한다."
        ],
        "physics": "헬기마다 주회전익과 꼬리회전익이 있으며 회전 흔적이 보여 비행을 지탱하는 동력이 설명된다. 낮아진 기수도 전진 비행으로 가능한 자세다. 무장은 기체에 부착되어 있고 누운 몸체들은 잔디에 지지되어 있다. 안테나와 스피커는 옥상 기초에 연결되어 있으며, 근거 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "보이는 헬기 다섯 대의 기수와 전방 무장은 왼쪽을 향한다. 일부는 기수가 약간 낮지만 전체적으로 수평 비행에 가까워 맹렬한 좌하향 돌진은 약하다. 편대는 우측 하늘에 모여 있으나 우상단에서 좌하단 옥상으로 이어지는 층진 대각선보다 넓게 흩어진 배열이다.",
        "built_space": "하단에 잔디 옥상, 난간, 통로와 여러 설비실을 넓게 보여준다. 왼쪽에는 다방향 스피커 기둥 하나와 구형 덮개를 가진 안테나탑 하나가 있다. 탑 중간의 큰 격자형 패널들은 구조 참조에서 보이는 탑 외형과 다르다. 옥상 시점 자체는 가능하지만, 좁은 옥상 가장자리만 보여야 하는 구도에서 벗어나 지붕 위 인물과 설비를 과도하게 포함한다.",
        "entities": "검은 무장 헬기 다섯 대와 흐린 낮 하늘, 바다, 옥상 잔디 및 방송 설비가 보인다. 가장 큰 헬기의 '유빅사' 글자는 선명하고 다른 기체에도 글자와 로고가 보인다. 잔디에는 여러 군복 차림 몸체가 있으며, 중앙의 두 인물은 상체를 세우고 앉거나 무릎을 꿇고 있다. 이는 모두 무력화되어 누워 있는 인공병사만 남은 상태나 인물이 등장하지 않는 숏으로 읽히지 않는다. 노출된 인물들의 국적·민족 정체성은 이 거리에서 확정할 수 없다.",
        "hard_violations": [
         "옥상 중앙에 상체를 세우고 앉거나 무릎을 꿇은 인물들이 등장하여, 앉은 인물도 금지한 명시적 무인 숏 조건을 위반한다.",
         "헬기에 선명하게 읽히는 '유빅사' 표기와 로고가 노출되어, 최종 지시의 문자·로고 금지를 위반한다."
        ],
        "physics": "헬기는 회전익으로 비행하며 무장과 착륙장치는 기체에 연결되어 있다. 누운 몸체는 잔디에, 상체를 세운 인물은 무릎이나 하체에 지지되어 있어 자세 자체가 물리적으로 불가능하지는 않다. 다만 그 자세는 요구된 무인 장면 및 쓰러진 상태와 맞지 않는다. 옥상 설비들도 바닥 기초에 지지되어 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "기수를 낮춘 좌하향 편대와 옥상 설비는 비교적 충실하지만, 헬기와 잔디 전경을 지나치게 크게 담았고 판독 가능한 글자·로고가 금지 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "큰 헬기와 넓은 옥상으로 지정 구도를 벗어난 데 더해, 앉아 있는 인물들과 선명한 글자·로고가 인물 및 문자 금지 조건을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "다섯 헬기의 기수는 모두 화면 왼쪽을 향하고 꼬리보다 낮다. 기수와 전방 무장의 방향은 대체로 좌하단 옥상 접근 방향에 맞는다. 편대도 우상단에서 좌하단으로 이어지지만, 우상단 선두기가 매우 커서 먼 배경의 접근 편대보다는 바로 머리 위를 지나는 헬기로 보인다.",
        "built_space": "옥상 잔디와 금속 난간이 하단 약 4분의 1 이상을 차지한다. 오른쪽에 구형 덮개가 있는 안테나탑 하나, 다방향 스피커 기둥 하나, 낮은 설비실 하나가 있으며 왼쪽 뒤에 사각 개구부 구조가 보인다. 옥상에서 바라보는 시점과 주요 설비 종류는 참조에 부합하지만, 옥상 가장자리만 좁게 남기라는 구도와 달리 잔디와 쓰러진 병사들까지 넓게 노출했다.",
        "entities": "검은 무장 헬기 다섯 대, 먹구름, 바다, 잔디 옥상, 안테나와 스피커가 보인다. 전경에는 장갑을 입고 누운 인공병사형 몸체 네 개와 오른쪽 탑 아래 부분적으로 보이는 몸체 하나가 있다. 서거나 앉아 있는 인물은 없다. 헬기 측면의 한글과 원형 로고는 식별 가능하여 최종 문자·로고 금지 조건에 어긋난다. 찰리와 센서는 보이지 않으며, 이 숏에서 이를 추가로 보여줄 필요는 없다.",
        "hard_violations": [
         "헬기 측면에 판독 가능한 한글 표기와 식별 가능한 로고가 노출되어, 최종 지시의 문자·로고 금지를 위반한다."
        ],
        "physics": "헬기마다 주회전익과 꼬리회전익이 있으며 회전 흔적이 보여 비행을 지탱하는 동력이 설명된다. 낮아진 기수도 전진 비행으로 가능한 자세다. 무장은 기체에 부착되어 있고 누운 몸체들은 잔디에 지지되어 있다. 안테나와 스피커는 옥상 기초에 연결되어 있으며, 근거 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "보이는 헬기 다섯 대의 기수와 전방 무장은 왼쪽을 향한다. 일부는 기수가 약간 낮지만 전체적으로 수평 비행에 가까워 맹렬한 좌하향 돌진은 약하다. 편대는 우측 하늘에 모여 있으나 우상단에서 좌하단 옥상으로 이어지는 층진 대각선보다 넓게 흩어진 배열이다.",
        "built_space": "하단에 잔디 옥상, 난간, 통로와 여러 설비실을 넓게 보여준다. 왼쪽에는 다방향 스피커 기둥 하나와 구형 덮개를 가진 안테나탑 하나가 있다. 탑 중간의 큰 격자형 패널들은 구조 참조에서 보이는 탑 외형과 다르다. 옥상 시점 자체는 가능하지만, 좁은 옥상 가장자리만 보여야 하는 구도에서 벗어나 지붕 위 인물과 설비를 과도하게 포함한다.",
        "entities": "검은 무장 헬기 다섯 대와 흐린 낮 하늘, 바다, 옥상 잔디 및 방송 설비가 보인다. 가장 큰 헬기의 '유빅사' 글자는 선명하고 다른 기체에도 글자와 로고가 보인다. 잔디에는 여러 군복 차림 몸체가 있으며, 중앙의 두 인물은 상체를 세우고 앉거나 무릎을 꿇고 있다. 이는 모두 무력화되어 누워 있는 인공병사만 남은 상태나 인물이 등장하지 않는 숏으로 읽히지 않는다. 노출된 인물들의 국적·민족 정체성은 이 거리에서 확정할 수 없다.",
        "hard_violations": [
         "옥상 중앙에 상체를 세우고 앉거나 무릎을 꿇은 인물들이 등장하여, 앉은 인물도 금지한 명시적 무인 숏 조건을 위반한다.",
         "헬기에 선명하게 읽히는 '유빅사' 표기와 로고가 노출되어, 최종 지시의 문자·로고 금지를 위반한다."
        ],
        "physics": "헬기는 회전익으로 비행하며 무장과 착륙장치는 기체에 연결되어 있다. 누운 몸체는 잔디에, 상체를 세운 인물은 무릎이나 하체에 지지되어 있어 자세 자체가 물리적으로 불가능하지는 않다. 다만 그 자세는 요구된 무인 장면 및 쓰러진 상태와 맞지 않는다. 옥상 설비들도 바닥 기초에 지지되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.762,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.512,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 지시문을 위반하여 살아 움직이고 앉아있는 사람들을 추가함 (invented people)",
     "[gpt-high] 옥상 중앙에 상체를 세우고 앉거나 무릎을 꿇은 인물들이 등장하여, 앉은 인물도 금지한 명시적 무인 숏 조건을 위반한다.",
     "[gpt-high] 헬기에 선명하게 읽히는 '유빅사' 표기와 로고가 노출되어, 최종 지시의 문자·로고 금지를 위반한다."
    ],
    "B": [
     "[gpt-high] 헬기 측면에 판독 가능한 한글 표기와 식별 가능한 로고가 노출되어, 최종 지시의 문자·로고 금지를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 512
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "우측 상단에서 좌측 하단으로 하강하는 헬기 편대의 배치와 비활성화되어 쓰러진 병사들의 모습 등 프레이밍과 피사체 상태를 지시문과 가장 정확하게 일치시켰습니다.  ★위반: [gpt-high] 헬기 측면에 판독 가능한 한글 표기와 식별 가능한 로고가 노출되어, 최종 지시의 문자·로고 금지를 위반한다."
   },
   {
    "label": "A",
    "score": 512,
    "verdict_ko": "화면에 살아있는 사람이 없어야 한다는 명시적 지시를 어기고 잔디밭에 앉아 활동하는 인물들을 포함해 치명적인 위반이 발생했습니다.  ★위반: [gemini-pro] 지시문을 위반하여 살아 움직이고 앉아있는 사람들을 추가함 (invented people) / [gpt-high] 옥상 중앙에 상체를 세우고 앉거나 무릎을 꿇은 인물들이 등장하여, 앉은 인물도 금지한 명시적 무인 숏 조건을 위반한다. / [gpt-high] 헬기에 선명하게 읽히는 '유빅사' 표기와 로고가 노출되어, 최종 지시의 문자·로고 금지를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L138B01.png",
    "asset_id": "942815d8-a414-4766-953f-ed74b867df83",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_island_research_complex_sel.png",
    "asset_id": "65007d02-9a82-4f55-987f-29e526ed7aee",
    "role": "structure_seed_look"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf75-854e-78fb-a165-8994da5a0a03",
  "ref_mode": "플레이트+seed만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S89sh58": {
  "input_fingerprint": "b43cbb547bb7d4df",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링에서 밤하늘을 찢을 듯한 거대한 새하얀 광선 기둥이 수직으로 치솟는 압도적인 찰나.\n\nLOCATION (lock): On the exposed rooftop garden of the island research institute, near the large antenna and all-direction speakers beneath a rising column of white light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Beam origin at 찰리's chest ring in the lower-center of the frame, midground; Vertical energy column in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Vertical energy beam (Erupting upward from 찰리's chest ring) — Its visible base connects to the chest before the column extends vertically beyond the upper edge; used as Forms the central vertical axis while leaving sky visible on both sides; Rooftop park ground (Visible beneath 찰리) — Seen at a shallow angle across the lower edge; used as Grounds the body beneath the otherwise overwhelming vertical event.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The intensely white beam creates overwhelming brightness and strong local contrast while preserving enough contour to read 찰리 at its source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The research center is heavily damaged by bombardment, and the command helicopter has landed on the rooftop. Charlie's rapidly spinning chest ring emits a towering white, star-bright beam while energy graphs surge and lights fluctuate in brightness.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링에서 밤하늘을 찢을 듯한 거대한 새하얀 광선 기둥이 수직으로 치솟는 압도적인 찰나.\n\nLOCATION (lock): On the exposed rooftop garden of the island research institute, near the large antenna and all-direction speakers beneath a rising column of white light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Beam origin at 찰리's chest ring in the lower-center of the frame, midground; Vertical energy column in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Vertical energy beam (Erupting upward from 찰리's chest ring) — Its visible base connects to the chest before the column extends vertically beyond the upper edge; used as Forms the central vertical axis while leaving sky visible on both sides; Rooftop park ground (Visible beneath 찰리) — Seen at a shallow angle across the lower edge; used as Grounds the body beneath the otherwise overwhelming vertical event.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The intensely white beam creates overwhelming brightness and strong local contrast while preserving enough contour to read 찰리 at its source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The research center is heavily damaged by bombardment, and the command helicopter has landed on the rooftop. Charlie's rapidly spinning chest ring emits a towering white, star-bright beam while energy graphs surge and lights fluctuate in brightness.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링에서 밤하늘을 찢을 듯한 거대한 새하얀 광선 기둥이 수직으로 치솟는 압도적인 찰나.\n\nLOCATION (lock): On the exposed rooftop garden of the island research institute, near the large antenna and all-direction speakers beneath a rising column of white light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Beam origin at 찰리's chest ring in the lower-center of the frame, midground; Vertical energy column in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Vertical energy beam (Erupting upward from 찰리's chest ring) — Its visible base connects to the chest before the column extends vertically beyond the upper edge; used as Forms the central vertical axis while leaving sky visible on both sides; Rooftop park ground (Visible beneath 찰리) — Seen at a shallow angle across the lower edge; used as Grounds the body beneath the otherwise overwhelming vertical event.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The intensely white beam creates overwhelming brightness and strong local contrast while preserving enough contour to read 찰리 at its source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The research center is heavily damaged by bombardment, and the command helicopter has landed on the rooftop. Charlie's rapidly spinning chest ring emits a towering white, star-bright beam while energy graphs surge and lights fluctuate in brightness.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 2,
  "readings": [
   {
    "label": "A",
    "direction": "수직으로 솟구치는 광선 기둥이 있으나, 가슴 링이 아닌 캐릭터 등 뒤의 바닥에서 뻗어나감.",
    "built_space": "안테나, 스피커, 벤치 등이 있는 옥상 구역과 파손된 구조물 위에 착륙한 헬리콥터가 배치됨.",
    "entities": "고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈을 가진 찰리가 정확히 묘사됨. 헬리콥터 존재.",
    "hard_violations": [
     "[gemini-pro] 허공에 떠 있는 홀로그램 UI 및 그래프 오버레이 (leaked markers/diagrams/text)",
     "[gemini-pro] 광선 기둥의 베이스가 가슴이 아닌 캐릭터 뒤쪽 바닥에 위치함 (스테이징 지시 위반)",
     "[gpt-high] 헬리콥터 주변의 에너지 그래프를 물리적 표시 장치 없이 공중에 뜬 평면 도표로 표현하여, 금지된 그래픽 오버레이가 화면에 노출되었습니다."
    ],
    "physics": "찰리는 바닥에 안정적으로 서 있으며, 발밑으로 떠오르는 흙먼지가 보임."
   },
   {
    "label": "B",
    "direction": "수직으로 솟구치는 광선이 가슴이 아닌 목/머리 위치의 링에서 위로 뻗어나감.",
    "built_space": "잔해와 파손된 태양광 패널이 널려 있는 옥상에 안테나, 스피커, 헬리콥터가 배치됨.",
    "entities": "찰리의 몸통과 헬리콥터는 존재하나, 프롬프트가 요구한 찰리의 세밀한 기계 머리가 완전히 누락됨.",
    "hard_violations": [
     "[gemini-pro] 캐릭터의 머리가 완전히 누락됨 (physically impossible anatomy / 필수 엔티티 위반)"
    ],
    "physics": "찰리는 파손된 옥상 바닥에 두 발로 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "캐릭터 외형과 배경은 잘 구현되었으나, 허공의 UI 그래픽 오버레이와 광선 시작점 오류가 치명적입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 머리가 완전히 누락되어 지정된 형태를 상실한 치명적인 해부학적 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수직으로 솟구치는 광선 기둥이 있으나, 가슴 링이 아닌 캐릭터 등 뒤의 바닥에서 뻗어나감.",
        "built_space": "안테나, 스피커, 벤치 등이 있는 옥상 구역과 파손된 구조물 위에 착륙한 헬리콥터가 배치됨.",
        "entities": "고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈을 가진 찰리가 정확히 묘사됨. 헬리콥터 존재.",
        "hard_violations": [
         "허공에 떠 있는 홀로그램 UI 및 그래프 오버레이 (leaked markers/diagrams/text)",
         "광선 기둥의 베이스가 가슴이 아닌 캐릭터 뒤쪽 바닥에 위치함 (스테이징 지시 위반)"
        ],
        "physics": "찰리는 바닥에 안정적으로 서 있으며, 발밑으로 떠오르는 흙먼지가 보임."
       },
       {
        "label": "B",
        "direction": "수직으로 솟구치는 광선이 가슴이 아닌 목/머리 위치의 링에서 위로 뻗어나감.",
        "built_space": "잔해와 파손된 태양광 패널이 널려 있는 옥상에 안테나, 스피커, 헬리콥터가 배치됨.",
        "entities": "찰리의 몸통과 헬리콥터는 존재하나, 프롬프트가 요구한 찰리의 세밀한 기계 머리가 완전히 누락됨.",
        "hard_violations": [
         "캐릭터의 머리가 완전히 누락됨 (physically impossible anatomy / 필수 엔티티 위반)"
        ],
        "physics": "찰리는 파손된 옥상 바닥에 두 발로 서 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "캐릭터 외형과 배경은 잘 구현되었으나, 허공의 UI 그래픽 오버레이와 광선 시작점 오류가 치명적입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 머리가 완전히 누락되어 지정된 형태를 상실한 치명적인 해부학적 오류가 있습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "수직으로 솟구치는 광선 기둥이 있으나, 가슴 링이 아닌 캐릭터 등 뒤의 바닥에서 뻗어나감.",
        "built_space": "안테나, 스피커, 벤치 등이 있는 옥상 구역과 파손된 구조물 위에 착륙한 헬리콥터가 배치됨.",
        "entities": "고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈을 가진 찰리가 정확히 묘사됨. 헬리콥터 존재.",
        "hard_violations": [
         "허공에 떠 있는 홀로그램 UI 및 그래프 오버레이 (leaked markers/diagrams/text)",
         "광선 기둥의 베이스가 가슴이 아닌 캐릭터 뒤쪽 바닥에 위치함 (스테이징 지시 위반)"
        ],
        "physics": "찰리는 바닥에 안정적으로 서 있으며, 발밑으로 떠오르는 흙먼지가 보임."
       },
       {
        "label": "B",
        "direction": "수직으로 솟구치는 광선이 가슴이 아닌 목/머리 위치의 링에서 위로 뻗어나감.",
        "built_space": "잔해와 파손된 태양광 패널이 널려 있는 옥상에 안테나, 스피커, 헬리콥터가 배치됨.",
        "entities": "찰리의 몸통과 헬리콥터는 존재하나, 프롬프트가 요구한 찰리의 세밀한 기계 머리가 완전히 누락됨.",
        "hard_violations": [
         "캐릭터의 머리가 완전히 누락됨 (physically impossible anatomy / 필수 엔티티 위반)"
        ],
        "physics": "찰리는 파손된 옥상 바닥에 두 발로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "하단 중앙의 찰리와 상단 밖으로 솟는 광선을 와이드 숏으로 구현했지만, 과도한 발광 때문에 가슴 링의 정확한 위치와 머리 정체성이 불분명합니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "찰리의 얼굴은 충실하지만 광선이 가슴 링이 아니라 몸 뒤 지면에서 솟고, 인물이 너무 크게 잡혔으며 공중 그래프가 삽입되어 핵심 연출을 위반합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "광선은 하단 중앙 찰리의 상체에 겹친 발광 링에서 수직으로 올라가 화면 상단 밖으로 이어집니다. 다만 링이 어깨와 목 높이까지 크게 번져 정확한 가슴 접속점은 불명확합니다. 얼굴과 눈은 빛에 가려 시선 방향을 확인할 수 없습니다. 오른쪽 스피커 네 개는 서로 다른 바깥 방향을 향합니다.",
        "built_space": "찰리는 손상된 옥상 중앙 통로에 서 있고, 양옆에 콘크리트 화단과 잔해가 있습니다. 왼쪽에는 파손된 설비실 하나, 패널형 안테나 탑 하나, 접시형 장비가 달린 탑 하나와 가는 마스트들이 보입니다. 오른쪽에는 네 개의 혼 스피커가 달린 기둥 하나와 착륙한 헬리콥터 한 대가 있습니다. 옥상 바닥이 하단에 드러나지만 요구한 얕은 각도보다 바닥 면적이 다소 큽니다. 참고 사진은 해안에서 본 연구소 외관이므로 이 옥상의 세부 배치까지 일치한다고 확인할 수는 없습니다.",
        "entities": "찰리로 제시된 기계형 개체 하나만 있으며 추가 인물은 없습니다. 금속 몸체와 양팔·양다리는 보이지만, 지정된 고릴라형 각진 머리, 흰 얼굴판, 점 눈과 선 입은 발광에 가려 확인되지 않습니다. 흰 광선, 발광 링, 옥상 화단, 안테나, 스피커, 폭격 손상과 헬리콥터는 보입니다. 낮의 하늘이며 읽을 수 있는 문구는 없습니다. 에너지 그래프와 조명 밝기의 변동은 확인되지 않습니다.",
        "hard_violations": [],
        "physics": "찰리의 두 발이 옥상 바닥에 닿아 체중을 지탱합니다. 양팔은 몸에 연결된 채 아래로 내려져 있습니다. 헬리콥터는 착륙 바퀴로 지붕에 놓여 있고, 안테나와 스피커는 기둥 및 지지 구조로 고정되어 있습니다. 잔해는 바닥이나 주변 구조물에 걸쳐 있습니다. 광선은 상체 발광부와 연속되어 있으며, 지지 없이 떠 있는 신체나 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴과 눈은 카메라 정면을 향합니다. 가슴 링도 정면으로 빛나지만, 거대한 수직 광선은 찰리 뒤에 있으며 다리 사이의 지면 부근까지 이어집니다. 따라서 가슴 링에서 위로 분출되는 연결 관계가 성립하지 않습니다. 왼쪽의 혼 스피커 여섯 개는 좌우 바깥쪽을 향합니다.",
        "built_space": "전경 난간 뒤 중앙에 찰리가 크게 서 있고 하체 일부가 화면 아래로 잘립니다. 좌우에는 식재 화단과 벤치들이 있으며, 왼쪽에는 접시 안테나 하나와 여섯 개의 혼 스피커를 단 마스트 하나가 있습니다. 오른쪽 뒤에는 심하게 파손된 설비 건물 하나와 그 지붕에 착륙한 헬리콥터 한 대가 있습니다. 옥상 정원이라는 공간은 읽히지만, 인물 크기와 전경 난간 때문에 요구된 중경 하단 중앙의 작은 광원 배치에서 벗어납니다. 참고 사진의 외관과 동일한 장소인지는 확증하기 어렵습니다.",
        "entities": "기계형 찰리 하나가 보이며 각진 고릴라형 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입이 지정과 잘 맞습니다. 원형 가슴 링도 명확합니다. 낮의 하늘, 옥상 정원, 안테나, 스피커, 폭격 손상, 헬리콥터가 있습니다. 헬리콥터 주변에는 물리적 화면 없이 흰 그래프 패널들이 공중에 삽입되어 있습니다. 읽을 수 있는 문장은 보이지 않습니다.",
        "hard_violations": [
         "헬리콥터 주변의 에너지 그래프를 물리적 표시 장치 없이 공중에 뜬 평면 도표로 표현하여, 금지된 그래픽 오버레이가 화면에 노출되었습니다."
        ],
        "physics": "찰리의 발은 전경 난간과 화면 경계에 가려져 접지를 직접 확인할 수 없지만, 몸은 옥상 통로에 선 수직 자세여서 부유한다고 볼 근거는 없습니다. 헬리콥터는 지붕에 착륙해 있고 설비는 지지대에 고정되어 있습니다. 주변의 작은 공중 잔해와 먼지는 광선 분출에 따른 비산으로 해석할 수 있습니다. 반면 공중 그래프에는 화면, 기둥 또는 다른 물리적 지지체가 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "하단 중앙의 찰리와 상단 밖으로 솟는 광선을 와이드 숏으로 구현했지만, 과도한 발광 때문에 가슴 링의 정확한 위치와 머리 정체성이 불분명합니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "찰리의 얼굴은 충실하지만 광선이 가슴 링이 아니라 몸 뒤 지면에서 솟고, 인물이 너무 크게 잡혔으며 공중 그래프가 삽입되어 핵심 연출을 위반합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "광선은 하단 중앙 찰리의 상체에 겹친 발광 링에서 수직으로 올라가 화면 상단 밖으로 이어집니다. 다만 링이 어깨와 목 높이까지 크게 번져 정확한 가슴 접속점은 불명확합니다. 얼굴과 눈은 빛에 가려 시선 방향을 확인할 수 없습니다. 오른쪽 스피커 네 개는 서로 다른 바깥 방향을 향합니다.",
        "built_space": "찰리는 손상된 옥상 중앙 통로에 서 있고, 양옆에 콘크리트 화단과 잔해가 있습니다. 왼쪽에는 파손된 설비실 하나, 패널형 안테나 탑 하나, 접시형 장비가 달린 탑 하나와 가는 마스트들이 보입니다. 오른쪽에는 네 개의 혼 스피커가 달린 기둥 하나와 착륙한 헬리콥터 한 대가 있습니다. 옥상 바닥이 하단에 드러나지만 요구한 얕은 각도보다 바닥 면적이 다소 큽니다. 참고 사진은 해안에서 본 연구소 외관이므로 이 옥상의 세부 배치까지 일치한다고 확인할 수는 없습니다.",
        "entities": "찰리로 제시된 기계형 개체 하나만 있으며 추가 인물은 없습니다. 금속 몸체와 양팔·양다리는 보이지만, 지정된 고릴라형 각진 머리, 흰 얼굴판, 점 눈과 선 입은 발광에 가려 확인되지 않습니다. 흰 광선, 발광 링, 옥상 화단, 안테나, 스피커, 폭격 손상과 헬리콥터는 보입니다. 낮의 하늘이며 읽을 수 있는 문구는 없습니다. 에너지 그래프와 조명 밝기의 변동은 확인되지 않습니다.",
        "hard_violations": [],
        "physics": "찰리의 두 발이 옥상 바닥에 닿아 체중을 지탱합니다. 양팔은 몸에 연결된 채 아래로 내려져 있습니다. 헬리콥터는 착륙 바퀴로 지붕에 놓여 있고, 안테나와 스피커는 기둥 및 지지 구조로 고정되어 있습니다. 잔해는 바닥이나 주변 구조물에 걸쳐 있습니다. 광선은 상체 발광부와 연속되어 있으며, 지지 없이 떠 있는 신체나 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴과 눈은 카메라 정면을 향합니다. 가슴 링도 정면으로 빛나지만, 거대한 수직 광선은 찰리 뒤에 있으며 다리 사이의 지면 부근까지 이어집니다. 따라서 가슴 링에서 위로 분출되는 연결 관계가 성립하지 않습니다. 왼쪽의 혼 스피커 여섯 개는 좌우 바깥쪽을 향합니다.",
        "built_space": "전경 난간 뒤 중앙에 찰리가 크게 서 있고 하체 일부가 화면 아래로 잘립니다. 좌우에는 식재 화단과 벤치들이 있으며, 왼쪽에는 접시 안테나 하나와 여섯 개의 혼 스피커를 단 마스트 하나가 있습니다. 오른쪽 뒤에는 심하게 파손된 설비 건물 하나와 그 지붕에 착륙한 헬리콥터 한 대가 있습니다. 옥상 정원이라는 공간은 읽히지만, 인물 크기와 전경 난간 때문에 요구된 중경 하단 중앙의 작은 광원 배치에서 벗어납니다. 참고 사진의 외관과 동일한 장소인지는 확증하기 어렵습니다.",
        "entities": "기계형 찰리 하나가 보이며 각진 고릴라형 머리, 흰 얼굴판, 점 형태의 눈, 선 형태의 입이 지정과 잘 맞습니다. 원형 가슴 링도 명확합니다. 낮의 하늘, 옥상 정원, 안테나, 스피커, 폭격 손상, 헬리콥터가 있습니다. 헬리콥터 주변에는 물리적 화면 없이 흰 그래프 패널들이 공중에 삽입되어 있습니다. 읽을 수 있는 문장은 보이지 않습니다.",
        "hard_violations": [
         "헬리콥터 주변의 에너지 그래프를 물리적 표시 장치 없이 공중에 뜬 평면 도표로 표현하여, 금지된 그래픽 오버레이가 화면에 노출되었습니다."
        ],
        "physics": "찰리의 발은 전경 난간과 화면 경계에 가려져 접지를 직접 확인할 수 없지만, 몸은 옥상 통로에 선 수직 자세여서 부유한다고 볼 근거는 없습니다. 헬리콥터는 지붕에 착륙해 있고 설비는 지지대에 고정되어 있습니다. 주변의 작은 공중 잔해와 먼지는 광선 분출에 따른 비산으로 해석할 수 있습니다. 반면 공중 그래프에는 화면, 기둥 또는 다른 물리적 지지체가 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.286,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.036,
    "B": 1.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 허공에 떠 있는 홀로그램 UI 및 그래프 오버레이 (leaked markers/diagrams/text)",
     "[gemini-pro] 광선 기둥의 베이스가 가슴이 아닌 캐릭터 뒤쪽 바닥에 위치함 (스테이징 지시 위반)",
     "[gpt-high] 헬리콥터 주변의 에너지 그래프를 물리적 표시 장치 없이 공중에 뜬 평면 도표로 표현하여, 금지된 그래픽 오버레이가 화면에 노출되었습니다."
    ],
    "B": [
     "[gemini-pro] 캐릭터의 머리가 완전히 누락됨 (physically impossible anatomy / 필수 엔티티 위반)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1036,
   "B": 1500
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1036,
    "verdict_ko": "캐릭터 외형과 배경은 잘 구현되었으나, 허공의 UI 그래픽 오버레이와 광선 시작점 오류가 치명적입니다.  ★위반: [gemini-pro] 허공에 떠 있는 홀로그램 UI 및 그래프 오버레이 (leaked markers/diagrams/text) / [gemini-pro] 광선 기둥의 베이스가 가슴이 아닌 캐릭터 뒤쪽 바닥에 위치함 (스테이징 지시 위반) / [gpt-high] 헬리콥터 주변의 에너지 그래프를 물리적 표시 장치 없이 공중에 뜬 평면 도표로 표현하여, 금지된 그래픽 오버레이가 화면에 노출되었습니다."
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "캐릭터의 머리가 완전히 누락되어 지정된 형태를 상실한 치명적인 해부학적 오류가 있습니다.  ★위반: [gemini-pro] 캐릭터의 머리가 완전히 누락됨 (physically impossible anatomy / 필수 엔티티 위반)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L138B02.png",
    "asset_id": "918aac9b-3638-42dc-acf3-037a906ea2ef",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf7a-3c22-7ea0-b3c2-d0d39618fa53",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "era_fail::9a7100a8da3f5060": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "era_fail::5561f9253f547356": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S90sh11": {
  "input_fingerprint": "70a9e74d6af58114",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 연구소 허공에 푸른빛으로 찰리의 거대한 홀로그램 형상이 떠오른 경이로운 전경.\n\nLOCATION (lock): In the open display area of the dark research room, where a blue robot hologram appears beside the large illuminated screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Large screen (Changed to an image of 제주도) — The image-bearing face is visible behind the hologram and shows 제주도; used as Establishes the changed visual environment without replacing the room itself; Occupied chairs (이현우 and 지소영 remain seated) — Their backs and left-side profiles are seen in the foreground; used as Provide a grounded scale reference beneath the holographic figure; Table bearing 찰리's parts (Parts remain laid out) — The tabletop is partially visible beyond the seated figures; used as Maintains the physical room's continuity beneath the projected encounter.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The blue holographic appearance provides selective luminous color within the otherwise dark room, with the seated figures still readable as physical observers.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The display has switched to Jeju and a holographic Charlie is present, with Amber and Raul appearing within the hologram sequence. Charlie's physical, disassembled components remain on the table in the dim research room.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (홀로그램) (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 연구소 허공에 푸른빛으로 찰리의 거대한 홀로그램 형상이 떠오른 경이로운 전경.\n\nLOCATION (lock): In the open display area of the dark research room, where a blue robot hologram appears beside the large illuminated screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Large screen (Changed to an image of 제주도) — The image-bearing face is visible behind the hologram and shows 제주도; used as Establishes the changed visual environment without replacing the room itself; Occupied chairs (이현우 and 지소영 remain seated) — Their backs and left-side profiles are seen in the foreground; used as Provide a grounded scale reference beneath the holographic figure; Table bearing 찰리's parts (Parts remain laid out) — The tabletop is partially visible beyond the seated figures; used as Maintains the physical room's continuity beneath the projected encounter.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The blue holographic appearance provides selective luminous color within the otherwise dark room, with the seated figures still readable as physical observers.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The display has switched to Jeju and a holographic Charlie is present, with Amber and Raul appearing within the hologram sequence. Charlie's physical, disassembled components remain on the table in the dim research room.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (홀로그램) (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 연구소 허공에 푸른빛으로 찰리의 거대한 홀로그램 형상이 떠오른 경이로운 전경.\n\nLOCATION (lock): In the open display area of the dark research room, where a blue robot hologram appears beside the large illuminated screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Large screen (Changed to an image of 제주도) — The image-bearing face is visible behind the hologram and shows 제주도; used as Establishes the changed visual environment without replacing the room itself; Occupied chairs (이현우 and 지소영 remain seated) — Their backs and left-side profiles are seen in the foreground; used as Provide a grounded scale reference beneath the holographic figure; Table bearing 찰리's parts (Parts remain laid out) — The tabletop is partially visible beyond the seated figures; used as Maintains the physical room's continuity beneath the projected encounter.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The blue holographic appearance provides selective luminous color within the otherwise dark room, with the seated figures still readable as physical observers.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The display has switched to Jeju and a holographic Charlie is present, with Amber and Raul appearing within the hologram sequence. Charlie's physical, disassembled components remain on the table in the dim research room.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (홀로그램) (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "두 인물이 중앙에 뜬 홀로그램과 대형 스크린을 향해 시선을 두고 있음.",
    "built_space": "연구실 내 금속 테이블, 대형 스크린, 좌측 수납장이 올바르게 배치됨. 전경에 2개의 의자와 인물이 위치함.",
    "entities": "제주도 이미지가 띄워진 스크린, 고릴라 형태의 푸른 기계 홀로그램(흰 얼굴판, 점 눈, 선 입), 앉아 있는 남녀 인물 2명.",
    "hard_violations": [],
    "physics": "홀로그램이 테이블 위 허공에 투사되어 떠 있음. 테이블 위에 부품들이 안정적으로 놓여 있고 인물들은 의자에 앉아 무게를 지탱함."
   },
   {
    "label": "B",
    "direction": "세 명의 인물이 중앙의 홀로그램과 스크린을 주시하고 있음.",
    "built_space": "연구실 배경에 대형 스크린과 중앙 테이블이 배치됨. 전경에 세 인물이 나란히 앉아 있음.",
    "entities": "고릴라 형태의 푸른 홀로그램, 제주도 지도 스크린, 인물 3명(레퍼런스와 동일한 오염된 흰 셔츠를 입은 인물 포함).",
    "hard_violations": [
     "[gemini-pro] 화면에 읽을 수 있는 'Jeju Island' 영어 텍스트 포함",
     "[gemini-pro] 레퍼런스 이미지의 인물(오염된 흰 셔츠)이 그대로 등장함",
     "[gemini-pro] 지정되지 않은 인물이 추가되어 총 3명이 배치됨",
     "[gpt-high] 두 명이어야 할 착석 관찰자에 여성을 한 명 추가하여 전경에 세 사람을 배치했습니다.",
     "[gpt-high] 배경 화면 중앙에 읽을 수 있는 영문 제주도 지명이 있어 읽히는 글자 금지 조건을 위반했습니다."
    ],
    "physics": "홀로그램이 허공에 떠 있으며, 테이블 위 부품들과 앉아 있는 인물들의 물리적 접촉이 정상적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프레이밍 지침과 찰리의 홀로그램 형태를 정확히 구현했으며, 읽을 수 있는 텍스트나 금지된 레퍼런스 인물 없이 지시사항을 잘 준수함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "화면에 금지된 텍스트('Jeju Island')가 노출되었고, 제외해야 할 레퍼런스의 인물을 그대로 포함했으며 불필요한 인물이 추가됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 중앙에 뜬 홀로그램과 대형 스크린을 향해 시선을 두고 있음.",
        "built_space": "연구실 내 금속 테이블, 대형 스크린, 좌측 수납장이 올바르게 배치됨. 전경에 2개의 의자와 인물이 위치함.",
        "entities": "제주도 이미지가 띄워진 스크린, 고릴라 형태의 푸른 기계 홀로그램(흰 얼굴판, 점 눈, 선 입), 앉아 있는 남녀 인물 2명.",
        "hard_violations": [],
        "physics": "홀로그램이 테이블 위 허공에 투사되어 떠 있음. 테이블 위에 부품들이 안정적으로 놓여 있고 인물들은 의자에 앉아 무게를 지탱함."
       },
       {
        "label": "B",
        "direction": "세 명의 인물이 중앙의 홀로그램과 스크린을 주시하고 있음.",
        "built_space": "연구실 배경에 대형 스크린과 중앙 테이블이 배치됨. 전경에 세 인물이 나란히 앉아 있음.",
        "entities": "고릴라 형태의 푸른 홀로그램, 제주도 지도 스크린, 인물 3명(레퍼런스와 동일한 오염된 흰 셔츠를 입은 인물 포함).",
        "hard_violations": [
         "화면에 읽을 수 있는 'Jeju Island' 영어 텍스트 포함",
         "레퍼런스 이미지의 인물(오염된 흰 셔츠)이 그대로 등장함",
         "지정되지 않은 인물이 추가되어 총 3명이 배치됨"
        ],
        "physics": "홀로그램이 허공에 떠 있으며, 테이블 위 부품들과 앉아 있는 인물들의 물리적 접촉이 정상적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프레이밍 지침과 찰리의 홀로그램 형태를 정확히 구현했으며, 읽을 수 있는 텍스트나 금지된 레퍼런스 인물 없이 지시사항을 잘 준수함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "화면에 금지된 텍스트('Jeju Island')가 노출되었고, 제외해야 할 레퍼런스의 인물을 그대로 포함했으며 불필요한 인물이 추가됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 중앙에 뜬 홀로그램과 대형 스크린을 향해 시선을 두고 있음.",
        "built_space": "연구실 내 금속 테이블, 대형 스크린, 좌측 수납장이 올바르게 배치됨. 전경에 2개의 의자와 인물이 위치함.",
        "entities": "제주도 이미지가 띄워진 스크린, 고릴라 형태의 푸른 기계 홀로그램(흰 얼굴판, 점 눈, 선 입), 앉아 있는 남녀 인물 2명.",
        "hard_violations": [],
        "physics": "홀로그램이 테이블 위 허공에 투사되어 떠 있음. 테이블 위에 부품들이 안정적으로 놓여 있고 인물들은 의자에 앉아 무게를 지탱함."
       },
       {
        "label": "B",
        "direction": "세 명의 인물이 중앙의 홀로그램과 스크린을 주시하고 있음.",
        "built_space": "연구실 배경에 대형 스크린과 중앙 테이블이 배치됨. 전경에 세 인물이 나란히 앉아 있음.",
        "entities": "고릴라 형태의 푸른 홀로그램, 제주도 지도 스크린, 인물 3명(레퍼런스와 동일한 오염된 흰 셔츠를 입은 인물 포함).",
        "hard_violations": [
         "화면에 읽을 수 있는 'Jeju Island' 영어 텍스트 포함",
         "레퍼런스 이미지의 인물(오염된 흰 셔츠)이 그대로 등장함",
         "지정되지 않은 인물이 추가되어 총 3명이 배치됨"
        ],
        "physics": "홀로그램이 허공에 떠 있으며, 테이블 위 부품들과 앉아 있는 인물들의 물리적 접촉이 정상적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "거대한 푸른 찰리와 연구실 전경은 구현했지만, 관찰자를 세 명으로 늘리고 화면에 읽히는 지명을 넣어 명시적 금지 조건을 위반했습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "두 관찰자가 앉은 전경, 부품 작업대, 찰리 뒤의 제주도 화면과 거대한 청색 홀로그램을 와이드 숏 안에 충실하게 배치했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 여성과 가운데 남성은 고개를 들어 찰리의 얼굴 쪽을 보고, 오른쪽 여성도 홀로그램 쪽으로 몸과 머리를 향합니다. 찰리는 관찰자 쪽을 내려다봅니다. 제주도 화면의 표시 면은 관찰자와 카메라를 향하며 홀로그램 뒤에서 보입니다.",
        "built_space": "전경에 사람이 앉은 의자가 세 개 있습니다. 요구된 두 관찰자보다 한 명 많습니다. 중앙 금속 작업대 한 개, 뒤쪽 대형 벽면 화면 한 개, 왼쪽 유리 수납장과 장비 단말, 오른쪽 선반과 보조 작업석은 참고 연구실의 배치를 대체로 유지합니다. 작업대 너머의 홀로그램과 그 뒤 화면이라는 앞뒤 관계는 맞습니다.",
        "entities": "찰리는 각진 머리 외장, 흰 고릴라형 얼굴판, 점 눈과 선 입을 지닌 거대한 푸른 로봇 홀로그램으로 보입니다. 전경에는 동아시아계로 보이는 성인 남성 한 명과 여성 두 명이 있어, 구도 지시의 이현우·지소영 두 명과 수가 맞지 않습니다. 남성의 얼룩진 흰옷은 이전 숏 인물의 옷을 그대로 이어받은 모습입니다. 금속 작업대에는 주황색 외장과 기계 부품들이 놓여 있습니다. 뒤 화면에는 제주도 형태의 섬과 읽을 수 있는 영문 지명이 표시됩니다.",
        "hard_violations": [
         "두 명이어야 할 착석 관찰자에 여성을 한 명 추가하여 전경에 세 사람을 배치했습니다.",
         "배경 화면 중앙에 읽을 수 있는 영문 제주도 지명이 있어 읽히는 글자 금지 조건을 위반했습니다."
        ],
        "physics": "세 사람은 각각 의자 좌판과 등받이로 지지되며, 기계 부품은 작업대 표면에 놓여 있습니다. 찰리는 반투명한 청색 발광 영상이고 작업대 중앙에서 올라오는 투사광이 보여, 물리적 몸이 지지 없이 떠 있는 상황이 아니라 명시된 가상 존재의 투영으로 읽힙니다."
       },
       {
        "label": "B",
        "direction": "왼쪽 여성은 오른쪽 위의 찰리 얼굴을 바라보고, 가운데 남성도 등과 머리를 홀로그램 쪽으로 향합니다. 여성의 왼쪽 옆얼굴은 분명하며 남성은 주로 뒷모습입니다. 찰리의 얼굴은 두 관찰자를 향해 약간 숙여져 있습니다. 제주도 화면의 표시 면은 그들 뒤가 아니라 찰리 뒤에 놓여 관찰자와 카메라 양쪽에서 볼 수 있습니다.",
        "built_space": "전경에 관찰자가 앉은 의자 두 개, 그 너머 금속 부품 작업대 한 개, 후면 대형 벽면 화면 한 개가 있습니다. 왼쪽 유리 수납장과 이동형 모니터·경사형 단말, 오른쪽 금속 선반과 빈 보조 작업석이 참고 장소의 구조와 재료를 유지합니다. 와이드 구도에서 관찰자, 부분적으로 보이는 작업대 상판, 대형 홀로그램, 후면 화면의 공간적 순서가 명확합니다.",
        "entities": "동아시아계로 보이는 성인 여성 한 명과 남성 한 명이 앉아 있어 지정된 두 관찰자 구성에 맞으며, 별도 인물 참고가 없어 정확한 신원은 확인할 수 없습니다. 찰리는 각진 기계 머리와 흰 고릴라형 얼굴판을 갖췄습니다. 눈은 단일 점보다 작은 점들의 원형 집합이고 입도 분절된 선이어서 설명과 약간 다릅니다. 화면에는 제주도 형태의 섬이 보이고, 작업대에는 주황색 외장과 분해된 기계 부품이 남아 있습니다. 명확하게 읽히는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "두 관찰자의 몸은 의자 좌판과 등받이에 지지되어 있고, 팔은 무릎이나 팔걸이 부근에 자연스럽게 놓여 있습니다. 분해 부품은 모두 금속 상판에 닿아 있습니다. 찰리 아래로 작업대 중앙에서 퍼지는 청색 투사광이 보이며 몸통도 반투명하여, 허공에 출현한 홀로그램이라는 설정에 맞습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "거대한 푸른 찰리와 연구실 전경은 구현했지만, 관찰자를 세 명으로 늘리고 화면에 읽히는 지명을 넣어 명시적 금지 조건을 위반했습니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "두 관찰자가 앉은 전경, 부품 작업대, 찰리 뒤의 제주도 화면과 거대한 청색 홀로그램을 와이드 숏 안에 충실하게 배치했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 여성과 가운데 남성은 고개를 들어 찰리의 얼굴 쪽을 보고, 오른쪽 여성도 홀로그램 쪽으로 몸과 머리를 향합니다. 찰리는 관찰자 쪽을 내려다봅니다. 제주도 화면의 표시 면은 관찰자와 카메라를 향하며 홀로그램 뒤에서 보입니다.",
        "built_space": "전경에 사람이 앉은 의자가 세 개 있습니다. 요구된 두 관찰자보다 한 명 많습니다. 중앙 금속 작업대 한 개, 뒤쪽 대형 벽면 화면 한 개, 왼쪽 유리 수납장과 장비 단말, 오른쪽 선반과 보조 작업석은 참고 연구실의 배치를 대체로 유지합니다. 작업대 너머의 홀로그램과 그 뒤 화면이라는 앞뒤 관계는 맞습니다.",
        "entities": "찰리는 각진 머리 외장, 흰 고릴라형 얼굴판, 점 눈과 선 입을 지닌 거대한 푸른 로봇 홀로그램으로 보입니다. 전경에는 동아시아계로 보이는 성인 남성 한 명과 여성 두 명이 있어, 구도 지시의 이현우·지소영 두 명과 수가 맞지 않습니다. 남성의 얼룩진 흰옷은 이전 숏 인물의 옷을 그대로 이어받은 모습입니다. 금속 작업대에는 주황색 외장과 기계 부품들이 놓여 있습니다. 뒤 화면에는 제주도 형태의 섬과 읽을 수 있는 영문 지명이 표시됩니다.",
        "hard_violations": [
         "두 명이어야 할 착석 관찰자에 여성을 한 명 추가하여 전경에 세 사람을 배치했습니다.",
         "배경 화면 중앙에 읽을 수 있는 영문 제주도 지명이 있어 읽히는 글자 금지 조건을 위반했습니다."
        ],
        "physics": "세 사람은 각각 의자 좌판과 등받이로 지지되며, 기계 부품은 작업대 표면에 놓여 있습니다. 찰리는 반투명한 청색 발광 영상이고 작업대 중앙에서 올라오는 투사광이 보여, 물리적 몸이 지지 없이 떠 있는 상황이 아니라 명시된 가상 존재의 투영으로 읽힙니다."
       },
       {
        "label": "A",
        "direction": "왼쪽 여성은 오른쪽 위의 찰리 얼굴을 바라보고, 가운데 남성도 등과 머리를 홀로그램 쪽으로 향합니다. 여성의 왼쪽 옆얼굴은 분명하며 남성은 주로 뒷모습입니다. 찰리의 얼굴은 두 관찰자를 향해 약간 숙여져 있습니다. 제주도 화면의 표시 면은 그들 뒤가 아니라 찰리 뒤에 놓여 관찰자와 카메라 양쪽에서 볼 수 있습니다.",
        "built_space": "전경에 관찰자가 앉은 의자 두 개, 그 너머 금속 부품 작업대 한 개, 후면 대형 벽면 화면 한 개가 있습니다. 왼쪽 유리 수납장과 이동형 모니터·경사형 단말, 오른쪽 금속 선반과 빈 보조 작업석이 참고 장소의 구조와 재료를 유지합니다. 와이드 구도에서 관찰자, 부분적으로 보이는 작업대 상판, 대형 홀로그램, 후면 화면의 공간적 순서가 명확합니다.",
        "entities": "동아시아계로 보이는 성인 여성 한 명과 남성 한 명이 앉아 있어 지정된 두 관찰자 구성에 맞으며, 별도 인물 참고가 없어 정확한 신원은 확인할 수 없습니다. 찰리는 각진 기계 머리와 흰 고릴라형 얼굴판을 갖췄습니다. 눈은 단일 점보다 작은 점들의 원형 집합이고 입도 분절된 선이어서 설명과 약간 다릅니다. 화면에는 제주도 형태의 섬이 보이고, 작업대에는 주황색 외장과 분해된 기계 부품이 남아 있습니다. 명확하게 읽히는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "두 관찰자의 몸은 의자 좌판과 등받이에 지지되어 있고, 팔은 무릎이나 팔걸이 부근에 자연스럽게 놓여 있습니다. 분해 부품은 모두 금속 상판에 닿아 있습니다. 찰리 아래로 작업대 중앙에서 퍼지는 청색 투사광이 보이며 몸통도 반투명하여, 허공에 출현한 홀로그램이라는 설정에 맞습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.651
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.401
   },
   "violations": {
    "B": [
     "[gemini-pro] 화면에 읽을 수 있는 'Jeju Island' 영어 텍스트 포함",
     "[gemini-pro] 레퍼런스 이미지의 인물(오염된 흰 셔츠)이 그대로 등장함",
     "[gemini-pro] 지정되지 않은 인물이 추가되어 총 3명이 배치됨",
     "[gpt-high] 두 명이어야 할 착석 관찰자에 여성을 한 명 추가하여 전경에 세 사람을 배치했습니다.",
     "[gpt-high] 배경 화면 중앙에 읽을 수 있는 영문 제주도 지명이 있어 읽히는 글자 금지 조건을 위반했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 401
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프레이밍 지침과 찰리의 홀로그램 형태를 정확히 구현했으며, 읽을 수 있는 텍스트나 금지된 레퍼런스 인물 없이 지시사항을 잘 준수함."
   },
   {
    "label": "B",
    "score": 401,
    "verdict_ko": "화면에 금지된 텍스트('Jeju Island')가 노출되었고, 제외해야 할 레퍼런스의 인물을 그대로 포함했으며 불필요한 인물이 추가됨.  ★위반: [gemini-pro] 화면에 읽을 수 있는 'Jeju Island' 영어 텍스트 포함 / [gemini-pro] 레퍼런스 이미지의 인물(오염된 흰 셔츠)이 그대로 등장함 / [gemini-pro] 지정되지 않은 인물이 추가되어 총 3명이 배치됨 / [gpt-high] 두 명이어야 할 착석 관찰자에 여성을 한 명 추가하여 전경에 세 사람을 배치했습니다. / [gpt-high] 배경 화면 중앙에 읽을 수 있는 영문 제주도 지명이 있어 읽히는 글자 금지 조건을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S90sh3_sel.png",
    "asset_id": "dadc1d1f-715b-4ec0-aa0d-a1015952df71",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf8c-bcfc-758a-a0e9-17ef31e4b517",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S90sh3"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S90sh18": {
  "input_fingerprint": "ee9bb910b1a3a14e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수술대 위에 온전한 몸으로 누운 찰리가 두 눈을 지그시 감은 평온한 얼굴 클로즈업.\n\nLOCATION (lock): On the reconstruction table inside the large dim research room, with the nearby display and holographic light illuminating the reassembled robot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Operating table (Supporting 찰리's assembled body) — A narrow top edge remains visible beside the head and recedes along the body; used as Confirms the physical reconstruction setting beneath the peaceful close-up.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Within the dark laboratory, gentle tonal separation reveals the closed eyelids and facial contours without carrying over the hologram's blue appearance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's broken physical components are being fitted back together on the operating table. His eyes are closed; successful reactivation has not yet been established.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (홀로그램) (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수술대 위에 온전한 몸으로 누운 찰리가 두 눈을 지그시 감은 평온한 얼굴 클로즈업.\n\nLOCATION (lock): On the reconstruction table inside the large dim research room, with the nearby display and holographic light illuminating the reassembled robot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Operating table (Supporting 찰리's assembled body) — A narrow top edge remains visible beside the head and recedes along the body; used as Confirms the physical reconstruction setting beneath the peaceful close-up.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Within the dark laboratory, gentle tonal separation reveals the closed eyelids and facial contours without carrying over the hologram's blue appearance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's broken physical components are being fitted back together on the operating table. His eyes are closed; successful reactivation has not yet been established.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (홀로그램) (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수술대 위에 온전한 몸으로 누운 찰리가 두 눈을 지그시 감은 평온한 얼굴 클로즈업.\n\nLOCATION (lock): On the reconstruction table inside the large dim research room, with the nearby display and holographic light illuminating the reassembled robot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Operating table (Supporting 찰리's assembled body) — A narrow top edge remains visible beside the head and recedes along the body; used as Confirms the physical reconstruction setting beneath the peaceful close-up.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Within the dark laboratory, gentle tonal separation reveals the closed eyelids and facial contours without carrying over the hologram's blue appearance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's broken physical components are being fitted back together on the operating table. His eyes are closed; successful reactivation has not yet been established.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (홀로그램) (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "지그시 감은 두 눈이 평온한 상태로 화면을 향함.",
    "built_space": "레퍼런스와 일치하는 스테인리스 금속 재질의 수술대 모서리가 머리 곁을 따라 뒤로 이어지며, 어두운 연구실 배경이 흐리게 보임.",
    "entities": "고릴라형 기계 머리, 각진 머리 외장, 흰색 얼굴판을 가진 찰리가 두 눈을 감은 모습이 정확히 묘사됨.",
    "hard_violations": [],
    "physics": "찰리의 머리와 조립된 몸체 일부가 단단한 금속 수술대 표면에 안정적으로 놓여 지지됨."
   },
   {
    "label": "B",
    "direction": "정면을 향해 반듯이 누운 상태로 눈을 감고 있음.",
    "built_space": "어두운 연구실 배경에 캐비닛이 보이나, 캐릭터가 누워있는 곳이 레퍼런스의 평평한 금속 작업대가 아닌 가죽/천 재질의 쿠션 침대로 묘사됨.",
    "entities": "고릴라 형태의 기계 머리와 흰 얼굴판 등 찰리의 외형과 감은 눈이 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 지정된 레퍼런스 장소(스테인리스 금속 작업대)와 완전히 다른 형태와 재질의 쿠션 침대가 등장함 (장소 일관성 위반)."
    ],
    "physics": "머리가 푹신한 침대 쿠션 위에 놓여 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 금속 작업대와 동일한 재질의 수술대 모서리를 정확히 구현했으며, 프롬프트의 찰리 외형 묘사와 지그시 눈을 감은 클로즈업 샷을 훌륭하게 반영함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 얼굴 형태는 잘 표현되었으나, 레퍼런스의 금속 작업대가 아닌 쿠션이 있는 침대를 그려 장소 일관성(Location Lock)을 크게 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지그시 감은 두 눈이 평온한 상태로 화면을 향함.",
        "built_space": "레퍼런스와 일치하는 스테인리스 금속 재질의 수술대 모서리가 머리 곁을 따라 뒤로 이어지며, 어두운 연구실 배경이 흐리게 보임.",
        "entities": "고릴라형 기계 머리, 각진 머리 외장, 흰색 얼굴판을 가진 찰리가 두 눈을 감은 모습이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "찰리의 머리와 조립된 몸체 일부가 단단한 금속 수술대 표면에 안정적으로 놓여 지지됨."
       },
       {
        "label": "B",
        "direction": "정면을 향해 반듯이 누운 상태로 눈을 감고 있음.",
        "built_space": "어두운 연구실 배경에 캐비닛이 보이나, 캐릭터가 누워있는 곳이 레퍼런스의 평평한 금속 작업대가 아닌 가죽/천 재질의 쿠션 침대로 묘사됨.",
        "entities": "고릴라 형태의 기계 머리와 흰 얼굴판 등 찰리의 외형과 감은 눈이 묘사됨.",
        "hard_violations": [
         "지정된 레퍼런스 장소(스테인리스 금속 작업대)와 완전히 다른 형태와 재질의 쿠션 침대가 등장함 (장소 일관성 위반)."
        ],
        "physics": "머리가 푹신한 침대 쿠션 위에 놓여 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 금속 작업대와 동일한 재질의 수술대 모서리를 정확히 구현했으며, 프롬프트의 찰리 외형 묘사와 지그시 눈을 감은 클로즈업 샷을 훌륭하게 반영함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리의 얼굴 형태는 잘 표현되었으나, 레퍼런스의 금속 작업대가 아닌 쿠션이 있는 침대를 그려 장소 일관성(Location Lock)을 크게 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "지그시 감은 두 눈이 평온한 상태로 화면을 향함.",
        "built_space": "레퍼런스와 일치하는 스테인리스 금속 재질의 수술대 모서리가 머리 곁을 따라 뒤로 이어지며, 어두운 연구실 배경이 흐리게 보임.",
        "entities": "고릴라형 기계 머리, 각진 머리 외장, 흰색 얼굴판을 가진 찰리가 두 눈을 감은 모습이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "찰리의 머리와 조립된 몸체 일부가 단단한 금속 수술대 표면에 안정적으로 놓여 지지됨."
       },
       {
        "label": "B",
        "direction": "정면을 향해 반듯이 누운 상태로 눈을 감고 있음.",
        "built_space": "어두운 연구실 배경에 캐비닛이 보이나, 캐릭터가 누워있는 곳이 레퍼런스의 평평한 금속 작업대가 아닌 가죽/천 재질의 쿠션 침대로 묘사됨.",
        "entities": "고릴라 형태의 기계 머리와 흰 얼굴판 등 찰리의 외형과 감은 눈이 묘사됨.",
        "hard_violations": [
         "지정된 레퍼런스 장소(스테인리스 금속 작업대)와 완전히 다른 형태와 재질의 쿠션 침대가 등장함 (장소 일관성 위반)."
        ],
        "physics": "머리가 푹신한 침대 쿠션 위에 놓여 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "감은 눈과 평온한 표정은 맞지만, 얼굴보다 넓은 패드와 흉갑이 두드러지고 금속 재조립 작업대의 장소 연속성이 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "평온하게 눈을 감은 얼굴의 밀착 클로즈업, 머리를 받치는 금속 작업대의 가장자리, 낡은 기계 몸체가 요청한 순간과 장소를 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴은 위쪽의 카메라를 정면으로 향하며 두 눈은 닫혀 있어 시선 대상이 없다. 입도 닫혀 있고 이동이나 다른 대상을 향한 동작은 없다.",
        "built_space": "찰리 아래에는 넓은 밝은색 패드와 양쪽 금속 테두리가 있는 받침대 하나가 보인다. 머리는 중앙에, 흉갑은 화면 하단 전경에 놓인다. 뒤에는 서랍장 하나와 유리문 수납장 구역이 보인다. 참고의 평평한 스테인리스 작업대와 달리 패드가 넓게 노출되어, 머리 옆에 좁은 상판 가장자리만 남기는 구도와 장소 재현이 약하다. 불가능한 반사는 보이지 않는다.",
        "entities": "등장 개체는 고릴라형 로봇 찰리 하나뿐이다. 각진 금속 머리 외장, 흰 얼굴판, 선형의 닫힌 입이 있다. 두 눈은 감겨 있지만 점 형태의 단순한 눈보다 입체적인 생물형 눈꺼풀로 표현되었다. 연결된 목과 흉갑 일부가 보이며, 나머지 몸은 프레임 밖이라 완전한 조립 여부는 판단할 수 없다. 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리 뒤쪽과 몸통은 패드가 있는 받침대에 기대어 지지된다. 목 아래의 기계 구조가 머리와 흉갑 사이를 연결한다. 떠 있는 부품이나 지지 없이 공중에 놓인 신체는 보이지 않으며, 정지한 자세는 물리적으로 가능하다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴은 화면 오른쪽으로 기울어진 채 위쪽과 카메라 쪽을 향한다. 양쪽 눈은 완전히 닫혀 있어 응시 대상이 없고, 입과 얼굴 근육 표현도 차분하다. 몸은 머리에서 화면 왼쪽 바깥으로 이어지며 이동 동작은 없다.",
        "built_space": "스테인리스 작업대 하나가 머리와 몸통을 받친다. 머리 옆의 금속 테두리가 오른쪽 뒤로 이어져 재조립 장소를 확인시킨다. 흐린 배경에는 수납 선반 구역, 세로형 모니터 하나, 경사진 화면을 가진 장비 하나와 작업대 위 소형 부품들이 보인다. 참고의 금속 작업대와 어두운 연구실 설비에 잘 부합하며, 중복된 작업대나 불가능한 반사는 보이지 않는다.",
        "entities": "고릴라형 기계 머리의 찰리 하나만 등장한다. 각진 회색 외장, 흰 얼굴판, 닫힌 선형 입이 확인된다. 눈은 감겨 있으나 점 형태보다는 주름이 있는 입체적 눈꺼풀이다. 화면 왼쪽에는 목 연결부와 낡은 주황색·금속색 몸체 일부가 이어져 참고의 부품 재질과 색을 계승한다. 전신은 클로즈업 밖에 있어 평가 대상이 아니며, 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리 외장 하단이 금속 상판에 닿아 있고, 목의 기계 연결부가 왼쪽 몸통과 이어진다. 몸통도 같은 작업대 위에 놓여 있어 기울어진 머리와 몸을 지지한다. 배경 부품들도 상판 위에 놓여 있으며, 지지 없는 부유나 불가능한 동작은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "감은 눈과 평온한 표정은 맞지만, 얼굴보다 넓은 패드와 흉갑이 두드러지고 금속 재조립 작업대의 장소 연속성이 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "평온하게 눈을 감은 얼굴의 밀착 클로즈업, 머리를 받치는 금속 작업대의 가장자리, 낡은 기계 몸체가 요청한 순간과 장소를 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴은 위쪽의 카메라를 정면으로 향하며 두 눈은 닫혀 있어 시선 대상이 없다. 입도 닫혀 있고 이동이나 다른 대상을 향한 동작은 없다.",
        "built_space": "찰리 아래에는 넓은 밝은색 패드와 양쪽 금속 테두리가 있는 받침대 하나가 보인다. 머리는 중앙에, 흉갑은 화면 하단 전경에 놓인다. 뒤에는 서랍장 하나와 유리문 수납장 구역이 보인다. 참고의 평평한 스테인리스 작업대와 달리 패드가 넓게 노출되어, 머리 옆에 좁은 상판 가장자리만 남기는 구도와 장소 재현이 약하다. 불가능한 반사는 보이지 않는다.",
        "entities": "등장 개체는 고릴라형 로봇 찰리 하나뿐이다. 각진 금속 머리 외장, 흰 얼굴판, 선형의 닫힌 입이 있다. 두 눈은 감겨 있지만 점 형태의 단순한 눈보다 입체적인 생물형 눈꺼풀로 표현되었다. 연결된 목과 흉갑 일부가 보이며, 나머지 몸은 프레임 밖이라 완전한 조립 여부는 판단할 수 없다. 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리 뒤쪽과 몸통은 패드가 있는 받침대에 기대어 지지된다. 목 아래의 기계 구조가 머리와 흉갑 사이를 연결한다. 떠 있는 부품이나 지지 없이 공중에 놓인 신체는 보이지 않으며, 정지한 자세는 물리적으로 가능하다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴은 화면 오른쪽으로 기울어진 채 위쪽과 카메라 쪽을 향한다. 양쪽 눈은 완전히 닫혀 있어 응시 대상이 없고, 입과 얼굴 근육 표현도 차분하다. 몸은 머리에서 화면 왼쪽 바깥으로 이어지며 이동 동작은 없다.",
        "built_space": "스테인리스 작업대 하나가 머리와 몸통을 받친다. 머리 옆의 금속 테두리가 오른쪽 뒤로 이어져 재조립 장소를 확인시킨다. 흐린 배경에는 수납 선반 구역, 세로형 모니터 하나, 경사진 화면을 가진 장비 하나와 작업대 위 소형 부품들이 보인다. 참고의 금속 작업대와 어두운 연구실 설비에 잘 부합하며, 중복된 작업대나 불가능한 반사는 보이지 않는다.",
        "entities": "고릴라형 기계 머리의 찰리 하나만 등장한다. 각진 회색 외장, 흰 얼굴판, 닫힌 선형 입이 확인된다. 눈은 감겨 있으나 점 형태보다는 주름이 있는 입체적 눈꺼풀이다. 화면 왼쪽에는 목 연결부와 낡은 주황색·금속색 몸체 일부가 이어져 참고의 부품 재질과 색을 계승한다. 전신은 클로즈업 밖에 있어 평가 대상이 아니며, 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리 외장 하단이 금속 상판에 닿아 있고, 목의 기계 연결부가 왼쪽 몸통과 이어진다. 몸통도 같은 작업대 위에 놓여 있어 기울어진 머리와 몸을 지지한다. 배경 부품들도 상판 위에 놓여 있으며, 지지 없는 부유나 불가능한 동작은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.095
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.845
   },
   "violations": {
    "B": [
     "[gemini-pro] 지정된 레퍼런스 장소(스테인리스 금속 작업대)와 완전히 다른 형태와 재질의 쿠션 침대가 등장함 (장소 일관성 위반)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 845
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스의 금속 작업대와 동일한 재질의 수술대 모서리를 정확히 구현했으며, 프롬프트의 찰리 외형 묘사와 지그시 눈을 감은 클로즈업 샷을 훌륭하게 반영함."
   },
   {
    "label": "B",
    "score": 845,
    "verdict_ko": "찰리의 얼굴 형태는 잘 표현되었으나, 레퍼런스의 금속 작업대가 아닌 쿠션이 있는 침대를 그려 장소 일관성(Location Lock)을 크게 위반함.  ★위반: [gemini-pro] 지정된 레퍼런스 장소(스테인리스 금속 작업대)와 완전히 다른 형태와 재질의 쿠션 침대가 등장함 (장소 일관성 위반)."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S90sh3_sel.png",
    "asset_id": "dadc1d1f-715b-4ec0-aa0d-a1015952df71",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf90-81ea-7a8e-8f5b-588db7f47176",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S90sh3"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "era_fail::caee92de16d80d21": {
  "stage": "assess",
  "error": "AppError('scene_image_pipeline 정지 요청 (2026-09-18 06:11:01.465475+00:00) — 유료 이미지 호출 직전에서 멈춘다.')",
  "attempts": 1
 },
 "S91sh7": {
  "input_fingerprint": "342f27e3b7091d5b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물 밖으로 텅 빈 두 손을 들어 올린 채 입가 금속 부품을 아래로 푹 떨군 시무룩한 찰리의 자태.\n\nLOCATION (lock): At the water's edge beside the rocky island outcrop, standing in the shallow sea after an unsuccessful catch. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sea surface (Below Charlie's raised, empty hands) — Seen from a gently elevated oblique angle beneath the hands; used as Provides a narrow lower-frame context for the unsuccessful fishing attempt without competing with the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps the empty palms and lowered mouth components legible, with restrained contrast supporting the quietly disappointed mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Clear seawater reveals several fish beneath the surface beside the island rock in daylight. Food is being arranged on the dining table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물 밖으로 텅 빈 두 손을 들어 올린 채 입가 금속 부품을 아래로 푹 떨군 시무룩한 찰리의 자태.\n\nLOCATION (lock): At the water's edge beside the rocky island outcrop, standing in the shallow sea after an unsuccessful catch. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sea surface (Below Charlie's raised, empty hands) — Seen from a gently elevated oblique angle beneath the hands; used as Provides a narrow lower-frame context for the unsuccessful fishing attempt without competing with the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps the empty palms and lowered mouth components legible, with restrained contrast supporting the quietly disappointed mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Clear seawater reveals several fish beneath the surface beside the island rock in daylight. Food is being arranged on the dining table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물 밖으로 텅 빈 두 손을 들어 올린 채 입가 금속 부품을 아래로 푹 떨군 시무룩한 찰리의 자태.\n\nLOCATION (lock): At the water's edge beside the rocky island outcrop, standing in the shallow sea after an unsuccessful catch. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sea surface (Below Charlie's raised, empty hands) — Seen from a gently elevated oblique angle beneath the hands; used as Provides a narrow lower-frame context for the unsuccessful fishing attempt without competing with the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light keeps the empty palms and lowered mouth components legible, with restrained contrast supporting the quietly disappointed mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Clear seawater reveals several fish beneath the surface beside the island rock in daylight. Food is being arranged on the dining table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "캐릭터는 화면 왼쪽 아래를 응시하고 있으며, 두 손을 들어 올리고 입가의 금속 부품 하나가 아래로 늘어져 있음.",
    "built_space": "레퍼런스와 동일한 바위섬과 얕은 바다가 배경으로 펼쳐져 있으며 인물은 물 속에 서 있음.",
    "entities": "고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈을 가진 찰리가 물 속에 서 있고 물 아래로 물고기들이 보임.",
    "hard_violations": [
     "[gpt-high] 화면 오른쪽 팔뚝의 노란 표찰에 판독 가능한 숫자 표기가 노출되어, 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반합니다."
    ],
    "physics": "물 속에 자연스럽게 서 있으며, 들어 올린 팔과 늘어진 금속 부품이 중력에 맞게 표현됨."
   },
   {
    "label": "B",
    "direction": "캐릭터는 정면을 향해 시선을 두고 있으며, 양손을 높이 들고 입가의 금속 부품들이 아래로 늘어져 있음.",
    "built_space": "레퍼런스와 일치하는 바위 해변 배경에 인물이 물 속에 배치됨.",
    "entities": "고릴라형 기계이지만 프롬프트가 요구한 흰 얼굴판과 점 형태의 눈이 아닌 전체가 회색인 조형적 얼굴을 가짐. 물 아래 물고기들이 존재함.",
    "hard_violations": [],
    "physics": "물 속에 서서 팔을 올린 자세가 안정적이며, 물이 뚝뚝 떨어지는 묘사와 부품의 매달림이 물리적으로 타당함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 캐릭터 디자인(흰 얼굴판, 점 형태의 눈)과 약간 비스듬한 카메라 앵글을 잘 구현하여 프롬프트에 충실합니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "입가 부품이 떨어진 모습은 표현되었으나, 캐릭터의 핵심 디자인(흰 얼굴판, 점 형태의 눈)이 누락되었고 정면 구도로 촬영되어 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캐릭터는 화면 왼쪽 아래를 응시하고 있으며, 두 손을 들어 올리고 입가의 금속 부품 하나가 아래로 늘어져 있음.",
        "built_space": "레퍼런스와 동일한 바위섬과 얕은 바다가 배경으로 펼쳐져 있으며 인물은 물 속에 서 있음.",
        "entities": "고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈을 가진 찰리가 물 속에 서 있고 물 아래로 물고기들이 보임.",
        "hard_violations": [],
        "physics": "물 속에 자연스럽게 서 있으며, 들어 올린 팔과 늘어진 금속 부품이 중력에 맞게 표현됨."
       },
       {
        "label": "B",
        "direction": "캐릭터는 정면을 향해 시선을 두고 있으며, 양손을 높이 들고 입가의 금속 부품들이 아래로 늘어져 있음.",
        "built_space": "레퍼런스와 일치하는 바위 해변 배경에 인물이 물 속에 배치됨.",
        "entities": "고릴라형 기계이지만 프롬프트가 요구한 흰 얼굴판과 점 형태의 눈이 아닌 전체가 회색인 조형적 얼굴을 가짐. 물 아래 물고기들이 존재함.",
        "hard_violations": [],
        "physics": "물 속에 서서 팔을 올린 자세가 안정적이며, 물이 뚝뚝 떨어지는 묘사와 부품의 매달림이 물리적으로 타당함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 캐릭터 디자인(흰 얼굴판, 점 형태의 눈)과 약간 비스듬한 카메라 앵글을 잘 구현하여 프롬프트에 충실합니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "입가 부품이 떨어진 모습은 표현되었으나, 캐릭터의 핵심 디자인(흰 얼굴판, 점 형태의 눈)이 누락되었고 정면 구도로 촬영되어 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "캐릭터는 화면 왼쪽 아래를 응시하고 있으며, 두 손을 들어 올리고 입가의 금속 부품 하나가 아래로 늘어져 있음.",
        "built_space": "레퍼런스와 동일한 바위섬과 얕은 바다가 배경으로 펼쳐져 있으며 인물은 물 속에 서 있음.",
        "entities": "고릴라형 기계 머리, 흰 얼굴판, 점 형태의 눈을 가진 찰리가 물 속에 서 있고 물 아래로 물고기들이 보임.",
        "hard_violations": [],
        "physics": "물 속에 자연스럽게 서 있으며, 들어 올린 팔과 늘어진 금속 부품이 중력에 맞게 표현됨."
       },
       {
        "label": "B",
        "direction": "캐릭터는 정면을 향해 시선을 두고 있으며, 양손을 높이 들고 입가의 금속 부품들이 아래로 늘어져 있음.",
        "built_space": "레퍼런스와 일치하는 바위 해변 배경에 인물이 물 속에 배치됨.",
        "entities": "고릴라형 기계이지만 프롬프트가 요구한 흰 얼굴판과 점 형태의 눈이 아닌 전체가 회색인 조형적 얼굴을 가짐. 물 아래 물고기들이 존재함.",
        "hard_violations": [],
        "physics": "물 속에 서서 팔을 올린 자세가 안정적이며, 물이 뚝뚝 떨어지는 묘사와 부품의 매달림이 물리적으로 타당함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "빈 두 손을 높이 들고 입가 부품을 아래로 늘어뜨린 미디엄 숏을 더 충실히 구현하지만, 흰 얼굴판·각진 외장·점 형태 눈이라는 정체성 묘사는 부족합니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "흰 얼굴판과 빈손을 내려다보는 연기는 적절하지만, 팔의 판독 가능한 숫자 표기가 무문자 조건을 위반하며 수면도 하단의 좁은 맥락 이상으로 넓게 차지합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개와 눈은 아래쪽 중앙의 수면을 향하며 시무룩한 인상을 줍니다. 양손은 머리 양옆으로 올라가 있고 빈 손바닥이 카메라 쪽으로 열려 있습니다. 입 양옆의 연결 금속 부품은 가슴 방향으로 축 늘어져 있어 지시된 하향 방향과 맞습니다.",
        "built_space": "인공 구조물이나 고정 설비는 없습니다. 오른쪽의 밝은 절벽과 소나무, 왼쪽 먼 바위섬, 맑은 물속 바위가 참고 장소의 배치를 유지합니다. 찰리는 중앙의 얕은 바다에 허리까지 잠겨 있으며 상체 중심의 미디엄 숏입니다. 아래쪽 수면을 내려다볼 수 있지만 전체 카메라 축은 거의 정면입니다.",
        "entities": "찰리 한 명만 있고 참고 사진의 아이들이나 그물은 없습니다. 고릴라형 기계 머리와 금속 몸체, 비어 있는 두 손, 아래로 내려온 입가 부품은 확인됩니다. 다만 머리 윗부분은 둥글고 얼굴은 회색 금속이며, 눈도 단순한 점보다는 주름진 고릴라 눈처럼 표현되어 지정된 각진 외장·흰 얼굴판·점 형태 눈과 다릅니다. 하단 수면 아래에는 여러 물고기가 보입니다. 식탁은 프레임에 없으며 읽을 수 있는 문자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "양팔은 어깨·팔꿈치·손목 관절로 연결되어 빈손을 들어 올린 자세를 지탱합니다. 입가 부품도 얼굴에 연결된 관절에서 매달려 있습니다. 하체와 발은 물에 가려져 직접적인 바닥 접촉은 보이지 않지만, 얕은 물에 선 몸의 수직 정렬이 자연스럽고 공중 부유 징후는 없습니다. 손과 팔에서 떨어지는 물방울, 수중 물고기, 수면의 일그러진 반사도 물리적으로 타당합니다."
       },
       {
        "label": "B",
        "direction": "찰리는 고개를 숙여 화면 왼쪽의 빈 손바닥 쪽을 바라봅니다. 양손은 가슴 앞에서 위로 열려 있고 물 밖에 나와 있습니다. 입 한쪽에 연결된 금속 부품은 턱 아래로 처져 있어 실패 후 낙담한 동작과 방향이 맞습니다.",
        "built_space": "인공 구조물이나 고정 설비는 없습니다. 오른쪽 바위 절벽과 소나무, 왼쪽 먼 바위섬이 참고 장소와 대응합니다. 찰리는 화면 오른쪽의 얕은 물에 서 있습니다. 상체 미디엄 숏이지만 왼쪽의 넓은 바다 여백과 아래쪽 수중 영역이 커서, 수면을 하단의 좁은 맥락으로 제한하라는 구도 지시에서는 멀어집니다.",
        "entities": "찰리 한 명만 보이며 고릴라형 기계 머리, 흰 얼굴판, 작은 눈과 선형 입이 확인됩니다. 얼굴판은 각지지만 머리 외피의 전체 윤곽은 둥급니다. 두 손은 비어 있고 아래쪽 맑은 물에는 여러 물고기가 있습니다. 식탁은 보이지 않습니다. 화면 오른쪽 팔뚝의 노란 표찰에는 읽을 수 있는 숫자 표기가 있어 무문자 조건과 충돌합니다.",
        "hard_violations": [
         "화면 오른쪽 팔뚝의 노란 표찰에 판독 가능한 숫자 표기가 노출되어, 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반합니다."
        ],
        "physics": "팔과 손은 연결된 기계 관절이 지탱하며, 팔꿈치를 굽혀 물 밖으로 손을 든 자세가 가능합니다. 처진 입가 부품에는 얼굴과의 연결점이 보입니다. 발은 수중에 가려졌지만 몸통은 곧게 서 있고 수면이 허리 부근을 감싸므로 지지 없이 떠 있는 몸으로 보이지 않습니다. 물고기는 수중에 있으며 반사와 굴절도 장면의 시점에 부합합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "빈 두 손을 높이 들고 입가 부품을 아래로 늘어뜨린 미디엄 숏을 더 충실히 구현하지만, 흰 얼굴판·각진 외장·점 형태 눈이라는 정체성 묘사는 부족합니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "흰 얼굴판과 빈손을 내려다보는 연기는 적절하지만, 팔의 판독 가능한 숫자 표기가 무문자 조건을 위반하며 수면도 하단의 좁은 맥락 이상으로 넓게 차지합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "고개와 눈은 아래쪽 중앙의 수면을 향하며 시무룩한 인상을 줍니다. 양손은 머리 양옆으로 올라가 있고 빈 손바닥이 카메라 쪽으로 열려 있습니다. 입 양옆의 연결 금속 부품은 가슴 방향으로 축 늘어져 있어 지시된 하향 방향과 맞습니다.",
        "built_space": "인공 구조물이나 고정 설비는 없습니다. 오른쪽의 밝은 절벽과 소나무, 왼쪽 먼 바위섬, 맑은 물속 바위가 참고 장소의 배치를 유지합니다. 찰리는 중앙의 얕은 바다에 허리까지 잠겨 있으며 상체 중심의 미디엄 숏입니다. 아래쪽 수면을 내려다볼 수 있지만 전체 카메라 축은 거의 정면입니다.",
        "entities": "찰리 한 명만 있고 참고 사진의 아이들이나 그물은 없습니다. 고릴라형 기계 머리와 금속 몸체, 비어 있는 두 손, 아래로 내려온 입가 부품은 확인됩니다. 다만 머리 윗부분은 둥글고 얼굴은 회색 금속이며, 눈도 단순한 점보다는 주름진 고릴라 눈처럼 표현되어 지정된 각진 외장·흰 얼굴판·점 형태 눈과 다릅니다. 하단 수면 아래에는 여러 물고기가 보입니다. 식탁은 프레임에 없으며 읽을 수 있는 문자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "양팔은 어깨·팔꿈치·손목 관절로 연결되어 빈손을 들어 올린 자세를 지탱합니다. 입가 부품도 얼굴에 연결된 관절에서 매달려 있습니다. 하체와 발은 물에 가려져 직접적인 바닥 접촉은 보이지 않지만, 얕은 물에 선 몸의 수직 정렬이 자연스럽고 공중 부유 징후는 없습니다. 손과 팔에서 떨어지는 물방울, 수중 물고기, 수면의 일그러진 반사도 물리적으로 타당합니다."
       },
       {
        "label": "A",
        "direction": "찰리는 고개를 숙여 화면 왼쪽의 빈 손바닥 쪽을 바라봅니다. 양손은 가슴 앞에서 위로 열려 있고 물 밖에 나와 있습니다. 입 한쪽에 연결된 금속 부품은 턱 아래로 처져 있어 실패 후 낙담한 동작과 방향이 맞습니다.",
        "built_space": "인공 구조물이나 고정 설비는 없습니다. 오른쪽 바위 절벽과 소나무, 왼쪽 먼 바위섬이 참고 장소와 대응합니다. 찰리는 화면 오른쪽의 얕은 물에 서 있습니다. 상체 미디엄 숏이지만 왼쪽의 넓은 바다 여백과 아래쪽 수중 영역이 커서, 수면을 하단의 좁은 맥락으로 제한하라는 구도 지시에서는 멀어집니다.",
        "entities": "찰리 한 명만 보이며 고릴라형 기계 머리, 흰 얼굴판, 작은 눈과 선형 입이 확인됩니다. 얼굴판은 각지지만 머리 외피의 전체 윤곽은 둥급니다. 두 손은 비어 있고 아래쪽 맑은 물에는 여러 물고기가 있습니다. 식탁은 보이지 않습니다. 화면 오른쪽 팔뚝의 노란 표찰에는 읽을 수 있는 숫자 표기가 있어 무문자 조건과 충돌합니다.",
        "hard_violations": [
         "화면 오른쪽 팔뚝의 노란 표찰에 판독 가능한 숫자 표기가 노출되어, 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반합니다."
        ],
        "physics": "팔과 손은 연결된 기계 관절이 지탱하며, 팔꿈치를 굽혀 물 밖으로 손을 든 자세가 가능합니다. 처진 입가 부품에는 얼굴과의 연결점이 보입니다. 발은 수중에 가려졌지만 몸통은 곧게 서 있고 수면이 허리 부근을 감싸므로 지지 없이 떠 있는 몸으로 보이지 않습니다. 물고기는 수중에 있으며 반사와 굴절도 장면의 시점에 부합합니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.321,
    "B": 1.571
   },
   "violations": {
    "A": [
     "[gpt-high] 화면 오른쪽 팔뚝의 노란 표찰에 판독 가능한 숫자 표기가 노출되어, 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1321,
   "B": 1571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1321,
    "verdict_ko": "지정된 캐릭터 디자인(흰 얼굴판, 점 형태의 눈)과 약간 비스듬한 카메라 앵글을 잘 구현하여 프롬프트에 충실합니다.  ★위반: [gpt-high] 화면 오른쪽 팔뚝의 노란 표찰에 판독 가능한 숫자 표기가 노출되어, 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반합니다."
   },
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "입가 부품이 떨어진 모습은 표현되었으나, 캐릭터의 핵심 디자인(흰 얼굴판, 점 형태의 눈)이 누락되었고 정면 구도로 촬영되어 아쉽습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S91sh3_sel.png",
    "asset_id": "9c6f33a7-494c-4543-b570-213c2406d780",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf9e-004c-739f-8b7b-bcaf0212aa11",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S91sh3"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S91sh10": {
  "input_fingerprint": "c575eef795c66f2c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈에 보이지 않을 만큼 맹렬히 도는 방향이 읽히게 한 위치에서 멈춘 찰리의 가슴 링이 눈부신 빛을 뿜어내는 긴장된 순간.\n\nLOCATION (lock): In the sunlit shallows beside the rocky island outcrop, at the robot's fishing spot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Charlie's chest ring (Attached to the chest and rotating too rapidly for individual positions to resolve) — The complete outward-facing ring is visible obliquely, preserving its circumferential path; used as Serves as the focal detail, with rotational blur contrasted against the surrounding torso.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The ring's dazzling emission creates concentrated brightness against the daytime exposure, while controlled highlights preserve its outline and surrounding chest detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Clear seawater reveals several fish beneath the surface beside the island rock in daylight. Food is being arranged on the dining table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈에 보이지 않을 만큼 맹렬히 도는 방향이 읽히게 한 위치에서 멈춘 찰리의 가슴 링이 눈부신 빛을 뿜어내는 긴장된 순간.\n\nLOCATION (lock): In the sunlit shallows beside the rocky island outcrop, at the robot's fishing spot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Charlie's chest ring (Attached to the chest and rotating too rapidly for individual positions to resolve) — The complete outward-facing ring is visible obliquely, preserving its circumferential path; used as Serves as the focal detail, with rotational blur contrasted against the surrounding torso.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The ring's dazzling emission creates concentrated brightness against the daytime exposure, while controlled highlights preserve its outline and surrounding chest detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Clear seawater reveals several fish beneath the surface beside the island rock in daylight. Food is being arranged on the dining table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈에 보이지 않을 만큼 맹렬히 도는 방향이 읽히게 한 위치에서 멈춘 찰리의 가슴 링이 눈부신 빛을 뿜어내는 긴장된 순간.\n\nLOCATION (lock): In the sunlit shallows beside the rocky island outcrop, at the robot's fishing spot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Charlie's chest ring (Attached to the chest and rotating too rapidly for individual positions to resolve) — The complete outward-facing ring is visible obliquely, preserving its circumferential path; used as Serves as the focal detail, with rotational blur contrasted against the surrounding torso.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The ring's dazzling emission creates concentrated brightness against the daytime exposure, while controlled highlights preserve its outline and surrounding chest detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Clear seawater reveals several fish beneath the surface beside the island rock in daylight. Food is being arranged on the dining table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 찰리의 가슴을 비스듬히 향하고 있으며, 링은 우측 전방을 향해 회전함.",
    "built_space": "자연 환경(바위섬과 바다). 인공 구조물 없음.",
    "entities": "기준 이미지와 외형이 일치하는 찰리의 가슴과 하관이 보이며, 가슴에 눈부시게 빛나는 회전 링이 있음.",
    "hard_violations": [],
    "physics": "로봇은 물속 바닥을 딛고 서 있으며, 가슴의 링은 몸체에 물리적으로 고정된 상태로 회전함."
   },
   {
    "label": "B",
    "direction": "카메라는 찰리를 정면에서 바라보고 있으며, 가슴의 링 역시 렌즈를 정면으로 향함.",
    "built_space": "자연 환경(바위섬과 바다). 인공 구조물 없음.",
    "entities": "기준 이미지와 일치하는 찰리의 상반신 전체가 보이며, 정면 가슴에 빛나는 회전 링이 부착됨.",
    "hard_violations": [],
    "physics": "로봇은 물속 바닥에 서 있으며, 가슴의 링은 안정적으로 몸체에 부착되어 작동함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 '비스듬한 시점(obliquely)'과 클로즈업 프레이밍 지침을 정확히 구현하여 촬영 텍스트의 의도를 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "가슴 링을 정면에서 촬영하여 '비스듬한 시점' 지침을 따르지 않았으며, 프레이밍 크기도 지정된 클로즈업보다 넓게 잡혔습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 찰리의 가슴을 비스듬히 향하고 있으며, 링은 우측 전방을 향해 회전함.",
        "built_space": "자연 환경(바위섬과 바다). 인공 구조물 없음.",
        "entities": "기준 이미지와 외형이 일치하는 찰리의 가슴과 하관이 보이며, 가슴에 눈부시게 빛나는 회전 링이 있음.",
        "hard_violations": [],
        "physics": "로봇은 물속 바닥을 딛고 서 있으며, 가슴의 링은 몸체에 물리적으로 고정된 상태로 회전함."
       },
       {
        "label": "B",
        "direction": "카메라는 찰리를 정면에서 바라보고 있으며, 가슴의 링 역시 렌즈를 정면으로 향함.",
        "built_space": "자연 환경(바위섬과 바다). 인공 구조물 없음.",
        "entities": "기준 이미지와 일치하는 찰리의 상반신 전체가 보이며, 정면 가슴에 빛나는 회전 링이 부착됨.",
        "hard_violations": [],
        "physics": "로봇은 물속 바닥에 서 있으며, 가슴의 링은 안정적으로 몸체에 부착되어 작동함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 '비스듬한 시점(obliquely)'과 클로즈업 프레이밍 지침을 정확히 구현하여 촬영 텍스트의 의도를 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "가슴 링을 정면에서 촬영하여 '비스듬한 시점' 지침을 따르지 않았으며, 프레이밍 크기도 지정된 클로즈업보다 넓게 잡혔습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 찰리의 가슴을 비스듬히 향하고 있으며, 링은 우측 전방을 향해 회전함.",
        "built_space": "자연 환경(바위섬과 바다). 인공 구조물 없음.",
        "entities": "기준 이미지와 외형이 일치하는 찰리의 가슴과 하관이 보이며, 가슴에 눈부시게 빛나는 회전 링이 있음.",
        "hard_violations": [],
        "physics": "로봇은 물속 바닥을 딛고 서 있으며, 가슴의 링은 몸체에 물리적으로 고정된 상태로 회전함."
       },
       {
        "label": "B",
        "direction": "카메라는 찰리를 정면에서 바라보고 있으며, 가슴의 링 역시 렌즈를 정면으로 향함.",
        "built_space": "자연 환경(바위섬과 바다). 인공 구조물 없음.",
        "entities": "기준 이미지와 일치하는 찰리의 상반신 전체가 보이며, 정면 가슴에 빛나는 회전 링이 부착됨.",
        "hard_violations": [],
        "physics": "로봇은 물속 바닥에 서 있으며, 가슴의 링은 안정적으로 몸체에 부착되어 작동함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "강한 발광과 원주형 회전 잔상은 구현했지만, 거의 정면인 링과 얼굴·팔까지 넓게 담은 구도가 비스듬한 가슴 링 중심 클로즈업 지시에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "가슴에 부착된 링의 전체 둘레를 비스듬한 클로즈업으로 보여주며, 선명한 흉갑과 고속 회전 잔상·눈부신 발광의 대비를 가장 충실하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 눈은 정면에서 약간 아래의 링 쪽을 향합니다. 링의 발광면은 거의 카메라 정면을 향하고 있어 요구된 사선 관찰각은 약합니다. 둘레를 따라 흐르는 잔상이 회전 경로를 보여주지만 시계·반시계 방향까지 구별되지는 않습니다.",
        "built_space": "인공 건축물이나 고정 설비는 보이지 않습니다. 오른쪽의 바위 해안과 소나무, 왼쪽 먼 바위섬, 햇빛을 받은 얕은 바다는 참조 장소와 대응합니다. 가슴 중앙에 링 장치 하나가 있으며 몸통 외에 얼굴 대부분과 양팔까지 포함되어 링 중심 클로즈업치고 범위가 넓습니다.",
        "entities": "찰리 한 대만 보이며 고릴라형 기계 얼굴, 낡은 회색 금속 외장, 어깨 장갑과 갈색 연결부가 참조와 대체로 일치합니다. 눈은 참조보다 갈색 홍채가 뚜렷합니다. 발광하는 가슴 링 하나의 전체 외곽이 보입니다. 수중 물고기 일부가 하단 가장자리에 보이고 식탁과 음식은 프레임 밖입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "링은 흉갑 중앙의 원형 하우징에 연결되어 있어 공중에 떠 있는 물체가 아닙니다. 주변 몸통은 선명하고 링 둘레에만 회전 흐림이 있어 고정된 몸체에서 회전하는 장치로 읽힙니다. 화면 오른쪽의 든 팔은 어깨와 굽힌 팔꿈치로 연결됩니다. 발과 지면 접촉은 화면 밖이며 비행하거나 떠 있는 자세는 아닙니다."
       },
       {
        "label": "B",
        "direction": "눈은 프레임 밖이라 시선은 확인할 수 없습니다. 링은 가슴에서 바깥쪽으로 돌출되어 있고 카메라는 발광면을 사선으로 봅니다. 타원으로 투영된 전체 둘레와 그 둘레를 따르는 잔상이 원주 방향의 회전을 보여줍니다. 정확한 시계·반시계 구분은 어렵습니다.",
        "built_space": "건축물이나 인공 고정 설비는 없습니다. 오른쪽 배경에 참조와 같은 밝은 바위 해안과 소나무, 아래에 투명한 얕은 바닷물이 보입니다. 링 장치는 가슴에 하나만 부착되어 있습니다. 머리 대부분을 제외하고 흉갑과 링에 밀착한 구도이며, 비스듬히 보이는 링의 전체 외곽이 잘리지 않습니다.",
        "entities": "찰리 한 대의 아래턱, 흉갑, 어깨와 팔 일부가 보입니다. 마모된 금속, 흉갑 체결부, 갈색 연결 소재와 물방울이 참조 외관에 잘 대응합니다. 눈과 머리 상부는 잘려 확인할 수 없습니다. 가슴 링 하나가 강하게 발광하며 읽을 수 있는 글자는 없습니다. 물고기와 식탁의 음식은 이 클로즈업에서 식별되지 않습니다.",
        "hard_violations": [],
        "physics": "돌출된 링 뒤쪽의 원통형 하우징이 흉갑에 이어져 물리적 지지와 회전축을 제공합니다. 링 주변의 원주형 흐림과 선명한 흉갑이 고정된 장착 위치에서의 고속 회전으로 읽힙니다. 팔은 몸통의 관절에 연결되어 아래로 내려갑니다. 하체 접촉점은 프레임 밖이며 지지 없이 떠 있는 물체나 신체는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "강한 발광과 원주형 회전 잔상은 구현했지만, 거의 정면인 링과 얼굴·팔까지 넓게 담은 구도가 비스듬한 가슴 링 중심 클로즈업 지시에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "가슴에 부착된 링의 전체 둘레를 비스듬한 클로즈업으로 보여주며, 선명한 흉갑과 고속 회전 잔상·눈부신 발광의 대비를 가장 충실하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 눈은 정면에서 약간 아래의 링 쪽을 향합니다. 링의 발광면은 거의 카메라 정면을 향하고 있어 요구된 사선 관찰각은 약합니다. 둘레를 따라 흐르는 잔상이 회전 경로를 보여주지만 시계·반시계 방향까지 구별되지는 않습니다.",
        "built_space": "인공 건축물이나 고정 설비는 보이지 않습니다. 오른쪽의 바위 해안과 소나무, 왼쪽 먼 바위섬, 햇빛을 받은 얕은 바다는 참조 장소와 대응합니다. 가슴 중앙에 링 장치 하나가 있으며 몸통 외에 얼굴 대부분과 양팔까지 포함되어 링 중심 클로즈업치고 범위가 넓습니다.",
        "entities": "찰리 한 대만 보이며 고릴라형 기계 얼굴, 낡은 회색 금속 외장, 어깨 장갑과 갈색 연결부가 참조와 대체로 일치합니다. 눈은 참조보다 갈색 홍채가 뚜렷합니다. 발광하는 가슴 링 하나의 전체 외곽이 보입니다. 수중 물고기 일부가 하단 가장자리에 보이고 식탁과 음식은 프레임 밖입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "링은 흉갑 중앙의 원형 하우징에 연결되어 있어 공중에 떠 있는 물체가 아닙니다. 주변 몸통은 선명하고 링 둘레에만 회전 흐림이 있어 고정된 몸체에서 회전하는 장치로 읽힙니다. 화면 오른쪽의 든 팔은 어깨와 굽힌 팔꿈치로 연결됩니다. 발과 지면 접촉은 화면 밖이며 비행하거나 떠 있는 자세는 아닙니다."
       },
       {
        "label": "A",
        "direction": "눈은 프레임 밖이라 시선은 확인할 수 없습니다. 링은 가슴에서 바깥쪽으로 돌출되어 있고 카메라는 발광면을 사선으로 봅니다. 타원으로 투영된 전체 둘레와 그 둘레를 따르는 잔상이 원주 방향의 회전을 보여줍니다. 정확한 시계·반시계 구분은 어렵습니다.",
        "built_space": "건축물이나 인공 고정 설비는 없습니다. 오른쪽 배경에 참조와 같은 밝은 바위 해안과 소나무, 아래에 투명한 얕은 바닷물이 보입니다. 링 장치는 가슴에 하나만 부착되어 있습니다. 머리 대부분을 제외하고 흉갑과 링에 밀착한 구도이며, 비스듬히 보이는 링의 전체 외곽이 잘리지 않습니다.",
        "entities": "찰리 한 대의 아래턱, 흉갑, 어깨와 팔 일부가 보입니다. 마모된 금속, 흉갑 체결부, 갈색 연결 소재와 물방울이 참조 외관에 잘 대응합니다. 눈과 머리 상부는 잘려 확인할 수 없습니다. 가슴 링 하나가 강하게 발광하며 읽을 수 있는 글자는 없습니다. 물고기와 식탁의 음식은 이 클로즈업에서 식별되지 않습니다.",
        "hard_violations": [],
        "physics": "돌출된 링 뒤쪽의 원통형 하우징이 흉갑에 이어져 물리적 지지와 회전축을 제공합니다. 링 주변의 원주형 흐림과 선명한 흉갑이 고정된 장착 위치에서의 고속 회전으로 읽힙니다. 팔은 몸통의 관절에 연결되어 아래로 내려갑니다. 하체 접촉점은 프레임 밖이며 지지 없이 떠 있는 물체나 신체는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.349
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.349
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1349
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 '비스듬한 시점(obliquely)'과 클로즈업 프레이밍 지침을 정확히 구현하여 촬영 텍스트의 의도를 충실히 반영했습니다."
   },
   {
    "label": "B",
    "score": 1349,
    "verdict_ko": "가슴 링을 정면에서 촬영하여 '비스듬한 시점' 지침을 따르지 않았으며, 프레이밍 크기도 지정된 클로즈업보다 넓게 잡혔습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S91sh7_sel.png",
    "asset_id": "334a7404-ddda-4f8d-9491-adc9e9cbacee",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacfa1-abf0-78ed-a25d-7e0b1d689dc1",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S91sh7"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S66sh40::cine::unresolved_log": [
  {
   "archived_at": "2026-09-18T06:54:28.769168+00:00",
   "attempted_at": "2026-09-18T06:11:16.032684+00:00",
   "submitted_at": "",
   "reason": "submission_unknown",
   "request_id": "",
   "provider": "grok",
   "endpoint": "openrouter/chat-completions",
   "status_url": "",
   "response_url": "",
   "fingerprint": "f78a84ccc8f9e50ad86afd3c678fca7714101ebb7abf07ab66cf85b565fe9d7e",
   "detail": "network_error_after_send:AppError"
  }
 ],
 "S67sh53::signage": {
  "fp": "76b9af873727da78",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S67sh53::bgfirst_bg": {
  "input_fingerprint": "0e66cb0998c68fc3",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박철진이 이현우의 관자놀이에 권총 총구를 바짝 들이민 매서운 찰나.\n\nLOCATION (lock): In the village's outdoor residential gathering area at night, illuminated by military vehicle headlights.\n\nTIME OF DAY (lock): night, bright full moon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Pistol (Muzzle pressed against 이현우's temple) — The side of the weapon is visible, with the muzzle directed into the temple rather than toward the camera; used as Small central threat detail held in the same readable plane as the face and gripping hand.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the established bright moonlit night and military-headlight contrast to give the faces and pistol contact a hard, legible separation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박철진이 이현우의 관자놀이에 권총 총구를 바짝 들이민 매서운 찰나.\n\nLOCATION (lock): In the village's outdoor residential gathering area at night, illuminated by military vehicle headlights.\n\nTIME OF DAY (lock): night, bright full moon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Pistol (Muzzle pressed against 이현우's temple) — The side of the weapon is visible, with the muzzle directed into the temple rather than toward the camera; used as Small central threat detail held in the same readable plane as the face and gripping hand.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the established bright moonlit night and military-headlight contrast to give the faces and pistol contact a hard, legible separation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S67sh53__bgfirst_bg.png",
  "asset_id": "5a521ded-b9d2-4d01-afe6-916481e4bc0f",
  "input_asset_ids": [
   "e4863aa4-a19a-4359-a3bb-20eee163d6d3",
   "0a5bd9fe-7778-4b7d-83a4-bd1f31ad9726"
  ]
 },
 "S67sh53": {
  "input_fingerprint": "9d61f53e86c957e5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 박철진이 이현우의 관자놀이에 권총 총구를 바짝 들이민 매서운 찰나.\n\nLOCATION (lock): In the village's outdoor residential gathering area at night, illuminated by military vehicle headlights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Pistol (Muzzle pressed against 이현우's temple) — The side of the weapon is visible, with the muzzle directed into the temple rather than toward the camera; used as Small central threat detail held in the same readable plane as the face and gripping hand.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the established bright moonlit night and military-headlight contrast to give the faces and pistol contact a hard, legible separation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Military headlights illuminate the occupied village under a bright full moon; the dump truck's rear door has been opened for the militia deployment. Charlie's dented, perforated body and faulty chest ring persist, while B-200 still has both gun arms and the baby birds have been released. 박철진: He stands in the village holding a raised pistol at head height. 이현우: He has been brought out as a captive and remains visibly injured from the earlier fighting.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 박철진이 이현우의 관자놀이에 권총 총구를 바짝 들이민 매서운 찰나.\n\nLOCATION (lock): In the village's outdoor residential gathering area at night, illuminated by military vehicle headlights. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Pistol (Muzzle pressed against 이현우's temple) — The side of the weapon is visible, with the muzzle directed into the temple rather than toward the camera; used as Small central threat detail held in the same readable plane as the face and gripping hand.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the established bright moonlit night and military-headlight contrast to give the faces and pistol contact a hard, legible separation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Military headlights illuminate the occupied village under a bright full moon; the dump truck's rear door has been opened for the militia deployment. Charlie's dented, perforated body and faulty chest ring persist, while B-200 still has both gun arms and the baby birds have been released. 박철진: He stands in the village holding a raised pistol at head height. 이현우: He has been brought out as a captive and remains visibly injured from the earlier fighting.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 박철진이 이현우의 관자놀이에 권총 총구를 바짝 들이민 매서운 찰나.\n\nLOCATION (lock): In the village's outdoor residential gathering area at night, illuminated by military vehicle headlights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Pistol (Muzzle pressed against 이현우's temple) — The side of the weapon is visible, with the muzzle directed into the temple rather than toward the camera; used as Small central threat detail held in the same readable plane as the face and gripping hand.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the established bright moonlit night and military-headlight contrast to give the faces and pistol contact a hard, legible separation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Military headlights illuminate the occupied village under a bright full moon; the dump truck's rear door has been opened for the militia deployment. Charlie's dented, perforated body and faulty chest ring persist, while B-200 still has both gun arms and the baby birds have been released. 박철진: He stands in the village holding a raised pistol at head height. 이현우: He has been brought out as a captive and remains visibly injured from the earlier fighting.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (44세 남성, 중년의 얼굴, 머리 모양 미상); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S67sh53__bgfirst_bg.png",
     "asset_id": "5a521ded-b9d2-4d01-afe6-916481e4bc0f",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S67sh53.png",
     "asset_id": "e4863aa4-a19a-4359-a3bb-20eee163d6d3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822047>",
     "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_village_common_7675fe.png",
     "asset_id": "0a5bd9fe-7778-4b7d-83a4-bd1f31ad9726",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822047>",
     "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "권총 총구가 이현우의 우측 관자놀이에 정확히 닿아 겨누어지고 있음.",
    "built_space": "배경에 트럭과 구조물이 있으나, 헤드라이트 주변과 우측에 다수의 인물이 서 있음.",
    "entities": "이현우와 권총은 존재하나, 박철진은 레퍼런스와 전혀 다른 외모의 인물로 그려졌으며 프롬프트에 없는 군인들이 다수 추가됨.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 다수의 추가 인물(배경의 군인들) 생성",
     "[gpt-high] 숏에 허용된 박철진과 이현우 외에 배경 인물 다섯 명을 추가했다."
    ],
    "physics": "오른손이 권총을 단단히 쥐고 허공에서 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "권총이 존재하지 않아 총구의 방향이 없으며, 박철진의 시선만 이현우를 향함.",
    "built_space": "후드가 열린 트럭, 의자 등 야외 공터의 지정된 요소들이 올바르게 배치됨.",
    "entities": "두 인물의 외모와 복장은 레퍼런스와 일치하나, 필수 소품인 권총이 완전히 누락됨.",
    "hard_violations": [],
    "physics": "박철진의 오른손이 이현우의 가슴 부위 옷에 닿아 지탱되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물의 외모와 배경은 레퍼런스에 충실하나, 샷의 핵심인 권총과 겨누는 동작이 완전히 누락되었습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "총을 겨누는 동작은 맞으나, 박철진의 외모가 다르고 지시되지 않은 배경 인물을 다수 생성하여 치명적 위반입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "권총 총구가 이현우의 우측 관자놀이에 정확히 닿아 겨누어지고 있음.",
        "built_space": "배경에 트럭과 구조물이 있으나, 헤드라이트 주변과 우측에 다수의 인물이 서 있음.",
        "entities": "이현우와 권총은 존재하나, 박철진은 레퍼런스와 전혀 다른 외모의 인물로 그려졌으며 프롬프트에 없는 군인들이 다수 추가됨.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 다수의 추가 인물(배경의 군인들) 생성"
        ],
        "physics": "오른손이 권총을 단단히 쥐고 허공에서 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "권총이 존재하지 않아 총구의 방향이 없으며, 박철진의 시선만 이현우를 향함.",
        "built_space": "후드가 열린 트럭, 의자 등 야외 공터의 지정된 요소들이 올바르게 배치됨.",
        "entities": "두 인물의 외모와 복장은 레퍼런스와 일치하나, 필수 소품인 권총이 완전히 누락됨.",
        "hard_violations": [],
        "physics": "박철진의 오른손이 이현우의 가슴 부위 옷에 닿아 지탱되고 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물의 외모와 배경은 레퍼런스에 충실하나, 샷의 핵심인 권총과 겨누는 동작이 완전히 누락되었습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "총을 겨누는 동작은 맞으나, 박철진의 외모가 다르고 지시되지 않은 배경 인물을 다수 생성하여 치명적 위반입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "권총 총구가 이현우의 우측 관자놀이에 정확히 닿아 겨누어지고 있음.",
        "built_space": "배경에 트럭과 구조물이 있으나, 헤드라이트 주변과 우측에 다수의 인물이 서 있음.",
        "entities": "이현우와 권총은 존재하나, 박철진은 레퍼런스와 전혀 다른 외모의 인물로 그려졌으며 프롬프트에 없는 군인들이 다수 추가됨.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 다수의 추가 인물(배경의 군인들) 생성"
        ],
        "physics": "오른손이 권총을 단단히 쥐고 허공에서 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "권총이 존재하지 않아 총구의 방향이 없으며, 박철진의 시선만 이현우를 향함.",
        "built_space": "후드가 열린 트럭, 의자 등 야외 공터의 지정된 요소들이 올바르게 배치됨.",
        "entities": "두 인물의 외모와 복장은 레퍼런스와 일치하나, 필수 소품인 권총이 완전히 누락됨.",
        "hard_violations": [],
        "physics": "박철진의 오른손이 이현우의 가슴 부위 옷에 닿아 지탱되고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "금지된 추가 인물은 없어 우선하지만, 권총과 관자놀이 위협이 아예 보이지 않고 구도도 넓어 핵심 순간을 구현하지 못했다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "권총을 든 손과 부상당한 얼굴은 선명하지만, 배경에 허용되지 않은 인물 다섯 명을 추가해 실격이며 총구 접점도 관자놀이보다 뒤쪽이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진과 이현우는 서로의 얼굴을 바라본다. 권총은 보이지 않으므로 이현우의 관자놀이를 겨누는 총구도 없다. 박철진의 보이는 손은 머리 높이가 아니라 화면 아래 이현우의 몸통 부근에 있다.",
        "built_space": "왼쪽에 지붕 있는 목조 공간, 중앙 뒤에 낮은 주택과 담장, 오른쪽에 주택 일부가 보인다. 열린 보닛의 낡은 트럭 한 대, 오른쪽 뒤 차량 한 대, 트럭 옆 타이어 더미 하나와 작은 의자 하나, 오른쪽 벤치 하나가 식별된다. 장소의 주요 재료와 배치는 참조를 상당 부분 따른다. 두 사람은 마당 전경에 있으나 몸통과 배경을 많이 포함해 요구된 접점 중심 클로즈업보다 넓다. 보름달은 있지만 군용 전조등 광원은 뚜렷하지 않다.",
        "entities": "보이는 사람은 두 명뿐이다. 박철진의 중년 남성 얼굴, 짧은 수염, 녹색 군모와 군복은 참조에 가깝다. 이현우는 젊은 한국계 남성의 외모, 헝클어진 검은 머리, 낡은 어두운 후드와 얼굴의 상처를 갖췄다. 핵심 소품인 권총은 없다. 덤프트럭 후문 상태는 확인할 수 없으며 찰리, B-200, 새들은 프레임 밖이다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 상체는 자연스럽게 세워져 있고 발은 프레임 밖이다. 박철진의 팔은 어깨에서 화면 아래 손까지 이어진다. 차량과 의자, 벤치는 지면에 놓여 있으며 트럭 보닛은 차체 연결부에서 열린 상태다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "박철진은 이현우의 머리 쪽을 보고, 이현우는 화면 오른쪽 바깥을 바라본다. 권총은 옆면을 드러낸 채 왼쪽에서 오른쪽으로 이현우의 머리를 향하며 카메라를 겨누지 않는다. 다만 총구는 눈 옆 관자놀이보다 귀 위쪽의 뒤편 머리카락에 닿아 보인다. 배경 오른쪽 인물이 든 긴 총은 아래쪽을 향한다.",
        "built_space": "왼쪽 목조 지붕, 중앙 기와 주택과 콘크리트 담장, 오른쪽 주택과 창문, 전신주와 전선이 참조 장소와 대응한다. 열린 보닛의 낡은 트럭 한 대, 중앙 뒤 전조등 두 개가 켜진 차량 한 대, 전경 뒤 벤치 하나, 오른쪽 공 하나와 드럼통 하나가 보인다. 두 주인공은 마당 전경에 있고 배경 차량 주변과 오른쪽 주택 앞에 추가 인물들이 서 있다. 얼굴·손·권총이 크게 잡히지만 배경도 상당히 드러난다. 보름달과 전조등은 모두 명확하다.",
        "entities": "전경에는 중년 남성 박철진과 젊은 한국계 남성 이현우가 있다. 박철진은 녹색 군복과 회색 섞인 수염을 갖췄지만 참조와 얼굴 인상이 다르고 군모 없이 빗어 올린 머리를 보인다. 이현우의 검은 머리와 얼굴 상처는 요구에 맞지만 후드는 참조의 풀오버와 달리 지퍼형이다. 권총 한 정과 이를 쥔 성인 남성의 손이 보인다. 배경에는 차량 부근 세 명, 그 오른쪽 한 명, 화면 맨 오른쪽 무장 인물 한 명으로 총 다섯 명이 추가되어 있다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "숏에 허용된 박철진과 이현우 외에 배경 인물 다섯 명을 추가했다."
        ],
        "physics": "권총 손잡이는 박철진의 손에 잡혀 있고 손목과 전완이 군복 소매 및 몸으로 자연스럽게 이어진다. 머리 높이로 총을 든 자세는 물리적으로 가능하다. 주인공들의 하체는 프레임 밖이며 배경 인물들은 지면에 서 있다. 차량과 벤치, 공도 땅에 놓여 있고 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "금지된 추가 인물은 없어 우선하지만, 권총과 관자놀이 위협이 아예 보이지 않고 구도도 넓어 핵심 순간을 구현하지 못했다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "권총을 든 손과 부상당한 얼굴은 선명하지만, 배경에 허용되지 않은 인물 다섯 명을 추가해 실격이며 총구 접점도 관자놀이보다 뒤쪽이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박철진과 이현우는 서로의 얼굴을 바라본다. 권총은 보이지 않으므로 이현우의 관자놀이를 겨누는 총구도 없다. 박철진의 보이는 손은 머리 높이가 아니라 화면 아래 이현우의 몸통 부근에 있다.",
        "built_space": "왼쪽에 지붕 있는 목조 공간, 중앙 뒤에 낮은 주택과 담장, 오른쪽에 주택 일부가 보인다. 열린 보닛의 낡은 트럭 한 대, 오른쪽 뒤 차량 한 대, 트럭 옆 타이어 더미 하나와 작은 의자 하나, 오른쪽 벤치 하나가 식별된다. 장소의 주요 재료와 배치는 참조를 상당 부분 따른다. 두 사람은 마당 전경에 있으나 몸통과 배경을 많이 포함해 요구된 접점 중심 클로즈업보다 넓다. 보름달은 있지만 군용 전조등 광원은 뚜렷하지 않다.",
        "entities": "보이는 사람은 두 명뿐이다. 박철진의 중년 남성 얼굴, 짧은 수염, 녹색 군모와 군복은 참조에 가깝다. 이현우는 젊은 한국계 남성의 외모, 헝클어진 검은 머리, 낡은 어두운 후드와 얼굴의 상처를 갖췄다. 핵심 소품인 권총은 없다. 덤프트럭 후문 상태는 확인할 수 없으며 찰리, B-200, 새들은 프레임 밖이다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 상체는 자연스럽게 세워져 있고 발은 프레임 밖이다. 박철진의 팔은 어깨에서 화면 아래 손까지 이어진다. 차량과 의자, 벤치는 지면에 놓여 있으며 트럭 보닛은 차체 연결부에서 열린 상태다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "박철진은 이현우의 머리 쪽을 보고, 이현우는 화면 오른쪽 바깥을 바라본다. 권총은 옆면을 드러낸 채 왼쪽에서 오른쪽으로 이현우의 머리를 향하며 카메라를 겨누지 않는다. 다만 총구는 눈 옆 관자놀이보다 귀 위쪽의 뒤편 머리카락에 닿아 보인다. 배경 오른쪽 인물이 든 긴 총은 아래쪽을 향한다.",
        "built_space": "왼쪽 목조 지붕, 중앙 기와 주택과 콘크리트 담장, 오른쪽 주택과 창문, 전신주와 전선이 참조 장소와 대응한다. 열린 보닛의 낡은 트럭 한 대, 중앙 뒤 전조등 두 개가 켜진 차량 한 대, 전경 뒤 벤치 하나, 오른쪽 공 하나와 드럼통 하나가 보인다. 두 주인공은 마당 전경에 있고 배경 차량 주변과 오른쪽 주택 앞에 추가 인물들이 서 있다. 얼굴·손·권총이 크게 잡히지만 배경도 상당히 드러난다. 보름달과 전조등은 모두 명확하다.",
        "entities": "전경에는 중년 남성 박철진과 젊은 한국계 남성 이현우가 있다. 박철진은 녹색 군복과 회색 섞인 수염을 갖췄지만 참조와 얼굴 인상이 다르고 군모 없이 빗어 올린 머리를 보인다. 이현우의 검은 머리와 얼굴 상처는 요구에 맞지만 후드는 참조의 풀오버와 달리 지퍼형이다. 권총 한 정과 이를 쥔 성인 남성의 손이 보인다. 배경에는 차량 부근 세 명, 그 오른쪽 한 명, 화면 맨 오른쪽 무장 인물 한 명으로 총 다섯 명이 추가되어 있다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "숏에 허용된 박철진과 이현우 외에 배경 인물 다섯 명을 추가했다."
        ],
        "physics": "권총 손잡이는 박철진의 손에 잡혀 있고 손목과 전완이 군복 소매 및 몸으로 자연스럽게 이어진다. 머리 높이로 총을 든 자세는 물리적으로 가능하다. 주인공들의 하체는 프레임 밖이며 배경 인물들은 지면에 서 있다. 차량과 벤치, 공도 땅에 놓여 있고 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.083,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.833,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 명시되지 않은 다수의 추가 인물(배경의 군인들) 생성",
     "[gpt-high] 숏에 허용된 박철진과 이현우 외에 배경 인물 다섯 명을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 833
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "인물의 외모와 배경은 레퍼런스에 충실하나, 샷의 핵심인 권총과 겨누는 동작이 완전히 누락되었습니다."
   },
   {
    "label": "A",
    "score": 833,
    "verdict_ko": "총을 겨누는 동작은 맞으나, 박철진의 외모가 다르고 지시되지 않은 배경 인물을 다수 생성하여 치명적 위반입니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 다수의 추가 인물(배경의 군인들) 생성 / [gpt-high] 숏에 허용된 박철진과 이현우 외에 배경 인물 다섯 명을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_village_common_7675fe.png",
    "asset_id": "0a5bd9fe-7778-4b7d-83a4-bd1f31ad9726",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822047>",
    "asset_id": "8a723b8b-2997-45c2-a2b4-2c5e862ae0ef",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace03-22d6-7872-a014-87eff857364e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S67sh53__bgfirst_bg.png",
   "bg_asset_id": "5a521ded-b9d2-4d01-afe6-916481e4bc0f",
   "bg_record_key": "S67sh53::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "village_common",
   "groupbg_asset_id": "0a5bd9fe-7778-4b7d-83a4-bd1f31ad9726"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S67sh53::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:57:41.175367+00:00",
  "fingerprint": "5a32e60e053ac71f093f8f89d28119b12be2fd36eb5058a651a61053b50e44a5",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S67sh53_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S67sh53_sel.png",
  "source_sha256": "c88a52cbcb56d43fbd9f4e806bc38a00ccbc94d6ad964cb76fe4d56377fb199d",
  "file": "S67sh53_cine.png",
  "staged_sha256": "8a0f67f9d52b57dddeb9945d6701ebccea9f4caae633194674b18dc6fe4fe7b7",
  "latency_ms": 11854
 },
 "S67sh75::signage": {
  "fp": "82175673e0c664e8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S67sh75::bgfirst_bg": {
  "input_fingerprint": "5b3253cd18b53b88",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): B-200이 민병대원들 한가운데서 거대한 폭발 화염에 휩싸이며 산산조각 나는 파괴적인 순간.\n\nLOCATION (lock): On the village departure road among overturned militia trucks, where the robot detonates amid the pursuing troops at night.\n\nTIME OF DAY (lock): night, bright full moon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Explosion and mechanical fragments (Expanding as B-200 self-detonates); used as Central destructive event with visible surrounding space and figure-scale references; Overturned trucks (Overturned during the preceding attack) — Oblique side and underside views remain at the margins; used as Peripheral mechanical scale references; no single truck occupies more than a minor portion of the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The explosion's intense brightness overwhelms the established nighttime contrast locally while preserving readable silhouettes around its perimeter.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): B-200이 민병대원들 한가운데서 거대한 폭발 화염에 휩싸이며 산산조각 나는 파괴적인 순간.\n\nLOCATION (lock): On the village departure road among overturned militia trucks, where the robot detonates amid the pursuing troops at night.\n\nTIME OF DAY (lock): night, bright full moon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Explosion and mechanical fragments (Expanding as B-200 self-detonates); used as Central destructive event with visible surrounding space and figure-scale references; Overturned trucks (Overturned during the preceding attack) — Oblique side and underside views remain at the margins; used as Peripheral mechanical scale references; no single truck occupies more than a minor portion of the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The explosion's intense brightness overwhelms the established nighttime contrast locally while preserving readable silhouettes around its perimeter.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S67sh75__bgfirst_bg.png",
  "asset_id": "c9d38229-36a1-4db8-8da6-460e5e83e09c",
  "input_asset_ids": [
   "ace14fda-75de-421d-af01-673e36b7d4c1",
   "476b4788-828b-41ae-bff5-dcf6666051ef"
  ]
 },
 "S67sh75": {
  "input_fingerprint": "aeb16f39540c0144",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): B-200이 민병대원들 한가운데서 거대한 폭발 화염에 휩싸이며 산산조각 나는 파괴적인 순간.\n\nLOCATION (lock): On the village departure road among overturned militia trucks, where the robot detonates amid the pursuing troops at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Explosion and mechanical fragments (Expanding as B-200 self-detonates); used as Central destructive event with visible surrounding space and figure-scale references; Overturned trucks (Overturned during the preceding attack) — Oblique side and underside views remain at the margins; used as Peripheral mechanical scale references; no single truck occupies more than a minor portion of the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The explosion's intense brightness overwhelms the established nighttime contrast locally while preserving readable silhouettes around its perimeter.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several militia trucks are overturned, including the transport with its cage opened, and Charlie's heavy restraints have been broken. B-200's torso is already broken and his gun hand destroyed as he self-destructs; Charlie retains fresh bullet damage over his earlier dents and holes, while a military helicopter's searchlight remains overhead.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): B-200이 민병대원들 한가운데서 거대한 폭발 화염에 휩싸이며 산산조각 나는 파괴적인 순간.\n\nLOCATION (lock): On the village departure road among overturned militia trucks, where the robot detonates amid the pursuing troops at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Explosion and mechanical fragments (Expanding as B-200 self-detonates); used as Central destructive event with visible surrounding space and figure-scale references; Overturned trucks (Overturned during the preceding attack) — Oblique side and underside views remain at the margins; used as Peripheral mechanical scale references; no single truck occupies more than a minor portion of the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The explosion's intense brightness overwhelms the established nighttime contrast locally while preserving readable silhouettes around its perimeter.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several militia trucks are overturned, including the transport with its cage opened, and Charlie's heavy restraints have been broken. B-200's torso is already broken and his gun hand destroyed as he self-destructs; Charlie retains fresh bullet damage over his earlier dents and holes, while a military helicopter's searchlight remains overhead.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): B-200이 민병대원들 한가운데서 거대한 폭발 화염에 휩싸이며 산산조각 나는 파괴적인 순간.\n\nLOCATION (lock): On the village departure road among overturned militia trucks, where the robot detonates amid the pursuing troops at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Explosion and mechanical fragments (Expanding as B-200 self-detonates); used as Central destructive event with visible surrounding space and figure-scale references; Overturned trucks (Overturned during the preceding attack) — Oblique side and underside views remain at the margins; used as Peripheral mechanical scale references; no single truck occupies more than a minor portion of the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The explosion's intense brightness overwhelms the established nighttime contrast locally while preserving readable silhouettes around its perimeter.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several militia trucks are overturned, including the transport with its cage opened, and Charlie's heavy restraints have been broken. B-200's torso is already broken and his gun hand destroyed as he self-destructs; Charlie retains fresh bullet damage over his earlier dents and holes, while a military helicopter's searchlight remains overhead.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S67sh75__bgfirst_bg.png",
     "asset_id": "c9d38229-36a1-4db8-8da6-460e5e83e09c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S67sh75.png",
     "asset_id": "ace14fda-75de-421d-af01-673e36b7d4c1",
     "role": "conti_light"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_village_bonfire_clearing_sel.png",
     "asset_id": "476b4788-828b-41ae-bff5-dcf6666051ef",
     "role": "location_seed_bg"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "헬기 탐조등과 인물들의 시선 및 총구가 중앙의 폭발 지점을 향함.",
    "built_space": "레퍼런스 이미지의 건물 구조(좌측 벽돌집, 우측 미닫이문 주택 등)가 전혀 반영되지 않은 임의의 가상 건물과 거리가 배치됨.",
    "entities": "산산조각 나는 로봇, 허공을 나는 민병대원들, 양쪽 가장자리의 전복된 트럭들, 헬리콥터, 보름달.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스 이미지와 전혀 다른 장소 및 건물 구조를 렌더링함",
     "[gemini-pro] 좌측에 날아가는 인물들이 물리적인 추진 궤적 없이 허공에 뻣뻣하게 수평으로 떠 있음"
    ],
    "physics": "폭발로 날아가는 인물들의 자세가 충격에 의한 곡선 궤적이 아니라 단단한 모형을 허공에 띄워 놓은 것처럼 부자연스러움."
   },
   {
    "label": "B",
    "direction": "헬기 탐조등이 지상을 비추고, 주변 인물들은 폭발을 바라보며 방어 자세를 취함.",
    "built_space": "레퍼런스 사진과 완벽히 일치함. 좌측의 녹색 창틀이 있는 벽돌집, 중앙의 슬레이트 지붕 창고, 우측의 미닫이문 주택 및 앞마당의 화분과 콘크리트 블록 배열이 모두 동일함.",
    "entities": "거대한 화염과 기계 파편들, 폭발을 피하는 민병대원들, 우측의 전복된 트럭, 좌측의 세워진 트럭, 헬리콥터, 보름달.",
    "hard_violations": [],
    "physics": "인물들은 모두 지면에 발을 딛고 폭발 충격에 밀려나는 자연스러운 무게 중심을 유지하며, 기계 파편들은 중심에서 밖으로 비산함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 장소 레퍼런스의 세부 요소를 완벽하게 재현했으며, 샷 텍스트에 명시된 파괴적 폭발과 인물들의 물리적 반응을 잘 담아냈습니다(단, 좌측 트럭이 전복되지 않은 것은 지시와 다름)."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "폭발하는 로봇의 형태는 알아보기 쉬우나, 유일한 권위인 장소 레퍼런스를 완전히 무시한 치명적인 위반이 있으며 공중에 뜬 인물의 물리적 묘사도 어색합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "헬기 탐조등과 인물들의 시선 및 총구가 중앙의 폭발 지점을 향함.",
        "built_space": "레퍼런스 이미지의 건물 구조(좌측 벽돌집, 우측 미닫이문 주택 등)가 전혀 반영되지 않은 임의의 가상 건물과 거리가 배치됨.",
        "entities": "산산조각 나는 로봇, 허공을 나는 민병대원들, 양쪽 가장자리의 전복된 트럭들, 헬리콥터, 보름달.",
        "hard_violations": [
         "레퍼런스 이미지와 전혀 다른 장소 및 건물 구조를 렌더링함",
         "좌측에 날아가는 인물들이 물리적인 추진 궤적 없이 허공에 뻣뻣하게 수평으로 떠 있음"
        ],
        "physics": "폭발로 날아가는 인물들의 자세가 충격에 의한 곡선 궤적이 아니라 단단한 모형을 허공에 띄워 놓은 것처럼 부자연스러움."
       },
       {
        "label": "B",
        "direction": "헬기 탐조등이 지상을 비추고, 주변 인물들은 폭발을 바라보며 방어 자세를 취함.",
        "built_space": "레퍼런스 사진과 완벽히 일치함. 좌측의 녹색 창틀이 있는 벽돌집, 중앙의 슬레이트 지붕 창고, 우측의 미닫이문 주택 및 앞마당의 화분과 콘크리트 블록 배열이 모두 동일함.",
        "entities": "거대한 화염과 기계 파편들, 폭발을 피하는 민병대원들, 우측의 전복된 트럭, 좌측의 세워진 트럭, 헬리콥터, 보름달.",
        "hard_violations": [],
        "physics": "인물들은 모두 지면에 발을 딛고 폭발 충격에 밀려나는 자연스러운 무게 중심을 유지하며, 기계 파편들은 중심에서 밖으로 비산함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 장소 레퍼런스의 세부 요소를 완벽하게 재현했으며, 샷 텍스트에 명시된 파괴적 폭발과 인물들의 물리적 반응을 잘 담아냈습니다(단, 좌측 트럭이 전복되지 않은 것은 지시와 다름)."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "폭발하는 로봇의 형태는 알아보기 쉬우나, 유일한 권위인 장소 레퍼런스를 완전히 무시한 치명적인 위반이 있으며 공중에 뜬 인물의 물리적 묘사도 어색합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "헬기 탐조등과 인물들의 시선 및 총구가 중앙의 폭발 지점을 향함.",
        "built_space": "레퍼런스 이미지의 건물 구조(좌측 벽돌집, 우측 미닫이문 주택 등)가 전혀 반영되지 않은 임의의 가상 건물과 거리가 배치됨.",
        "entities": "산산조각 나는 로봇, 허공을 나는 민병대원들, 양쪽 가장자리의 전복된 트럭들, 헬리콥터, 보름달.",
        "hard_violations": [
         "레퍼런스 이미지와 전혀 다른 장소 및 건물 구조를 렌더링함",
         "좌측에 날아가는 인물들이 물리적인 추진 궤적 없이 허공에 뻣뻣하게 수평으로 떠 있음"
        ],
        "physics": "폭발로 날아가는 인물들의 자세가 충격에 의한 곡선 궤적이 아니라 단단한 모형을 허공에 띄워 놓은 것처럼 부자연스러움."
       },
       {
        "label": "B",
        "direction": "헬기 탐조등이 지상을 비추고, 주변 인물들은 폭발을 바라보며 방어 자세를 취함.",
        "built_space": "레퍼런스 사진과 완벽히 일치함. 좌측의 녹색 창틀이 있는 벽돌집, 중앙의 슬레이트 지붕 창고, 우측의 미닫이문 주택 및 앞마당의 화분과 콘크리트 블록 배열이 모두 동일함.",
        "entities": "거대한 화염과 기계 파편들, 폭발을 피하는 민병대원들, 우측의 전복된 트럭, 좌측의 세워진 트럭, 헬리콥터, 보름달.",
        "hard_violations": [],
        "physics": "인물들은 모두 지면에 발을 딛고 폭발 충격에 밀려나는 자연스러운 무게 중심을 유지하며, 기계 파편들은 중심에서 밖으로 비산함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "기준 장소의 건물과 마당을 유지하면서 중앙의 거대한 화염 속 기계 파쇄를 구현했지만, 수송차가 전복되지 않았고 가장자리 트럭이 다소 크게 보입니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "민병대 한가운데서 로봇이 폭발하는 순간은 명확하지만, 고정된 기준 장소를 넓은 돌담길로 바꾼 점이 결정적인 불일치입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "기계 부품과 불꽃은 중앙 폭발점에서 사방으로 퍼집니다. 왼쪽 병사들은 폭발 쪽을 바라보며 물러나고, 오른쪽 세 사람은 폭발을 향해 손을 들거나 얼굴을 가립니다. 분명하게 특정 인물을 겨누는 총구는 보이지 않습니다. 헬리콥터 탐조등은 아래쪽 폭발 현장으로 향합니다.",
        "built_space": "왼쪽 붉은 벽돌집과 낮은 골지붕 부속 건물, 중앙의 긴 골지붕 주택, 뒤쪽의 개방형 금속 지붕 구조물, 오른쪽 출입구와 화분 줄, 앞쪽 낮은 블록 경계가 기준 사진과 대응합니다. 중앙 창문과 작업 공간 일부는 폭발에 가려집니다. 출입구 앞 흰 판은 기준과 위치가 조금 다릅니다. 병사들은 길과 흙마당에 서 있으며 건물이나 차량과 불가능하게 겹치지 않습니다.",
        "entities": "중앙에는 몸통으로 보이는 금속 덩어리, 관절, 배선과 다수의 분리된 부품이 있어 로봇의 산산조각 나는 상태로 읽힙니다. 식별 가능한 외형 참조가 없어 B-200의 개별 정체성까지 확인할 수는 없습니다. 주변 인물은 성인 남성으로 보이며 일부는 동아시아인처럼 보이지만 정확한 민족성은 판별하기 어렵습니다. 왼쪽 전경의 철창 수송차는 문이 열려 있으나 바퀴로 서 있고, 오른쪽 트럭은 뒤집혀 하부가 드러납니다. 왼쪽 뒤에도 기울어진 차량이 보입니다. 전경 두 차량은 주변부 소품치고 존재감이 큽니다. 보름달과 군용 헬리콥터, 탐조등이 있으며 읽을 수 있는 글자는 없습니다. 찰리와 끊어진 구속구는 식별되지 않습니다.",
        "hard_violations": [],
        "physics": "공중의 금속 조각과 몸통 잔해는 바로 뒤의 폭발이 발사 원인으로 보이고 방사형으로 흩어집니다. 오른쪽 인물들은 무릎을 굽히거나 뒤로 젖혀도 발이 지면에 닿아 있습니다. 왼쪽 병사들도 지면을 딛고 있습니다. 전복 차량은 차체가 땅에 닿고, 수송차는 바퀴가 받칩니다. 헬리콥터에는 회전날개가 보입니다. 원인 없이 떠 있는 몸이나 물체는 확인되지 않습니다."
       },
       {
        "label": "B",
        "direction": "로봇의 머리와 팔다리, 금속 파편이 중앙 화염에서 바깥으로 벌어집니다. 가까운 병사들은 폭발 반대쪽으로 몸이 밀리거나 웅크리고, 양옆 병사들은 중심에서 멀어지는 방향으로 움직입니다. 손에 든 소총들은 대체로 아래나 옆을 향하며 명확한 조준 대상은 없습니다. 탐조등은 오른쪽 위 헬리콥터에서 폭발 지점으로 내려옵니다.",
        "built_space": "화면은 양쪽 돌담 사이로 길게 이어지는 넓은 도로이며, 왼쪽 전경 지붕과 오른쪽 담장 너머 주택들이 보입니다. 기준 사진의 핵심인 긴 정면 주택, 그 뒤의 개방형 금속 구조물, 출입구 앞 화분 줄과 흙마당의 관계는 재현되지 않았습니다. 전복 트럭은 왼쪽 전경, 오른쪽 전경, 오른쪽 뒤에 각각 한 대씩 보입니다. 인물들은 도로와 가장자리 담장 부근에 배치되어 있으며 차량 중복이나 불가능한 반사는 보이지 않습니다.",
        "entities": "중앙 로봇은 분리된 머리와 벌어진 몸통, 긴 기계 팔다리를 가진 명확한 인간형 기계입니다. 다만 손가락 달린 손들이 비교적 온전하게 보여 파괴된 총 손이라는 이전 상태는 확인하기 어렵습니다. 주변에는 약 열 명의 성인 남성형 인물이 보이며 다수는 군복과 소총을 갖췄습니다. 왼쪽 아래 인물은 민간복처럼 보여 민병대 소속인지 불명확합니다. 얼굴 크기상 민족성은 판별하기 어렵습니다. 세 트럭은 전복되어 있고 왼쪽 차량의 격자문은 열려 있습니다. 보름달, 군용 헬리콥터와 탐조등이 있으며 읽을 수 있는 글자는 없습니다. 찰리와 구속구의 상태는 식별되지 않습니다.",
        "hard_violations": [],
        "physics": "중앙 로봇 부품과 손에서 떨어진 총들은 폭발에 의해 날아가는 것으로 읽힙니다. 왼쪽 가까운 병사와 뒤쪽 병사, 화염 오른쪽 인물은 공중에 있지만, 중심 폭발에서 멀어지는 자세와 위치가 충격에 의한 비산을 설명합니다. 이들은 근거 없이 정지 부유하는 인물은 아닙니다. 나머지 병사들은 발로 도로를 딛거나 차량 옆에서 몸을 지탱하고, 왼쪽 아래 인물은 낮은 담장 부근에 발을 둡니다. 트럭은 옆면으로 지면에 놓여 있으며 헬리콥터는 회전날개로 지지됩니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "기준 장소의 건물과 마당을 유지하면서 중앙의 거대한 화염 속 기계 파쇄를 구현했지만, 수송차가 전복되지 않았고 가장자리 트럭이 다소 크게 보입니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "민병대 한가운데서 로봇이 폭발하는 순간은 명확하지만, 고정된 기준 장소를 넓은 돌담길로 바꾼 점이 결정적인 불일치입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "기계 부품과 불꽃은 중앙 폭발점에서 사방으로 퍼집니다. 왼쪽 병사들은 폭발 쪽을 바라보며 물러나고, 오른쪽 세 사람은 폭발을 향해 손을 들거나 얼굴을 가립니다. 분명하게 특정 인물을 겨누는 총구는 보이지 않습니다. 헬리콥터 탐조등은 아래쪽 폭발 현장으로 향합니다.",
        "built_space": "왼쪽 붉은 벽돌집과 낮은 골지붕 부속 건물, 중앙의 긴 골지붕 주택, 뒤쪽의 개방형 금속 지붕 구조물, 오른쪽 출입구와 화분 줄, 앞쪽 낮은 블록 경계가 기준 사진과 대응합니다. 중앙 창문과 작업 공간 일부는 폭발에 가려집니다. 출입구 앞 흰 판은 기준과 위치가 조금 다릅니다. 병사들은 길과 흙마당에 서 있으며 건물이나 차량과 불가능하게 겹치지 않습니다.",
        "entities": "중앙에는 몸통으로 보이는 금속 덩어리, 관절, 배선과 다수의 분리된 부품이 있어 로봇의 산산조각 나는 상태로 읽힙니다. 식별 가능한 외형 참조가 없어 B-200의 개별 정체성까지 확인할 수는 없습니다. 주변 인물은 성인 남성으로 보이며 일부는 동아시아인처럼 보이지만 정확한 민족성은 판별하기 어렵습니다. 왼쪽 전경의 철창 수송차는 문이 열려 있으나 바퀴로 서 있고, 오른쪽 트럭은 뒤집혀 하부가 드러납니다. 왼쪽 뒤에도 기울어진 차량이 보입니다. 전경 두 차량은 주변부 소품치고 존재감이 큽니다. 보름달과 군용 헬리콥터, 탐조등이 있으며 읽을 수 있는 글자는 없습니다. 찰리와 끊어진 구속구는 식별되지 않습니다.",
        "hard_violations": [],
        "physics": "공중의 금속 조각과 몸통 잔해는 바로 뒤의 폭발이 발사 원인으로 보이고 방사형으로 흩어집니다. 오른쪽 인물들은 무릎을 굽히거나 뒤로 젖혀도 발이 지면에 닿아 있습니다. 왼쪽 병사들도 지면을 딛고 있습니다. 전복 차량은 차체가 땅에 닿고, 수송차는 바퀴가 받칩니다. 헬리콥터에는 회전날개가 보입니다. 원인 없이 떠 있는 몸이나 물체는 확인되지 않습니다."
       },
       {
        "label": "A",
        "direction": "로봇의 머리와 팔다리, 금속 파편이 중앙 화염에서 바깥으로 벌어집니다. 가까운 병사들은 폭발 반대쪽으로 몸이 밀리거나 웅크리고, 양옆 병사들은 중심에서 멀어지는 방향으로 움직입니다. 손에 든 소총들은 대체로 아래나 옆을 향하며 명확한 조준 대상은 없습니다. 탐조등은 오른쪽 위 헬리콥터에서 폭발 지점으로 내려옵니다.",
        "built_space": "화면은 양쪽 돌담 사이로 길게 이어지는 넓은 도로이며, 왼쪽 전경 지붕과 오른쪽 담장 너머 주택들이 보입니다. 기준 사진의 핵심인 긴 정면 주택, 그 뒤의 개방형 금속 구조물, 출입구 앞 화분 줄과 흙마당의 관계는 재현되지 않았습니다. 전복 트럭은 왼쪽 전경, 오른쪽 전경, 오른쪽 뒤에 각각 한 대씩 보입니다. 인물들은 도로와 가장자리 담장 부근에 배치되어 있으며 차량 중복이나 불가능한 반사는 보이지 않습니다.",
        "entities": "중앙 로봇은 분리된 머리와 벌어진 몸통, 긴 기계 팔다리를 가진 명확한 인간형 기계입니다. 다만 손가락 달린 손들이 비교적 온전하게 보여 파괴된 총 손이라는 이전 상태는 확인하기 어렵습니다. 주변에는 약 열 명의 성인 남성형 인물이 보이며 다수는 군복과 소총을 갖췄습니다. 왼쪽 아래 인물은 민간복처럼 보여 민병대 소속인지 불명확합니다. 얼굴 크기상 민족성은 판별하기 어렵습니다. 세 트럭은 전복되어 있고 왼쪽 차량의 격자문은 열려 있습니다. 보름달, 군용 헬리콥터와 탐조등이 있으며 읽을 수 있는 글자는 없습니다. 찰리와 구속구의 상태는 식별되지 않습니다.",
        "hard_violations": [],
        "physics": "중앙 로봇 부품과 손에서 떨어진 총들은 폭발에 의해 날아가는 것으로 읽힙니다. 왼쪽 가까운 병사와 뒤쪽 병사, 화염 오른쪽 인물은 공중에 있지만, 중심 폭발에서 멀어지는 자세와 위치가 충격에 의한 비산을 설명합니다. 이들은 근거 없이 정지 부유하는 인물은 아닙니다. 나머지 병사들은 발로 도로를 딛거나 차량 옆에서 몸을 지탱하고, 왼쪽 아래 인물은 낮은 담장 부근에 발을 둡니다. 트럭은 옆면으로 지면에 놓여 있으며 헬리콥터는 회전날개로 지지됩니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.054,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.804,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스 이미지와 전혀 다른 장소 및 건물 구조를 렌더링함",
     "[gemini-pro] 좌측에 날아가는 인물들이 물리적인 추진 궤적 없이 허공에 뻣뻣하게 수평으로 떠 있음"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 804
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 장소 레퍼런스의 세부 요소를 완벽하게 재현했으며, 샷 텍스트에 명시된 파괴적 폭발과 인물들의 물리적 반응을 잘 담아냈습니다(단, 좌측 트럭이 전복되지 않은 것은 지시와 다름)."
   },
   {
    "label": "A",
    "score": 804,
    "verdict_ko": "폭발하는 로봇의 형태는 알아보기 쉬우나, 유일한 권위인 장소 레퍼런스를 완전히 무시한 치명적인 위반이 있으며 공중에 뜬 인물의 물리적 묘사도 어색합니다.  ★위반: [gemini-pro] 레퍼런스 이미지와 전혀 다른 장소 및 건물 구조를 렌더링함 / [gemini-pro] 좌측에 날아가는 인물들이 물리적인 추진 궤적 없이 허공에 뻣뻣하게 수평으로 떠 있음"
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_village_bonfire_clearing_sel.png",
    "asset_id": "476b4788-828b-41ae-bff5-dcf6666051ef",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace0f-15d7-7c0b-8c3c-9de27b60ba73",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S67sh75__bgfirst_bg.png",
   "bg_asset_id": "c9d38229-36a1-4db8-8da6-460e5e83e09c",
   "bg_record_key": "S67sh75::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S67sh75::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T06:59:31.476534+00:00",
  "fingerprint": "784a2817a33cd0b02e40d1df0b8d1ac6c6a9f86630778e9ef5de18b7bb68d25f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S67sh75_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S67sh75_sel.png",
  "source_sha256": "4483925dd4985234aecf06fb0491ebdf154ad09810f9668547ad5312628ddacb",
  "file": "S67sh75_cine.png",
  "staged_sha256": "41628c521b173eed692b08c0dc0db143c1b0bdc42d2ae9f4f02275f4e7907e15",
  "latency_ms": 18365
 },
 "S67sh80::signage": {
  "fp": "a7283cc61666202e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S67sh80::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:00:44.808751+00:00",
  "fingerprint": "d801f9420c693384418e40d742be8d9582e2c7835000d0f9e09f2c4ae0be4c4b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S67sh80_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S67sh80_sel.png",
  "source_sha256": "e4479d188b08257a02966a56287770f569bbc0913fa199041f4fca00d2f0b1b0",
  "file": "S67sh80_cine.png",
  "staged_sha256": "158295689246b187127259b21a0c0ed9cc2b6244ae10878df95f6c24a1247e3a",
  "latency_ms": 11254
 },
 "S68sh4::signage": {
  "fp": "c19ab40ddddc05a5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S68sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:01:51.414224+00:00",
  "fingerprint": "74c572fbc0d0dc83425a43785eb52e2ce242a66e8d1f38dd1072188723d218e8",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S68sh4_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S68sh4_sel.png",
  "source_sha256": "a643b92e4c70a1a08949dd8710c359a0818a27296678c871bcd3ada4a3f90260",
  "file": "S68sh4_cine.png",
  "staged_sha256": "0e9cc335e0328216255f93e340acce47aa75d09016bba4748a8bc4ed4433989a",
  "latency_ms": 14340
 },
 "S68sh10::signage": {
  "fp": "f34fe92a13195e83",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S68sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:02:56.607518+00:00",
  "fingerprint": "f8fbbdcb9c05eef4f882c6071d27fc68151399485d0c18eea1b9f1ec375cc883",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S68sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S68sh10_sel.png",
  "source_sha256": "a9172da8cf17ff856455e7fef4ca7219e6b5a4806e3fe6c02acffe85106372bb",
  "file": "S68sh10_cine.png",
  "staged_sha256": "61358fa8c2a088c5ef687e23b18b506c7069845c6052017871ffe45f30a52167",
  "latency_ms": 14091
 },
 "S68sh16::signage": {
  "fp": "0bc058a11afccc88",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::b3e5ca9e4df6bba2": {
  "subjects": [],
  "subject_text": "익산 식량 운반 도로와 갈림길\n마을 바깥으로 길게 이어지다 두 방향으로 나뉘는 도로. 넓게 트인 길 주변과 갈라지는 두 노면이 낮빛 아래 보인다.",
  "identity": "canonical",
  "scope_id": "L96",
  "scope_role": "location_exterior",
  "scope_sha": "9ad9b2d43e0ea442"
 },
 "S68sh16::bgfirst_bg": {
  "input_fingerprint": "9b36605517b6d961",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 보라색 제비꽃이 만발한 들판 너머 낡은 연구소 건물을 향해 다리를 뻗고 질주 중인 지소영, 찰리, B-200의 mid-action 뒷모습.\n\nLOCATION (lock): Across an imagined field of purple violets leading toward a distant coastal research building in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant laboratory, unobstructed by runners in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Violet field (Purple violets bloom throughout and move in the spring breeze); used as Provides foreground depth and a continuous route beneath the running figures; 지동현's Haenam laboratory (An old building visible in the distance beyond the flowers) — The building is seen obliquely beyond the runners; used as An unobstructed destination that gives their movement purpose.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and the purple violets retain clear, gentle tonal separation within the dream, without added haze or optical distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 보라색 제비꽃이 만발한 들판 너머 낡은 연구소 건물을 향해 다리를 뻗고 질주 중인 지소영, 찰리, B-200의 mid-action 뒷모습.\n\nLOCATION (lock): Across an imagined field of purple violets leading toward a distant coastal research building in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant laboratory, unobstructed by runners in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Violet field (Purple violets bloom throughout and move in the spring breeze); used as Provides foreground depth and a continuous route beneath the running figures; 지동현's Haenam laboratory (An old building visible in the distance beyond the flowers) — The building is seen obliquely beyond the runners; used as An unobstructed destination that gives their movement purpose.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and the purple violets retain clear, gentle tonal separation within the dream, without added haze or optical distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S68sh16__bgfirst_bg.png",
  "asset_id": "091e330e-f559-4fa3-92da-673f5c8a0492",
  "input_asset_ids": [
   "f6f6ef3d-619a-4f6f-b87c-f23dbde5392b",
   "3353276e-8593-44a0-8507-e9efa6680c2a"
  ]
 },
 "S68sh16": {
  "input_fingerprint": "886000acdc55779a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 보라색 제비꽃이 만발한 들판 너머 낡은 연구소 건물을 향해 다리를 뻗고 질주 중인 지소영, 찰리, B-200의 mid-action 뒷모습.\n\nLOCATION (lock): Across an imagined field of purple violets leading toward a distant coastal research building in daylight. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant laboratory, unobstructed by runners in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Violet field (Purple violets bloom throughout and move in the spring breeze); used as Provides foreground depth and a continuous route beneath the running figures; 지동현's Haenam laboratory (An old building visible in the distance beyond the flowers) — The building is seen obliquely beyond the runners; used as An unobstructed destination that gives their movement purpose.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and the purple violets retain clear, gentle tonal separation within the dream, without added haze or optical distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Purple violets fill the field, with the Haenam research laboratory visible beyond them in daylight. A gentle spring breeze moves the flowers. 지소영: She appears as her younger self within the imagined Haenam scene, running toward the laboratory with one hand extended. 찰리: He appears whole and mobile within the imagined flower field, running toward the laboratory. B-200: He appears whole and mobile within the imagined scene, running toward the laboratory.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 보라색 제비꽃이 만발한 들판 너머 낡은 연구소 건물을 향해 다리를 뻗고 질주 중인 지소영, 찰리, B-200의 mid-action 뒷모습.\n\nLOCATION (lock): Across an imagined field of purple violets leading toward a distant coastal research building in daylight. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant laboratory, unobstructed by runners in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Violet field (Purple violets bloom throughout and move in the spring breeze); used as Provides foreground depth and a continuous route beneath the running figures; 지동현's Haenam laboratory (An old building visible in the distance beyond the flowers) — The building is seen obliquely beyond the runners; used as An unobstructed destination that gives their movement purpose.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and the purple violets retain clear, gentle tonal separation within the dream, without added haze or optical distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Purple violets fill the field, with the Haenam research laboratory visible beyond them in daylight. A gentle spring breeze moves the flowers. 지소영: She appears as her younger self within the imagined Haenam scene, running toward the laboratory with one hand extended. 찰리: He appears whole and mobile within the imagined flower field, running toward the laboratory. B-200: He appears whole and mobile within the imagined scene, running toward the laboratory.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 보라색 제비꽃이 만발한 들판 너머 낡은 연구소 건물을 향해 다리를 뻗고 질주 중인 지소영, 찰리, B-200의 mid-action 뒷모습.\n\nLOCATION (lock): Across an imagined field of purple violets leading toward a distant coastal research building in daylight. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant laboratory, unobstructed by runners in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Violet field (Purple violets bloom throughout and move in the spring breeze); used as Provides foreground depth and a continuous route beneath the running figures; 지동현's Haenam laboratory (An old building visible in the distance beyond the flowers) — The building is seen obliquely beyond the runners; used as An unobstructed destination that gives their movement purpose.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and the purple violets retain clear, gentle tonal separation within the dream, without added haze or optical distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Purple violets fill the field, with the Haenam research laboratory visible beyond them in daylight. A gentle spring breeze moves the flowers. 지소영: She appears as her younger self within the imagined Haenam scene, running toward the laboratory with one hand extended. 찰리: He appears whole and mobile within the imagined flower field, running toward the laboratory. B-200: He appears whole and mobile within the imagined scene, running toward the laboratory.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); B-200 (기계식 머리, 짙은 회색 철제 외장, 크기가 변하는 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S68sh16__bgfirst_bg.png",
     "asset_id": "091e330e-f559-4fa3-92da-673f5c8a0492",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S68sh16.png",
     "asset_id": "f6f6ef3d-619a-4f6f-b87c-f23dbde5392b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1384571>",
     "asset_id": "9df5307f-9b0a-4f43-989e-44588e3d0ba9",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:885644>",
     "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1173926>",
     "asset_id": "a096b85a-c67b-4c9d-8b4a-b9b1caef0d76",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_food_transport_fork_sel.png",
     "asset_id": "3353276e-8593-44a0-8507-e9efa6680c2a",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1384571>",
     "asset_id": "9df5307f-9b0a-4f43-989e-44588e3d0ba9",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:885644>",
     "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1173926>",
     "asset_id": "a096b85a-c67b-4c9d-8b4a-b9b1caef0d76",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "세 인물이 카메라를 등지고 멀리 해안가에 있는 흰색 레이더 건물을 향해 달려가고 있음.",
    "built_space": "배경에 해안 절벽과 흰색 건물이 있으나, 지정된 장소 레퍼런스(정자, 나무, 포장도로)가 완전히 누락됨.",
    "entities": "지소영(회색 숏컷, 어두운 상의)과 B-200(회색 금속체)은 뒷모습이 유사하나, 찰리는 회색 후드를 입은 엉뚱한 인물로 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 찰리가 아닌 발명된 인물 생성",
     "[gemini-pro] 지정된 장소 레퍼런스 누락"
    ],
    "physics": "세 인물 모두 달리는 자세를 취하고 있으며 디딤발이 땅에 닿아 물리적으로 자연스러움. 왼쪽 두 명은 손을 잡고 있음."
   },
   {
    "label": "B",
    "direction": "세 인물이 카메라를 등지고 멀리 있는 콘크리트 건물을 향해 달려가고 있음.",
    "built_space": "배경에 낡은 콘크리트 건물과 바다가 있으나, 지정된 장소 레퍼런스(정자, 도로)가 없음.",
    "entities": "찰리는 재킷을 입었으나 인간의 검은 머리카락이 보이고, 지소영은 긴 검은 머리의 젊은 여성이며, B-200은 전혀 다른 디자인의 로봇임.",
    "hard_violations": [
     "[gemini-pro] 지소영이 아닌 발명된 인물 생성",
     "[gemini-pro] B-200과 다른 발명된 로봇 생성",
     "[gemini-pro] 지정된 장소 레퍼런스 누락"
    ],
    "physics": "달리는 동작에 맞춰 발이 들려 있고 체중 이동과 지지 상태가 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지소영과 B-200의 뒷모습은 일부 일치하지만, 찰리의 외형과 지정된 장소 레퍼런스를 완전히 무시함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "세 캐릭터의 외형이 모두 레퍼런스와 불일치하며 장소 조건 역시 무시되어 크게 실패함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 인물이 카메라를 등지고 멀리 해안가에 있는 흰색 레이더 건물을 향해 달려가고 있음.",
        "built_space": "배경에 해안 절벽과 흰색 건물이 있으나, 지정된 장소 레퍼런스(정자, 나무, 포장도로)가 완전히 누락됨.",
        "entities": "지소영(회색 숏컷, 어두운 상의)과 B-200(회색 금속체)은 뒷모습이 유사하나, 찰리는 회색 후드를 입은 엉뚱한 인물로 묘사됨.",
        "hard_violations": [
         "찰리가 아닌 발명된 인물 생성",
         "지정된 장소 레퍼런스 누락"
        ],
        "physics": "세 인물 모두 달리는 자세를 취하고 있으며 디딤발이 땅에 닿아 물리적으로 자연스러움. 왼쪽 두 명은 손을 잡고 있음."
       },
       {
        "label": "B",
        "direction": "세 인물이 카메라를 등지고 멀리 있는 콘크리트 건물을 향해 달려가고 있음.",
        "built_space": "배경에 낡은 콘크리트 건물과 바다가 있으나, 지정된 장소 레퍼런스(정자, 도로)가 없음.",
        "entities": "찰리는 재킷을 입었으나 인간의 검은 머리카락이 보이고, 지소영은 긴 검은 머리의 젊은 여성이며, B-200은 전혀 다른 디자인의 로봇임.",
        "hard_violations": [
         "지소영이 아닌 발명된 인물 생성",
         "B-200과 다른 발명된 로봇 생성",
         "지정된 장소 레퍼런스 누락"
        ],
        "physics": "달리는 동작에 맞춰 발이 들려 있고 체중 이동과 지지 상태가 자연스러움."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지소영과 B-200의 뒷모습은 일부 일치하지만, 찰리의 외형과 지정된 장소 레퍼런스를 완전히 무시함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "세 캐릭터의 외형이 모두 레퍼런스와 불일치하며 장소 조건 역시 무시되어 크게 실패함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "세 인물이 카메라를 등지고 멀리 해안가에 있는 흰색 레이더 건물을 향해 달려가고 있음.",
        "built_space": "배경에 해안 절벽과 흰색 건물이 있으나, 지정된 장소 레퍼런스(정자, 나무, 포장도로)가 완전히 누락됨.",
        "entities": "지소영(회색 숏컷, 어두운 상의)과 B-200(회색 금속체)은 뒷모습이 유사하나, 찰리는 회색 후드를 입은 엉뚱한 인물로 묘사됨.",
        "hard_violations": [
         "찰리가 아닌 발명된 인물 생성",
         "지정된 장소 레퍼런스 누락"
        ],
        "physics": "세 인물 모두 달리는 자세를 취하고 있으며 디딤발이 땅에 닿아 물리적으로 자연스러움. 왼쪽 두 명은 손을 잡고 있음."
       },
       {
        "label": "B",
        "direction": "세 인물이 카메라를 등지고 멀리 있는 콘크리트 건물을 향해 달려가고 있음.",
        "built_space": "배경에 낡은 콘크리트 건물과 바다가 있으나, 지정된 장소 레퍼런스(정자, 도로)가 없음.",
        "entities": "찰리는 재킷을 입었으나 인간의 검은 머리카락이 보이고, 지소영은 긴 검은 머리의 젊은 여성이며, B-200은 전혀 다른 디자인의 로봇임.",
        "hard_violations": [
         "지소영이 아닌 발명된 인물 생성",
         "B-200과 다른 발명된 로봇 생성",
         "지정된 장소 레퍼런스 누락"
        ],
        "physics": "달리는 동작에 맞춰 발이 들려 있고 체중 이동과 지지 상태가 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "상단 중앙의 연구소와 달리는 두 인물은 맞지만, B-200이 카메라를 향해 세 인물의 질주 뒷모습 조건을 깨며 장소 구조와 캐릭터 외형도 다릅니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "세 인물이 먼 연구소를 향해 달리는 뒷모습과 와이드 구도는 더 충실하지만, 고정 장소 구조와 찰리의 외형, 젊어진 지소영 조건은 충족하지 못합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 찰리와 중앙 지소영은 머리와 몸을 화면 상단 중앙 연구소 쪽으로 돌리고 달린다. 오른쪽 B-200은 두 눈과 가슴 전면이 카메라를 향해 있어 연구소로 질주하는 뒷모습으로 읽히지 않는다. 지소영의 오른손은 어깨 옆으로 들려 있으며 연구소를 향해 길게 내민 동작은 아니다. 무기나 별도의 지시 물체는 없다.",
        "built_space": "꽃밭 너머 상단 중앙에 비대칭 콘크리트 연구소 한 동이 있고, 뒤로 바다가 보인다. 연구소는 인물에게 가려지지 않는다. 그러나 장소 참조의 두 갈래 아스팔트 도로, 중앙의 큰 낙엽수 한 그루와 목조 정자 한 채, 오른쪽 철길, 전신주와 마을 배치는 보이지 않는다. 참조 장소의 고정 구조를 유지한 공간으로 식별하기 어렵다. 반사면은 없다.",
        "entities": "등장 개체는 세 명이다. 지소영은 긴 검은 묶음머리, 밝은 상의와 청바지를 입은 젊은 여성으로 보여 젊어진 상태에는 맞지만 참조의 짧은 머리와 남색 상의는 다르다. 뒷모습이라 얼굴과 민족적 정체성은 확인할 수 없다. 찰리는 밀짚모자와 조각무늬 겉옷을 입었지만 색감, 청바지, 갈색 신발이 참조의 선명한 우의·녹색 바지·노란 장화와 다르고, 드러난 손도 기계 손처럼 보이지 않는다. B-200은 밝은 회색의 가느다란 로봇으로, 참조의 짙은 철제 외장과 머리 형태가 다르다. 보라색 꽃밭과 낡은 해안 건물, 낮 풍경은 있다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "지소영과 찰리는 한쪽 다리를 뒤로 접고 반대쪽 다리를 꽃밭 지면으로 내려 디디는 달리기 자세다. 접지점 일부는 꽃에 가려지지만 몸을 지탱할 다리와 지면의 관계는 성립한다. B-200도 한쪽 다리를 아래로 뻗고 다른 발을 들어 올려 지상 보행 동작으로 볼 수 있다. 지지 없이 공중에 매달린 몸이나 물체는 없다. 다만 B-200의 몸 방향은 요구된 이동 방향과 맞지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 지소영, 중앙 찰리, 오른쪽 B-200 모두 등과 뒤통수를 보이며 화면 상단 중앙의 연구소 쪽으로 달린다. 눈은 보이지 않지만 머리와 몸의 진행축은 목적지 방향이다. 지소영은 오른손을 찰리 쪽으로 뻗어 손을 잡고, 왼팔은 옆뒤로 벌리고 있다. 연구소를 향해 손을 내미는 동작은 아니지만 한 손을 뻗은 상태는 보인다.",
        "built_space": "상단 중앙의 먼 해안 언덕에 낮은 건물 날개와 탑형 부분, 흰 구형 옥상 설비를 갖춘 연구소 한 채가 보인다. 주요 건물 윤곽은 주자들 위로 드러나며 인물보다 충분히 멀다. 오른쪽에는 바다와 해안 절벽이 있다. 그러나 장소 참조의 도로 분기, 큰 나무와 정자, 오른쪽 철길 및 마을의 고정 배치는 재현되지 않았다. 인물들은 모두 꽃밭 지면 위에 있고 건물 내부 설비나 반사는 없다.",
        "entities": "세 인물만 등장한다. 지소영은 짧은 회색 머리와 어두운 상의로 참조의 보이는 특징에 가깝지만, 젊어진 모습보다는 중년 이후 여성으로 읽힌다. 얼굴이 숨겨져 얼굴 동일성과 민족적 정체성은 확인할 수 없다. 찰리는 흰 후드형 상의와 청바지를 입었고 머리 뒤에 작은 원형 부품 두 개가 보인다. 참조의 밀짚모자, 다색 우의, 녹색 바지, 노란 장화가 없으며 각진 고릴라형 기계 머리도 확인되지 않는다. B-200은 짙은 회색 금속 머리와 장갑형 몸체로 A보다 참조의 재질과 색에 가깝다. 보라색 꽃밭, 먼 연구소와 밝은 낮은 구현되어 있고, 판독 가능한 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "지소영은 한쪽 발을 뒤로 접고 반대쪽 다리를 지면으로 뻗는다. 찰리는 뒤쪽 신발 밑창을 보이며 다른 다리를 몸 아래로 내려 다음 보폭을 받친다. B-200도 큰 보폭으로 한 다리를 뒤로 뻗고 다른 다리를 몸 아래에 두며, 지면 가까운 발과 굽힌 관절이 달리기의 추진과 착지를 설명한다. 발 일부는 꽃에 가려져 있지만 근거 없는 공중 부양은 보이지 않는다. 지소영과 찰리의 맞잡은 손도 팔의 연속성과 접촉이 자연스럽다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "상단 중앙의 연구소와 달리는 두 인물은 맞지만, B-200이 카메라를 향해 세 인물의 질주 뒷모습 조건을 깨며 장소 구조와 캐릭터 외형도 다릅니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "세 인물이 먼 연구소를 향해 달리는 뒷모습과 와이드 구도는 더 충실하지만, 고정 장소 구조와 찰리의 외형, 젊어진 지소영 조건은 충족하지 못합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 찰리와 중앙 지소영은 머리와 몸을 화면 상단 중앙 연구소 쪽으로 돌리고 달린다. 오른쪽 B-200은 두 눈과 가슴 전면이 카메라를 향해 있어 연구소로 질주하는 뒷모습으로 읽히지 않는다. 지소영의 오른손은 어깨 옆으로 들려 있으며 연구소를 향해 길게 내민 동작은 아니다. 무기나 별도의 지시 물체는 없다.",
        "built_space": "꽃밭 너머 상단 중앙에 비대칭 콘크리트 연구소 한 동이 있고, 뒤로 바다가 보인다. 연구소는 인물에게 가려지지 않는다. 그러나 장소 참조의 두 갈래 아스팔트 도로, 중앙의 큰 낙엽수 한 그루와 목조 정자 한 채, 오른쪽 철길, 전신주와 마을 배치는 보이지 않는다. 참조 장소의 고정 구조를 유지한 공간으로 식별하기 어렵다. 반사면은 없다.",
        "entities": "등장 개체는 세 명이다. 지소영은 긴 검은 묶음머리, 밝은 상의와 청바지를 입은 젊은 여성으로 보여 젊어진 상태에는 맞지만 참조의 짧은 머리와 남색 상의는 다르다. 뒷모습이라 얼굴과 민족적 정체성은 확인할 수 없다. 찰리는 밀짚모자와 조각무늬 겉옷을 입었지만 색감, 청바지, 갈색 신발이 참조의 선명한 우의·녹색 바지·노란 장화와 다르고, 드러난 손도 기계 손처럼 보이지 않는다. B-200은 밝은 회색의 가느다란 로봇으로, 참조의 짙은 철제 외장과 머리 형태가 다르다. 보라색 꽃밭과 낡은 해안 건물, 낮 풍경은 있다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "지소영과 찰리는 한쪽 다리를 뒤로 접고 반대쪽 다리를 꽃밭 지면으로 내려 디디는 달리기 자세다. 접지점 일부는 꽃에 가려지지만 몸을 지탱할 다리와 지면의 관계는 성립한다. B-200도 한쪽 다리를 아래로 뻗고 다른 발을 들어 올려 지상 보행 동작으로 볼 수 있다. 지지 없이 공중에 매달린 몸이나 물체는 없다. 다만 B-200의 몸 방향은 요구된 이동 방향과 맞지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 지소영, 중앙 찰리, 오른쪽 B-200 모두 등과 뒤통수를 보이며 화면 상단 중앙의 연구소 쪽으로 달린다. 눈은 보이지 않지만 머리와 몸의 진행축은 목적지 방향이다. 지소영은 오른손을 찰리 쪽으로 뻗어 손을 잡고, 왼팔은 옆뒤로 벌리고 있다. 연구소를 향해 손을 내미는 동작은 아니지만 한 손을 뻗은 상태는 보인다.",
        "built_space": "상단 중앙의 먼 해안 언덕에 낮은 건물 날개와 탑형 부분, 흰 구형 옥상 설비를 갖춘 연구소 한 채가 보인다. 주요 건물 윤곽은 주자들 위로 드러나며 인물보다 충분히 멀다. 오른쪽에는 바다와 해안 절벽이 있다. 그러나 장소 참조의 도로 분기, 큰 나무와 정자, 오른쪽 철길 및 마을의 고정 배치는 재현되지 않았다. 인물들은 모두 꽃밭 지면 위에 있고 건물 내부 설비나 반사는 없다.",
        "entities": "세 인물만 등장한다. 지소영은 짧은 회색 머리와 어두운 상의로 참조의 보이는 특징에 가깝지만, 젊어진 모습보다는 중년 이후 여성으로 읽힌다. 얼굴이 숨겨져 얼굴 동일성과 민족적 정체성은 확인할 수 없다. 찰리는 흰 후드형 상의와 청바지를 입었고 머리 뒤에 작은 원형 부품 두 개가 보인다. 참조의 밀짚모자, 다색 우의, 녹색 바지, 노란 장화가 없으며 각진 고릴라형 기계 머리도 확인되지 않는다. B-200은 짙은 회색 금속 머리와 장갑형 몸체로 A보다 참조의 재질과 색에 가깝다. 보라색 꽃밭, 먼 연구소와 밝은 낮은 구현되어 있고, 판독 가능한 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "지소영은 한쪽 발을 뒤로 접고 반대쪽 다리를 지면으로 뻗는다. 찰리는 뒤쪽 신발 밑창을 보이며 다른 다리를 몸 아래로 내려 다음 보폭을 받친다. B-200도 큰 보폭으로 한 다리를 뒤로 뻗고 다른 다리를 몸 아래에 두며, 지면 가까운 발과 굽힌 관절이 달리기의 추진과 착지를 설명한다. 발 일부는 꽃에 가려져 있지만 근거 없는 공중 부양은 보이지 않는다. 지소영과 찰리의 맞잡은 손도 팔의 연속성과 접촉이 자연스럽다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.417
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.167
   },
   "violations": {
    "A": [
     "[gemini-pro] 찰리가 아닌 발명된 인물 생성",
     "[gemini-pro] 지정된 장소 레퍼런스 누락"
    ],
    "B": [
     "[gemini-pro] 지소영이 아닌 발명된 인물 생성",
     "[gemini-pro] B-200과 다른 발명된 로봇 생성",
     "[gemini-pro] 지정된 장소 레퍼런스 누락"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1167
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지소영과 B-200의 뒷모습은 일부 일치하지만, 찰리의 외형과 지정된 장소 레퍼런스를 완전히 무시함.  ★위반: [gemini-pro] 찰리가 아닌 발명된 인물 생성 / [gemini-pro] 지정된 장소 레퍼런스 누락"
   },
   {
    "label": "B",
    "score": 1167,
    "verdict_ko": "세 캐릭터의 외형이 모두 레퍼런스와 불일치하며 장소 조건 역시 무시되어 크게 실패함.  ★위반: [gemini-pro] 지소영이 아닌 발명된 인물 생성 / [gemini-pro] B-200과 다른 발명된 로봇 생성 / [gemini-pro] 지정된 장소 레퍼런스 누락"
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_food_transport_fork_sel.png",
    "asset_id": "3353276e-8593-44a0-8507-e9efa6680c2a",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1384571>",
    "asset_id": "9df5307f-9b0a-4f43-989e-44588e3d0ba9",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:885644>",
    "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1173926>",
    "asset_id": "a096b85a-c67b-4c9d-8b4a-b9b1caef0d76",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace22-f0a0-774e-bb61-09ae26da1168",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S68sh16__bgfirst_bg.png",
   "bg_asset_id": "091e330e-f559-4fa3-92da-673f5c8a0492",
   "bg_record_key": "S68sh16::bgfirst_bg",
   "chain_winner": true,
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S68sh16::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:05:07.038984+00:00",
  "fingerprint": "76579e9d6be046ee17805eee02d8d6f4aa0bcb7a17f02e0a3e560dad97b4c032",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S68sh16_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S68sh16_sel.png",
  "source_sha256": "ac980a33430a695944bd5beb8a6590ea64ceaa079780bc30eecd7eb865d9d2e8",
  "file": "S68sh16_cine.png",
  "staged_sha256": "ce0b806b6e8ba565856b1e3ee17f6c55c9494c0636b0e305a5a4ad24819e7200",
  "latency_ms": 13844
 },
 "S69sh2::signage": {
  "fp": "8839d9308dd1b1d2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::60132f46c1f2b04a": {
  "subjects": [],
  "subject_text": "해남 바다가 보이는 언덕과 트럭 주차 지점\n도로 옆 언덕의 흙 공터. 겹친 능선 사이로 햇빛에 반짝이는 바다가 보이고 주변 지형을 넓게 내려다볼 수 있다.",
  "identity": "canonical",
  "scope_id": "L108",
  "scope_role": "location_exterior",
  "scope_sha": "5275d6f438e37994"
 },
 "S69sh2::bgfirst_bg": {
  "input_fingerprint": "5111abeb580bd1ec",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 트럭 짐칸 구석에서 이현우, 앰버, 라울이 머리를 맞대고 깊이 잠든 평온한 구도.\n\nLOCATION (lock): In the open cargo bed of a truck parked in the shade of a rocky outcrop, beneath a movable shade cloth.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Truck-bed corner (Occupied by the three sleeping companions in the parked truck) — The inner corner is visible diagonally from above the outer side; used as Encloses the sleeping cluster without arranging the bodies into a formal row; Truck-bed edge (Visible along the near margin) — Its inner face recedes diagonally toward the sleepers; used as A narrow foreground boundary and continuity line for the lateral track.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight subdued by the sheltering rocks gives the sleeping faces soft contrast and a peaceful, unforced appearance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 트럭 짐칸 구석에서 이현우, 앰버, 라울이 머리를 맞대고 깊이 잠든 평온한 구도.\n\nLOCATION (lock): In the open cargo bed of a truck parked in the shade of a rocky outcrop, beneath a movable shade cloth.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Truck-bed corner (Occupied by the three sleeping companions in the parked truck) — The inner corner is visible diagonally from above the outer side; used as Encloses the sleeping cluster without arranging the bodies into a formal row; Truck-bed edge (Visible along the near margin) — Its inner face recedes diagonally toward the sleepers; used as A narrow foreground boundary and continuity line for the lateral track.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight subdued by the sheltering rocks gives the sleeping faces soft contrast and a peaceful, unforced appearance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S69sh2__bgfirst_bg.png",
  "asset_id": "d14c0d0e-d685-4178-b74c-02a372218bdc",
  "input_asset_ids": [
   "772d313a-cec1-42dd-9d7c-7d2e9bd4f9bc",
   "3e51b6e9-6c91-4500-b642-35d55d0f8d22"
  ]
 },
 "S69sh2": {
  "input_fingerprint": "88b31ca155bbf630",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 트럭 짐칸 구석에서 이현우, 앰버, 라울이 머리를 맞대고 깊이 잠든 평온한 구도.\n\nLOCATION (lock): In the open cargo bed of a truck parked in the shade of a rocky outcrop, beneath a movable shade cloth. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Truck-bed corner (Occupied by the three sleeping companions in the parked truck) — The inner corner is visible diagonally from above the outer side; used as Encloses the sleeping cluster without arranging the bodies into a formal row; Truck-bed edge (Visible along the near margin) — Its inner face recedes diagonally toward the sleepers; used as A narrow foreground boundary and continuity line for the lateral track.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight subdued by the sheltering rocks gives the sleeping faces soft contrast and a peaceful, unforced appearance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's cargo bed with his head gathered together with Amber's and Raul's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's cargo bed with her head gathered together with Hyunwoo's and Raul's beneath the shade. Her exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's cargo bed with his head gathered together with Hyunwoo's and Amber's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old open-topped truck is parked in the shade of rocks. A movable shade cover is present over the cargo-bed resting area. 이현우: He is asleep in the rear cargo bed, with his head toward the center of the sleeping cluster. His accumulated injuries remain, including the treated, bandaged leg. 앰버: She is asleep in the rear cargo bed, with her head toward the center of the sleeping cluster. Her recent head injury remains untreated on the page. 라울: He is asleep in the rear cargo bed, with his head toward the center of the sleeping cluster.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 트럭 짐칸 구석에서 이현우, 앰버, 라울이 머리를 맞대고 깊이 잠든 평온한 구도.\n\nLOCATION (lock): In the open cargo bed of a truck parked in the shade of a rocky outcrop, beneath a movable shade cloth. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Truck-bed corner (Occupied by the three sleeping companions in the parked truck) — The inner corner is visible diagonally from above the outer side; used as Encloses the sleeping cluster without arranging the bodies into a formal row; Truck-bed edge (Visible along the near margin) — Its inner face recedes diagonally toward the sleepers; used as A narrow foreground boundary and continuity line for the lateral track.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight subdued by the sheltering rocks gives the sleeping faces soft contrast and a peaceful, unforced appearance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's cargo bed with his head gathered together with Amber's and Raul's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's cargo bed with her head gathered together with Hyunwoo's and Raul's beneath the shade. Her exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's cargo bed with his head gathered together with Hyunwoo's and Amber's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old open-topped truck is parked in the shade of rocks. A movable shade cover is present over the cargo-bed resting area. 이현우: He is asleep in the rear cargo bed, with his head toward the center of the sleeping cluster. His accumulated injuries remain, including the treated, bandaged leg. 앰버: She is asleep in the rear cargo bed, with her head toward the center of the sleeping cluster. Her recent head injury remains untreated on the page. 라울: He is asleep in the rear cargo bed, with his head toward the center of the sleeping cluster.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 트럭 짐칸 구석에서 이현우, 앰버, 라울이 머리를 맞대고 깊이 잠든 평온한 구도.\n\nLOCATION (lock): In the open cargo bed of a truck parked in the shade of a rocky outcrop, beneath a movable shade cloth. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Truck-bed corner (Occupied by the three sleeping companions in the parked truck) — The inner corner is visible diagonally from above the outer side; used as Encloses the sleeping cluster without arranging the bodies into a formal row; Truck-bed edge (Visible along the near margin) — Its inner face recedes diagonally toward the sleepers; used as A narrow foreground boundary and continuity line for the lateral track.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight subdued by the sheltering rocks gives the sleeping faces soft contrast and a peaceful, unforced appearance.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's cargo bed with his head gathered together with Amber's and Raul's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's cargo bed with her head gathered together with Hyunwoo's and Raul's beneath the shade. Her exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's cargo bed with his head gathered together with Hyunwoo's and Amber's beneath the shade. His exact torso orientation, head support, arm and leg positions, and contact with the cargo-bed floor are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old open-topped truck is parked in the shade of rocks. A movable shade cover is present over the cargo-bed resting area. 이현우: He is asleep in the rear cargo bed, with his head toward the center of the sleeping cluster. His accumulated injuries remain, including the treated, bandaged leg. 앰버: She is asleep in the rear cargo bed, with her head toward the center of the sleeping cluster. Her recent head injury remains untreated on the page. 라울: He is asleep in the rear cargo bed, with his head toward the center of the sleeping cluster.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S69sh2__bgfirst_bg.png",
     "asset_id": "d14c0d0e-d685-4178-b74c-02a372218bdc",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S69sh2.png",
     "asset_id": "772d313a-cec1-42dd-9d7c-7d2e9bd4f9bc",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_coastal_hill_overlook_sel.png",
     "asset_id": "3e51b6e9-6c91-4500-b642-35d55d0f8d22",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "세 인물의 머리가 짐칸 중앙으로 모여 서로 맞대고 있는 구도를 형성함.",
    "built_space": "대각선 위에서 바라본 트럭 짐칸 모서리와 차광막, 배경의 바위 지형이 프롬프트대로 배치됨.",
    "entities": "세 인물의 인종 및 연령 특징은 일치하나, 이현우의 후드티와 앰버의 재킷이 누락됨.",
    "hard_violations": [
     "[gpt-high] 라울 뒤에 지시문이나 참고 이미지에 없는 녹색 및 흰색 수납함 두 개를 추가했다."
    ],
    "physics": "세 사람 모두 짐칸 바닥에 자연스럽게 몸을 뉘고 무게를 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "세 인물이 한 방향을 향해 일렬로 누워 있어 머리가 한데 모이지 않음.",
    "built_space": "트럭 짐칸과 차광막, 배경 바위가 나타남. 인물 배치가 모서리 구도를 제대로 활용하지 못함.",
    "entities": "이현우, 앰버, 라울의 외양과 의상 디테일이 참조 이미지와 매우 정확히 일치함.",
    "hard_violations": [
     "[gemini-pro] 트럭 좌측 하단 뒷문에 식별 가능한 텍스트('HYUN') 누출",
     "[gpt-high] 화면 하단 왼쪽 차체에 식별 가능한 영문 일부가 남아 있어, 읽을 수 있는 문자나 로고를 두지 말라는 조건을 위반한다."
    ],
    "physics": "바닥에 눕혀져 서로에게 기대며 중력을 자연스럽게 따르고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "일부 의상이 누락되었으나, 요구된 프레이밍과 '머리를 맞댄 클러스터 구도'(우선순위 2)를 정확히 연출하여 가장 우수합니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "캐릭터 의상 고증은 완벽하나, 금지된 일렬 배치(formal row) 구도를 사용하여 핵심 병치 지시를 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 인물의 머리가 짐칸 중앙으로 모여 서로 맞대고 있는 구도를 형성함.",
        "built_space": "대각선 위에서 바라본 트럭 짐칸 모서리와 차광막, 배경의 바위 지형이 프롬프트대로 배치됨.",
        "entities": "세 인물의 인종 및 연령 특징은 일치하나, 이현우의 후드티와 앰버의 재킷이 누락됨.",
        "hard_violations": [],
        "physics": "세 사람 모두 짐칸 바닥에 자연스럽게 몸을 뉘고 무게를 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "세 인물이 한 방향을 향해 일렬로 누워 있어 머리가 한데 모이지 않음.",
        "built_space": "트럭 짐칸과 차광막, 배경 바위가 나타남. 인물 배치가 모서리 구도를 제대로 활용하지 못함.",
        "entities": "이현우, 앰버, 라울의 외양과 의상 디테일이 참조 이미지와 매우 정확히 일치함.",
        "hard_violations": [
         "트럭 좌측 하단 뒷문에 식별 가능한 텍스트('HYUN') 누출"
        ],
        "physics": "바닥에 눕혀져 서로에게 기대며 중력을 자연스럽게 따르고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "일부 의상이 누락되었으나, 요구된 프레이밍과 '머리를 맞댄 클러스터 구도'(우선순위 2)를 정확히 연출하여 가장 우수합니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "캐릭터 의상 고증은 완벽하나, 금지된 일렬 배치(formal row) 구도를 사용하여 핵심 병치 지시를 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "세 인물의 머리가 짐칸 중앙으로 모여 서로 맞대고 있는 구도를 형성함.",
        "built_space": "대각선 위에서 바라본 트럭 짐칸 모서리와 차광막, 배경의 바위 지형이 프롬프트대로 배치됨.",
        "entities": "세 인물의 인종 및 연령 특징은 일치하나, 이현우의 후드티와 앰버의 재킷이 누락됨.",
        "hard_violations": [],
        "physics": "세 사람 모두 짐칸 바닥에 자연스럽게 몸을 뉘고 무게를 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "세 인물이 한 방향을 향해 일렬로 누워 있어 머리가 한데 모이지 않음.",
        "built_space": "트럭 짐칸과 차광막, 배경 바위가 나타남. 인물 배치가 모서리 구도를 제대로 활용하지 못함.",
        "entities": "이현우, 앰버, 라울의 외양과 의상 디테일이 참조 이미지와 매우 정확히 일치함.",
        "hard_violations": [
         "트럭 좌측 하단 뒷문에 식별 가능한 텍스트('HYUN') 누출"
        ],
        "physics": "바닥에 눕혀져 서로에게 기대며 중력을 자연스럽게 따르고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "의상과 수면 자세는 비교적 충실하지만, 전경 차체가 지나치게 넓은 구도이며 하단의 식별 가능한 영문 표기가 무문자 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "머리를 모은 비정렬 수면 구도와 중간 크기 프레이밍은 더 충실하지만, 지시되지 않은 수납함 두 개가 추가되어 최종 사용에는 부적합하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 사람 모두 눈을 감고 있다. 왼쪽 현우는 오른쪽의 앰버 쪽으로, 오른쪽 라울은 왼쪽의 앰버 쪽으로 머리를 기울인다. 앰버의 머리는 현우의 머리에 닿고 라울은 앰버의 어깨에 기대어, 머리들이 한곳으로 모인다. 겨누는 물체나 이동 중인 몸은 없다.",
        "built_space": "적재함 바닥 하나와 좌우 측벽, 가까운 후면 판, 운전석 뒤쪽 격벽이 보인다. 위에는 차광막 한 장이 걸려 있고 뒤에는 운전석 후면 유리 하나가 있다. 세 사람은 격벽 가까이에 모여 있다. 외측 위에서 비스듬히 내려다보지만, 가까운 적재함 판의 외면이 화면 하단을 크게 차지해 요구한 좁은 전경 경계보다 두껍다. 배경에는 참고 사진과 유사한 풍화 구멍의 바위와 해안이 있으나 왼쪽의 큰 돌출 바위는 참고 사진에서 확인되지 않는 형태다.",
        "entities": "인물은 정확히 세 명이다. 현우는 헝클어진 검은 머리의 젊은 한국계 남성으로 보이며 짙은 후드와 낡은 청바지가 참고와 가깝고 무릎 부근에 붕대가 보인다. 앰버는 금발 여자아이이며 연녹색 겉옷과 회색 원피스가 맞지만, 참고의 마스크는 없고 미처치 머리 상처도 확인되지 않는다. 라울은 갈색 피부의 어린 남자아이이며 묶은 곱슬머리, 올리브색 티셔츠와 청반바지가 맞는다. 두 아이의 정확한 혼혈 배경은 외형만으로 확정할 수 없다. 낡은 개방형 트럭과 차광막은 표현되었으며 하단 왼쪽 차체에는 영문 일부가 식별된다.",
        "hard_violations": [
         "화면 하단 왼쪽 차체에 식별 가능한 영문 일부가 남아 있어, 읽을 수 있는 문자나 로고를 두지 말라는 조건을 위반한다."
        ],
        "physics": "현우의 몸통과 굽힌 다리는 적재함 바닥에 놓이고 머리는 접은 팔과 손에 받쳐져 있다. 앰버의 몸과 다리는 바닥에 놓이고 머리는 팔과 현우 쪽에 기대어 있다. 라울의 머리는 앰버의 어깨에, 팔과 손은 자기 다리와 동료의 몸에 내려앉아 있다. 지지 없이 뜬 신체 부위는 보이지 않는다. 차광막은 가장자리 결속점과 차량 쪽 지지 구조 사이에서 처져 있다."
       },
       {
        "label": "B",
        "direction": "세 사람 모두 눈을 감고 잠들어 있다. 현우는 오른쪽 아래의 앰버 쪽으로, 라울은 왼쪽 아래의 앰버 쪽으로 머리를 낮췄다. 앰버는 두 사람보다 앞쪽에 누워 머리를 중앙으로 향한다. 세 머리가 삼각형으로 밀집하여 일렬 배치보다 자연스럽게 머리를 맞댄 관계가 보인다. 조준하거나 이동하는 물체는 없다.",
        "built_space": "적재함의 바닥 하나와 뒤쪽에서 모이는 두 벽면, 가까운 측면 판, 오른쪽 운전석 외장이 보인다. 차광막 한 장은 보이는 두 수직 지지대와 결속줄에 걸려 있다. 카메라는 적재함 바깥 위에서 안쪽 모서리를 대각선으로 내려다보고, 가까운 가장자리가 수면 집단 쪽으로 이어진다. 인물 비중이 A보다 커 중간 크기 촬영에 더 가깝다. 바위의 풍화 구멍, 소나무와 해안 풍경은 참고 장소의 특징을 반영하지만 바위 배열이 정확히 같지는 않다.",
        "entities": "인물은 정확히 세 명이다. 현우의 젊은 한국계 얼굴과 검은 머리, 앰버의 금발과 여자아이 체격, 라울의 갈색 피부와 묶은 곱슬머리는 지정 인물에 대체로 부합한다. 현우는 참고의 후드 대신 반소매 티셔츠를 입었고, 앰버도 연녹색 겉옷과 마스크가 없다. 라울의 티셔츠와 청반바지는 참고에 가깝다. 현우의 다리에는 붕대가 있고 앰버의 이마에는 붕대로 가리지 않은 상처가 있다. 낡은 트럭과 차광막 외에 오른쪽에 녹색 수납함 하나와 흰색 수납함 하나가 추가되어 있다. 읽을 수 있는 문자는 보이지 않는다.",
        "hard_violations": [
         "라울 뒤에 지시문이나 참고 이미지에 없는 녹색 및 흰색 수납함 두 개를 추가했다."
        ],
        "physics": "세 사람의 골반과 굽힌 다리는 적재함 바닥에 놓여 있다. 현우의 머리는 중앙의 겹친 팔과 동료 쪽에 기대고, 앰버의 머리는 바닥에 내려놓은 접힌 팔에 받쳐져 있으며, 라울의 머리도 중앙의 팔과 동료 쪽으로 기대어 있다. 손과 팔이 허공에 들린 모습은 없다. 수납함은 바닥에 놓여 있고 차광막은 기둥과 줄에 고정되어 있어, 지지 없는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "의상과 수면 자세는 비교적 충실하지만, 전경 차체가 지나치게 넓은 구도이며 하단의 식별 가능한 영문 표기가 무문자 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "머리를 모은 비정렬 수면 구도와 중간 크기 프레이밍은 더 충실하지만, 지시되지 않은 수납함 두 개가 추가되어 최종 사용에는 부적합하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "세 사람 모두 눈을 감고 있다. 왼쪽 현우는 오른쪽의 앰버 쪽으로, 오른쪽 라울은 왼쪽의 앰버 쪽으로 머리를 기울인다. 앰버의 머리는 현우의 머리에 닿고 라울은 앰버의 어깨에 기대어, 머리들이 한곳으로 모인다. 겨누는 물체나 이동 중인 몸은 없다.",
        "built_space": "적재함 바닥 하나와 좌우 측벽, 가까운 후면 판, 운전석 뒤쪽 격벽이 보인다. 위에는 차광막 한 장이 걸려 있고 뒤에는 운전석 후면 유리 하나가 있다. 세 사람은 격벽 가까이에 모여 있다. 외측 위에서 비스듬히 내려다보지만, 가까운 적재함 판의 외면이 화면 하단을 크게 차지해 요구한 좁은 전경 경계보다 두껍다. 배경에는 참고 사진과 유사한 풍화 구멍의 바위와 해안이 있으나 왼쪽의 큰 돌출 바위는 참고 사진에서 확인되지 않는 형태다.",
        "entities": "인물은 정확히 세 명이다. 현우는 헝클어진 검은 머리의 젊은 한국계 남성으로 보이며 짙은 후드와 낡은 청바지가 참고와 가깝고 무릎 부근에 붕대가 보인다. 앰버는 금발 여자아이이며 연녹색 겉옷과 회색 원피스가 맞지만, 참고의 마스크는 없고 미처치 머리 상처도 확인되지 않는다. 라울은 갈색 피부의 어린 남자아이이며 묶은 곱슬머리, 올리브색 티셔츠와 청반바지가 맞는다. 두 아이의 정확한 혼혈 배경은 외형만으로 확정할 수 없다. 낡은 개방형 트럭과 차광막은 표현되었으며 하단 왼쪽 차체에는 영문 일부가 식별된다.",
        "hard_violations": [
         "화면 하단 왼쪽 차체에 식별 가능한 영문 일부가 남아 있어, 읽을 수 있는 문자나 로고를 두지 말라는 조건을 위반한다."
        ],
        "physics": "현우의 몸통과 굽힌 다리는 적재함 바닥에 놓이고 머리는 접은 팔과 손에 받쳐져 있다. 앰버의 몸과 다리는 바닥에 놓이고 머리는 팔과 현우 쪽에 기대어 있다. 라울의 머리는 앰버의 어깨에, 팔과 손은 자기 다리와 동료의 몸에 내려앉아 있다. 지지 없이 뜬 신체 부위는 보이지 않는다. 차광막은 가장자리 결속점과 차량 쪽 지지 구조 사이에서 처져 있다."
       },
       {
        "label": "A",
        "direction": "세 사람 모두 눈을 감고 잠들어 있다. 현우는 오른쪽 아래의 앰버 쪽으로, 라울은 왼쪽 아래의 앰버 쪽으로 머리를 낮췄다. 앰버는 두 사람보다 앞쪽에 누워 머리를 중앙으로 향한다. 세 머리가 삼각형으로 밀집하여 일렬 배치보다 자연스럽게 머리를 맞댄 관계가 보인다. 조준하거나 이동하는 물체는 없다.",
        "built_space": "적재함의 바닥 하나와 뒤쪽에서 모이는 두 벽면, 가까운 측면 판, 오른쪽 운전석 외장이 보인다. 차광막 한 장은 보이는 두 수직 지지대와 결속줄에 걸려 있다. 카메라는 적재함 바깥 위에서 안쪽 모서리를 대각선으로 내려다보고, 가까운 가장자리가 수면 집단 쪽으로 이어진다. 인물 비중이 A보다 커 중간 크기 촬영에 더 가깝다. 바위의 풍화 구멍, 소나무와 해안 풍경은 참고 장소의 특징을 반영하지만 바위 배열이 정확히 같지는 않다.",
        "entities": "인물은 정확히 세 명이다. 현우의 젊은 한국계 얼굴과 검은 머리, 앰버의 금발과 여자아이 체격, 라울의 갈색 피부와 묶은 곱슬머리는 지정 인물에 대체로 부합한다. 현우는 참고의 후드 대신 반소매 티셔츠를 입었고, 앰버도 연녹색 겉옷과 마스크가 없다. 라울의 티셔츠와 청반바지는 참고에 가깝다. 현우의 다리에는 붕대가 있고 앰버의 이마에는 붕대로 가리지 않은 상처가 있다. 낡은 트럭과 차광막 외에 오른쪽에 녹색 수납함 하나와 흰색 수납함 하나가 추가되어 있다. 읽을 수 있는 문자는 보이지 않는다.",
        "hard_violations": [
         "라울 뒤에 지시문이나 참고 이미지에 없는 녹색 및 흰색 수납함 두 개를 추가했다."
        ],
        "physics": "세 사람의 골반과 굽힌 다리는 적재함 바닥에 놓여 있다. 현우의 머리는 중앙의 겹친 팔과 동료 쪽에 기대고, 앰버의 머리는 바닥에 내려놓은 접힌 팔에 받쳐져 있으며, 라울의 머리도 중앙의 팔과 동료 쪽으로 기대어 있다. 손과 팔이 허공에 들린 모습은 없다. 수납함은 바닥에 놓여 있고 차광막은 기둥과 줄에 고정되어 있어, 지지 없는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.464
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.214
   },
   "violations": {
    "B": [
     "[gemini-pro] 트럭 좌측 하단 뒷문에 식별 가능한 텍스트('HYUN') 누출",
     "[gpt-high] 화면 하단 왼쪽 차체에 식별 가능한 영문 일부가 남아 있어, 읽을 수 있는 문자나 로고를 두지 말라는 조건을 위반한다."
    ],
    "A": [
     "[gpt-high] 라울 뒤에 지시문이나 참고 이미지에 없는 녹색 및 흰색 수납함 두 개를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1214
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "일부 의상이 누락되었으나, 요구된 프레이밍과 '머리를 맞댄 클러스터 구도'(우선순위 2)를 정확히 연출하여 가장 우수합니다.  ★위반: [gpt-high] 라울 뒤에 지시문이나 참고 이미지에 없는 녹색 및 흰색 수납함 두 개를 추가했다."
   },
   {
    "label": "B",
    "score": 1214,
    "verdict_ko": "캐릭터 의상 고증은 완벽하나, 금지된 일렬 배치(formal row) 구도를 사용하여 핵심 병치 지시를 위반했습니다.  ★위반: [gemini-pro] 트럭 좌측 하단 뒷문에 식별 가능한 텍스트('HYUN') 누출 / [gpt-high] 화면 하단 왼쪽 차체에 식별 가능한 영문 일부가 남아 있어, 읽을 수 있는 문자나 로고를 두지 말라는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_coastal_hill_overlook_sel.png",
    "asset_id": "3e51b6e9-6c91-4500-b642-35d55d0f8d22",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:943979>",
    "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace2b-1352-75cc-b871-39389c1ff927",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S69sh2__bgfirst_bg.png",
   "bg_asset_id": "d14c0d0e-d685-4178-b74c-02a372218bdc",
   "bg_record_key": "S69sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S69sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:07:24.668934+00:00",
  "fingerprint": "44668c0d8fcb7974e433718ac3dfb083bd6e7d1dfffd33e880af58466043bd07",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S69sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S69sh2_sel.png",
  "source_sha256": "343e401d94d54e0a2c5720d246a4ce15f8984772bf9749fa383e902d4db335b3",
  "file": "S69sh2_cine.png",
  "staged_sha256": "8f764ea450f05aba6c027e847d2716d142223879e03170b929cf5004ba9ea667",
  "latency_ms": 11606
 },
 "S69sh4::signage": {
  "fp": "c286c98b24151ad6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S69sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:08:34.397407+00:00",
  "fingerprint": "c011381ab7bc874dd71ed3b8cd5605748ea9035b83a4f37d5e558cd0598aeebb",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S69sh4_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S69sh4_sel.png",
  "source_sha256": "80c973dd50bde1fbeabcf3edb664598605ba78aced09089b2edf9aad43ea921c",
  "file": "S69sh4_cine.png",
  "staged_sha256": "4d4ee86c5b23b2d0a958553aeb47a9f813261561b58260d7a5a248dd7010b656",
  "latency_ms": 11423
 },
 "S70sh8::signage": {
  "fp": "2c30e9a7efb1d098",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::c136306f7326421e": {
  "subjects": [],
  "subject_text": "군 병원 박철진의 병실\n병상과 침구가 놓인 병원 입원실. 출입문이 복도와 연결되고 병상 주변으로 좁은 여유 공간이 남아 있다.",
  "identity": "canonical",
  "scope_id": "L106",
  "scope_role": "location_interior",
  "scope_sha": "2d3fe8035d44fd92"
 },
 "groupbg::military_hospital_room": {
  "input_fingerprint": "dce3261764df162e",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "military_hospital_room",
    "tags": [
     "S70sh10",
     "S70sh8"
    ]
   },
   "context_sig": "4f9f230376464a2d"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the bedside near the doorway of a military hospital patient room, under nighttime hospital lighting.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n군 병원 박철진의 병실: 한쪽 눈에 안대를 하고 팔다리에 무거운 깁스를 한 부상병이 누워 있는 병상. (특징: 하얀 붕대와 육중한 깁스 장치; 수액 링거가 걸린 철제 병대)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 병실에 누워있는 사람은... 한 쪽 눈, 팔과 다리 깁스를 하고 치료 중인 박철진이다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the bedside near the doorway of a military hospital patient room, under nighttime hospital lighting.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n군 병원 박철진의 병실: 한쪽 눈에 안대를 하고 팔다리에 무거운 깁스를 한 부상병이 누워 있는 병상. (특징: 하얀 붕대와 육중한 깁스 장치; 수액 링거가 걸린 철제 병대)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 병실에 누워있는 사람은... 한 쪽 눈, 팔과 다리 깁스를 하고 치료 중인 박철진이다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_military_hospital_room_03987e.png",
  "asset_id": "6bc0dc72-529b-4213-9d79-2ddb693b45c3",
  "input_asset_ids": [
   "b32800d2-82f1-4492-ba48-7cb92e1273ba"
  ],
  "origin_tag": "S70sh8",
  "place_text": "At the bedside near the doorway of a military hospital patient room, under nighttime hospital lighting.",
  "origin_inputs": {
   "place_text": "At the bedside near the doorway of a military hospital patient room, under nighttime hospital lighting.",
   "time_of_day_en": "night",
   "conti_asset_id": "b32800d2-82f1-4492-ba48-7cb92e1273ba"
  }
 },
 "S70sh8::bgfirst_bg": {
  "input_fingerprint": "f337bd4ffffe8520",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 국방장관이 병상 위의 박철진을 외면한 채 복도 쪽으로 고개를 확 돌린 매몰찬 찰나.\n\nLOCATION (lock): At the bedside near the doorway of a military hospital patient room, under nighttime hospital lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Hospital bed (Occupied by 박철진 during treatment) — Viewed along its side from beside the patient's shoulder; used as Anchors the patient low in the frame beneath the minister; Ward doorway and corridor (The passage toward the corridor is visible) — Seen obliquely to the right of the minister; used as Gives the minister's averted attention a concrete exit direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination appropriate to the nighttime hospital maintains legible faces and a severe contrast between authority and vulnerability.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 국방장관이 병상 위의 박철진을 외면한 채 복도 쪽으로 고개를 확 돌린 매몰찬 찰나.\n\nLOCATION (lock): At the bedside near the doorway of a military hospital patient room, under nighttime hospital lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Hospital bed (Occupied by 박철진 during treatment) — Viewed along its side from beside the patient's shoulder; used as Anchors the patient low in the frame beneath the minister; Ward doorway and corridor (The passage toward the corridor is visible) — Seen obliquely to the right of the minister; used as Gives the minister's averted attention a concrete exit direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination appropriate to the nighttime hospital maintains legible faces and a severe contrast between authority and vulnerability.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S70sh8__bgfirst_bg.png",
  "asset_id": "b137fe69-2861-4d1f-9ee1-fc4be2d579ad",
  "input_asset_ids": [
   "b32800d2-82f1-4492-ba48-7cb92e1273ba",
   "6bc0dc72-529b-4213-9d79-2ddb693b45c3"
  ]
 },
 "S70sh8": {
  "input_fingerprint": "4739884a03083250",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 국방장관이 병상 위의 박철진을 외면한 채 복도 쪽으로 고개를 확 돌린 매몰찬 찰나.\n\nLOCATION (lock): At the bedside near the doorway of a military hospital patient room, under nighttime hospital lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Hospital bed (Occupied by 박철진 during treatment) — Viewed along its side from beside the patient's shoulder; used as Anchors the patient low in the frame beneath the minister; Ward doorway and corridor (The passage toward the corridor is visible) — Seen obliquely to the right of the minister; used as Gives the minister's averted attention a concrete exit direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination appropriate to the nighttime hospital maintains legible faces and a severe contrast between authority and vulnerability.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Cheoljin is reclining on a hospital bed, severely injured, with one eye covered and an arm and a leg immobilized in casts; his uncovered eye looks upward. His exact head support, torso inclination, limb placement, and the sides of the covered eye and cast limbs are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A military-hospital patient room opens onto the corridor. 국방장관: He is leaving the bedside for the corridor.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 국방장관이 병상 위의 박철진을 외면한 채 복도 쪽으로 고개를 확 돌린 매몰찬 찰나.\n\nLOCATION (lock): At the bedside near the doorway of a military hospital patient room, under nighttime hospital lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Hospital bed (Occupied by 박철진 during treatment) — Viewed along its side from beside the patient's shoulder; used as Anchors the patient low in the frame beneath the minister; Ward doorway and corridor (The passage toward the corridor is visible) — Seen obliquely to the right of the minister; used as Gives the minister's averted attention a concrete exit direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination appropriate to the nighttime hospital maintains legible faces and a severe contrast between authority and vulnerability.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Cheoljin is reclining on a hospital bed, severely injured, with one eye covered and an arm and a leg immobilized in casts; his uncovered eye looks upward. His exact head support, torso inclination, limb placement, and the sides of the covered eye and cast limbs are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A military-hospital patient room opens onto the corridor. 국방장관: He is leaving the bedside for the corridor.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 국방장관이 병상 위의 박철진을 외면한 채 복도 쪽으로 고개를 확 돌린 매몰찬 찰나.\n\nLOCATION (lock): At the bedside near the doorway of a military hospital patient room, under nighttime hospital lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Hospital bed (Occupied by 박철진 during treatment) — Viewed along its side from beside the patient's shoulder; used as Anchors the patient low in the frame beneath the minister; Ward doorway and corridor (The passage toward the corridor is visible) — Seen obliquely to the right of the minister; used as Gives the minister's averted attention a concrete exit direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination appropriate to the nighttime hospital maintains legible faces and a severe contrast between authority and vulnerability.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Cheoljin is reclining on a hospital bed, severely injured, with one eye covered and an arm and a leg immobilized in casts; his uncovered eye looks upward. His exact head support, torso inclination, limb placement, and the sides of the covered eye and cast limbs are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A military-hospital patient room opens onto the corridor. 국방장관: He is leaving the bedside for the corridor.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S70sh8__bgfirst_bg.png",
     "asset_id": "b137fe69-2861-4d1f-9ee1-fc4be2d579ad",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S70sh8.png",
     "asset_id": "b32800d2-82f1-4492-ba48-7cb92e1273ba",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:684284>",
     "asset_id": "b8636b60-6a4d-4eb8-bb28-08231e30fcd5",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_military_hospital_room_03987e.png",
     "asset_id": "6bc0dc72-529b-4213-9d79-2ddb693b45c3",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:684284>",
     "asset_id": "b8636b60-6a4d-4eb8-bb28-08231e30fcd5",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "장관은 침대에서 시선을 거두고 열린 문 너머 복도 쪽을 향해 고개를 돌리고 있으며, 환자는 위를 응시함.",
    "built_space": "병실 내부. 위치 참조 사진의 침대, 벽면 의료 기기, 커튼, 열린 문과 복도의 조명 및 구조가 정확히 배치됨.",
    "entities": "박철진은 한쪽 눈에 안대를 하고 팔다리에 깁스를 한 채 누워 있음. 장관은 참조 이미지의 외형(인종, 수염, 모자)을 전혀 반영하지 않은 완전히 다른 아시아계 남성으로 렌더링됨.",
    "hard_violations": [
     "[gemini-pro] 지정된 캐릭터 레퍼런스와 완전히 다른 인물(창조된 인물)이 국방장관으로 등장함."
    ],
    "physics": "침대에 누운 환자의 몸과 바닥에 서 있는 장관의 자세 모두 중력에 맞게 자연스럽게 지지됨."
   },
   {
    "label": "B",
    "direction": "장관은 환자를 외면한 채 복도를 바라보고 있으며, 침대에 누운 환자의 드러난 눈은 위쪽을 향함.",
    "built_space": "병실 내부. 참조 사진과 동일한 병상, 의료용 패널, 반쯤 열린 병실 문, 그리고 이어지는 복도의 디테일이 충실히 재현됨.",
    "entities": "박철진은 프롬프트대로 깁스를 하고 한쪽 눈이 가려진 상태임. 장관은 레퍼런스 이미지의 수염, 이목구비, 정장, 페도라 모자를 정확하게 착용하고 있음.",
    "hard_violations": [],
    "physics": "환자의 체중이 병상에 자연스럽게 실려 있으며, 서 있는 장관의 자세도 바닥에 온전히 지지되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "국방장관의 캐릭터 레퍼런스(외모, 의상, 모자)를 정확히 일치시켰으며, 복도로 고개를 돌린 매몰찬 찰나의 연출과 환자의 캐논 포즈를 충실하게 구현함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "병실 배경과 환자의 상태는 잘 표현되었으나, 국방장관이 캐릭터 레퍼런스와 전혀 다른 인물로 렌더링되어 핵심 지침을 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장관은 침대에서 시선을 거두고 열린 문 너머 복도 쪽을 향해 고개를 돌리고 있으며, 환자는 위를 응시함.",
        "built_space": "병실 내부. 위치 참조 사진의 침대, 벽면 의료 기기, 커튼, 열린 문과 복도의 조명 및 구조가 정확히 배치됨.",
        "entities": "박철진은 한쪽 눈에 안대를 하고 팔다리에 깁스를 한 채 누워 있음. 장관은 참조 이미지의 외형(인종, 수염, 모자)을 전혀 반영하지 않은 완전히 다른 아시아계 남성으로 렌더링됨.",
        "hard_violations": [
         "지정된 캐릭터 레퍼런스와 완전히 다른 인물(창조된 인물)이 국방장관으로 등장함."
        ],
        "physics": "침대에 누운 환자의 몸과 바닥에 서 있는 장관의 자세 모두 중력에 맞게 자연스럽게 지지됨."
       },
       {
        "label": "B",
        "direction": "장관은 환자를 외면한 채 복도를 바라보고 있으며, 침대에 누운 환자의 드러난 눈은 위쪽을 향함.",
        "built_space": "병실 내부. 참조 사진과 동일한 병상, 의료용 패널, 반쯤 열린 병실 문, 그리고 이어지는 복도의 디테일이 충실히 재현됨.",
        "entities": "박철진은 프롬프트대로 깁스를 하고 한쪽 눈이 가려진 상태임. 장관은 레퍼런스 이미지의 수염, 이목구비, 정장, 페도라 모자를 정확하게 착용하고 있음.",
        "hard_violations": [],
        "physics": "환자의 체중이 병상에 자연스럽게 실려 있으며, 서 있는 장관의 자세도 바닥에 온전히 지지되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "국방장관의 캐릭터 레퍼런스(외모, 의상, 모자)를 정확히 일치시켰으며, 복도로 고개를 돌린 매몰찬 찰나의 연출과 환자의 캐논 포즈를 충실하게 구현함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "병실 배경과 환자의 상태는 잘 표현되었으나, 국방장관이 캐릭터 레퍼런스와 전혀 다른 인물로 렌더링되어 핵심 지침을 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "장관은 침대에서 시선을 거두고 열린 문 너머 복도 쪽을 향해 고개를 돌리고 있으며, 환자는 위를 응시함.",
        "built_space": "병실 내부. 위치 참조 사진의 침대, 벽면 의료 기기, 커튼, 열린 문과 복도의 조명 및 구조가 정확히 배치됨.",
        "entities": "박철진은 한쪽 눈에 안대를 하고 팔다리에 깁스를 한 채 누워 있음. 장관은 참조 이미지의 외형(인종, 수염, 모자)을 전혀 반영하지 않은 완전히 다른 아시아계 남성으로 렌더링됨.",
        "hard_violations": [
         "지정된 캐릭터 레퍼런스와 완전히 다른 인물(창조된 인물)이 국방장관으로 등장함."
        ],
        "physics": "침대에 누운 환자의 몸과 바닥에 서 있는 장관의 자세 모두 중력에 맞게 자연스럽게 지지됨."
       },
       {
        "label": "B",
        "direction": "장관은 환자를 외면한 채 복도를 바라보고 있으며, 침대에 누운 환자의 드러난 눈은 위쪽을 향함.",
        "built_space": "병실 내부. 참조 사진과 동일한 병상, 의료용 패널, 반쯤 열린 병실 문, 그리고 이어지는 복도의 디테일이 충실히 재현됨.",
        "entities": "박철진은 프롬프트대로 깁스를 하고 한쪽 눈이 가려진 상태임. 장관은 레퍼런스 이미지의 수염, 이목구비, 정장, 페도라 모자를 정확하게 착용하고 있음.",
        "hard_violations": [],
        "physics": "환자의 체중이 병상에 자연스럽게 실려 있으며, 서 있는 장관의 자세도 바닥에 온전히 지지되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "환자를 화면 아래에 두고 장관이 오른쪽 복도로 고개만 돌린 관계와 참조 인물의 정체성을 잘 살렸지만, 미디엄 숏보다 다소 넓고 떠나는 동작은 약하다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "복도로 떠나는 움직임은 자연스럽지만, 고개를 확 돌린 순간보다 몸 전체가 돌아선 모습이며 환자의 화면 점유가 크고 장관의 얼굴·수염·모자가 참조와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장관의 몸통은 병상 쪽으로 열려 있고 얼굴과 시선은 화면 오른쪽 열린 출입문과 복도를 향한다. 환자와 눈을 맞추지 않아 외면하는 관계가 분명하다. 박철진의 가리지 않은 눈은 위쪽을 향한다. 무기나 겨냥하는 소품은 없다.",
        "built_space": "전경의 병상 하나, 왼쪽 수액대와 수액 용기 각 하나, 벽면 의료 설비대 하나, 모니터 하나, 협탁 하나, 뒤쪽 커튼 하나, 오른쪽 열린 문 하나가 보인다. 참조의 재료와 배치, 복도의 손잡이와 천장 조명이 유지된다. 장관은 침대와 출입문 사이에 서고 환자는 그 아래에 놓인다. 병상 옆을 보는 구도지만 발치까지 포함해 요구한 미디엄 숏보다 다소 넓다. 불가능한 거울 반사는 없다.",
        "entities": "두 인물은 장관과 병상의 박철진으로 읽히며 추가 인물은 없다. 장관은 참조처럼 검은 페도라, 짧은 수염, 어두운 더블브레스트 정장, 흰 셔츠와 회색 넥타이를 착용하고 얼굴과 체격도 가깝다. 박철진은 성인 남성으로 한쪽 눈의 드레싱, 얼굴 상처, 팔 하나와 다리 하나의 깁스가 보인다. 인종·국적은 외양만으로 확정할 수 없다. 판독 가능한 글자나 그래픽 오버레이는 보이지 않는다.",
        "hard_violations": [],
        "physics": "환자의 머리와 상체는 베개와 기울어진 침상에 받쳐지고, 깁스한 팔과 손은 침구 위에 놓인다. 깁스한 다리도 매트리스에 길게 지지된다. 장관은 하체가 일부 가려진 직립 자세이며 공중에 뜬 징후는 없다. 수액 용기는 수액대에 매달려 있다. 목을 옆으로 돌린 자세는 가능하지만 걸음을 떼는 움직임은 뚜렷하지 않다."
       },
       {
        "label": "B",
        "direction": "장관의 얼굴과 시선은 오른쪽 출입문과 복도를 향하고 환자를 보지 않는다. 다만 어깨와 몸통도 같은 방향으로 돌아가 있어 고개만 급히 돌린 순간보다는 이미 떠나는 모습에 가깝다. 환자의 얼굴은 오른쪽 위로 향하지만 가리지 않은 눈은 이 각도에서 명확히 확인되지 않는다.",
        "built_space": "병상 하나, 왼쪽 수액대와 용기 각 하나, 의료 설비대 하나, 모니터 하나, 협탁 하나, 커튼 하나, 오른쪽 열린 문 하나가 보인다. 참조 병실의 벽 마감과 복도 구조는 잘 유지되며 장관은 출입문 바로 안쪽 병상 곁에 있다. 환자 어깨 가까이에서 병상 옆을 보는 시점은 맞지만, 크게 잡힌 환자의 머리와 상체가 화면 중간까지 올라와 장관 아래 낮게 배치하는 효과가 약하다. 복도 바닥의 조명 반사는 자연스럽다.",
        "entities": "장관과 박철진으로 읽히는 성인 남성 두 명만 보인다. 장관은 어두운 더블브레스트 정장과 넥타이는 맞지만 참조의 모자와 수염이 없고 얼굴형과 인상이 뚜렷하게 다르다. 박철진에게는 한쪽 눈을 덮은 패드와 머리 붕대, 팔 하나와 다리 하나의 깁스가 있다. 국적이나 민족적 정체성은 외양만으로 확정할 수 없다. 읽을 수 있는 글자나 오버레이는 보이지 않는다.",
        "hard_violations": [],
        "physics": "환자의 머리와 등은 베개와 올려진 침상에 지지되고, 깁스한 팔과 손은 복부 쪽 침구에 놓여 있다. 다리 깁스는 침상 위에 닿아 있으며 위로 향한 발은 깁스로 고정된 자세로 설명된다. 장관은 몸을 출구 쪽으로 옮기며 팔을 자연스럽게 벌린 모습이고, 발은 화면 밖이지만 부유하는 징후는 없다. 수액 용기는 수액대가 지지한다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "환자를 화면 아래에 두고 장관이 오른쪽 복도로 고개만 돌린 관계와 참조 인물의 정체성을 잘 살렸지만, 미디엄 숏보다 다소 넓고 떠나는 동작은 약하다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "복도로 떠나는 움직임은 자연스럽지만, 고개를 확 돌린 순간보다 몸 전체가 돌아선 모습이며 환자의 화면 점유가 크고 장관의 얼굴·수염·모자가 참조와 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "장관의 몸통은 병상 쪽으로 열려 있고 얼굴과 시선은 화면 오른쪽 열린 출입문과 복도를 향한다. 환자와 눈을 맞추지 않아 외면하는 관계가 분명하다. 박철진의 가리지 않은 눈은 위쪽을 향한다. 무기나 겨냥하는 소품은 없다.",
        "built_space": "전경의 병상 하나, 왼쪽 수액대와 수액 용기 각 하나, 벽면 의료 설비대 하나, 모니터 하나, 협탁 하나, 뒤쪽 커튼 하나, 오른쪽 열린 문 하나가 보인다. 참조의 재료와 배치, 복도의 손잡이와 천장 조명이 유지된다. 장관은 침대와 출입문 사이에 서고 환자는 그 아래에 놓인다. 병상 옆을 보는 구도지만 발치까지 포함해 요구한 미디엄 숏보다 다소 넓다. 불가능한 거울 반사는 없다.",
        "entities": "두 인물은 장관과 병상의 박철진으로 읽히며 추가 인물은 없다. 장관은 참조처럼 검은 페도라, 짧은 수염, 어두운 더블브레스트 정장, 흰 셔츠와 회색 넥타이를 착용하고 얼굴과 체격도 가깝다. 박철진은 성인 남성으로 한쪽 눈의 드레싱, 얼굴 상처, 팔 하나와 다리 하나의 깁스가 보인다. 인종·국적은 외양만으로 확정할 수 없다. 판독 가능한 글자나 그래픽 오버레이는 보이지 않는다.",
        "hard_violations": [],
        "physics": "환자의 머리와 상체는 베개와 기울어진 침상에 받쳐지고, 깁스한 팔과 손은 침구 위에 놓인다. 깁스한 다리도 매트리스에 길게 지지된다. 장관은 하체가 일부 가려진 직립 자세이며 공중에 뜬 징후는 없다. 수액 용기는 수액대에 매달려 있다. 목을 옆으로 돌린 자세는 가능하지만 걸음을 떼는 움직임은 뚜렷하지 않다."
       },
       {
        "label": "A",
        "direction": "장관의 얼굴과 시선은 오른쪽 출입문과 복도를 향하고 환자를 보지 않는다. 다만 어깨와 몸통도 같은 방향으로 돌아가 있어 고개만 급히 돌린 순간보다는 이미 떠나는 모습에 가깝다. 환자의 얼굴은 오른쪽 위로 향하지만 가리지 않은 눈은 이 각도에서 명확히 확인되지 않는다.",
        "built_space": "병상 하나, 왼쪽 수액대와 용기 각 하나, 의료 설비대 하나, 모니터 하나, 협탁 하나, 커튼 하나, 오른쪽 열린 문 하나가 보인다. 참조 병실의 벽 마감과 복도 구조는 잘 유지되며 장관은 출입문 바로 안쪽 병상 곁에 있다. 환자 어깨 가까이에서 병상 옆을 보는 시점은 맞지만, 크게 잡힌 환자의 머리와 상체가 화면 중간까지 올라와 장관 아래 낮게 배치하는 효과가 약하다. 복도 바닥의 조명 반사는 자연스럽다.",
        "entities": "장관과 박철진으로 읽히는 성인 남성 두 명만 보인다. 장관은 어두운 더블브레스트 정장과 넥타이는 맞지만 참조의 모자와 수염이 없고 얼굴형과 인상이 뚜렷하게 다르다. 박철진에게는 한쪽 눈을 덮은 패드와 머리 붕대, 팔 하나와 다리 하나의 깁스가 있다. 국적이나 민족적 정체성은 외양만으로 확정할 수 없다. 읽을 수 있는 글자나 오버레이는 보이지 않는다.",
        "hard_violations": [],
        "physics": "환자의 머리와 등은 베개와 올려진 침상에 지지되고, 깁스한 팔과 손은 복부 쪽 침구에 놓여 있다. 다리 깁스는 침상 위에 닿아 있으며 위로 향한 발은 깁스로 고정된 자세로 설명된다. 장관은 몸을 출구 쪽으로 옮기며 팔을 자연스럽게 벌린 모습이고, 발은 화면 밖이지만 부유하는 징후는 없다. 수액 용기는 수액대가 지지한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.179,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.929,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지정된 캐릭터 레퍼런스와 완전히 다른 인물(창조된 인물)이 국방장관으로 등장함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 929
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "국방장관의 캐릭터 레퍼런스(외모, 의상, 모자)를 정확히 일치시켰으며, 복도로 고개를 돌린 매몰찬 찰나의 연출과 환자의 캐논 포즈를 충실하게 구현함."
   },
   {
    "label": "A",
    "score": 929,
    "verdict_ko": "병실 배경과 환자의 상태는 잘 표현되었으나, 국방장관이 캐릭터 레퍼런스와 전혀 다른 인물로 렌더링되어 핵심 지침을 위반함.  ★위반: [gemini-pro] 지정된 캐릭터 레퍼런스와 완전히 다른 인물(창조된 인물)이 국방장관으로 등장함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_military_hospital_room_03987e.png",
    "asset_id": "6bc0dc72-529b-4213-9d79-2ddb693b45c3",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:684284>",
    "asset_id": "b8636b60-6a4d-4eb8-bb28-08231e30fcd5",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace37-e295-7cfa-9aa6-ba127cf6e1f4",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S70sh8__bgfirst_bg.png",
   "bg_asset_id": "b137fe69-2861-4d1f-9ee1-fc4be2d579ad",
   "bg_record_key": "S70sh8::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "military_hospital_room",
   "groupbg_asset_id": "6bc0dc72-529b-4213-9d79-2ddb693b45c3"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S70sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:10:44.443719+00:00",
  "fingerprint": "2704fd5264269d122bf6c884203051115104c4d1dca62f58de82afcd97a586fb",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S70sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S70sh8_sel.png",
  "source_sha256": "467fafbaae5f0d793bda88614ac38902d9cb30268d755b6f20800f2d0c406c56",
  "file": "S70sh8_cine.png",
  "staged_sha256": "27e889626803dcdb4fe2bb7a43cdac8f1f7f410db128708533d8484632828c82",
  "latency_ms": 12209
 },
 "S70sh10::signage": {
  "fp": "1f2a8212f17112f8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S70sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:11:52.443709+00:00",
  "fingerprint": "1b00850fed3c3ec8fcc92536b73e12fed017ff273b24b4dfe24ab1ee39189c1a",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S70sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S70sh10_sel.png",
  "source_sha256": "7f8eab0e73f92476c7064a87750c39759465bf5d4e322d9bb388e62253fae28a",
  "file": "S70sh10_cine.png",
  "staged_sha256": "0b0e2ce6f6e9f3055e0f752a9668fef0a15b7d0292eb371c25d58edd914ce24b",
  "latency_ms": 12487
 },
 "S71sh4::signage": {
  "fp": "db41f9e289142bc5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S71sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:12:58.987665+00:00",
  "fingerprint": "8e00898d44bc15997cf0da4a6de75afbcb817df9135655330d7d7382038eb2ce",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S71sh4_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S71sh4_sel.png",
  "source_sha256": "10f577bf29ab9eed892cf094f5af2036f0f5a31d5d18db33a25f45bf46892927",
  "file": "S71sh4_cine.png",
  "staged_sha256": "4c925ca8cc6df410231266b8ba1bad2ee89858d31ca9737c49be272d3b742ffd",
  "latency_ms": 12462
 },
 "S71sh13::signage": {
  "fp": "5d6416885ece332f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::e5474b9c2b17cdaf": {
  "subjects": [],
  "subject_text": "지동현의 연구소 인근 절벽 끝\n보라색 제비꽃밭 끝에서 지면이 급격히 끊기는 절벽. 꽃이 핀 평탄한 상부와 가파른 낭떠러지가 맞닿아 있다.",
  "identity": "canonical",
  "scope_id": "L112",
  "scope_role": "location_exterior",
  "scope_sha": "898043f30deed560"
 },
 "groupbg::coastal_search_hill": {
  "input_fingerprint": "f679056138151279",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "coastal_search_hill",
    "tags": [
     "S71sh13"
    ]
   },
   "context_sig": "4f84ea2de543e5ff"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: On an open roadside hill overlooking coastal ridges, beside the parked truck during the search for a violet field.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n지동현의 연구소 인근 절벽 끝: 꽃밭을 가로질러 도망치다 마주하는, 아래로 바다가 출렁이는 깎아지른 아찔한 해안 낭떠러지. (특징: 발끝에서 부서져 내리는 흙과 돌멩이; 포위망을 좁혀오는 병사들이 겨눈 총구들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 언덕 위에 서서 능선을 쳐다보는 현우. 저 멀리 능선 사이사이로 보이는 바다윤슬.\n- 현우가 능선 너머로 제비꽃을 찾는 사이, 앰버와 찰리, 라울은 흙으로 장난을 치고 있다.\n\nTIME OF DAY (lock): dawn.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: On an open roadside hill overlooking coastal ridges, beside the parked truck during the search for a violet field.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n지동현의 연구소 인근 절벽 끝: 꽃밭을 가로질러 도망치다 마주하는, 아래로 바다가 출렁이는 깎아지른 아찔한 해안 낭떠러지. (특징: 발끝에서 부서져 내리는 흙과 돌멩이; 포위망을 좁혀오는 병사들이 겨눈 총구들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 언덕 위에 서서 능선을 쳐다보는 현우. 저 멀리 능선 사이사이로 보이는 바다윤슬.\n- 현우가 능선 너머로 제비꽃을 찾는 사이, 앰버와 찰리, 라울은 흙으로 장난을 치고 있다.\n\nTIME OF DAY (lock): dawn.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_coastal_search_hill_7092b9.png",
  "asset_id": "3b958113-b2b7-416f-84ac-bf77de635bd9",
  "input_asset_ids": [
   "3845cba1-4eec-425a-b748-144e27b06465"
  ],
  "origin_tag": "S71sh13",
  "place_text": "On an open roadside hill overlooking coastal ridges, beside the parked truck during the search for a violet field.",
  "origin_inputs": {
   "place_text": "On an open roadside hill overlooking coastal ridges, beside the parked truck during the search for a violet field.",
   "time_of_day_en": "dawn",
   "conti_asset_id": "3845cba1-4eec-425a-b748-144e27b06465"
  }
 },
 "S71sh13::bgfirst_bg": {
  "input_fingerprint": "97f594f2b617a8ef",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 찰리를 올려다보며 눈가에 눈물이 맺힌 애절하고 슬픈 표정 클로즈업.\n\nLOCATION (lock): On an open roadside hill overlooking coastal ridges, beside the parked truck during the search for a violet field.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Hillside ground (The ground where the companions have been playing with dirt remains behind 앰버); used as A subdued background that retains the outdoor location without competing with her eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight gives the tears modest highlights while preserving natural facial detail and the quiet sadness of the exchange.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 찰리를 올려다보며 눈가에 눈물이 맺힌 애절하고 슬픈 표정 클로즈업.\n\nLOCATION (lock): On an open roadside hill overlooking coastal ridges, beside the parked truck during the search for a violet field.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Hillside ground (The ground where the companions have been playing with dirt remains behind 앰버); used as A subdued background that retains the outdoor location without competing with her eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight gives the tears modest highlights while preserving natural facial detail and the quiet sadness of the exchange.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S71sh13__bgfirst_bg.png",
  "asset_id": "72945e5f-679f-4ba0-aeb7-142d9df70439",
  "input_asset_ids": [
   "3845cba1-4eec-425a-b748-144e27b06465",
   "3b958113-b2b7-416f-84ac-bf77de635bd9"
  ]
 },
 "S71sh13": {
  "input_fingerprint": "8adad82bf2fefbee",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 앰버가 찰리를 올려다보며 눈가에 눈물이 맺힌 애절하고 슬픈 표정 클로즈업.\n\nLOCATION (lock): On an open roadside hill overlooking coastal ridges, beside the parked truck during the search for a violet field. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Hillside ground (The ground where the companions have been playing with dirt remains behind 앰버); used as A subdued background that retains the outdoor location without competing with her eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight gives the tears modest highlights while preserving natural facial detail and the quiet sadness of the exchange.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight hillside overlooks ridges and glittering seawater, with no violets visible here. The flower sketch remains on the reverse of its sheet. 앰버: She is at the dirt-play spot on the hillside, visibly tearful, and retains the received drawing. Her earlier head injury remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 앰버가 찰리를 올려다보며 눈가에 눈물이 맺힌 애절하고 슬픈 표정 클로즈업.\n\nLOCATION (lock): On an open roadside hill overlooking coastal ridges, beside the parked truck during the search for a violet field. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Hillside ground (The ground where the companions have been playing with dirt remains behind 앰버); used as A subdued background that retains the outdoor location without competing with her eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight gives the tears modest highlights while preserving natural facial detail and the quiet sadness of the exchange.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight hillside overlooks ridges and glittering seawater, with no violets visible here. The flower sketch remains on the reverse of its sheet. 앰버: She is at the dirt-play spot on the hillside, visibly tearful, and retains the received drawing. Her earlier head injury remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 앰버가 찰리를 올려다보며 눈가에 눈물이 맺힌 애절하고 슬픈 표정 클로즈업.\n\nLOCATION (lock): On an open roadside hill overlooking coastal ridges, beside the parked truck during the search for a violet field. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Hillside ground (The ground where the companions have been playing with dirt remains behind 앰버); used as A subdued background that retains the outdoor location without competing with her eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight gives the tears modest highlights while preserving natural facial detail and the quiet sadness of the exchange.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight hillside overlooks ridges and glittering seawater, with no violets visible here. The flower sketch remains on the reverse of its sheet. 앰버: She is at the dirt-play spot on the hillside, visibly tearful, and retains the received drawing. Her earlier head injury remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S71sh13__bgfirst_bg.png",
     "asset_id": "72945e5f-679f-4ba0-aeb7-142d9df70439",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S71sh13.png",
     "asset_id": "3845cba1-4eec-425a-b748-144e27b06465",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_coastal_search_hill_7092b9.png",
     "asset_id": "3b958113-b2b7-416f-84ac-bf77de635bd9",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "화면 왼쪽 전경의 인물(어깨)을 올려다봄.",
    "built_space": "해안가 언덕, 우측 배경에 트럭 위치.",
    "entities": "앰버의 외모는 일치하나 재킷 누락. 꽃 그림이 앞면으로 잘못 노출됨. 금지된 추가 인물 등장.",
    "hard_violations": [
     "[gemini-pro] 허용되지 않은 인물 등장(명시된 앰버 외의 인물 노출)",
     "[gpt-high] 등장 가능한 인물을 앰버로 한정한 조건과 달리, 성인의 턱·목·어깨를 전경에 추가했다."
    ],
    "physics": "서서 두 손으로 종이를 지탱하며 들고 있음."
   },
   {
    "label": "B",
    "direction": "프레임 밖 좌측 상단을 애절하게 올려다봄.",
    "built_space": "해안가 언덕, 능선과 바다가 보이는 야외.",
    "entities": "앰버의 외모와 복장(재킷 포함) 일치. 눈물 연기 구현됨. 머리 상처는 미구현.",
    "hard_violations": [],
    "physics": "자연스럽게 서서 지면에 안정적으로 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "허용되지 않은 인물의 신체가 등장하여 치명적 위반이며, 앰버의 재킷 누락 및 종이 방향 오류가 있음."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레임 밖을 보는 시선으로 제약을 잘 지켰고, 앰버의 복장과 애절한 눈물 연기를 정확히 재현함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 왼쪽 전경의 인물(어깨)을 올려다봄.",
        "built_space": "해안가 언덕, 우측 배경에 트럭 위치.",
        "entities": "앰버의 외모는 일치하나 재킷 누락. 꽃 그림이 앞면으로 잘못 노출됨. 금지된 추가 인물 등장.",
        "hard_violations": [
         "허용되지 않은 인물 등장(명시된 앰버 외의 인물 노출)"
        ],
        "physics": "서서 두 손으로 종이를 지탱하며 들고 있음."
       },
       {
        "label": "B",
        "direction": "프레임 밖 좌측 상단을 애절하게 올려다봄.",
        "built_space": "해안가 언덕, 능선과 바다가 보이는 야외.",
        "entities": "앰버의 외모와 복장(재킷 포함) 일치. 눈물 연기 구현됨. 머리 상처는 미구현.",
        "hard_violations": [],
        "physics": "자연스럽게 서서 지면에 안정적으로 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "허용되지 않은 인물의 신체가 등장하여 치명적 위반이며, 앰버의 재킷 누락 및 종이 방향 오류가 있음."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레임 밖을 보는 시선으로 제약을 잘 지켰고, 앰버의 복장과 애절한 눈물 연기를 정확히 재현함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "화면 왼쪽 전경의 인물(어깨)을 올려다봄.",
        "built_space": "해안가 언덕, 우측 배경에 트럭 위치.",
        "entities": "앰버의 외모는 일치하나 재킷 누락. 꽃 그림이 앞면으로 잘못 노출됨. 금지된 추가 인물 등장.",
        "hard_violations": [
         "허용되지 않은 인물 등장(명시된 앰버 외의 인물 노출)"
        ],
        "physics": "서서 두 손으로 종이를 지탱하며 들고 있음."
       },
       {
        "label": "B",
        "direction": "프레임 밖 좌측 상단을 애절하게 올려다봄.",
        "built_space": "해안가 언덕, 능선과 바다가 보이는 야외.",
        "entities": "앰버의 외모와 복장(재킷 포함) 일치. 눈물 연기 구현됨. 머리 상처는 미구현.",
        "hard_violations": [],
        "physics": "자연스럽게 서서 지면에 안정적으로 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "허용된 앰버만 등장하고 눈물과 슬픈 상향 시선, 의상을 비교적 충실히 재현하지만, 얼굴 클로즈업보다 넓고 기존 머리 부상이 드러나지 않는다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "올려다보는 시선과 애절한 표정은 명확하지만, 허용되지 않은 성인 신체를 전경에 추가해 어깨너머 구도로 바꾸었고 앰버의 재킷도 누락했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 눈은 화면 왼쪽 위의 프레임 밖 상대를 향한다. 찰리는 보이지 않지만 그를 올려다보는 관계로 해석 가능한 방향이며, 렌즈를 직접 응시하지 않는다. 다만 턱과 눈의 상승 각도는 크지 않다.",
        "built_space": "흙과 돌, 낮은 식생이 있는 언덕 뒤로 해안 능선과 바다가 보인다. 건축물이나 고정 설비는 없으며 트럭은 프레임에 없다. 장소의 지형과 재질은 참조에 가깝지만, 상체까지 담은 넓은 구도와 상당한 해안 풍경 때문에 흙바닥을 억제된 배경으로 삼는 얼굴 클로즈업에는 미흡하다.",
        "entities": "금발과 큰 눈을 가진 약 10세 여자아이 한 명이 보인다. 얼굴은 참조와 대체로 유사하지만 한국계·백인 혼혈 정체성을 외모만으로 확정할 수는 없다. 회녹색 재킷과 때 묻은 회색 안옷은 일치하고 참조의 마스크는 없다. 눈가의 눈물과 슬픈 표정은 명확하나 기존 머리 부상은 확인되지 않는다. 손과 그림은 화면 밖이므로 소지 여부는 평가할 수 없다. 배경에 보라색 꽃처럼 보이는 식생이 있어 꽃이 보이지 않아야 한다는 조건에 불리하며, 정확한 종은 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 목, 어깨와 몸통은 자연스럽게 연결되어 있고 서 있는 상체로 읽힌다. 하체의 접지는 크롭 밖이므로 확인할 수 없지만 공중에 떠 있다는 징후는 없다. 눈물은 눈 아래에서 볼을 따라 내려가며 피부에 붙어 있어 물리적으로 타당하다."
       },
       {
        "label": "B",
        "direction": "앰버는 턱을 들고 화면 왼쪽 위 성인의 얼굴 쪽을 뚜렷하게 올려다본다. 시선의 목표가 전경 인물로 명확해 찰리를 올려다보는 동작 자체는 잘 드러난다. 성인의 눈은 잘려 있어 그의 시선은 확인할 수 없다.",
        "built_space": "흙과 돌이 있는 언덕, 낮은 관목, 해안 능선과 바다가 보이고 오른쪽에는 주차된 트럭 한 대의 일부가 있다. 보이는 바퀴와 사이드미러는 각각 하나이며 중복된 설비는 없다. 장소는 참조와 가깝지만, 성인의 어깨를 크게 넣은 어깨너머 구도와 선명한 해안·트럭 배경이 요구된 앰버 중심의 차분한 얼굴 클로즈업을 바꾼다.",
        "entities": "금발과 큰 눈을 가진 약 10세 여자아이의 눈가와 볼에 눈물이 보이며 표정은 애절하다. 참조와 얼굴의 세부 인상은 다르고, 혼혈 정체성을 외모만으로 확정할 수는 없다. 회색 옷은 있으나 회녹색 재킷과 마스크는 없다. 기존 머리 부상도 명확하지 않다. 양손으로 종이 한 장을 잡고 있으며 카메라 쪽 면에 꽃 선화가 보인다. 왼쪽에는 앰버가 아닌 성인의 목·턱·어깨가 크게 등장한다. 배경에는 보라색 꽃무리가 보이나 종은 확정할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "등장 가능한 인물을 앰버로 한정한 조건과 달리, 성인의 턱·목·어깨를 전경에 추가했다."
        ],
        "physics": "앰버의 두 손이 종이 양쪽을 실제로 잡고 있어 종이는 지지된다. 목을 젖혀 가까운 성인을 바라보는 자세도 가능하다. 성인의 목과 어깨는 이어져 있으며, 두 인물 모두 하체가 프레임 밖일 뿐 공중 부양의 징후는 없다. 눈물은 볼의 표면을 따라 아래로 흐른다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "허용된 앰버만 등장하고 눈물과 슬픈 상향 시선, 의상을 비교적 충실히 재현하지만, 얼굴 클로즈업보다 넓고 기존 머리 부상이 드러나지 않는다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "올려다보는 시선과 애절한 표정은 명확하지만, 허용되지 않은 성인 신체를 전경에 추가해 어깨너머 구도로 바꾸었고 앰버의 재킷도 누락했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 눈은 화면 왼쪽 위의 프레임 밖 상대를 향한다. 찰리는 보이지 않지만 그를 올려다보는 관계로 해석 가능한 방향이며, 렌즈를 직접 응시하지 않는다. 다만 턱과 눈의 상승 각도는 크지 않다.",
        "built_space": "흙과 돌, 낮은 식생이 있는 언덕 뒤로 해안 능선과 바다가 보인다. 건축물이나 고정 설비는 없으며 트럭은 프레임에 없다. 장소의 지형과 재질은 참조에 가깝지만, 상체까지 담은 넓은 구도와 상당한 해안 풍경 때문에 흙바닥을 억제된 배경으로 삼는 얼굴 클로즈업에는 미흡하다.",
        "entities": "금발과 큰 눈을 가진 약 10세 여자아이 한 명이 보인다. 얼굴은 참조와 대체로 유사하지만 한국계·백인 혼혈 정체성을 외모만으로 확정할 수는 없다. 회녹색 재킷과 때 묻은 회색 안옷은 일치하고 참조의 마스크는 없다. 눈가의 눈물과 슬픈 표정은 명확하나 기존 머리 부상은 확인되지 않는다. 손과 그림은 화면 밖이므로 소지 여부는 평가할 수 없다. 배경에 보라색 꽃처럼 보이는 식생이 있어 꽃이 보이지 않아야 한다는 조건에 불리하며, 정확한 종은 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 목, 어깨와 몸통은 자연스럽게 연결되어 있고 서 있는 상체로 읽힌다. 하체의 접지는 크롭 밖이므로 확인할 수 없지만 공중에 떠 있다는 징후는 없다. 눈물은 눈 아래에서 볼을 따라 내려가며 피부에 붙어 있어 물리적으로 타당하다."
       },
       {
        "label": "A",
        "direction": "앰버는 턱을 들고 화면 왼쪽 위 성인의 얼굴 쪽을 뚜렷하게 올려다본다. 시선의 목표가 전경 인물로 명확해 찰리를 올려다보는 동작 자체는 잘 드러난다. 성인의 눈은 잘려 있어 그의 시선은 확인할 수 없다.",
        "built_space": "흙과 돌이 있는 언덕, 낮은 관목, 해안 능선과 바다가 보이고 오른쪽에는 주차된 트럭 한 대의 일부가 있다. 보이는 바퀴와 사이드미러는 각각 하나이며 중복된 설비는 없다. 장소는 참조와 가깝지만, 성인의 어깨를 크게 넣은 어깨너머 구도와 선명한 해안·트럭 배경이 요구된 앰버 중심의 차분한 얼굴 클로즈업을 바꾼다.",
        "entities": "금발과 큰 눈을 가진 약 10세 여자아이의 눈가와 볼에 눈물이 보이며 표정은 애절하다. 참조와 얼굴의 세부 인상은 다르고, 혼혈 정체성을 외모만으로 확정할 수는 없다. 회색 옷은 있으나 회녹색 재킷과 마스크는 없다. 기존 머리 부상도 명확하지 않다. 양손으로 종이 한 장을 잡고 있으며 카메라 쪽 면에 꽃 선화가 보인다. 왼쪽에는 앰버가 아닌 성인의 목·턱·어깨가 크게 등장한다. 배경에는 보라색 꽃무리가 보이나 종은 확정할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "등장 가능한 인물을 앰버로 한정한 조건과 달리, 성인의 턱·목·어깨를 전경에 추가했다."
        ],
        "physics": "앰버의 두 손이 종이 양쪽을 실제로 잡고 있어 종이는 지지된다. 목을 젖혀 가까운 성인을 바라보는 자세도 가능하다. 성인의 목과 어깨는 이어져 있으며, 두 인물 모두 하체가 프레임 밖일 뿐 공중 부양의 징후는 없다. 눈물은 볼의 표면을 따라 아래로 흐른다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.929,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.679,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 허용되지 않은 인물 등장(명시된 앰버 외의 인물 노출)",
     "[gpt-high] 등장 가능한 인물을 앰버로 한정한 조건과 달리, 성인의 턱·목·어깨를 전경에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 679,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 679,
    "verdict_ko": "허용되지 않은 인물의 신체가 등장하여 치명적 위반이며, 앰버의 재킷 누락 및 종이 방향 오류가 있음.  ★위반: [gemini-pro] 허용되지 않은 인물 등장(명시된 앰버 외의 인물 노출) / [gpt-high] 등장 가능한 인물을 앰버로 한정한 조건과 달리, 성인의 턱·목·어깨를 전경에 추가했다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프레임 밖을 보는 시선으로 제약을 잘 지켰고, 앰버의 복장과 애절한 눈물 연기를 정확히 재현함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_coastal_search_hill_7092b9.png",
    "asset_id": "3b958113-b2b7-416f-84ac-bf77de635bd9",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace48-7cde-7636-a001-203410eebd47",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S71sh13__bgfirst_bg.png",
   "bg_asset_id": "72945e5f-679f-4ba0-aeb7-142d9df70439",
   "bg_record_key": "S71sh13::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "coastal_search_hill",
   "groupbg_asset_id": "3b958113-b2b7-416f-84ac-bf77de635bd9"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S71sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:16:33.473212+00:00",
  "fingerprint": "73a18f0bd0a142b6d364423957ad268a716015a5692030a23f27a674e70d1421",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S71sh13_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S71sh13_sel.png",
  "source_sha256": "c24719c31d9c38d85b81d28f5230667110883ef3299739630f0668daf1eb0c27",
  "file": "S71sh13_cine.png",
  "staged_sha256": "2bccf4ba2186b4353a89bf5275c27f67357c0968c12d7162231e28e935896587",
  "latency_ms": 12440
 },
 "S71sh25::signage": {
  "fp": "011aafd33c7247f1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S71sh25::bgfirst_bg": {
  "input_fingerprint": "4ac228702c35540a",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 차에서 내린 이현우, 앰버, 라울, 찰리가 제비꽃밭을 내려다보며 입을 반쯤 벌린 경이로운 구도.\n\nLOCATION (lock): At a roadside stopping point above a broad violet field, with a research building visible in the distance.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Violet field (An immense abundance of purple violets is in bloom); used as Occupies the ground beneath their collective gaze and opens depth toward the laboratory; 지동현's laboratory (Visible rising in the distance beyond the flowers) — Seen obliquely in the upper-right background; used as Connects the discovery of the flowers with the destination they have sought.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight preserves the purple flowers as the principal vivid color while keeping the companions' astonished faces naturally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 차에서 내린 이현우, 앰버, 라울, 찰리가 제비꽃밭을 내려다보며 입을 반쯤 벌린 경이로운 구도.\n\nLOCATION (lock): At a roadside stopping point above a broad violet field, with a research building visible in the distance.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Violet field (An immense abundance of purple violets is in bloom); used as Occupies the ground beneath their collective gaze and opens depth toward the laboratory; 지동현's laboratory (Visible rising in the distance beyond the flowers) — Seen obliquely in the upper-right background; used as Connects the discovery of the flowers with the destination they have sought.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight preserves the purple flowers as the principal vivid color while keeping the companions' astonished faces naturally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S71sh25__bgfirst_bg.png",
  "asset_id": "75771a95-b6e7-4b86-a207-a70d7c058772",
  "input_asset_ids": [
   "741e4934-89bb-4fb9-b217-afac2d51a808",
   "028b3ce5-f561-4770-8eaa-0b953b434726"
  ]
 },
 "S71sh25": {
  "input_fingerprint": "ee325068ca59f970",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 차에서 내린 이현우, 앰버, 라울, 찰리가 제비꽃밭을 내려다보며 입을 반쯤 벌린 경이로운 구도.\n\nLOCATION (lock): At a roadside stopping point above a broad violet field, with a research building visible in the distance. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Violet field (An immense abundance of purple violets is in bloom); used as Occupies the ground beneath their collective gaze and opens depth toward the laboratory; 지동현's laboratory (Visible rising in the distance beyond the flowers) — Seen obliquely in the upper-right background; used as Connects the discovery of the flowers with the destination they have sought.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight preserves the purple flowers as the principal vivid color while keeping the companions' astonished faces naturally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An extensive field of violets fills the daylight landscape, with the research laboratory visible in the distance. The old truck has stopped beside the viewing point. 이현우: He has stepped out of the truck and is looking over the flower field, still fatigued and carrying his accumulated injuries. 앰버: She has stepped out to look over the flower field and retains the flower drawing. Her recent head injury remains. 라울: He has stepped out to look over the flower field. 찰리: He has stepped out to look over the flower field. His unrepaired battle damage, exposed chest opening and salvaged chest component remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 차에서 내린 이현우, 앰버, 라울, 찰리가 제비꽃밭을 내려다보며 입을 반쯤 벌린 경이로운 구도.\n\nLOCATION (lock): At a roadside stopping point above a broad violet field, with a research building visible in the distance. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Violet field (An immense abundance of purple violets is in bloom); used as Occupies the ground beneath their collective gaze and opens depth toward the laboratory; 지동현's laboratory (Visible rising in the distance beyond the flowers) — Seen obliquely in the upper-right background; used as Connects the discovery of the flowers with the destination they have sought.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight preserves the purple flowers as the principal vivid color while keeping the companions' astonished faces naturally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An extensive field of violets fills the daylight landscape, with the research laboratory visible in the distance. The old truck has stopped beside the viewing point. 이현우: He has stepped out of the truck and is looking over the flower field, still fatigued and carrying his accumulated injuries. 앰버: She has stepped out to look over the flower field and retains the flower drawing. Her recent head injury remains. 라울: He has stepped out to look over the flower field. 찰리: He has stepped out to look over the flower field. His unrepaired battle damage, exposed chest opening and salvaged chest component remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 차에서 내린 이현우, 앰버, 라울, 찰리가 제비꽃밭을 내려다보며 입을 반쯤 벌린 경이로운 구도.\n\nLOCATION (lock): At a roadside stopping point above a broad violet field, with a research building visible in the distance. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Violet field (An immense abundance of purple violets is in bloom); used as Occupies the ground beneath their collective gaze and opens depth toward the laboratory; 지동현's laboratory (Visible rising in the distance beyond the flowers) — Seen obliquely in the upper-right background; used as Connects the discovery of the flowers with the destination they have sought.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight preserves the purple flowers as the principal vivid color while keeping the companions' astonished faces naturally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An extensive field of violets fills the daylight landscape, with the research laboratory visible in the distance. The old truck has stopped beside the viewing point. 이현우: He has stepped out of the truck and is looking over the flower field, still fatigued and carrying his accumulated injuries. 앰버: She has stepped out to look over the flower field and retains the flower drawing. Her recent head injury remains. 라울: He has stepped out to look over the flower field. 찰리: He has stepped out to look over the flower field. His unrepaired battle damage, exposed chest opening and salvaged chest component remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S71sh25__bgfirst_bg.png",
     "asset_id": "75771a95-b6e7-4b86-a207-a70d7c058772",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S71sh25.png",
     "asset_id": "741e4934-89bb-4fb9-b217-afac2d51a808",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:885644>",
     "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_coastal_cliff_edge_sel.png",
     "asset_id": "028b3ce5-f561-4770-8eaa-0b953b434726",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:885644>",
     "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "네 명의 인물 모두 입을 반쯤 벌린 채 시선을 아래로 향해 제비꽃밭을 정확히 바라보고 있음.",
    "built_space": "해안가 절벽 위의 제비꽃밭을 전경으로, 좌측에 정차된 차량과 우측 원경에 연구소 건물이 프롬프트 지시대로 배치됨.",
    "entities": "이현우(피로와 상처), 앰버(이마 밴드, 꽃 그림 소지), 라울(꽁지머리), 찰리(흰 얼굴판의 고릴라 기계 머리, 가슴 부품 노출) 모두 참조와 일치함.",
    "hard_violations": [],
    "physics": "모든 인물이 지면을 단단히 딛고 몸을 낮추거나 구부려 경이로움을 느끼는 자연스러운 자세를 유지함."
   },
   {
    "label": "B",
    "direction": "인물들이 입을 벌리고 앞쪽 제비꽃밭을 향해 시선을 두고 있으나, 앰버가 든 종이가 카메라 렌즈를 향해 정면으로 부자연스럽게 펼쳐져 있음.",
    "built_space": "전경에 제비꽃밭, 우측에 정차된 차량, 중앙 원경에 연구소 건물이 배치되어 지형 조건은 대체로 충족함.",
    "entities": "이현우, 앰버, 라울은 프롬프트에 부합하나, 찰리가 기계가 아닌 실제 털을 가진 생물학적 고릴라로 완전히 잘못 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 발명된 대상: 찰리가 지시된 기계 얼굴판이 아닌 생물학적 털을 가진 실제 고릴라로 생성됨."
    ],
    "physics": "네 인물이 모두 지면에 똑바로 서서 몸을 지탱하고 있어 물리적인 오류는 보이지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "새벽의 조명과 와이드 샷 구도를 훌륭하게 살렸으며, 찰리의 기계 외장을 포함한 모든 캐릭터의 디테일과 시선 방향이 프롬프트와 참조 이미지에 정확히 부합합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리가 기계 머리가 아닌 털이 있는 생물학적 고릴라로 생성되는 심각한 오류가 발생했으며, 앰버가 든 그림이 카메라 정면을 향해 어색하게 노출되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "네 명의 인물 모두 입을 반쯤 벌린 채 시선을 아래로 향해 제비꽃밭을 정확히 바라보고 있음.",
        "built_space": "해안가 절벽 위의 제비꽃밭을 전경으로, 좌측에 정차된 차량과 우측 원경에 연구소 건물이 프롬프트 지시대로 배치됨.",
        "entities": "이현우(피로와 상처), 앰버(이마 밴드, 꽃 그림 소지), 라울(꽁지머리), 찰리(흰 얼굴판의 고릴라 기계 머리, 가슴 부품 노출) 모두 참조와 일치함.",
        "hard_violations": [],
        "physics": "모든 인물이 지면을 단단히 딛고 몸을 낮추거나 구부려 경이로움을 느끼는 자연스러운 자세를 유지함."
       },
       {
        "label": "B",
        "direction": "인물들이 입을 벌리고 앞쪽 제비꽃밭을 향해 시선을 두고 있으나, 앰버가 든 종이가 카메라 렌즈를 향해 정면으로 부자연스럽게 펼쳐져 있음.",
        "built_space": "전경에 제비꽃밭, 우측에 정차된 차량, 중앙 원경에 연구소 건물이 배치되어 지형 조건은 대체로 충족함.",
        "entities": "이현우, 앰버, 라울은 프롬프트에 부합하나, 찰리가 기계가 아닌 실제 털을 가진 생물학적 고릴라로 완전히 잘못 묘사됨.",
        "hard_violations": [
         "발명된 대상: 찰리가 지시된 기계 얼굴판이 아닌 생물학적 털을 가진 실제 고릴라로 생성됨."
        ],
        "physics": "네 인물이 모두 지면에 똑바로 서서 몸을 지탱하고 있어 물리적인 오류는 보이지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "새벽의 조명과 와이드 샷 구도를 훌륭하게 살렸으며, 찰리의 기계 외장을 포함한 모든 캐릭터의 디테일과 시선 방향이 프롬프트와 참조 이미지에 정확히 부합합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "찰리가 기계 머리가 아닌 털이 있는 생물학적 고릴라로 생성되는 심각한 오류가 발생했으며, 앰버가 든 그림이 카메라 정면을 향해 어색하게 노출되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "네 명의 인물 모두 입을 반쯤 벌린 채 시선을 아래로 향해 제비꽃밭을 정확히 바라보고 있음.",
        "built_space": "해안가 절벽 위의 제비꽃밭을 전경으로, 좌측에 정차된 차량과 우측 원경에 연구소 건물이 프롬프트 지시대로 배치됨.",
        "entities": "이현우(피로와 상처), 앰버(이마 밴드, 꽃 그림 소지), 라울(꽁지머리), 찰리(흰 얼굴판의 고릴라 기계 머리, 가슴 부품 노출) 모두 참조와 일치함.",
        "hard_violations": [],
        "physics": "모든 인물이 지면을 단단히 딛고 몸을 낮추거나 구부려 경이로움을 느끼는 자연스러운 자세를 유지함."
       },
       {
        "label": "B",
        "direction": "인물들이 입을 벌리고 앞쪽 제비꽃밭을 향해 시선을 두고 있으나, 앰버가 든 종이가 카메라 렌즈를 향해 정면으로 부자연스럽게 펼쳐져 있음.",
        "built_space": "전경에 제비꽃밭, 우측에 정차된 차량, 중앙 원경에 연구소 건물이 배치되어 지형 조건은 대체로 충족함.",
        "entities": "이현우, 앰버, 라울은 프롬프트에 부합하나, 찰리가 기계가 아닌 실제 털을 가진 생물학적 고릴라로 완전히 잘못 묘사됨.",
        "hard_violations": [
         "발명된 대상: 찰리가 지시된 기계 얼굴판이 아닌 생물학적 털을 가진 실제 고릴라로 생성됨."
        ],
        "physics": "네 인물이 모두 지면에 똑바로 서서 몸을 지탱하고 있어 물리적인 오류는 보이지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "꽃밭 앞의 놀란 일행은 보이지만, 앰버의 정면 시선과 좁은 꽃밭 구도, 생물 고릴라로 바뀐 찰리가 핵심 지시와 어긋납니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "새벽 해안의 넓은 제비꽃밭과 우측 원경 연구소, 꽃을 살피는 일행 및 손상된 기계 찰리를 더 충실히 구현하지만, 이현우의 시선은 아래보다 앞을 향합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 라울은 고개와 눈을 전경 꽃 쪽으로 내리고 있으며 입이 벌어져 있다. 앰버는 꽃밭 아래가 아니라 카메라 쪽을 거의 정면으로 바라본다. 찰리는 얼굴을 약간 숙여 앞쪽 꽃밭을 보는 듯하나 입은 닫혀 있다. 앰버의 그림은 카메라를 향하지만 읽는 동작은 아니므로 사용 방향의 모순은 아니다.",
        "built_space": "오른쪽 뒤에 낡은 트럭 한 대, 상단 중앙에 낮은 수평동과 세로 탑이 결합된 연구소 한 곳이 있다. 소나무와 바위 절벽은 참조 장소에 부합하지만 바다는 거의 가려진다. 네 인물은 트럭 앞 흙길에 나란히 서 있고 꽃은 전경의 비교적 좁은 띠를 채운다. 연구소가 요구된 우측 상단보다 중앙에 가깝고, 꽃밭에서 연구소로 이어지는 깊이도 약하다. 중복 건물이나 불가능한 반사는 보이지 않는다.",
        "entities": "추가 인물 없이 네 등장 개체가 있다. 이현우는 젊은 한국계 남성 외모, 헝클어진 검은 머리, 낡은 검정 후드와 청바지 및 얼굴 상처가 보인다. 앰버는 금발의 어린 여자아이이며 녹색 겉옷과 회색 옷, 꽃 그림을 유지하지만 혼혈 정체성은 외모만으로 확정하기 어렵고 머리 부상은 드러나지 않는다. 라울은 짙은 피부의 어린아이로 녹색 셔츠와 청반바지를 입었으나 뒤로 묶은 꽁지머리는 뚜렷하지 않다. 찰리는 흰 기계 얼굴판을 가진 로봇이 아니라 털과 생물학적 얼굴을 가진 고릴라처럼 보이며, 참조의 모자와 다색 외투도 없다. 가슴 상처는 있지만 요구된 노출 기계 흉부와 회수 부품으로 읽히지 않는다. 보라색 꽃, 트럭, 원경 연구소와 꽃 그림은 존재하고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물들의 발은 전경 꽃에 가려지지만 몸통과 다리는 뒤쪽 흙길 위에 서 있는 자세로 연결된다. 공중에 떠 있는 몸은 없다. 앰버는 종이 양옆을 두 손으로 잡고 있다. 트럭은 지면에 놓여 있고 꽃은 땅에서 자란다. 자세는 다소 일렬로 경직되어 있으나 물리적으로 불가능하지는 않다."
       },
       {
        "label": "B",
        "direction": "라울은 몸을 굽혀 발밑 꽃을 직접 내려다본다. 찰리도 숙인 얼굴을 왼쪽 아래 꽃밭으로 향한다. 앰버는 오른쪽 앞 꽃밭 쪽을 보며 입을 벌리고 있다. 이현우는 입을 벌렸지만 시선이 바로 아래 꽃보다 오른쪽 앞의 먼 지점으로 향해 네 명의 집단적인 내려다보기가 완전히 일치하지는 않는다. 그림은 손에 들린 채 비스듬히 바깥을 향하며 누구도 그것을 읽고 있지 않다.",
        "built_space": "왼쪽 도로에 정차 차량 한 대와 그 뒤 목재 난간 일부가 있고, 네 인물은 도로 아래쪽 꽃이 핀 경사면에 모여 있다. 넓은 전경 꽃밭 너머로 소나무 해안, 바위 절벽, 오른쪽 바다가 이어진다. 연구소 한 곳은 우측 상단 원경에 작게 자리하며 수평동, 세로 탑, 안테나와 흰 돔이 참조와 대체로 일치한다. 목재 난간은 참조 사진에서 확인되지 않는 배경 차이다. 와이드 구도와 꽃밭에서 연구소로 이어지는 공간 깊이는 A보다 명확하며 불가능한 반사는 없다.",
        "entities": "이현우, 앰버, 라울, 찰리만 등장한다. 이현우는 젊은 한국계 남성 외모와 검은 헝클어진 머리, 낡은 후드·청바지 및 얼굴 상처를 유지한다. 앰버는 금발의 어린 여자아이로 녹색 겉옷과 회색 옷을 입고, 이마의 붕대와 꽃 그림을 보유한다. 라울은 짙은 피부의 어린아이로 머리를 뒤로 묶고 녹색 셔츠와 청반바지를 입었다. 두 아이의 세부 혼혈 계통은 영상만으로 확정할 수 없다. 찰리는 각진 기계 머리와 흰 얼굴판, 점 눈과 선 입, 밀짚모자와 다색 외투, 녹색 바지와 노란 장화를 유지한다. 열린 가슴에는 기계 부품이 보이나 그것이 회수 부품인지는 구분되지 않는다. 차량은 앞부분만 보여 낡은 트럭인지 확정하기 어렵다. 풍성한 보라색 꽃과 원경 연구소가 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우는 한쪽 무릎을 땅에 대고 반대쪽 발로 몸을 지탱하며 손을 다리에 얹는다. 앰버는 발을 땅에 댄 쪼그린 자세로 종이를 두 손에 잡고 있다. 라울은 양발로 서서 양손을 무릎에 짚고, 찰리도 장화를 지면에 두고 무릎을 굽힌 채 허벅지에 손을 댄다. 차량 바퀴는 도로에 닿아 있다. 꽃을 가까이 살피는 자세 모두에 지지점이 있으며 떠 있는 몸이나 소품은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "꽃밭 앞의 놀란 일행은 보이지만, 앰버의 정면 시선과 좁은 꽃밭 구도, 생물 고릴라로 바뀐 찰리가 핵심 지시와 어긋납니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "새벽 해안의 넓은 제비꽃밭과 우측 원경 연구소, 꽃을 살피는 일행 및 손상된 기계 찰리를 더 충실히 구현하지만, 이현우의 시선은 아래보다 앞을 향합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우와 라울은 고개와 눈을 전경 꽃 쪽으로 내리고 있으며 입이 벌어져 있다. 앰버는 꽃밭 아래가 아니라 카메라 쪽을 거의 정면으로 바라본다. 찰리는 얼굴을 약간 숙여 앞쪽 꽃밭을 보는 듯하나 입은 닫혀 있다. 앰버의 그림은 카메라를 향하지만 읽는 동작은 아니므로 사용 방향의 모순은 아니다.",
        "built_space": "오른쪽 뒤에 낡은 트럭 한 대, 상단 중앙에 낮은 수평동과 세로 탑이 결합된 연구소 한 곳이 있다. 소나무와 바위 절벽은 참조 장소에 부합하지만 바다는 거의 가려진다. 네 인물은 트럭 앞 흙길에 나란히 서 있고 꽃은 전경의 비교적 좁은 띠를 채운다. 연구소가 요구된 우측 상단보다 중앙에 가깝고, 꽃밭에서 연구소로 이어지는 깊이도 약하다. 중복 건물이나 불가능한 반사는 보이지 않는다.",
        "entities": "추가 인물 없이 네 등장 개체가 있다. 이현우는 젊은 한국계 남성 외모, 헝클어진 검은 머리, 낡은 검정 후드와 청바지 및 얼굴 상처가 보인다. 앰버는 금발의 어린 여자아이이며 녹색 겉옷과 회색 옷, 꽃 그림을 유지하지만 혼혈 정체성은 외모만으로 확정하기 어렵고 머리 부상은 드러나지 않는다. 라울은 짙은 피부의 어린아이로 녹색 셔츠와 청반바지를 입었으나 뒤로 묶은 꽁지머리는 뚜렷하지 않다. 찰리는 흰 기계 얼굴판을 가진 로봇이 아니라 털과 생물학적 얼굴을 가진 고릴라처럼 보이며, 참조의 모자와 다색 외투도 없다. 가슴 상처는 있지만 요구된 노출 기계 흉부와 회수 부품으로 읽히지 않는다. 보라색 꽃, 트럭, 원경 연구소와 꽃 그림은 존재하고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물들의 발은 전경 꽃에 가려지지만 몸통과 다리는 뒤쪽 흙길 위에 서 있는 자세로 연결된다. 공중에 떠 있는 몸은 없다. 앰버는 종이 양옆을 두 손으로 잡고 있다. 트럭은 지면에 놓여 있고 꽃은 땅에서 자란다. 자세는 다소 일렬로 경직되어 있으나 물리적으로 불가능하지는 않다."
       },
       {
        "label": "A",
        "direction": "라울은 몸을 굽혀 발밑 꽃을 직접 내려다본다. 찰리도 숙인 얼굴을 왼쪽 아래 꽃밭으로 향한다. 앰버는 오른쪽 앞 꽃밭 쪽을 보며 입을 벌리고 있다. 이현우는 입을 벌렸지만 시선이 바로 아래 꽃보다 오른쪽 앞의 먼 지점으로 향해 네 명의 집단적인 내려다보기가 완전히 일치하지는 않는다. 그림은 손에 들린 채 비스듬히 바깥을 향하며 누구도 그것을 읽고 있지 않다.",
        "built_space": "왼쪽 도로에 정차 차량 한 대와 그 뒤 목재 난간 일부가 있고, 네 인물은 도로 아래쪽 꽃이 핀 경사면에 모여 있다. 넓은 전경 꽃밭 너머로 소나무 해안, 바위 절벽, 오른쪽 바다가 이어진다. 연구소 한 곳은 우측 상단 원경에 작게 자리하며 수평동, 세로 탑, 안테나와 흰 돔이 참조와 대체로 일치한다. 목재 난간은 참조 사진에서 확인되지 않는 배경 차이다. 와이드 구도와 꽃밭에서 연구소로 이어지는 공간 깊이는 A보다 명확하며 불가능한 반사는 없다.",
        "entities": "이현우, 앰버, 라울, 찰리만 등장한다. 이현우는 젊은 한국계 남성 외모와 검은 헝클어진 머리, 낡은 후드·청바지 및 얼굴 상처를 유지한다. 앰버는 금발의 어린 여자아이로 녹색 겉옷과 회색 옷을 입고, 이마의 붕대와 꽃 그림을 보유한다. 라울은 짙은 피부의 어린아이로 머리를 뒤로 묶고 녹색 셔츠와 청반바지를 입었다. 두 아이의 세부 혼혈 계통은 영상만으로 확정할 수 없다. 찰리는 각진 기계 머리와 흰 얼굴판, 점 눈과 선 입, 밀짚모자와 다색 외투, 녹색 바지와 노란 장화를 유지한다. 열린 가슴에는 기계 부품이 보이나 그것이 회수 부품인지는 구분되지 않는다. 차량은 앞부분만 보여 낡은 트럭인지 확정하기 어렵다. 풍성한 보라색 꽃과 원경 연구소가 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우는 한쪽 무릎을 땅에 대고 반대쪽 발로 몸을 지탱하며 손을 다리에 얹는다. 앰버는 발을 땅에 댄 쪼그린 자세로 종이를 두 손에 잡고 있다. 라울은 양발로 서서 양손을 무릎에 짚고, 찰리도 장화를 지면에 두고 무릎을 굽힌 채 허벅지에 손을 댄다. 차량 바퀴는 도로에 닿아 있다. 꽃을 가까이 살피는 자세 모두에 지지점이 있으며 떠 있는 몸이나 소품은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.875
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.625
   },
   "violations": {
    "B": [
     "[gemini-pro] 발명된 대상: 찰리가 지시된 기계 얼굴판이 아닌 생물학적 털을 가진 실제 고릴라로 생성됨."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 625
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "새벽의 조명과 와이드 샷 구도를 훌륭하게 살렸으며, 찰리의 기계 외장을 포함한 모든 캐릭터의 디테일과 시선 방향이 프롬프트와 참조 이미지에 정확히 부합합니다."
   },
   {
    "label": "B",
    "score": 625,
    "verdict_ko": "찰리가 기계 머리가 아닌 털이 있는 생물학적 고릴라로 생성되는 심각한 오류가 발생했으며, 앰버가 든 그림이 카메라 정면을 향해 어색하게 노출되었습니다.  ★위반: [gemini-pro] 발명된 대상: 찰리가 지시된 기계 얼굴판이 아닌 생물학적 털을 가진 실제 고릴라로 생성됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_coastal_cliff_edge_sel.png",
    "asset_id": "028b3ce5-f561-4770-8eaa-0b953b434726",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:943979>",
    "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:885644>",
    "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace55-e441-78d6-b58a-248c56b87791",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S71sh25__bgfirst_bg.png",
   "bg_asset_id": "75771a95-b6e7-4b86-a207-a70d7c058772",
   "bg_record_key": "S71sh25::bgfirst_bg",
   "chain_winner": true,
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S71sh25::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:18:40.615162+00:00",
  "fingerprint": "b9e873c9f3d2c2ef7ea876b1cf26ff7705c2e2184af61bc513b8c2d8ba0a8e11",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S71sh25_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S71sh25_sel.png",
  "source_sha256": "117290ef8a23418b625423848a23fb871903be44e2689aba1ac902b241fc1210",
  "file": "S71sh25_cine.png",
  "staged_sha256": "7ddfd2e24d1f6810daea41db5205a891f9727d702e321dd2919a73c79df202ce",
  "latency_ms": 13260
 },
 "S72sh38::signage": {
  "fp": "c4101354f930a280",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::b3b950dfa6afdda8": {
  "subjects": [],
  "subject_text": "지동현의 연구소 건물 밖과 트럭 정차 지점\n오래된 연구소 외벽과 철문 앞에 펼쳐진 공터. 일부 지면은 진흙이고 건물 바깥으로 넓은 보라색 꽃밭이 이어진다.",
  "identity": "canonical",
  "scope_id": "L111",
  "scope_role": "location_exterior",
  "scope_sha": "aba54169912e7144"
 },
 "S72sh38::bgfirst_bg": {
  "input_fingerprint": "33836471f9522217",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리가 이현우의 손을 향해 굵은 금속 손바닥을 거칠게 뻗은 채 허공에 떠 있는 역동적인 찰나.\n\nLOCATION (lock): Beside the open door of a rescue truck at the cliff-edge boundary of a violet field near the abandoned research building.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open truck doorway sill viewed from inside in the lower-center of the frame, foreground; Cliff edge outside behind the boarding figure in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open truck doorway (Open for 찰리 to board) — Viewed from inside, with the sill low in frame and one jamb along the side; used as Defines the crossing threshold without blocking the reaching hands; Cliff edge (Visible behind the airborne 찰리) — Seen obliquely beyond the doorway; used as Establishes the dangerous exterior from which he is boarding.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established sunset gives the boarding action clear directional contrast without adding the rain or flashlight illumination that belongs to the later nighttime passage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리가 이현우의 손을 향해 굵은 금속 손바닥을 거칠게 뻗은 채 허공에 떠 있는 역동적인 찰나.\n\nLOCATION (lock): Beside the open door of a rescue truck at the cliff-edge boundary of a violet field near the abandoned research building.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open truck doorway sill viewed from inside in the lower-center of the frame, foreground; Cliff edge outside behind the boarding figure in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open truck doorway (Open for 찰리 to board) — Viewed from inside, with the sill low in frame and one jamb along the side; used as Defines the crossing threshold without blocking the reaching hands; Cliff edge (Visible behind the airborne 찰리) — Seen obliquely beyond the doorway; used as Establishes the dangerous exterior from which he is boarding.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established sunset gives the boarding action clear directional contrast without adding the rain or flashlight illumination that belongs to the later nighttime passage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S72sh38__bgfirst_bg.png",
  "asset_id": "9e3c08c6-c287-46a0-9d3b-3198ee8e073d",
  "input_asset_ids": [
   "dbf5c1c0-8b76-4383-a15c-2e9faa7c9029",
   "f09ab42d-aa2d-42e7-a39b-f9e2bb5fd406",
   "c9eeddba-70be-466d-b8a1-f5b11b99ba19"
  ]
 },
 "S72sh38": {
  "input_fingerprint": "3d49de614c4d0fcc",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리가 이현우의 손을 향해 굵은 금속 손바닥을 거칠게 뻗은 채 허공에 떠 있는 역동적인 찰나.\n\nLOCATION (lock): Beside the open door of a rescue truck at the cliff-edge boundary of a violet field near the abandoned research building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open truck doorway sill viewed from inside in the lower-center of the frame, foreground; Cliff edge outside behind the boarding figure in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open truck doorway (Open for 찰리 to board) — Viewed from inside, with the sill low in frame and one jamb along the side; used as Defines the crossing threshold without blocking the reaching hands; Cliff edge (Visible behind the airborne 찰리) — Seen obliquely beyond the doorway; used as Establishes the dangerous exterior from which he is boarding.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established sunset gives the boarding action clear directional contrast without adding the rain or flashlight illumination that belongs to the later nighttime passage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has reached the cliff-edge violet field, and its access door is open. The pursuit takes place in deep sunset light. 이현우: He is at the truck's driving position with the door open and one hand extended. His earlier injuries remain. 찰리: He is boarding the truck from the cliff edge. His chest already has an opening exposing its interior, alongside the earlier dents and holes, and he retains the salvaged chest component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리가 이현우의 손을 향해 굵은 금속 손바닥을 거칠게 뻗은 채 허공에 떠 있는 역동적인 찰나.\n\nLOCATION (lock): Beside the open door of a rescue truck at the cliff-edge boundary of a violet field near the abandoned research building. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open truck doorway sill viewed from inside in the lower-center of the frame, foreground; Cliff edge outside behind the boarding figure in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open truck doorway (Open for 찰리 to board) — Viewed from inside, with the sill low in frame and one jamb along the side; used as Defines the crossing threshold without blocking the reaching hands; Cliff edge (Visible behind the airborne 찰리) — Seen obliquely beyond the doorway; used as Establishes the dangerous exterior from which he is boarding.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established sunset gives the boarding action clear directional contrast without adding the rain or flashlight illumination that belongs to the later nighttime passage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has reached the cliff-edge violet field, and its access door is open. The pursuit takes place in deep sunset light. 이현우: He is at the truck's driving position with the door open and one hand extended. His earlier injuries remain. 찰리: He is boarding the truck from the cliff edge. His chest already has an opening exposing its interior, alongside the earlier dents and holes, and he retains the salvaged chest component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리가 이현우의 손을 향해 굵은 금속 손바닥을 거칠게 뻗은 채 허공에 떠 있는 역동적인 찰나.\n\nLOCATION (lock): Beside the open door of a rescue truck at the cliff-edge boundary of a violet field near the abandoned research building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open truck doorway sill viewed from inside in the lower-center of the frame, foreground; Cliff edge outside behind the boarding figure in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Open truck doorway (Open for 찰리 to board) — Viewed from inside, with the sill low in frame and one jamb along the side; used as Defines the crossing threshold without blocking the reaching hands; Cliff edge (Visible behind the airborne 찰리) — Seen obliquely beyond the doorway; used as Establishes the dangerous exterior from which he is boarding.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established sunset gives the boarding action clear directional contrast without adding the rain or flashlight illumination that belongs to the later nighttime passage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has reached the cliff-edge violet field, and its access door is open. The pursuit takes place in deep sunset light. 이현우: He is at the truck's driving position with the door open and one hand extended. His earlier injuries remain. 찰리: He is boarding the truck from the cliff edge. His chest already has an opening exposing its interior, alongside the earlier dents and holes, and he retains the salvaged chest component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S72sh38__bgfirst_bg.png",
     "asset_id": "9e3c08c6-c287-46a0-9d3b-3198ee8e073d",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S72sh38.png",
     "asset_id": "dbf5c1c0-8b76-4383-a15c-2e9faa7c9029",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:885644>",
     "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L111B01.png",
     "asset_id": "f09ab42d-aa2d-42e7-a39b-f9e2bb5fd406",
     "role": "location_plate"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_coastal_abandoned_laboratory_sel.png",
     "asset_id": "c9eeddba-70be-466d-b8a1-f5b11b99ba19",
     "role": "structure_seed_look"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:885644>",
     "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "찰리와 이현우가 서로의 손을 향해 팔을 뻗고 있음.",
    "built_space": "트럭 내부에서 우측(조수석) 열린 문 밖을 내다보는 구도로, 문턱이 프레임 하단에 위치함.",
    "entities": "찰리는 밀짚모자와 지정된 우비를 입었으나 노출되어야 할 가슴 내부 기계가 셔츠 위 평면 그림으로 묘사됨. 왼손에 부품을 들고 있음. 이현우는 레퍼런스와 일치하는 후드티를 착용함.",
    "hard_violations": [
     "[gemini-pro] 찰리가 지지대나 도약 징후 없이 허공에 수평으로 떠 있음 (nothing supports it)",
     "[gemini-pro] 이현우가 운전석이 아닌 조수석(우측) 문에 배치됨"
    ],
    "physics": "찰리가 수평으로 비행하는 자세로 허공에 떠 있으나, 도약이나 지탱하는 물리적 힘의 원천이 전혀 없음 (nothing supports it)."
   },
   {
    "label": "B",
    "direction": "찰리와 이현우가 서로를 향해 손을 뻗고 있음.",
    "built_space": "트럭 내부에서 좌측(운전석) 열린 문을 바라보는 구도로, 이현우가 운전석 위치에 있음.",
    "entities": "찰리는 가슴 내부 기계가 실제로 노출되어 있으나 밀짚모자가 생략됨. 부품을 들고 있지 않음. 이현우는 후드티 대신 반팔 티셔츠를 착용함.",
    "hard_violations": [
     "[gemini-pro] 찰리가 지지대나 도약 없이 허공에 수직으로 떠 있음 (nothing supports it)"
    ],
    "physics": "찰리가 공중에 수직으로 떠 있으나, 발을 지탱하거나 도약하는 요소가 전혀 없음 (nothing supports it)."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지지대 없이 허공에 떠 있는 물리 법칙 위반과 조수석 배치 오류가 있으나, 지정된 의상과 부품 등 시각적 요소는 더 많이 반영됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지지대 없이 허공에 떠 있는 물리 법칙 위반과 함께 캐릭터의 핵심 의상(밀짚모자, 후드티)과 들고 있어야 할 부품이 완전히 누락됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리와 이현우가 서로의 손을 향해 팔을 뻗고 있음.",
        "built_space": "트럭 내부에서 우측(조수석) 열린 문 밖을 내다보는 구도로, 문턱이 프레임 하단에 위치함.",
        "entities": "찰리는 밀짚모자와 지정된 우비를 입었으나 노출되어야 할 가슴 내부 기계가 셔츠 위 평면 그림으로 묘사됨. 왼손에 부품을 들고 있음. 이현우는 레퍼런스와 일치하는 후드티를 착용함.",
        "hard_violations": [
         "찰리가 지지대나 도약 징후 없이 허공에 수평으로 떠 있음 (nothing supports it)",
         "이현우가 운전석이 아닌 조수석(우측) 문에 배치됨"
        ],
        "physics": "찰리가 수평으로 비행하는 자세로 허공에 떠 있으나, 도약이나 지탱하는 물리적 힘의 원천이 전혀 없음 (nothing supports it)."
       },
       {
        "label": "B",
        "direction": "찰리와 이현우가 서로를 향해 손을 뻗고 있음.",
        "built_space": "트럭 내부에서 좌측(운전석) 열린 문을 바라보는 구도로, 이현우가 운전석 위치에 있음.",
        "entities": "찰리는 가슴 내부 기계가 실제로 노출되어 있으나 밀짚모자가 생략됨. 부품을 들고 있지 않음. 이현우는 후드티 대신 반팔 티셔츠를 착용함.",
        "hard_violations": [
         "찰리가 지지대나 도약 없이 허공에 수직으로 떠 있음 (nothing supports it)"
        ],
        "physics": "찰리가 공중에 수직으로 떠 있으나, 발을 지탱하거나 도약하는 요소가 전혀 없음 (nothing supports it)."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지지대 없이 허공에 떠 있는 물리 법칙 위반과 조수석 배치 오류가 있으나, 지정된 의상과 부품 등 시각적 요소는 더 많이 반영됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지지대 없이 허공에 떠 있는 물리 법칙 위반과 함께 캐릭터의 핵심 의상(밀짚모자, 후드티)과 들고 있어야 할 부품이 완전히 누락됨."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "찰리와 이현우가 서로의 손을 향해 팔을 뻗고 있음.",
        "built_space": "트럭 내부에서 우측(조수석) 열린 문 밖을 내다보는 구도로, 문턱이 프레임 하단에 위치함.",
        "entities": "찰리는 밀짚모자와 지정된 우비를 입었으나 노출되어야 할 가슴 내부 기계가 셔츠 위 평면 그림으로 묘사됨. 왼손에 부품을 들고 있음. 이현우는 레퍼런스와 일치하는 후드티를 착용함.",
        "hard_violations": [
         "찰리가 지지대나 도약 징후 없이 허공에 수평으로 떠 있음 (nothing supports it)",
         "이현우가 운전석이 아닌 조수석(우측) 문에 배치됨"
        ],
        "physics": "찰리가 수평으로 비행하는 자세로 허공에 떠 있으나, 도약이나 지탱하는 물리적 힘의 원천이 전혀 없음 (nothing supports it)."
       },
       {
        "label": "B",
        "direction": "찰리와 이현우가 서로를 향해 손을 뻗고 있음.",
        "built_space": "트럭 내부에서 좌측(운전석) 열린 문을 바라보는 구도로, 이현우가 운전석 위치에 있음.",
        "entities": "찰리는 가슴 내부 기계가 실제로 노출되어 있으나 밀짚모자가 생략됨. 부품을 들고 있지 않음. 이현우는 후드티 대신 반팔 티셔츠를 착용함.",
        "hard_violations": [
         "찰리가 지지대나 도약 없이 허공에 수직으로 떠 있음 (nothing supports it)"
        ],
        "physics": "찰리가 공중에 수직으로 떠 있으나, 발을 지탱하거나 도약하는 요소가 전혀 없음 (nothing supports it)."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "실내에서 본 문턱과 도약은 구현했지만, 금속 손이 이현우의 손보다 위를 향하고 찰리의 배치와 두 인물의 복장이 지정에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "하단 문턱·중앙 오른쪽 절벽의 와이드 구도에서 찰리가 이현우의 손으로 도약하며, 의상·가슴 손상·회수 부품까지 가장 충실하게 유지합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴은 오른쪽 이현우 쪽으로 돌아가 있지만, 펼친 금속 손은 이현우의 손보다 상당히 높은 문기둥 쪽을 향한다. 이현우의 손가락도 금속 손이 아니라 문기둥 부근에 뻗어 있어 두 손이 같은 접촉점을 향하는 관계가 약하다. 찰리의 몸은 외부에서 문턱 안쪽으로 진입하는 방향이다.",
        "built_space": "왼쪽에 열린 문짝 하나와 그 안쪽 손잡이·원형 스피커가 각각 하나 보이고, 아래에는 문턱 하나, 오른쪽에는 좌석 하나와 대시보드·핸들이 보인다. 이현우는 좌석에 앉아 문 쪽으로 몸을 기울인다. 실내에서 밖을 보는 와이드 시점은 맞지만 찰리는 중앙 왼쪽에 있고, 절벽은 그 뒤 중앙에 놓여 지정된 중앙 오른쪽 배치와 다르다. 연구동은 오른쪽 배경의 낡은 2층 콘크리트 건물로, 참조와 측면 및 창 배치가 다르게 보인다. 불가능한 반사나 중복 설비는 보이지 않는다.",
        "entities": "찰리는 흰 얼굴판과 작은 눈을 가진 금속 고릴라형 머리, 색동 외투, 녹색 바지, 노란 장화를 갖췄다. 다만 머리 외장이 참조보다 둥글고 밀짚모자가 없으며, 줄무늬 상의 대신 넓게 드러난 기계 흉부가 보인다. 가슴 내부 노출과 금속 손상은 확인되지만 회수 부품은 확인되지 않는다. 이현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보이고 얼굴 상처와 찢어진 청바지는 유지하지만, 참조의 후드 상의가 검은 반소매로 바뀌었다. 보라색 꽃밭·해안 절벽·낡은 트럭·연구동과 따뜻한 해 질 녘 자연광이 보인다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "찰리는 두 발이 지면에서 떨어지고 무릎을 뒤로 접은 상태다. 바로 아래와 뒤의 흙 가장자리가 도약 지점, 앞의 트럭 문턱이 착지 지점으로 읽히므로 단순한 무근거 부유로 보이지는 않는다. 금속 손은 팔에 연결되어 있고 다른 팔은 문짝 뒤로 가려져 접촉 여부를 확인할 수 없다. 이현우는 엉덩이를 좌석에, 보이는 신발을 차내 바닥에 두어 앞으로 기울인 몸을 지지한다."
       },
       {
        "label": "B",
        "direction": "찰리의 펼친 금속 손과 뻗은 팔은 왼쪽 전경 이현우의 열린 손을 직접 향하며 두 손 사이에 짧은 간격이 남아 있다. 찰리의 얼굴도 그 손과 이현우 쪽을 향한다. 이현우는 찰리 쪽으로 고개를 돌리고 손을 뻗는다. 찰리의 몸은 절벽 쪽 외부에서 카메라가 있는 트럭 내부로 비스듬히 날아드는 방향이다.",
        "built_space": "왼쪽에 열린 문짝 하나와 손잡이·원형 스피커가 각각 하나, 하단 중앙에 문턱 하나, 오른쪽 문기둥에 세로 보조 손잡이 하나가 보인다. 오른쪽 가장자리에는 좌석 일부가 보인다. 실내에서 문밖을 보는 와이드 구도이며 찰리는 중앙 오른쪽, 절벽은 그의 뒤 오른쪽에 놓여 지정된 배치에 부합한다. 이현우는 왼쪽 실내 전경에 있지만 운전석과 핸들은 가려져 정확한 착석 위치는 확인되지 않는다. 왼쪽 배경 연구동은 참조의 낡은 2층 콘크리트 외벽, 평지붕, 옥상 돌출부, 정면 큰 창과 출입구 구성을 대체로 유지한다. 설비 중복이나 불가능한 반사는 없다.",
        "entities": "찰리의 각진 기계 머리, 흰 얼굴판, 점 형태 눈과 선 형태 입, 밀짚모자, 색동 외투, 줄무늬 상의, 녹색 바지와 노란 장화가 참조에 가깝다. 가슴의 찢어진 개구부 안에 기계 구조가 보이고 옷에도 구멍과 마모가 있으며, 반대 손에는 회수한 것으로 읽히는 금속 부품을 쥐고 있다. 이현우는 검은 헝클어진 머리와 낡은 어두운 후드 상의를 유지한다. 얼굴 대부분이 보이지 않아 정확한 얼굴 일치와 기존 얼굴 상처는 판단하기 어렵다. 보라색 꽃밭, 바다를 낀 절벽, 열린 낡은 트럭 문과 연구동이 있고, 낮은 해의 따뜻한 자연광이 장면을 비춘다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "찰리는 몸통을 전방으로 기울이고 한쪽 무릎을 뒤로 접으며 다른 다리를 아래뒤로 뻗은 도약 자세다. 뒤쪽 꽃밭의 절벽 가장자리가 발구름 지점으로, 앞쪽 열린 문턱과 차내 바닥이 진입·착지 공간으로 읽힌다. 두 발이 떠 있어도 추진 방향과 착지 경로가 이어져 정지한 공중 부유가 아니다. 회수 부품은 반대쪽 금속 손가락으로 잡고 있어 지지가 명확하다. 이현우의 하체는 화면 밖이므로 좌석 접촉은 확인할 수 없지만, 보이는 상체와 팔에는 불가능한 관절이나 독립적으로 떠 있는 부분이 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "실내에서 본 문턱과 도약은 구현했지만, 금속 손이 이현우의 손보다 위를 향하고 찰리의 배치와 두 인물의 복장이 지정에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "하단 문턱·중앙 오른쪽 절벽의 와이드 구도에서 찰리가 이현우의 손으로 도약하며, 의상·가슴 손상·회수 부품까지 가장 충실하게 유지합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴은 오른쪽 이현우 쪽으로 돌아가 있지만, 펼친 금속 손은 이현우의 손보다 상당히 높은 문기둥 쪽을 향한다. 이현우의 손가락도 금속 손이 아니라 문기둥 부근에 뻗어 있어 두 손이 같은 접촉점을 향하는 관계가 약하다. 찰리의 몸은 외부에서 문턱 안쪽으로 진입하는 방향이다.",
        "built_space": "왼쪽에 열린 문짝 하나와 그 안쪽 손잡이·원형 스피커가 각각 하나 보이고, 아래에는 문턱 하나, 오른쪽에는 좌석 하나와 대시보드·핸들이 보인다. 이현우는 좌석에 앉아 문 쪽으로 몸을 기울인다. 실내에서 밖을 보는 와이드 시점은 맞지만 찰리는 중앙 왼쪽에 있고, 절벽은 그 뒤 중앙에 놓여 지정된 중앙 오른쪽 배치와 다르다. 연구동은 오른쪽 배경의 낡은 2층 콘크리트 건물로, 참조와 측면 및 창 배치가 다르게 보인다. 불가능한 반사나 중복 설비는 보이지 않는다.",
        "entities": "찰리는 흰 얼굴판과 작은 눈을 가진 금속 고릴라형 머리, 색동 외투, 녹색 바지, 노란 장화를 갖췄다. 다만 머리 외장이 참조보다 둥글고 밀짚모자가 없으며, 줄무늬 상의 대신 넓게 드러난 기계 흉부가 보인다. 가슴 내부 노출과 금속 손상은 확인되지만 회수 부품은 확인되지 않는다. 이현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보이고 얼굴 상처와 찢어진 청바지는 유지하지만, 참조의 후드 상의가 검은 반소매로 바뀌었다. 보라색 꽃밭·해안 절벽·낡은 트럭·연구동과 따뜻한 해 질 녘 자연광이 보인다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "찰리는 두 발이 지면에서 떨어지고 무릎을 뒤로 접은 상태다. 바로 아래와 뒤의 흙 가장자리가 도약 지점, 앞의 트럭 문턱이 착지 지점으로 읽히므로 단순한 무근거 부유로 보이지는 않는다. 금속 손은 팔에 연결되어 있고 다른 팔은 문짝 뒤로 가려져 접촉 여부를 확인할 수 없다. 이현우는 엉덩이를 좌석에, 보이는 신발을 차내 바닥에 두어 앞으로 기울인 몸을 지지한다."
       },
       {
        "label": "A",
        "direction": "찰리의 펼친 금속 손과 뻗은 팔은 왼쪽 전경 이현우의 열린 손을 직접 향하며 두 손 사이에 짧은 간격이 남아 있다. 찰리의 얼굴도 그 손과 이현우 쪽을 향한다. 이현우는 찰리 쪽으로 고개를 돌리고 손을 뻗는다. 찰리의 몸은 절벽 쪽 외부에서 카메라가 있는 트럭 내부로 비스듬히 날아드는 방향이다.",
        "built_space": "왼쪽에 열린 문짝 하나와 손잡이·원형 스피커가 각각 하나, 하단 중앙에 문턱 하나, 오른쪽 문기둥에 세로 보조 손잡이 하나가 보인다. 오른쪽 가장자리에는 좌석 일부가 보인다. 실내에서 문밖을 보는 와이드 구도이며 찰리는 중앙 오른쪽, 절벽은 그의 뒤 오른쪽에 놓여 지정된 배치에 부합한다. 이현우는 왼쪽 실내 전경에 있지만 운전석과 핸들은 가려져 정확한 착석 위치는 확인되지 않는다. 왼쪽 배경 연구동은 참조의 낡은 2층 콘크리트 외벽, 평지붕, 옥상 돌출부, 정면 큰 창과 출입구 구성을 대체로 유지한다. 설비 중복이나 불가능한 반사는 없다.",
        "entities": "찰리의 각진 기계 머리, 흰 얼굴판, 점 형태 눈과 선 형태 입, 밀짚모자, 색동 외투, 줄무늬 상의, 녹색 바지와 노란 장화가 참조에 가깝다. 가슴의 찢어진 개구부 안에 기계 구조가 보이고 옷에도 구멍과 마모가 있으며, 반대 손에는 회수한 것으로 읽히는 금속 부품을 쥐고 있다. 이현우는 검은 헝클어진 머리와 낡은 어두운 후드 상의를 유지한다. 얼굴 대부분이 보이지 않아 정확한 얼굴 일치와 기존 얼굴 상처는 판단하기 어렵다. 보라색 꽃밭, 바다를 낀 절벽, 열린 낡은 트럭 문과 연구동이 있고, 낮은 해의 따뜻한 자연광이 장면을 비춘다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "찰리는 몸통을 전방으로 기울이고 한쪽 무릎을 뒤로 접으며 다른 다리를 아래뒤로 뻗은 도약 자세다. 뒤쪽 꽃밭의 절벽 가장자리가 발구름 지점으로, 앞쪽 열린 문턱과 차내 바닥이 진입·착지 공간으로 읽힌다. 두 발이 떠 있어도 추진 방향과 착지 경로가 이어져 정지한 공중 부유가 아니다. 회수 부품은 반대쪽 금속 손가락으로 잡고 있어 지지가 명확하다. 이현우의 하체는 화면 밖이므로 좌석 접촉은 확인할 수 없지만, 보이는 상체와 팔에는 불가능한 관절이나 독립적으로 떠 있는 부분이 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 찰리가 지지대나 도약 징후 없이 허공에 수평으로 떠 있음 (nothing supports it)",
     "[gemini-pro] 이현우가 운전석이 아닌 조수석(우측) 문에 배치됨"
    ],
    "B": [
     "[gemini-pro] 찰리가 지지대나 도약 없이 허공에 수직으로 떠 있음 (nothing supports it)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1417
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지지대 없이 허공에 떠 있는 물리 법칙 위반과 조수석 배치 오류가 있으나, 지정된 의상과 부품 등 시각적 요소는 더 많이 반영됨.  ★위반: [gemini-pro] 찰리가 지지대나 도약 징후 없이 허공에 수평으로 떠 있음 (nothing supports it) / [gemini-pro] 이현우가 운전석이 아닌 조수석(우측) 문에 배치됨"
   },
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "지지대 없이 허공에 떠 있는 물리 법칙 위반과 함께 캐릭터의 핵심 의상(밀짚모자, 후드티)과 들고 있어야 할 부품이 완전히 누락됨.  ★위반: [gemini-pro] 찰리가 지지대나 도약 없이 허공에 수직으로 떠 있음 (nothing supports it)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L111B01.png",
    "asset_id": "f09ab42d-aa2d-42e7-a39b-f9e2bb5fd406",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/background_chain/seed_bg_coastal_abandoned_laboratory_sel.png",
    "asset_id": "c9eeddba-70be-466d-b8a1-f5b11b99ba19",
    "role": "structure_seed_look"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:885644>",
    "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace5d-e385-760c-8e66-966dc967ca7c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S72sh38__bgfirst_bg.png",
   "bg_asset_id": "9e3c08c6-c287-46a0-9d3b-3198ee8e073d",
   "bg_record_key": "S72sh38::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S72sh38::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:21:01.384395+00:00",
  "fingerprint": "e0e12cf5218d67f5d7eb6d31175a146e19b0b6d7c4548027dacbc066d00bb796",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S72sh38_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S72sh38_sel.png",
  "source_sha256": "5f58269fb70dd6894e5c4a0e86f98e419f285d3421c1865d47f8f9a9bce097aa",
  "file": "S72sh38_cine.png",
  "staged_sha256": "831b76b0f6cf7ab66ab2a24793dc1d14b838a09b0fcef95c2adb431b43d37259",
  "latency_ms": 14241
 },
 "S72sh61::signage": {
  "fp": "c67771cf45e951fd",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S72sh61::bgfirst_bg": {
  "input_fingerprint": "cbc5f5d6e4882439",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 짐칸에서 축 늘어진 앰버의 상체를 두 팔로 빈틈없이 꽉 껴안은 채 어깨를 잔뜩 움츠린 구도.\n\nLOCATION (lock): In the rear cargo bed of a truck stranded on a road at night, exposed to the sound and weather of the surrounding rain.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Truck cargo-area interior (The truck has stopped, and the companions remain inside) — Floor and inner side are seen obliquely around the seated embrace; used as Forms a restrained enclosing margin around the bodies without introducing additional visible occupants.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime interior keeps the embrace readable without anticipating the later flashlight beams.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 짐칸에서 축 늘어진 앰버의 상체를 두 팔로 빈틈없이 꽉 껴안은 채 어깨를 잔뜩 움츠린 구도.\n\nLOCATION (lock): In the rear cargo bed of a truck stranded on a road at night, exposed to the sound and weather of the surrounding rain.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Truck cargo-area interior (The truck has stopped, and the companions remain inside) — Floor and inner side are seen obliquely around the seated embrace; used as Forms a restrained enclosing margin around the bodies without introducing additional visible occupants.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime interior keeps the embrace readable without anticipating the later flashlight beams.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S72sh61__bgfirst_bg.png",
  "asset_id": "c0a62b79-7f9e-4a70-8d1d-1517b798c992",
  "input_asset_ids": [
   "04b44537-a6cb-4e63-82f9-1511af98f1ef",
   "49292695-5612-447b-9aa2-62b343de2c5d"
  ]
 },
 "S72sh61": {
  "input_fingerprint": "3d32d0d441ae43a9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 짐칸에서 축 늘어진 앰버의 상체를 두 팔로 빈틈없이 꽉 껴안은 채 어깨를 잔뜩 움츠린 구도.\n\nLOCATION (lock): In the rear cargo bed of a truck stranded on a road at night, exposed to the sound and weather of the surrounding rain. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Truck cargo-area interior (The truck has stopped, and the companions remain inside) — Floor and inner side are seen obliquely around the seated embrace; used as Forms a restrained enclosing margin around the bodies without introducing additional visible occupants.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime interior keeps the embrace readable without anticipating the later flashlight beams.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is limp in the truck's cargo bed, with her upper body tightly cradled against Hyunwoo in both of his arms and her eyes closed after her side injury. Her exact head support, arm and leg positions, and lower-body contact with the cargo bed are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has stopped on the road with its fuel completely exhausted. Rain is falling outside at night. 이현우: He is hunched inside the stopped truck, holding his arms tightly around a supported weight and sobbing. His earlier injuries remain. 앰버: She has a serious bleeding wound in her side and is limp, with her eyes closed. Her earlier head injury remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 짐칸에서 축 늘어진 앰버의 상체를 두 팔로 빈틈없이 꽉 껴안은 채 어깨를 잔뜩 움츠린 구도.\n\nLOCATION (lock): In the rear cargo bed of a truck stranded on a road at night, exposed to the sound and weather of the surrounding rain. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Truck cargo-area interior (The truck has stopped, and the companions remain inside) — Floor and inner side are seen obliquely around the seated embrace; used as Forms a restrained enclosing margin around the bodies without introducing additional visible occupants.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime interior keeps the embrace readable without anticipating the later flashlight beams.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is limp in the truck's cargo bed, with her upper body tightly cradled against Hyunwoo in both of his arms and her eyes closed after her side injury. Her exact head support, arm and leg positions, and lower-body contact with the cargo bed are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has stopped on the road with its fuel completely exhausted. Rain is falling outside at night. 이현우: He is hunched inside the stopped truck, holding his arms tightly around a supported weight and sobbing. His earlier injuries remain. 앰버: She has a serious bleeding wound in her side and is limp, with her eyes closed. Her earlier head injury remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 짐칸에서 축 늘어진 앰버의 상체를 두 팔로 빈틈없이 꽉 껴안은 채 어깨를 잔뜩 움츠린 구도.\n\nLOCATION (lock): In the rear cargo bed of a truck stranded on a road at night, exposed to the sound and weather of the surrounding rain. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Truck cargo-area interior (The truck has stopped, and the companions remain inside) — Floor and inner side are seen obliquely around the seated embrace; used as Forms a restrained enclosing margin around the bodies without introducing additional visible occupants.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime interior keeps the embrace readable without anticipating the later flashlight beams.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is limp in the truck's cargo bed, with her upper body tightly cradled against Hyunwoo in both of his arms and her eyes closed after her side injury. Her exact head support, arm and leg positions, and lower-body contact with the cargo bed are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has stopped on the road with its fuel completely exhausted. Rain is falling outside at night. 이현우: He is hunched inside the stopped truck, holding his arms tightly around a supported weight and sobbing. His earlier injuries remain. 앰버: She has a serious bleeding wound in her side and is limp, with her eyes closed. Her earlier head injury remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S72sh61__bgfirst_bg.png",
     "asset_id": "c0a62b79-7f9e-4a70-8d1d-1517b798c992",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S72sh61.png",
     "asset_id": "04b44537-a6cb-4e63-82f9-1511af98f1ef",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1599151>",
     "asset_id": "572581b2-472b-4a1e-98a8-663ea4c87d61",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L111B03.png",
     "asset_id": "49292695-5612-447b-9aa2-62b343de2c5d",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1599151>",
     "asset_id": "572581b2-472b-4a1e-98a8-663ea4c87d61",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "현우는 어깨를 움츠리고 고개를 숙인 채 품에 안긴 앰버를 향해 있으며, 앰버는 눈을 감은 상태임.",
    "built_space": "어두운 트럭 짐칸 내부. 벽면의 금속 패널 구조와 창문 밖 비 내리는 풍경이 레퍼런스와 정확히 일치하며, 두 사람은 맨바닥에 직접 앉아 있음.",
    "entities": "현우(검은 머리, 낡고 어두운 후드티, 아시아계 청년)와 앰버(금발 머리, 회색 셔츠, 어린 소녀) 모두 캐릭터 레퍼런스의 특징과 의상을 잘 반영함.",
    "hard_violations": [],
    "physics": "현우는 바닥에 앉아 체중을 지탱하고 있으며, 앰버의 축 늘어진 몸은 현우의 두 팔에 의해 단단히 지탱되어 중력에 자연스럽게 순응하고 있음."
   },
   {
    "label": "B",
    "direction": "현우는 고개를 꺾어 아래를 향하고 있고, 앰버는 눈을 감고 현우의 가슴에 기대어 있음.",
    "built_space": "트럭 짐칸 내부 구조는 레퍼런스와 유사하나 창밖으로 외부 차량이 보임. 두 사람이 앉은 바닥에 프롬프트에 없는 두꺼운 매트리스 패드와 덮는 담요가 깔려 있음.",
    "entities": "현우와 앰버 모두 레퍼런스에 제시된 인상착의와 의상 특징을 적절히 갖추고 있음.",
    "hard_violations": [
     "[gemini-pro] 프롬프트 및 공간 레퍼런스에 존재하지 않는 사물(매트리스 패드 및 덮은 담요)을 바닥에 임의로 추가함.",
     "[gpt-high] 지정되지 않은 두꺼운 바닥 깔개와 큰 담요를 추가하여 인물의 착석면과 앰버 하체를 덮는 별도 침구 구성을 만들었습니다."
    ],
    "physics": "현우와 앰버 모두 매트리스 위에서 안정적으로 포즈를 취하고 중력을 받고 있으나, 앰버의 하반신은 담요에 덮여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시된 미디엄 샷(medium shot) 프레이밍을 정확하게 준수하여 인물을 화면에 꽉 차게 담아냈으며, 추가적인 사물 없이 지시된 자세와 배경 레퍼런스를 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미디엄 샷 지시를 어기고 공간이 넓게 보이는 와이드 샷으로 연출하였으며, 프롬프트와 레퍼런스에 없는 매트리스와 담요를 바닥에 임의로 추가하여 우선순위 규정을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 어깨를 움츠리고 고개를 숙인 채 품에 안긴 앰버를 향해 있으며, 앰버는 눈을 감은 상태임.",
        "built_space": "어두운 트럭 짐칸 내부. 벽면의 금속 패널 구조와 창문 밖 비 내리는 풍경이 레퍼런스와 정확히 일치하며, 두 사람은 맨바닥에 직접 앉아 있음.",
        "entities": "현우(검은 머리, 낡고 어두운 후드티, 아시아계 청년)와 앰버(금발 머리, 회색 셔츠, 어린 소녀) 모두 캐릭터 레퍼런스의 특징과 의상을 잘 반영함.",
        "hard_violations": [],
        "physics": "현우는 바닥에 앉아 체중을 지탱하고 있으며, 앰버의 축 늘어진 몸은 현우의 두 팔에 의해 단단히 지탱되어 중력에 자연스럽게 순응하고 있음."
       },
       {
        "label": "B",
        "direction": "현우는 고개를 꺾어 아래를 향하고 있고, 앰버는 눈을 감고 현우의 가슴에 기대어 있음.",
        "built_space": "트럭 짐칸 내부 구조는 레퍼런스와 유사하나 창밖으로 외부 차량이 보임. 두 사람이 앉은 바닥에 프롬프트에 없는 두꺼운 매트리스 패드와 덮는 담요가 깔려 있음.",
        "entities": "현우와 앰버 모두 레퍼런스에 제시된 인상착의와 의상 특징을 적절히 갖추고 있음.",
        "hard_violations": [
         "프롬프트 및 공간 레퍼런스에 존재하지 않는 사물(매트리스 패드 및 덮은 담요)을 바닥에 임의로 추가함."
        ],
        "physics": "현우와 앰버 모두 매트리스 위에서 안정적으로 포즈를 취하고 중력을 받고 있으나, 앰버의 하반신은 담요에 덮여 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시된 미디엄 샷(medium shot) 프레이밍을 정확하게 준수하여 인물을 화면에 꽉 차게 담아냈으며, 추가적인 사물 없이 지시된 자세와 배경 레퍼런스를 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미디엄 샷 지시를 어기고 공간이 넓게 보이는 와이드 샷으로 연출하였으며, 프롬프트와 레퍼런스에 없는 매트리스와 담요를 바닥에 임의로 추가하여 우선순위 규정을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "현우는 어깨를 움츠리고 고개를 숙인 채 품에 안긴 앰버를 향해 있으며, 앰버는 눈을 감은 상태임.",
        "built_space": "어두운 트럭 짐칸 내부. 벽면의 금속 패널 구조와 창문 밖 비 내리는 풍경이 레퍼런스와 정확히 일치하며, 두 사람은 맨바닥에 직접 앉아 있음.",
        "entities": "현우(검은 머리, 낡고 어두운 후드티, 아시아계 청년)와 앰버(금발 머리, 회색 셔츠, 어린 소녀) 모두 캐릭터 레퍼런스의 특징과 의상을 잘 반영함.",
        "hard_violations": [],
        "physics": "현우는 바닥에 앉아 체중을 지탱하고 있으며, 앰버의 축 늘어진 몸은 현우의 두 팔에 의해 단단히 지탱되어 중력에 자연스럽게 순응하고 있음."
       },
       {
        "label": "B",
        "direction": "현우는 고개를 꺾어 아래를 향하고 있고, 앰버는 눈을 감고 현우의 가슴에 기대어 있음.",
        "built_space": "트럭 짐칸 내부 구조는 레퍼런스와 유사하나 창밖으로 외부 차량이 보임. 두 사람이 앉은 바닥에 프롬프트에 없는 두꺼운 매트리스 패드와 덮는 담요가 깔려 있음.",
        "entities": "현우와 앰버 모두 레퍼런스에 제시된 인상착의와 의상 특징을 적절히 갖추고 있음.",
        "hard_violations": [
         "프롬프트 및 공간 레퍼런스에 존재하지 않는 사물(매트리스 패드 및 덮은 담요)을 바닥에 임의로 추가함."
        ],
        "physics": "현우와 앰버 모두 매트리스 위에서 안정적으로 포즈를 취하고 중력을 받고 있으나, 앰버의 하반신은 담요에 덮여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "낮과 포옹의 지지는 맞지만, 전신과 넓은 짐칸을 보여주는 와이드 구도로 미디엄 숏을 어기고 지정되지 않은 침구를 추가했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "미디엄 숏에서 움츠린 현우가 앰버의 상체를 두 팔로 밀착해 받치는 핵심 연출이 우세하지만, 명시된 낮 시간과 머리 부상 표현은 맞지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 고개와 시선을 품에 안긴 앰버 쪽으로 내립니다. 앰버는 눈을 감고 머리를 화면 오른쪽의 현우 팔에 기댑니다. 두 팔은 앰버의 어깨와 몸통을 감싸며, 카메라를 바라보는 사람은 없습니다.",
        "built_space": "왼쪽 벽에 중앙 분할된 창 한 개, 정면과 양옆의 낡은 패널 벽, 벽을 따라 이어지는 화물 고정 레일과 어두운 바닥이 보입니다. 주요 구조와 재질은 장소 참조에 부합합니다. 현우는 오른쪽 뒤편 바닥의 깔개 위에 앉아 있습니다. 인물 전신과 넓은 빈 바닥이 드러나 공간이 제한적인 테두리가 아니라 화면의 큰 비중을 차지합니다. 창밖은 비 오는 낮으로 보입니다.",
        "entities": "검은 헝클어진 머리의 젊은 한국계 남성과 금발 여자아이 두 명만 짐칸 안에 보입니다. 현우의 낡은 짙은 후드와 찢어진 청바지는 참조와 가깝고, 앰버의 회색 상의도 대체로 맞습니다. 얼굴의 세부 동일성은 작은 인물 크기 때문에 확인이 제한됩니다. 앰버는 눈을 감았지만 노출된 머리에 기존 부상이 뚜렷하지 않습니다. 옆구리는 팔과 천에 가려 출혈 상태를 확인할 수 없습니다. 바닥 깔개와 큰 담요가 추가되었으며 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "지정되지 않은 두꺼운 바닥 깔개와 큰 담요를 추가하여 인물의 착석면과 앰버 하체를 덮는 별도 침구 구성을 만들었습니다."
        ],
        "physics": "현우의 엉덩이와 접힌 다리는 바닥에 놓인 깔개가 받칩니다. 앰버의 상체는 현우의 양팔과 무릎에 지지되고 머리는 감싼 팔에 기댑니다. 보이는 손은 몸통 아래쪽에 내려와 받쳐져 있으며 허공에 떠 있지 않습니다. 담요 아래 하체의 정확한 자세는 확인할 수 없지만, 보이는 신체에 무지지 부유는 없습니다."
       },
       {
        "label": "B",
        "direction": "현우는 앰버의 정수리 쪽으로 고개를 깊이 숙이고 얼굴을 가까이 댑니다. 앰버는 눈을 감은 채 현우의 가슴 쪽으로 머리가 기울어져 있습니다. 현우의 두 팔은 앰버의 어깨와 몸통을 향해 안쪽으로 감겨 있어 밀착 포옹의 대상과 방향이 명확합니다.",
        "built_space": "왼쪽에 분할 창 한 개, 정면과 오른쪽의 낡은 패널 벽, 화물 고정 레일, 젖은 듯한 어두운 바닥이 보이며 장소 참조의 구조를 유지합니다. 두 사람은 짐칸 바닥에 앉아 있고 인물이 화면 중심과 전경을 크게 채웁니다. 하체는 화면 아래에서 잘리며 바닥과 내벽이 포옹 주변에 비스듬히 보여 미디엄 숏 요구에 더 가깝습니다. 창밖의 검은 하늘과 가로등은 밤을 나타내므로 명시된 낮 시간 잠금과 충돌합니다.",
        "entities": "검은 헝클어진 머리의 젊은 한국계 남성과 금발 여자아이만 등장합니다. 현우의 얼굴은 숙여 일부 가려지지만 머리와 체격, 낡은 짙은 후드 및 청바지는 참조에 가깝습니다. 앰버는 어린 여자아이의 얼굴과 금발을 유지하며 눈을 감고 있습니다. 회색 상의는 참조의 니트보다 얇고 소매가 짧아 보입니다. 노출된 머리에 기존 상처가 확인되지 않고, 옆구리는 현우의 팔에 가려 심한 출혈 여부를 판단할 수 없습니다. 추가 인물이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "현우는 바닥에 앉아 다리를 굽힌 상태로 체중을 지탱합니다. 앰버의 몸통은 현우의 가슴과 양팔에 받쳐지고 머리는 그의 가슴에 옆으로 기댑니다. 앰버의 두 팔은 아래로 떨어져 팔뚝이 무릎 쪽에 놓이며 능동적으로 들려 있지 않습니다. 머리와 상체, 보이는 팔에 실제 지지와 자연스러운 중력 방향이 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "낮과 포옹의 지지는 맞지만, 전신과 넓은 짐칸을 보여주는 와이드 구도로 미디엄 숏을 어기고 지정되지 않은 침구를 추가했습니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미디엄 숏에서 움츠린 현우가 앰버의 상체를 두 팔로 밀착해 받치는 핵심 연출이 우세하지만, 명시된 낮 시간과 머리 부상 표현은 맞지 않습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 고개와 시선을 품에 안긴 앰버 쪽으로 내립니다. 앰버는 눈을 감고 머리를 화면 오른쪽의 현우 팔에 기댑니다. 두 팔은 앰버의 어깨와 몸통을 감싸며, 카메라를 바라보는 사람은 없습니다.",
        "built_space": "왼쪽 벽에 중앙 분할된 창 한 개, 정면과 양옆의 낡은 패널 벽, 벽을 따라 이어지는 화물 고정 레일과 어두운 바닥이 보입니다. 주요 구조와 재질은 장소 참조에 부합합니다. 현우는 오른쪽 뒤편 바닥의 깔개 위에 앉아 있습니다. 인물 전신과 넓은 빈 바닥이 드러나 공간이 제한적인 테두리가 아니라 화면의 큰 비중을 차지합니다. 창밖은 비 오는 낮으로 보입니다.",
        "entities": "검은 헝클어진 머리의 젊은 한국계 남성과 금발 여자아이 두 명만 짐칸 안에 보입니다. 현우의 낡은 짙은 후드와 찢어진 청바지는 참조와 가깝고, 앰버의 회색 상의도 대체로 맞습니다. 얼굴의 세부 동일성은 작은 인물 크기 때문에 확인이 제한됩니다. 앰버는 눈을 감았지만 노출된 머리에 기존 부상이 뚜렷하지 않습니다. 옆구리는 팔과 천에 가려 출혈 상태를 확인할 수 없습니다. 바닥 깔개와 큰 담요가 추가되었으며 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "지정되지 않은 두꺼운 바닥 깔개와 큰 담요를 추가하여 인물의 착석면과 앰버 하체를 덮는 별도 침구 구성을 만들었습니다."
        ],
        "physics": "현우의 엉덩이와 접힌 다리는 바닥에 놓인 깔개가 받칩니다. 앰버의 상체는 현우의 양팔과 무릎에 지지되고 머리는 감싼 팔에 기댑니다. 보이는 손은 몸통 아래쪽에 내려와 받쳐져 있으며 허공에 떠 있지 않습니다. 담요 아래 하체의 정확한 자세는 확인할 수 없지만, 보이는 신체에 무지지 부유는 없습니다."
       },
       {
        "label": "A",
        "direction": "현우는 앰버의 정수리 쪽으로 고개를 깊이 숙이고 얼굴을 가까이 댑니다. 앰버는 눈을 감은 채 현우의 가슴 쪽으로 머리가 기울어져 있습니다. 현우의 두 팔은 앰버의 어깨와 몸통을 향해 안쪽으로 감겨 있어 밀착 포옹의 대상과 방향이 명확합니다.",
        "built_space": "왼쪽에 분할 창 한 개, 정면과 오른쪽의 낡은 패널 벽, 화물 고정 레일, 젖은 듯한 어두운 바닥이 보이며 장소 참조의 구조를 유지합니다. 두 사람은 짐칸 바닥에 앉아 있고 인물이 화면 중심과 전경을 크게 채웁니다. 하체는 화면 아래에서 잘리며 바닥과 내벽이 포옹 주변에 비스듬히 보여 미디엄 숏 요구에 더 가깝습니다. 창밖의 검은 하늘과 가로등은 밤을 나타내므로 명시된 낮 시간 잠금과 충돌합니다.",
        "entities": "검은 헝클어진 머리의 젊은 한국계 남성과 금발 여자아이만 등장합니다. 현우의 얼굴은 숙여 일부 가려지지만 머리와 체격, 낡은 짙은 후드 및 청바지는 참조에 가깝습니다. 앰버는 어린 여자아이의 얼굴과 금발을 유지하며 눈을 감고 있습니다. 회색 상의는 참조의 니트보다 얇고 소매가 짧아 보입니다. 노출된 머리에 기존 상처가 확인되지 않고, 옆구리는 현우의 팔에 가려 심한 출혈 여부를 판단할 수 없습니다. 추가 인물이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "현우는 바닥에 앉아 다리를 굽힌 상태로 체중을 지탱합니다. 앰버의 몸통은 현우의 가슴과 양팔에 받쳐지고 머리는 그의 가슴에 옆으로 기댑니다. 앰버의 두 팔은 아래로 떨어져 팔뚝이 무릎 쪽에 놓이며 능동적으로 들려 있지 않습니다. 머리와 상체, 보이는 팔에 실제 지지와 자연스러운 중력 방향이 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.905
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.655
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트 및 공간 레퍼런스에 존재하지 않는 사물(매트리스 패드 및 덮은 담요)을 바닥에 임의로 추가함.",
     "[gpt-high] 지정되지 않은 두꺼운 바닥 깔개와 큰 담요를 추가하여 인물의 착석면과 앰버 하체를 덮는 별도 침구 구성을 만들었습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 655
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 미디엄 샷(medium shot) 프레이밍을 정확하게 준수하여 인물을 화면에 꽉 차게 담아냈으며, 추가적인 사물 없이 지시된 자세와 배경 레퍼런스를 충실하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 655,
    "verdict_ko": "미디엄 샷 지시를 어기고 공간이 넓게 보이는 와이드 샷으로 연출하였으며, 프롬프트와 레퍼런스에 없는 매트리스와 담요를 바닥에 임의로 추가하여 우선순위 규정을 위반했습니다.  ★위반: [gemini-pro] 프롬프트 및 공간 레퍼런스에 존재하지 않는 사물(매트리스 패드 및 덮은 담요)을 바닥에 임의로 추가함. / [gpt-high] 지정되지 않은 두꺼운 바닥 깔개와 큰 담요를 추가하여 인물의 착석면과 앰버 하체를 덮는 별도 침구 구성을 만들었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L111B03.png",
    "asset_id": "49292695-5612-447b-9aa2-62b343de2c5d",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1599151>",
    "asset_id": "572581b2-472b-4a1e-98a8-663ea4c87d61",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace66-bfb4-7f61-b760-afda274ae7d8",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S72sh61__bgfirst_bg.png",
   "bg_asset_id": "c0a62b79-7f9e-4a70-8d1d-1517b798c992",
   "bg_record_key": "S72sh61::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S72sh61::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:24:08.376333+00:00",
  "fingerprint": "7cff152ba92da23941e3e4541bc48084a4aa86b2aca938e11fd41bb0b3ebff83",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S72sh61_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S72sh61_sel.png",
  "source_sha256": "8102c2784d3027f9ce714fb681ecf8b127ad4ea4d663b115d17065001681bf2f",
  "file": "S72sh61_cine.png",
  "staged_sha256": "1c2c79d696f6ae884a6db36f8aa01a7b9c61cf4f270e2d89b1857db56dd4abe7",
  "latency_ms": 13220
 },
 "S72sh65::signage": {
  "fp": "86751044773f65c5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S72sh65": {
  "input_fingerprint": "fca20c80f728afe5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 트럭 문 너머로, 손전등을 든 성인 남성들의 짙은 실루엣이 위압적으로 서 있는 구도.\n\nLOCATION (lock): At the open door of the stranded truck on a rainy nighttime road, with rescuers standing outside and shining flashlights toward the occupants. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck doorway (Fully open) — Both sides of the opening are seen obliquely from inside the truck; used as Frames the silhouettes and preserves the low, sheltered viewing position; Handheld flashlights (Held by the men and directed into the truck) — Seen at differing oblique angles beside the men's hands; used as Small practical objects establish the source of the occupants' discomfort.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Powerful flashlight glare entering the nighttime truck leaves the men in dense silhouette and makes the opening feel threatening without adding visible atmospheric beams.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the stopped truck's interior surfaces, existing damage, and rainy nighttime setting. EXCLUDE a closed door panel where the rescuers have now opened the entrance.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is limp in the truck's cargo bed, with her upper body tightly cradled against Hyunwoo in both of his arms and her eyes closed after her side injury. Her exact head support, arm and leg positions, and lower-body contact with the cargo bed are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The fuel-empty truck remains stopped on the rainy nighttime road, now with its door open. Strong flashlight beams flood the truck interior, producing intense glare and silhouetting the figures beyond the doorway.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 성인 남성들 right now, so 성인 남성들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 성인 남성들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 트럭 문 너머로, 손전등을 든 성인 남성들의 짙은 실루엣이 위압적으로 서 있는 구도.\n\nLOCATION (lock): At the open door of the stranded truck on a rainy nighttime road, with rescuers standing outside and shining flashlights toward the occupants. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck doorway (Fully open) — Both sides of the opening are seen obliquely from inside the truck; used as Frames the silhouettes and preserves the low, sheltered viewing position; Handheld flashlights (Held by the men and directed into the truck) — Seen at differing oblique angles beside the men's hands; used as Small practical objects establish the source of the occupants' discomfort.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Powerful flashlight glare entering the nighttime truck leaves the men in dense silhouette and makes the opening feel threatening without adding visible atmospheric beams.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the stopped truck's interior surfaces, existing damage, and rainy nighttime setting. EXCLUDE a closed door panel where the rescuers have now opened the entrance.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is limp in the truck's cargo bed, with her upper body tightly cradled against Hyunwoo in both of his arms and her eyes closed after her side injury. Her exact head support, arm and leg positions, and lower-body contact with the cargo bed are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The fuel-empty truck remains stopped on the rainy nighttime road, now with its door open. Strong flashlight beams flood the truck interior, producing intense glare and silhouetting the figures beyond the doorway.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 성인 남성들 right now, so 성인 남성들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 성인 남성들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 트럭 문 너머로, 손전등을 든 성인 남성들의 짙은 실루엣이 위압적으로 서 있는 구도.\n\nLOCATION (lock): At the open door of the stranded truck on a rainy nighttime road, with rescuers standing outside and shining flashlights toward the occupants. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck doorway (Fully open) — Both sides of the opening are seen obliquely from inside the truck; used as Frames the silhouettes and preserves the low, sheltered viewing position; Handheld flashlights (Held by the men and directed into the truck) — Seen at differing oblique angles beside the men's hands; used as Small practical objects establish the source of the occupants' discomfort.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Powerful flashlight glare entering the nighttime truck leaves the men in dense silhouette and makes the opening feel threatening without adding visible atmospheric beams.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): TAKE the stopped truck's interior surfaces, existing damage, and rainy nighttime setting. EXCLUDE a closed door panel where the rescuers have now opened the entrance.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is limp in the truck's cargo bed, with her upper body tightly cradled against Hyunwoo in both of his arms and her eyes closed after her side injury. Her exact head support, arm and leg positions, and lower-body contact with the cargo bed are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The fuel-empty truck remains stopped on the rainy nighttime road, now with its door open. Strong flashlight beams flood the truck interior, producing intense glare and silhouetting the figures beyond the doorway.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 성인 남성들 right now, so 성인 남성들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 성인 남성들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "남성들의 손전등 불빛이 트럭 내부인 카메라 방향을 정확히 향하고 있음.",
    "built_space": "카메라는 트럭 내부에서 열린 문 밖을 바라보는 구도이나, 이전 샷에서 트럭 안에 고정되어야 할 현우와 앰버가 트럭 외부 도로로 이동해 있음.",
    "entities": "성인 남성들의 실루엣이 보임. 현우와 앰버가 등장하지만 지정된 위치와 자세를 따르지 않음.",
    "hard_violations": [
     "[gemini-pro] person placed where the staging does not put them (현우와 앰버가 지정된 트럭 내부가 아닌 외부로 이동됨)",
     "[gpt-high] 권위 있는 숏 텍스트가 보여 주도록 지정하지 않은 현우와 앰버를 화면 중앙에 추가했다.",
     "[gpt-high] 움직일 수 없는 앰버를 적재함 안에서 문밖 도로로 옮기고, 현우에게 밀착해 안긴 기존 자세를 무릎과 등을 받쳐 든 자세로 변경했다.",
     "[gpt-high] 이전 이미지에서 창이 없던 오른쪽 벽에 창문을 추가하여 고정 구조를 변경했다."
    ],
    "physics": "현우가 젖은 도로 위에 무릎을 꿇고 앰버를 안고 있으며 자세의 물리적 지지는 정상적임."
   },
   {
    "label": "B",
    "direction": "손전등 불빛들이 트럭 내부를 강하게 비추고 있음.",
    "built_space": "트럭 내부에서 문 밖을 바라보는 구도이며, 현우와 앰버는 이전 샷과 동일한 바닥 위치에 있음. 오른쪽 트럭 문짝 안쪽에 인물의 실루엣이 2D 평면처럼 맺혀 있음.",
    "entities": "현우와 앰버의 자세가 레퍼런스와 일치함. 남성 실루엣이 외부에 서 있음. 샷 텍스트에 없는 찰리(고릴라 형태)가 화면에 추가됨.",
    "hard_violations": [
     "[gemini-pro] invented people (지시문에 명시되지 않은 찰리가 프레임에 추가됨)",
     "[gemini-pro] physically impossible staging (오른쪽 문에 맺힌 2D 평면 그림자가 3D 입체의 손전등 불빛을 발산하는 물리적 오류)",
     "[gpt-high] 권위 있는 숏 텍스트가 보여 주도록 지정하지 않은 현우·앰버·찰리를 화면에 추가했다."
    ],
    "physics": "현우와 앰버는 트럭 바닥에 자연스럽게 앉아 지지받고 있음. 우측 문의 인물 형태는 물리적으로 불가능한 조명 오류를 보임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이전 샷의 레퍼런스 자세와 위치는 유지했으나, 지시문에 없는 캐릭터(찰리)를 임의로 추가하고 우측 문짝에 물리적으로 불가능한 그림자 형태의 인물이 렌더링되어 실격입니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이전 샷의 위치와 고정된 자세를 유지해야 한다는 지침을 완전히 무시하고 인물들을 트럭 밖으로 임의 배치하여 치명적인 구도 위반을 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남성들의 손전등 불빛이 트럭 내부인 카메라 방향을 정확히 향하고 있음.",
        "built_space": "카메라는 트럭 내부에서 열린 문 밖을 바라보는 구도이나, 이전 샷에서 트럭 안에 고정되어야 할 현우와 앰버가 트럭 외부 도로로 이동해 있음.",
        "entities": "성인 남성들의 실루엣이 보임. 현우와 앰버가 등장하지만 지정된 위치와 자세를 따르지 않음.",
        "hard_violations": [
         "person placed where the staging does not put them (현우와 앰버가 지정된 트럭 내부가 아닌 외부로 이동됨)"
        ],
        "physics": "현우가 젖은 도로 위에 무릎을 꿇고 앰버를 안고 있으며 자세의 물리적 지지는 정상적임."
       },
       {
        "label": "B",
        "direction": "손전등 불빛들이 트럭 내부를 강하게 비추고 있음.",
        "built_space": "트럭 내부에서 문 밖을 바라보는 구도이며, 현우와 앰버는 이전 샷과 동일한 바닥 위치에 있음. 오른쪽 트럭 문짝 안쪽에 인물의 실루엣이 2D 평면처럼 맺혀 있음.",
        "entities": "현우와 앰버의 자세가 레퍼런스와 일치함. 남성 실루엣이 외부에 서 있음. 샷 텍스트에 없는 찰리(고릴라 형태)가 화면에 추가됨.",
        "hard_violations": [
         "invented people (지시문에 명시되지 않은 찰리가 프레임에 추가됨)",
         "physically impossible staging (오른쪽 문에 맺힌 2D 평면 그림자가 3D 입체의 손전등 불빛을 발산하는 물리적 오류)"
        ],
        "physics": "현우와 앰버는 트럭 바닥에 자연스럽게 앉아 지지받고 있음. 우측 문의 인물 형태는 물리적으로 불가능한 조명 오류를 보임."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이전 샷의 레퍼런스 자세와 위치는 유지했으나, 지시문에 없는 캐릭터(찰리)를 임의로 추가하고 우측 문짝에 물리적으로 불가능한 그림자 형태의 인물이 렌더링되어 실격입니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이전 샷의 위치와 고정된 자세를 유지해야 한다는 지침을 완전히 무시하고 인물들을 트럭 밖으로 임의 배치하여 치명적인 구도 위반을 범했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "남성들의 손전등 불빛이 트럭 내부인 카메라 방향을 정확히 향하고 있음.",
        "built_space": "카메라는 트럭 내부에서 열린 문 밖을 바라보는 구도이나, 이전 샷에서 트럭 안에 고정되어야 할 현우와 앰버가 트럭 외부 도로로 이동해 있음.",
        "entities": "성인 남성들의 실루엣이 보임. 현우와 앰버가 등장하지만 지정된 위치와 자세를 따르지 않음.",
        "hard_violations": [
         "person placed where the staging does not put them (현우와 앰버가 지정된 트럭 내부가 아닌 외부로 이동됨)"
        ],
        "physics": "현우가 젖은 도로 위에 무릎을 꿇고 앰버를 안고 있으며 자세의 물리적 지지는 정상적임."
       },
       {
        "label": "B",
        "direction": "손전등 불빛들이 트럭 내부를 강하게 비추고 있음.",
        "built_space": "트럭 내부에서 문 밖을 바라보는 구도이며, 현우와 앰버는 이전 샷과 동일한 바닥 위치에 있음. 오른쪽 트럭 문짝 안쪽에 인물의 실루엣이 2D 평면처럼 맺혀 있음.",
        "entities": "현우와 앰버의 자세가 레퍼런스와 일치함. 남성 실루엣이 외부에 서 있음. 샷 텍스트에 없는 찰리(고릴라 형태)가 화면에 추가됨.",
        "hard_violations": [
         "invented people (지시문에 명시되지 않은 찰리가 프레임에 추가됨)",
         "physically impossible staging (오른쪽 문에 맺힌 2D 평면 그림자가 3D 입체의 손전등 불빛을 발산하는 물리적 오류)"
        ],
        "physics": "현우와 앰버는 트럭 바닥에 자연스럽게 앉아 지지받고 있음. 우측 문의 인물 형태는 물리적으로 불가능한 조명 오류를 보임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "불필요한 인물 추가와 이전 숏 구도 답습으로 부적합하지만, B와 달리 앰버의 지지 자세와 적재함 내부 위치·구조는 유지한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "문 양쪽으로 실루엣을 감싸는 와이드 구도는 더 정확하지만, 움직일 수 없는 앰버를 도로로 옮기고 오른쪽 창문까지 추가하여 연속성을 크게 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "문밖 남성들은 트럭 안쪽을 향해 서 있고, 보이는 손전등 발광면도 실내와 카메라 쪽을 향한다. 강한 눈부심은 있으나 뚜렷한 원뿔형 광선은 없다. 현우는 앰버를 내려다보고, 앰버는 눈을 감고 있으며, 찰리는 카메라 쪽을 향한다.",
        "built_space": "적재함 안에서 출입구를 보는 시점이다. 좌측의 두 구획으로 나뉜 창 하나, 양쪽 벽의 고정 레일, 닳고 젖은 바닥이 이전 이미지와 대체로 이어진다. 출입구 하나에 문짝 두 개가 열려 있고 남성들은 문턱 너머에 있다. 다만 전경의 현우와 앰버가 화면 대부분을 차지하여, 출입구와 남성 실루엣 중심의 와이드 숏보다 이전 숏 구도를 답습한다.",
        "entities": "문밖에는 성인 남성으로 보이는 인물 세 명이 있으며, 어두운 옷과 머리·팔·몸통의 형태가 보인다. 짙은 역광 때문에 민족적 정체성은 판별하기 어렵고, 이들에게 지정된 정체성도 없다. 전경에는 검은 머리의 젊은 한국계 남성 현우와 금발 여자아이 앰버가 이전 이미지에 가까운 의상과 모습으로 등장한다. 그 뒤에는 고릴라형 기계 머리와 흰 얼굴판을 가진 찰리도 추가되었다. 읽을 수 있는 글자는 보이지 않는다. 비와 어둠은 야간 연속성 지시에는 맞지만 명시된 낮 시간 잠금에는 맞지 않는다.",
        "hard_violations": [
         "권위 있는 숏 텍스트가 보여 주도록 지정하지 않은 현우·앰버·찰리를 화면에 추가했다."
        ],
        "physics": "현우는 적재함 바닥에 앉아 두 팔로 앰버의 상체를 감싸며, 앰버의 머리와 몸은 현우에게 기대고 하체는 바닥과 겹친 다리 위에 놓여 있다. 이전 이미지의 지지 관계가 대체로 유지된다. 문밖 남성들의 발은 문턱과 전경에 가려져 접지점을 확인할 수 없으나, 서 있는 몸통에서 공중 부유는 보이지 않는다. 손전등은 들어 올린 손 주변에 있고, 일부 손잡이는 눈부심에 가린다. 찰리의 하체도 전경 인물에 가려 지지점을 확인할 수 없다."
       },
       {
        "label": "B",
        "direction": "네 남성 모두 트럭 안을 향하고, 각 손전등의 발광면도 카메라와 적재함 안쪽을 향한다. 높이가 서로 달라 손전등의 운용 각도 차이가 보인다. 현우는 아래쪽을 바라보며, 앰버는 눈을 감고 머리를 뒤로 늘어뜨린다. 손전등 눈부심은 강하지만 뚜렷한 대기 중 광선 기둥은 없다.",
        "built_space": "낮은 적재함 내부에서 출입구 전체를 보는 와이드 숏으로, 양쪽 문틀과 열린 문짝 두 개가 남성들을 감싼다. 벽면 고정 레일과 젖은 바닥은 이어지지만 좌우에 창이 하나씩 보여, 이전 이미지의 창 없는 오른쪽 벽과 다르다. 네 남성은 젖은 도로 위에 서 있고, 현우와 앰버도 문턱 밖 도로로 옮겨졌다.",
        "entities": "문밖에는 어두운 옷을 입은 성인 남성 네 명과 손전등 네 개가 보인다. 얼굴은 어둡지만 구체적인 사람의 체형과 팔다리는 유지되며, 민족적 정체성은 확인하기 어렵다. 중앙에는 검은 머리의 젊은 남성과 금발 여자아이가 있어 현우·앰버로 읽히고 의상도 대체로 이어진다. 찰리는 없으며 이 숏에서 빠진 것은 문제가 아니다. 읽을 수 있는 글자는 없다. 배경은 비 오는 야간으로, 명시된 낮 시간 잠금과는 불일치한다.",
        "hard_violations": [
         "권위 있는 숏 텍스트가 보여 주도록 지정하지 않은 현우와 앰버를 화면 중앙에 추가했다.",
         "움직일 수 없는 앰버를 적재함 안에서 문밖 도로로 옮기고, 현우에게 밀착해 안긴 기존 자세를 무릎과 등을 받쳐 든 자세로 변경했다.",
         "이전 이미지에서 창이 없던 오른쪽 벽에 창문을 추가하여 고정 구조를 변경했다."
        ],
        "physics": "네 남성의 신발은 도로에 접지하고 손전등은 각자의 손에 들려 있다. 현우는 도로에 무릎을 대고 앰버의 무릎 쪽과 등을 받치며, 앰버의 머리와 한쪽 팔은 중력 방향으로 처져 있다. 공중에 지지 없이 떠 있는 몸은 아니지만, 앰버의 하체가 적재함에 닿지 않고 상체도 두 팔로 밀착해 감싸지 않아 지정된 지지 자세와 다르다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "불필요한 인물 추가와 이전 숏 구도 답습으로 부적합하지만, B와 달리 앰버의 지지 자세와 적재함 내부 위치·구조는 유지한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "문 양쪽으로 실루엣을 감싸는 와이드 구도는 더 정확하지만, 움직일 수 없는 앰버를 도로로 옮기고 오른쪽 창문까지 추가하여 연속성을 크게 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "문밖 남성들은 트럭 안쪽을 향해 서 있고, 보이는 손전등 발광면도 실내와 카메라 쪽을 향한다. 강한 눈부심은 있으나 뚜렷한 원뿔형 광선은 없다. 현우는 앰버를 내려다보고, 앰버는 눈을 감고 있으며, 찰리는 카메라 쪽을 향한다.",
        "built_space": "적재함 안에서 출입구를 보는 시점이다. 좌측의 두 구획으로 나뉜 창 하나, 양쪽 벽의 고정 레일, 닳고 젖은 바닥이 이전 이미지와 대체로 이어진다. 출입구 하나에 문짝 두 개가 열려 있고 남성들은 문턱 너머에 있다. 다만 전경의 현우와 앰버가 화면 대부분을 차지하여, 출입구와 남성 실루엣 중심의 와이드 숏보다 이전 숏 구도를 답습한다.",
        "entities": "문밖에는 성인 남성으로 보이는 인물 세 명이 있으며, 어두운 옷과 머리·팔·몸통의 형태가 보인다. 짙은 역광 때문에 민족적 정체성은 판별하기 어렵고, 이들에게 지정된 정체성도 없다. 전경에는 검은 머리의 젊은 한국계 남성 현우와 금발 여자아이 앰버가 이전 이미지에 가까운 의상과 모습으로 등장한다. 그 뒤에는 고릴라형 기계 머리와 흰 얼굴판을 가진 찰리도 추가되었다. 읽을 수 있는 글자는 보이지 않는다. 비와 어둠은 야간 연속성 지시에는 맞지만 명시된 낮 시간 잠금에는 맞지 않는다.",
        "hard_violations": [
         "권위 있는 숏 텍스트가 보여 주도록 지정하지 않은 현우·앰버·찰리를 화면에 추가했다."
        ],
        "physics": "현우는 적재함 바닥에 앉아 두 팔로 앰버의 상체를 감싸며, 앰버의 머리와 몸은 현우에게 기대고 하체는 바닥과 겹친 다리 위에 놓여 있다. 이전 이미지의 지지 관계가 대체로 유지된다. 문밖 남성들의 발은 문턱과 전경에 가려져 접지점을 확인할 수 없으나, 서 있는 몸통에서 공중 부유는 보이지 않는다. 손전등은 들어 올린 손 주변에 있고, 일부 손잡이는 눈부심에 가린다. 찰리의 하체도 전경 인물에 가려 지지점을 확인할 수 없다."
       },
       {
        "label": "A",
        "direction": "네 남성 모두 트럭 안을 향하고, 각 손전등의 발광면도 카메라와 적재함 안쪽을 향한다. 높이가 서로 달라 손전등의 운용 각도 차이가 보인다. 현우는 아래쪽을 바라보며, 앰버는 눈을 감고 머리를 뒤로 늘어뜨린다. 손전등 눈부심은 강하지만 뚜렷한 대기 중 광선 기둥은 없다.",
        "built_space": "낮은 적재함 내부에서 출입구 전체를 보는 와이드 숏으로, 양쪽 문틀과 열린 문짝 두 개가 남성들을 감싼다. 벽면 고정 레일과 젖은 바닥은 이어지지만 좌우에 창이 하나씩 보여, 이전 이미지의 창 없는 오른쪽 벽과 다르다. 네 남성은 젖은 도로 위에 서 있고, 현우와 앰버도 문턱 밖 도로로 옮겨졌다.",
        "entities": "문밖에는 어두운 옷을 입은 성인 남성 네 명과 손전등 네 개가 보인다. 얼굴은 어둡지만 구체적인 사람의 체형과 팔다리는 유지되며, 민족적 정체성은 확인하기 어렵다. 중앙에는 검은 머리의 젊은 남성과 금발 여자아이가 있어 현우·앰버로 읽히고 의상도 대체로 이어진다. 찰리는 없으며 이 숏에서 빠진 것은 문제가 아니다. 읽을 수 있는 글자는 없다. 배경은 비 오는 야간으로, 명시된 낮 시간 잠금과는 불일치한다.",
        "hard_violations": [
         "권위 있는 숏 텍스트가 보여 주도록 지정하지 않은 현우와 앰버를 화면 중앙에 추가했다.",
         "움직일 수 없는 앰버를 적재함 안에서 문밖 도로로 옮기고, 현우에게 밀착해 안긴 기존 자세를 무릎과 등을 받쳐 든 자세로 변경했다.",
         "이전 이미지에서 창이 없던 오른쪽 벽에 창문을 추가하여 고정 구조를 변경했다."
        ],
        "physics": "네 남성의 신발은 도로에 접지하고 손전등은 각자의 손에 들려 있다. 현우는 도로에 무릎을 대고 앰버의 무릎 쪽과 등을 받치며, 앰버의 머리와 한쪽 팔은 중력 방향으로 처져 있다. 공중에 지지 없이 떠 있는 몸은 아니지만, 앰버의 하체가 적재함에 닿지 않고 상체도 두 팔로 밀착해 감싸지 않아 지정된 지지 자세와 다르다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.417,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.167,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] person placed where the staging does not put them (현우와 앰버가 지정된 트럭 내부가 아닌 외부로 이동됨)",
     "[gpt-high] 권위 있는 숏 텍스트가 보여 주도록 지정하지 않은 현우와 앰버를 화면 중앙에 추가했다.",
     "[gpt-high] 움직일 수 없는 앰버를 적재함 안에서 문밖 도로로 옮기고, 현우에게 밀착해 안긴 기존 자세를 무릎과 등을 받쳐 든 자세로 변경했다.",
     "[gpt-high] 이전 이미지에서 창이 없던 오른쪽 벽에 창문을 추가하여 고정 구조를 변경했다."
    ],
    "B": [
     "[gemini-pro] invented people (지시문에 명시되지 않은 찰리가 프레임에 추가됨)",
     "[gemini-pro] physically impossible staging (오른쪽 문에 맺힌 2D 평면 그림자가 3D 입체의 손전등 불빛을 발산하는 물리적 오류)",
     "[gpt-high] 권위 있는 숏 텍스트가 보여 주도록 지정하지 않은 현우·앰버·찰리를 화면에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 1167
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "이전 샷의 레퍼런스 자세와 위치는 유지했으나, 지시문에 없는 캐릭터(찰리)를 임의로 추가하고 우측 문짝에 물리적으로 불가능한 그림자 형태의 인물이 렌더링되어 실격입니다.  ★위반: [gemini-pro] invented people (지시문에 명시되지 않은 찰리가 프레임에 추가됨) / [gemini-pro] physically impossible staging (오른쪽 문에 맺힌 2D 평면 그림자가 3D 입체의 손전등 불빛을 발산하는 물리적 오류) / [gpt-high] 권위 있는 숏 텍스트가 보여 주도록 지정하지 않은 현우·앰버·찰리를 화면에 추가했다."
   },
   {
    "label": "A",
    "score": 1167,
    "verdict_ko": "이전 샷의 위치와 고정된 자세를 유지해야 한다는 지침을 완전히 무시하고 인물들을 트럭 밖으로 임의 배치하여 치명적인 구도 위반을 범했습니다.  ★위반: [gemini-pro] person placed where the staging does not put them (현우와 앰버가 지정된 트럭 내부가 아닌 외부로 이동됨) / [gpt-high] 권위 있는 숏 텍스트가 보여 주도록 지정하지 않은 현우와 앰버를 화면 중앙에 추가했다. / [gpt-high] 움직일 수 없는 앰버를 적재함 안에서 문밖 도로로 옮기고, 현우에게 밀착해 안긴 기존 자세를 무릎과 등을 받쳐 든 자세로 변경했다. / [gpt-high] 이전 이미지에서 창이 없던 오른쪽 벽에 창문을 추가하여 고정 구조를 변경했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S72sh61_sel.png",
    "asset_id": "b5c28acc-70cd-4fe8-9e98-5e4f35494b97",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1260199>",
    "asset_id": "ed57a831-f148-48bd-a65c-5b72dfdc874a",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1313681>",
    "asset_id": "3de05311-7b80-409b-bc84-fe8042dbb3a1",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace72-5f38-70ba-9d41-7e86e61da6ca",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S72sh61"
  },
  "lane_policy": "ab_select_bypass:prev",
  "locked_char_refs_excluded": [
   "앰버(C05)"
  ]
 },
 "S72sh65::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:25:28.853205+00:00",
  "fingerprint": "77790d3a7fdd29f2ab4d85013e566414d94d4aa635aedbcca19d8dd9db370acf",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S72sh65_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S72sh65_sel.png",
  "source_sha256": "a8b76060ba98461b90b7038b5f1d73b8fb08cd49f9fa0be76127b35b63979f1d",
  "file": "S72sh65_cine.png",
  "staged_sha256": "a2a2c8bc61173937326725f1f6f2b0d313747b64e0afc120e782a25da3c07bc3",
  "latency_ms": 14093
 },
 "S73sh2::signage": {
  "fp": "a7a096e7a07627e1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::14712d4fd10d329f": {
  "subjects": [],
  "subject_text": "목포 보건소 병실\n침대와 기본 진료 비품이 놓인 소규모 보건소 병실. 실내 조명 아래 침구와 병상 주변의 단순한 공간이 드러난다.",
  "identity": "canonical",
  "scope_id": "L115",
  "scope_role": "location_interior",
  "scope_sha": "3f8d5b64d32d1212"
 },
 "S73sh2::bgfirst_bg": {
  "input_fingerprint": "88f61877ebf79dc2",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 침대에 누운 앰버를 내려다보며 미간을 살짝 찌푸린 걱정스러운 표정.\n\nLOCATION (lock): Beside a sleeping patient's bed inside a harbor-town health clinic, under nighttime clinical lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Health-center bed (Occupied by 앰버 after emergency treatment) — The near side runs diagonally beneath the two faces; used as Connects the worried observer above with the sleeping patient below without exaggerating their height difference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime health center keeps his worried expression and her resting face legible without a stylized perceptual effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 침대에 누운 앰버를 내려다보며 미간을 살짝 찌푸린 걱정스러운 표정.\n\nLOCATION (lock): Beside a sleeping patient's bed inside a harbor-town health clinic, under nighttime clinical lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Health-center bed (Occupied by 앰버 after emergency treatment) — The near side runs diagonally beneath the two faces; used as Connects the worried observer above with the sleeping patient below without exaggerating their height difference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime health center keeps his worried expression and her resting face legible without a stylized perceptual effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S73sh2__bgfirst_bg.png",
  "asset_id": "3c2d60cc-e5ee-406d-bc22-ac9bab4b356b",
  "input_asset_ids": [
   "8c030b7e-b4bd-422b-a85e-837a4edf1d19",
   "688d6723-0899-4764-9cab-c02e19df9eff"
  ]
 },
 "S73sh2": {
  "input_fingerprint": "246d1feb9e38098c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 침대에 누운 앰버를 내려다보며 미간을 살짝 찌푸린 걱정스러운 표정.\n\nLOCATION (lock): Beside a sleeping patient's bed inside a harbor-town health clinic, under nighttime clinical lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Health-center bed (Occupied by 앰버 after emergency treatment) — The near side runs diagonally beneath the two faces; used as Connects the worried observer above with the sleeping patient below without exaggerating their height difference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime health center keeps his worried expression and her resting face legible without a stylized perceptual effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is reclining on the clinic bed after emergency treatment, with her body supported by the bed and her eyes closed in sleep. Her exact head support, torso orientation, and arm and leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic contains the bed used for emergency recovery. 앰버: She is reclining on the clinic bed after emergency treatment for her side wound, with her body supported by the bed and her eyes closed in sleep. Her condition has stabilized for the moment. 이현우: He stands at the bedside with a worried expression. His skin and clothing are grimy, with accumulated injuries still evident.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 침대에 누운 앰버를 내려다보며 미간을 살짝 찌푸린 걱정스러운 표정.\n\nLOCATION (lock): Beside a sleeping patient's bed inside a harbor-town health clinic, under nighttime clinical lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Health-center bed (Occupied by 앰버 after emergency treatment) — The near side runs diagonally beneath the two faces; used as Connects the worried observer above with the sleeping patient below without exaggerating their height difference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime health center keeps his worried expression and her resting face legible without a stylized perceptual effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is reclining on the clinic bed after emergency treatment, with her body supported by the bed and her eyes closed in sleep. Her exact head support, torso orientation, and arm and leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic contains the bed used for emergency recovery. 앰버: She is reclining on the clinic bed after emergency treatment for her side wound, with her body supported by the bed and her eyes closed in sleep. Her condition has stabilized for the moment. 이현우: He stands at the bedside with a worried expression. His skin and clothing are grimy, with accumulated injuries still evident.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 침대에 누운 앰버를 내려다보며 미간을 살짝 찌푸린 걱정스러운 표정.\n\nLOCATION (lock): Beside a sleeping patient's bed inside a harbor-town health clinic, under nighttime clinical lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Health-center bed (Occupied by 앰버 after emergency treatment) — The near side runs diagonally beneath the two faces; used as Connects the worried observer above with the sleeping patient below without exaggerating their height difference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime health center keeps his worried expression and her resting face legible without a stylized perceptual effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is reclining on the clinic bed after emergency treatment, with her body supported by the bed and her eyes closed in sleep. Her exact head support, torso orientation, and arm and leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic contains the bed used for emergency recovery. 앰버: She is reclining on the clinic bed after emergency treatment for her side wound, with her body supported by the bed and her eyes closed in sleep. Her condition has stabilized for the moment. 이현우: He stands at the bedside with a worried expression. His skin and clothing are grimy, with accumulated injuries still evident.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S73sh2__bgfirst_bg.png",
     "asset_id": "3c2d60cc-e5ee-406d-bc22-ac9bab4b356b",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S73sh2.png",
     "asset_id": "8c030b7e-b4bd-422b-a85e-837a4edf1d19",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L115B02.png",
     "asset_id": "688d6723-0899-4764-9cab-c02e19df9eff",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선이 침대에 누운 앰버의 얼굴을 향하며, 앰버는 눈을 감고 있음.",
    "built_space": "레퍼런스의 병실 구조(왼쪽 벽의 패널, 정면 수납장, 우측 창문)가 올바른 위치와 비례로 구현됨.",
    "entities": "이현우(헝클어진 머리, 오염된 후드티)와 앰버(금발, 회색 상의) 모두 참조 이미지의 정체성과 복장에 일치함.",
    "hard_violations": [],
    "physics": "이현우는 바닥을 딛고 서서 상체를 숙이고 있으며, 앰버는 침대 위에 누워 안정적으로 지탱됨."
   },
   {
    "label": "B",
    "direction": "이현우의 시선이 앰버를 향하며, 앰버는 눈을 감고 있음.",
    "built_space": "공간 구조가 왜곡됨. 벽면의 의료 패널이 레퍼런스에 비해 비정상적으로 길게 묘사되었고 창문과의 배치 비율이 다름.",
    "entities": "이현우와 앰버 모두 참조 이미지의 외형과 의상(앰버의 겉옷과 신발 포함)에 일치함.",
    "hard_violations": [],
    "physics": "이현우는 바닥에 서 있고, 앰버는 침대 위에 누워 체중이 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 미디엄 샷 스케일과 침대의 대각선 구도를 정확히 구현했으며, 장소 레퍼런스의 공간 구조를 충실히 재현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 미디엄 샷보다 프레임이 너무 넓고 침대가 가로로 배치되어 구도 지침을 실패했으며, 배경의 병실 구조가 왜곡됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 침대에 누운 앰버의 얼굴을 향하며, 앰버는 눈을 감고 있음.",
        "built_space": "레퍼런스의 병실 구조(왼쪽 벽의 패널, 정면 수납장, 우측 창문)가 올바른 위치와 비례로 구현됨.",
        "entities": "이현우(헝클어진 머리, 오염된 후드티)와 앰버(금발, 회색 상의) 모두 참조 이미지의 정체성과 복장에 일치함.",
        "hard_violations": [],
        "physics": "이현우는 바닥을 딛고 서서 상체를 숙이고 있으며, 앰버는 침대 위에 누워 안정적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 앰버를 향하며, 앰버는 눈을 감고 있음.",
        "built_space": "공간 구조가 왜곡됨. 벽면의 의료 패널이 레퍼런스에 비해 비정상적으로 길게 묘사되었고 창문과의 배치 비율이 다름.",
        "entities": "이현우와 앰버 모두 참조 이미지의 외형과 의상(앰버의 겉옷과 신발 포함)에 일치함.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 서 있고, 앰버는 침대 위에 누워 체중이 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 미디엄 샷 스케일과 침대의 대각선 구도를 정확히 구현했으며, 장소 레퍼런스의 공간 구조를 충실히 재현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 미디엄 샷보다 프레임이 너무 넓고 침대가 가로로 배치되어 구도 지침을 실패했으며, 배경의 병실 구조가 왜곡됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 침대에 누운 앰버의 얼굴을 향하며, 앰버는 눈을 감고 있음.",
        "built_space": "레퍼런스의 병실 구조(왼쪽 벽의 패널, 정면 수납장, 우측 창문)가 올바른 위치와 비례로 구현됨.",
        "entities": "이현우(헝클어진 머리, 오염된 후드티)와 앰버(금발, 회색 상의) 모두 참조 이미지의 정체성과 복장에 일치함.",
        "hard_violations": [],
        "physics": "이현우는 바닥을 딛고 서서 상체를 숙이고 있으며, 앰버는 침대 위에 누워 안정적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 앰버를 향하며, 앰버는 눈을 감고 있음.",
        "built_space": "공간 구조가 왜곡됨. 벽면의 의료 패널이 레퍼런스에 비해 비정상적으로 길게 묘사되었고 창문과의 배치 비율이 다름.",
        "entities": "이현우와 앰버 모두 참조 이미지의 외형과 의상(앰버의 겉옷과 신발 포함)에 일치함.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 서 있고, 앰버는 침대 위에 누워 체중이 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "침대와 병실 배치는 참조에 가깝지만, 앰버의 발끝까지 담은 넓은 구도로 지정된 미디엄 쇼트와 두 얼굴 중심의 연결이 약해졌다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "침대 위치와 일부 의상은 참조에서 달라졌으나, 미디엄 쇼트에서 앰버를 걱정스럽게 내려다보는 현우와 두 얼굴 아래의 대각선 난간을 더 충실하게 구현했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 화면 왼쪽 아래의 앰버 얼굴 쪽으로 고개와 눈을 내리고 있다. 앰버는 눈을 감고 얼굴을 약간 위로 향한 채 누워 있다. 카메라를 바라보는 인물이나 별도의 지향성 소품은 없다.",
        "built_space": "침대 1개가 왼쪽 머리맡에서 오른쪽 아래로 뻗는다. 머리맡 벽에는 의료용 연결 패널 1개, 가로등 1개, 벽 수납장 1개가 있고, 침상 옆 수납장 1개와 오른쪽 창문 및 커튼이 보인다. 참조의 주요 시설 관계가 대체로 유지된다. 현우는 침대 건너편 옆에 서 있으며 가구와 충돌하지 않는다. 가까운 침대 가장자리는 두 얼굴 아래를 지나지만 화면을 넓게 잡아 앰버의 전신과 침상 상당 부분까지 보인다. 창의 실내 조명 반사는 가능한 배치다.",
        "entities": "두 사람만 보인다. 현우는 헝클어진 검은 머리의 젊은 한국계 남성으로 보이며, 참조와 유사한 얼굴·체격, 낡은 짙은 후드와 청바지, 얼굴의 오염과 상처가 확인된다. 금발의 앰버는 약 10세 여자아이로 보이고 참조와 유사한 얼굴과 체격이다. 혼혈 정체성을 외모만으로 확정할 수는 없다. 앰버의 녹색 재킷, 회색 원피스와 갈색 부츠는 참조와 맞지만 마스크는 없다. 눈은 감겨 있어 큰 눈이라는 특징은 확인할 수 없다. 밤의 창밖과 절제된 병실 조명이 보이며, 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "앰버의 머리와 목은 베개, 몸통과 다리는 매트리스에 지지되어 있다. 보이는 손과 팔도 침상 위에 놓여 있고 부츠를 신은 발도 매트리스에 닿는다. 현우의 발은 프레임 밖이지만 하체가 자연스럽게 아래로 이어져 서 있는 자세로 읽힌다. 부유하는 신체나 소품은 없다."
       },
       {
        "label": "B",
        "direction": "현우는 상체를 살짝 숙이고 오른쪽 아래에 있는 앰버의 얼굴을 내려다본다. 눈썹 사이의 긴장과 다문 입이 걱정스러운 표정을 만든다. 앰버는 눈을 감고 베개에 기대어 있으며 시선 교환은 없다.",
        "built_space": "침대 1개와 가까운 쪽 난간 1개가 전경에 보인다. 난간은 왼쪽 아래에서 두 인물 사이로 대각선을 이루며 두 얼굴을 연결한다. 배경에는 출입문 1개, 의료용 연결 패널 1개, 가로등 1개, 벽 수납장 1개, 침상 옆 수납장 1개, 오른쪽 창문과 커튼이 보인다. 벽의 두 색 도장과 야간 창밖도 참조에 대응한다. 다만 침대 머리맡이 참조보다 창문 쪽 전경으로 옮겨져 후면 의료 패널과의 위치 관계는 덜 충실하다. 현우가 선 통로에는 공간이 있고 난간과 몸이 충돌하지 않는다. 현우의 상체와 앰버의 상반신이 중심인 미디엄 구도다.",
        "entities": "현우와 앰버 두 사람만 있다. 현우는 참조에 가까운 젊은 한국계 남성 얼굴, 헝클어진 검은 머리, 마른 체격과 낡은 후드·청바지를 갖추었고 얼굴의 때와 상처가 보인다. 앰버는 금발의 어린 여자아이이며 참조와 유사한 얼굴이지만 혼혈 여부는 외모만으로 확정할 수 없다. 눈은 감겨 있다. 녹색 재킷과 마스크는 보이지 않고 회색 반소매 옷을 입었으며, 몸에는 담요가 덮여 있다. 팔의 의료용 테이프는 치료 후 상태에 부합한다. 창밖은 밤이고 글자는 판독되지 않는다.",
        "hard_violations": [],
        "physics": "앰버의 머리는 베개에, 몸통은 침상에 완전히 기대어 있다. 보이는 팔은 담요와 매트리스 위에 내려놓여 있고 손은 아래쪽에서 가려진다. 담요도 몸과 침대 표면을 따라 놓인다. 현우는 하체를 유지한 채 허리에서 조금 숙인 자세로, 발은 화면 밖이지만 지지 없는 공중 자세로 보이지 않는다. 난간과 의료용 테이프에도 명확한 물리적 지지가 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "침대와 병실 배치는 참조에 가깝지만, 앰버의 발끝까지 담은 넓은 구도로 지정된 미디엄 쇼트와 두 얼굴 중심의 연결이 약해졌다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "침대 위치와 일부 의상은 참조에서 달라졌으나, 미디엄 쇼트에서 앰버를 걱정스럽게 내려다보는 현우와 두 얼굴 아래의 대각선 난간을 더 충실하게 구현했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 화면 왼쪽 아래의 앰버 얼굴 쪽으로 고개와 눈을 내리고 있다. 앰버는 눈을 감고 얼굴을 약간 위로 향한 채 누워 있다. 카메라를 바라보는 인물이나 별도의 지향성 소품은 없다.",
        "built_space": "침대 1개가 왼쪽 머리맡에서 오른쪽 아래로 뻗는다. 머리맡 벽에는 의료용 연결 패널 1개, 가로등 1개, 벽 수납장 1개가 있고, 침상 옆 수납장 1개와 오른쪽 창문 및 커튼이 보인다. 참조의 주요 시설 관계가 대체로 유지된다. 현우는 침대 건너편 옆에 서 있으며 가구와 충돌하지 않는다. 가까운 침대 가장자리는 두 얼굴 아래를 지나지만 화면을 넓게 잡아 앰버의 전신과 침상 상당 부분까지 보인다. 창의 실내 조명 반사는 가능한 배치다.",
        "entities": "두 사람만 보인다. 현우는 헝클어진 검은 머리의 젊은 한국계 남성으로 보이며, 참조와 유사한 얼굴·체격, 낡은 짙은 후드와 청바지, 얼굴의 오염과 상처가 확인된다. 금발의 앰버는 약 10세 여자아이로 보이고 참조와 유사한 얼굴과 체격이다. 혼혈 정체성을 외모만으로 확정할 수는 없다. 앰버의 녹색 재킷, 회색 원피스와 갈색 부츠는 참조와 맞지만 마스크는 없다. 눈은 감겨 있어 큰 눈이라는 특징은 확인할 수 없다. 밤의 창밖과 절제된 병실 조명이 보이며, 판독 가능한 글자는 없다.",
        "hard_violations": [],
        "physics": "앰버의 머리와 목은 베개, 몸통과 다리는 매트리스에 지지되어 있다. 보이는 손과 팔도 침상 위에 놓여 있고 부츠를 신은 발도 매트리스에 닿는다. 현우의 발은 프레임 밖이지만 하체가 자연스럽게 아래로 이어져 서 있는 자세로 읽힌다. 부유하는 신체나 소품은 없다."
       },
       {
        "label": "A",
        "direction": "현우는 상체를 살짝 숙이고 오른쪽 아래에 있는 앰버의 얼굴을 내려다본다. 눈썹 사이의 긴장과 다문 입이 걱정스러운 표정을 만든다. 앰버는 눈을 감고 베개에 기대어 있으며 시선 교환은 없다.",
        "built_space": "침대 1개와 가까운 쪽 난간 1개가 전경에 보인다. 난간은 왼쪽 아래에서 두 인물 사이로 대각선을 이루며 두 얼굴을 연결한다. 배경에는 출입문 1개, 의료용 연결 패널 1개, 가로등 1개, 벽 수납장 1개, 침상 옆 수납장 1개, 오른쪽 창문과 커튼이 보인다. 벽의 두 색 도장과 야간 창밖도 참조에 대응한다. 다만 침대 머리맡이 참조보다 창문 쪽 전경으로 옮겨져 후면 의료 패널과의 위치 관계는 덜 충실하다. 현우가 선 통로에는 공간이 있고 난간과 몸이 충돌하지 않는다. 현우의 상체와 앰버의 상반신이 중심인 미디엄 구도다.",
        "entities": "현우와 앰버 두 사람만 있다. 현우는 참조에 가까운 젊은 한국계 남성 얼굴, 헝클어진 검은 머리, 마른 체격과 낡은 후드·청바지를 갖추었고 얼굴의 때와 상처가 보인다. 앰버는 금발의 어린 여자아이이며 참조와 유사한 얼굴이지만 혼혈 여부는 외모만으로 확정할 수 없다. 눈은 감겨 있다. 녹색 재킷과 마스크는 보이지 않고 회색 반소매 옷을 입었으며, 몸에는 담요가 덮여 있다. 팔의 의료용 테이프는 치료 후 상태에 부합한다. 창밖은 밤이고 글자는 판독되지 않는다.",
        "hard_violations": [],
        "physics": "앰버의 머리는 베개에, 몸통은 침상에 완전히 기대어 있다. 보이는 팔은 담요와 매트리스 위에 내려놓여 있고 손은 아래쪽에서 가려진다. 담요도 몸과 침대 표면을 따라 놓인다. 현우는 하체를 유지한 채 허리에서 조금 숙인 자세로, 발은 화면 밖이지만 지지 없는 공중 자세로 보이지 않는다. 난간과 의료용 테이프에도 명확한 물리적 지지가 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.304
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.304
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1304
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "요구된 미디엄 샷 스케일과 침대의 대각선 구도를 정확히 구현했으며, 장소 레퍼런스의 공간 구조를 충실히 재현함."
   },
   {
    "label": "B",
    "score": 1304,
    "verdict_ko": "요구된 미디엄 샷보다 프레임이 너무 넓고 침대가 가로로 배치되어 구도 지침을 실패했으며, 배경의 병실 구조가 왜곡됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L115B02.png",
    "asset_id": "688d6723-0899-4764-9cab-c02e19df9eff",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace77-74e9-70ff-8519-3bc17bcb94fe",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S73sh2__bgfirst_bg.png",
   "bg_asset_id": "3c2d60cc-e5ee-406d-bc22-ac9bab4b356b",
   "bg_record_key": "S73sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S73sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:27:38.003915+00:00",
  "fingerprint": "427db5bbbaa9ee63d3672bfbfb8f739e32f72a6bcc29f75f8d52dd07a9b999a1",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S73sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S73sh2_sel.png",
  "source_sha256": "846774d6c6afc0dd6aaee83618000779c3c334bae7bcd96817a8fbe2f09f2754",
  "file": "S73sh2_cine.png",
  "staged_sha256": "8f97b394d217d9b503daf72ac521545fadce638376a1d9cbb0891ebe1bda499e",
  "latency_ms": 14514
 },
 "S73sh8::signage": {
  "fp": "5426066f37d32cc9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S73sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:29:22.213488+00:00",
  "fingerprint": "c118c088b44bf26075ecfb073751e2f35eb3ef54635126290c835d1dcfc65bac",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S73sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S73sh8_sel.png",
  "source_sha256": "81deff9d30b5e36b10b45d97d1f28123fa2abcc94552618b9d2b97d9ed4eb5e4",
  "file": "S73sh8_cine.png",
  "staged_sha256": "4f90d7efc9a6a4a12569c4c8c76a64626b68afba8f7a697d02ca56f4a1360eee",
  "latency_ms": 13663
 },
 "S74sh3::signage": {
  "fp": "3c982032915629d6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::3d37fa42e862f1ba": {
  "subjects": [],
  "subject_text": "목포 항구마을 거리\n생선 판매대와 길거리 식당이 이어지는 항구마을 거리. 어두운 길 양옆으로 생선이 놓인 좌판과 작은 상점들이 보인다.",
  "identity": "canonical",
  "scope_id": "L116",
  "scope_role": "location_exterior",
  "scope_sha": "ed8b6122bb721244"
 },
 "groupbg::harbor_street_diner": {
  "input_fingerprint": "1473f7c8360ed380",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "harbor_street_diner",
    "tags": [
     "S74sh3",
     "S74sh6"
    ]
   },
   "context_sig": "aedc13247a6936c6"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At an outdoor street-food table in the subdued harbor market at night, among fish traders.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구마을 거리: 어부들이 기형적인 생선을 나르고 모포를 뒤집어쓴 사람들이 오가는 어두컴컴하고 활기 없는 바닷가 길거리. (특징: 플라스틱 바구니에 담긴 돌연변이 생선들; 축축하고 비린내가 밴 항구 골목 바닥)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 이곳은 목포 항구마을 밤. 길거리 식당..\n- 한 생선 장수가 갖다 주는 생선회를 내려다보는 현우.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At an outdoor street-food table in the subdued harbor market at night, among fish traders.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구마을 거리: 어부들이 기형적인 생선을 나르고 모포를 뒤집어쓴 사람들이 오가는 어두컴컴하고 활기 없는 바닷가 길거리. (특징: 플라스틱 바구니에 담긴 돌연변이 생선들; 축축하고 비린내가 밴 항구 골목 바닥)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 이곳은 목포 항구마을 밤. 길거리 식당..\n- 한 생선 장수가 갖다 주는 생선회를 내려다보는 현우.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_harbor_street_diner_053abf.png",
  "asset_id": "b6a644ba-eb88-4ecd-8ecd-14b83870dbc0",
  "input_asset_ids": [
   "c1c3a0ce-fea6-46ce-9e47-2e14fa7238b0"
  ],
  "origin_tag": "S74sh3",
  "place_text": "At an outdoor street-food table in the subdued harbor market at night, among fish traders.",
  "origin_inputs": {
   "place_text": "At an outdoor street-food table in the subdued harbor market at night, among fish traders.",
   "time_of_day_en": "night",
   "conti_asset_id": "c1c3a0ce-fea6-46ce-9e47-2e14fa7238b0"
  }
 },
 "S74sh3::bgfirst_bg": {
  "input_fingerprint": "d2a287b3b61d0bb0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 노상 식당 의자에 마주 앉은 쿠마가 이현우를 향해 씁쓸한 표정으로 고개를 떨군 구도.\n\nLOCATION (lock): At an outdoor street-food table in the subdued harbor market at night, among fish traders.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Street-restaurant table (Between the seated men, before the fish tray arrives) — The near edge is diagonal to the camera, with 쿠마 beyond it; used as Maintains the conversation axis and provides the destination of 쿠마's lowered gaze; Restaurant chairs (Occupied by the two men) — Only portions around the seated bodies are visible; used as Grounds their seated relationship without adding distracting restaurant activity.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime street restaurant preserves facial detail while keeping the exchange emotionally muted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 노상 식당 의자에 마주 앉은 쿠마가 이현우를 향해 씁쓸한 표정으로 고개를 떨군 구도.\n\nLOCATION (lock): At an outdoor street-food table in the subdued harbor market at night, among fish traders.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Street-restaurant table (Between the seated men, before the fish tray arrives) — The near edge is diagonal to the camera, with 쿠마 beyond it; used as Maintains the conversation axis and provides the destination of 쿠마's lowered gaze; Restaurant chairs (Occupied by the two men) — Only portions around the seated bodies are visible; used as Grounds their seated relationship without adding distracting restaurant activity.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime street restaurant preserves facial detail while keeping the exchange emotionally muted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S74sh3__bgfirst_bg.png",
  "asset_id": "4ed7e42b-8f2d-47f4-8075-89a8d0fd5e18",
  "input_asset_ids": [
   "c1c3a0ce-fea6-46ce-9e47-2e14fa7238b0",
   "b6a644ba-eb88-4ecd-8ecd-14b83870dbc0"
  ]
 },
 "S74sh3": {
  "input_fingerprint": "8fcd27b1f768f39e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 노상 식당 의자에 마주 앉은 쿠마가 이현우를 향해 씁쓸한 표정으로 고개를 떨군 구도.\n\nLOCATION (lock): At an outdoor street-food table in the subdued harbor market at night, among fish traders. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Street-restaurant table (Between the seated men, before the fish tray arrives) — The near edge is diagonal to the camera, with 쿠마 beyond it; used as Maintains the conversation axis and provides the destination of 쿠마's lowered gaze; Restaurant chairs (Occupied by the two men) — Only portions around the seated bodies are visible; used as Grounds their seated relationship without adding distracting restaurant activity.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime street restaurant preserves facial detail while keeping the exchange emotionally muted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime harbor street contains a roadside eating place and fish-trading stalls, with little bustle. 이현우: He is at the roadside eating place, still grimy and carrying his accumulated injuries. 쿠마: He is at the roadside eating place.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 노상 식당 의자에 마주 앉은 쿠마가 이현우를 향해 씁쓸한 표정으로 고개를 떨군 구도.\n\nLOCATION (lock): At an outdoor street-food table in the subdued harbor market at night, among fish traders. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Street-restaurant table (Between the seated men, before the fish tray arrives) — The near edge is diagonal to the camera, with 쿠마 beyond it; used as Maintains the conversation axis and provides the destination of 쿠마's lowered gaze; Restaurant chairs (Occupied by the two men) — Only portions around the seated bodies are visible; used as Grounds their seated relationship without adding distracting restaurant activity.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime street restaurant preserves facial detail while keeping the exchange emotionally muted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime harbor street contains a roadside eating place and fish-trading stalls, with little bustle. 이현우: He is at the roadside eating place, still grimy and carrying his accumulated injuries. 쿠마: He is at the roadside eating place.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 노상 식당 의자에 마주 앉은 쿠마가 이현우를 향해 씁쓸한 표정으로 고개를 떨군 구도.\n\nLOCATION (lock): At an outdoor street-food table in the subdued harbor market at night, among fish traders. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Street-restaurant table (Between the seated men, before the fish tray arrives) — The near edge is diagonal to the camera, with 쿠마 beyond it; used as Maintains the conversation axis and provides the destination of 쿠마's lowered gaze; Restaurant chairs (Occupied by the two men) — Only portions around the seated bodies are visible; used as Grounds their seated relationship without adding distracting restaurant activity.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the nighttime street restaurant preserves facial detail while keeping the exchange emotionally muted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime harbor street contains a roadside eating place and fish-trading stalls, with little bustle. 이현우: He is at the roadside eating place, still grimy and carrying his accumulated injuries. 쿠마: He is at the roadside eating place.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 쿠마 (중국계 혼혈 얼굴, 20대 남성, 젊은 성인의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S74sh3__bgfirst_bg.png",
     "asset_id": "4ed7e42b-8f2d-47f4-8075-89a8d0fd5e18",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S74sh3.png",
     "asset_id": "c1c3a0ce-fea6-46ce-9e47-2e14fa7238b0",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904970>",
     "asset_id": "afafdeae-9f05-477d-8a3f-6cedcc1096d3",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_harbor_street_diner_053abf.png",
     "asset_id": "b6a644ba-eb88-4ecd-8ecd-14b83870dbc0",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904970>",
     "asset_id": "afafdeae-9f05-477d-8a3f-6cedcc1096d3",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "쿠마의 시선은 테이블 쪽으로 낮게 향해 있고, 등을 보인 이현우는 쿠마를 향해 마주 앉아 있음.",
    "built_space": "레퍼런스와 동일한 파란색 원형 플라스틱 테이블과 파란색 의자가 배치되었고, 카메라는 지시된 대각선 구도를 취함.",
    "entities": "쿠마는 얼굴과 의상은 일치하나 모자가 없음. 이현우는 뒷모습이 찍혔으나 더러운 검은 후드티를 입고 있음.",
    "hard_violations": [],
    "physics": "두 인물 모두 의자에 체중을 싣고 자연스럽게 앉아 있음."
   },
   {
    "label": "B",
    "direction": "쿠마는 고개를 숙여 앞을 향하고 있고, 이현우는 쿠마의 얼굴을 응시하고 있음.",
    "built_space": "레퍼런스와 전혀 다른 재질과 형태의 긴 직사각형 테이블과 회색 팔걸이 의자가 배치됨.",
    "entities": "쿠마는 모자를 포함해 레퍼런스와 일치함. 이현우는 레퍼런스의 복장과 함께 얼굴의 상처가 묘사됨.",
    "hard_violations": [
     "[gpt-high] 쟁반 도착 전 장면인데 두 사람 사이 식탁 위에 큰 금속 쟁반이 이미 놓여 있다."
    ],
    "physics": "두 인물 모두 의자에 앉아 테이블에 팔을 올리고 지탱하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 카메라 구도(테이블 너머로 고개를 숙인 쿠마)와 장소 레퍼런스(파란색 원형 테이블)를 매우 정확하게 구현했으나, 쿠마의 모자가 누락되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "카메라 구도 지시를 어기고 인물들을 측면으로 배치했으며, 장소 레퍼런스에 없는 직사각형 테이블과 의자, 빈 식판을 임의로 등장시켜 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "쿠마의 시선은 테이블 쪽으로 낮게 향해 있고, 등을 보인 이현우는 쿠마를 향해 마주 앉아 있음.",
        "built_space": "레퍼런스와 동일한 파란색 원형 플라스틱 테이블과 파란색 의자가 배치되었고, 카메라는 지시된 대각선 구도를 취함.",
        "entities": "쿠마는 얼굴과 의상은 일치하나 모자가 없음. 이현우는 뒷모습이 찍혔으나 더러운 검은 후드티를 입고 있음.",
        "hard_violations": [],
        "physics": "두 인물 모두 의자에 체중을 싣고 자연스럽게 앉아 있음."
       },
       {
        "label": "B",
        "direction": "쿠마는 고개를 숙여 앞을 향하고 있고, 이현우는 쿠마의 얼굴을 응시하고 있음.",
        "built_space": "레퍼런스와 전혀 다른 재질과 형태의 긴 직사각형 테이블과 회색 팔걸이 의자가 배치됨.",
        "entities": "쿠마는 모자를 포함해 레퍼런스와 일치함. 이현우는 레퍼런스의 복장과 함께 얼굴의 상처가 묘사됨.",
        "hard_violations": [],
        "physics": "두 인물 모두 의자에 앉아 테이블에 팔을 올리고 지탱하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 카메라 구도(테이블 너머로 고개를 숙인 쿠마)와 장소 레퍼런스(파란색 원형 테이블)를 매우 정확하게 구현했으나, 쿠마의 모자가 누락되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "카메라 구도 지시를 어기고 인물들을 측면으로 배치했으며, 장소 레퍼런스에 없는 직사각형 테이블과 의자, 빈 식판을 임의로 등장시켜 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "쿠마의 시선은 테이블 쪽으로 낮게 향해 있고, 등을 보인 이현우는 쿠마를 향해 마주 앉아 있음.",
        "built_space": "레퍼런스와 동일한 파란색 원형 플라스틱 테이블과 파란색 의자가 배치되었고, 카메라는 지시된 대각선 구도를 취함.",
        "entities": "쿠마는 얼굴과 의상은 일치하나 모자가 없음. 이현우는 뒷모습이 찍혔으나 더러운 검은 후드티를 입고 있음.",
        "hard_violations": [],
        "physics": "두 인물 모두 의자에 체중을 싣고 자연스럽게 앉아 있음."
       },
       {
        "label": "B",
        "direction": "쿠마는 고개를 숙여 앞을 향하고 있고, 이현우는 쿠마의 얼굴을 응시하고 있음.",
        "built_space": "레퍼런스와 전혀 다른 재질과 형태의 긴 직사각형 테이블과 회색 팔걸이 의자가 배치됨.",
        "entities": "쿠마는 모자를 포함해 레퍼런스와 일치함. 이현우는 레퍼런스의 복장과 함께 얼굴의 상처가 묘사됨.",
        "hard_violations": [],
        "physics": "두 인물 모두 의자에 앉아 테이블에 팔을 올리고 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "쿠마의 고개 숙임과 두 사람의 대화 방향은 맞지만, 도착 전이어야 하는 쟁반이 이미 놓였고 장소의 파란 원형 식탁도 다른 식탁으로 바뀌었다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "쿠마의 씁쓸한 고개 숙임을 미디엄 숏으로 담고, 쟁반 없는 원래 식탁과 마주 앉은 관계를 충실히 구현했으나 참조의 모자가 빠졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 쿠마는 이현우 쪽으로 몸을 둔 채 고개와 눈을 식탁 위 쟁반 방향으로 내린다. 오른쪽 이현우는 쿠마의 얼굴을 바라본다. 두 사람의 대화 방향은 맞지만 내려다보는 자리에 아직 없어야 할 쟁반이 있다.",
        "built_space": "직사각형 식탁 하나와 회색 플라스틱 의자 두 개가 보이며 두 사람은 식탁 양쪽에 앉아 있다. 가까운 식탁 가장자리는 대각선이다. 뒤에는 주황색 천막, 매달린 전구 네 개, 수산물 판매대와 상자들, 부두와 어선이 보여 항구 장소는 잘 이어진다. 다만 참조의 파란 원형 식탁과 파란 의자가 교체됐으며, 무릎까지 보이는 구도는 요구한 미디엄 숏보다 넓은 편이다.",
        "entities": "주요 인물은 두 명이다. 이현우는 어린 동아시아계 남성 외모, 헝클어진 검은 머리, 더러운 회색 후드와 얼굴 상처가 보인다. 쿠마는 젊은 성인 남성으로 참조와 유사한 얼굴, 검은 모자, 장식 가죽 조끼와 무늬 셔츠를 갖췄다. 구체적인 민족적 배경은 외모만으로 확정할 수 없다. 식탁 위에는 큰 빈 금속 쟁반 하나가 있어 쟁반 도착 전이라는 조건과 충돌한다.",
        "hard_violations": [
         "쟁반 도착 전 장면인데 두 사람 사이 식탁 위에 큰 금속 쟁반이 이미 놓여 있다."
        ],
        "physics": "두 사람은 각각 의자 좌판에 앉고 등받이가 등 뒤에 있어 지지 관계가 자연스럽다. 이현우의 손은 무릎 부근에 있고 쿠마의 팔은 식탁 가장자리 아래로 내려간다. 금속 쟁반은 식탁 표면에 놓여 지지된다. 공중에 뜬 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "식탁 너머 쿠마는 이현우를 향해 몸을 둔 상태에서 턱을 내리고 두 사람 사이의 식탁을 내려다본다. 찌푸린 미간과 다문 입이 씁쓸한 감정을 전달한다. 전경 이현우의 머리는 쿠마를 향하며, 눈은 카메라에서 가려져 있다.",
        "built_space": "파란 원형 식탁 하나와 각 인물이 차지한 파란 의자 두 개의 일부가 보인다. 이현우는 가까운 왼쪽, 쿠마는 식탁 건너편에 앉는다. 원형 상판의 가까운 테두리가 비스듬히 흐르며 대화 축을 형성한다. 주황색 천막과 전구 네 개, 금속 기둥, 수산물 판매대, 상자, 젖은 부두와 어선이 참조 장소를 유지한다. 쿠마의 상반신을 중심으로 한 미디엄 숏이고 의자는 몸 주변의 일부만 드러난다.",
        "entities": "이현우는 뒤쪽에서 보이는 젊은 남성으로 헝클어진 검은 머리와 때 묻은 회색 후드가 참조에 부합한다. 얼굴은 대부분 가려져 정확한 얼굴 및 상처 상태는 판단하기 어렵다. 쿠마는 참조와 유사한 젊은 성인 남성 얼굴과 장식 가죽 조끼, 다색 무늬 셔츠를 유지하지만 검은 모자가 없다. 민족적 배경 자체는 외모만으로 확정할 수 없다. 식탁에는 참조의 젓가락통, 양념병, 작은 그릇이 있으며 생선 쟁반은 없다. 명확히 읽히는 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 하체와 좌판 접촉점은 식탁과 프레임에 가려졌지만 각자의 등 뒤로 의자 등받이가 보이고 상체 높이와 자세가 착석에 부합한다. 쿠마의 숙인 머리와 처진 어깨는 자연스러운 정지 동작이다. 젓가락은 통에 담겨 있고 통, 병, 그릇은 상판에 안정적으로 놓여 있다. 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "쿠마의 고개 숙임과 두 사람의 대화 방향은 맞지만, 도착 전이어야 하는 쟁반이 이미 놓였고 장소의 파란 원형 식탁도 다른 식탁으로 바뀌었다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "쿠마의 씁쓸한 고개 숙임을 미디엄 숏으로 담고, 쟁반 없는 원래 식탁과 마주 앉은 관계를 충실히 구현했으나 참조의 모자가 빠졌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 쿠마는 이현우 쪽으로 몸을 둔 채 고개와 눈을 식탁 위 쟁반 방향으로 내린다. 오른쪽 이현우는 쿠마의 얼굴을 바라본다. 두 사람의 대화 방향은 맞지만 내려다보는 자리에 아직 없어야 할 쟁반이 있다.",
        "built_space": "직사각형 식탁 하나와 회색 플라스틱 의자 두 개가 보이며 두 사람은 식탁 양쪽에 앉아 있다. 가까운 식탁 가장자리는 대각선이다. 뒤에는 주황색 천막, 매달린 전구 네 개, 수산물 판매대와 상자들, 부두와 어선이 보여 항구 장소는 잘 이어진다. 다만 참조의 파란 원형 식탁과 파란 의자가 교체됐으며, 무릎까지 보이는 구도는 요구한 미디엄 숏보다 넓은 편이다.",
        "entities": "주요 인물은 두 명이다. 이현우는 어린 동아시아계 남성 외모, 헝클어진 검은 머리, 더러운 회색 후드와 얼굴 상처가 보인다. 쿠마는 젊은 성인 남성으로 참조와 유사한 얼굴, 검은 모자, 장식 가죽 조끼와 무늬 셔츠를 갖췄다. 구체적인 민족적 배경은 외모만으로 확정할 수 없다. 식탁 위에는 큰 빈 금속 쟁반 하나가 있어 쟁반 도착 전이라는 조건과 충돌한다.",
        "hard_violations": [
         "쟁반 도착 전 장면인데 두 사람 사이 식탁 위에 큰 금속 쟁반이 이미 놓여 있다."
        ],
        "physics": "두 사람은 각각 의자 좌판에 앉고 등받이가 등 뒤에 있어 지지 관계가 자연스럽다. 이현우의 손은 무릎 부근에 있고 쿠마의 팔은 식탁 가장자리 아래로 내려간다. 금속 쟁반은 식탁 표면에 놓여 지지된다. 공중에 뜬 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "식탁 너머 쿠마는 이현우를 향해 몸을 둔 상태에서 턱을 내리고 두 사람 사이의 식탁을 내려다본다. 찌푸린 미간과 다문 입이 씁쓸한 감정을 전달한다. 전경 이현우의 머리는 쿠마를 향하며, 눈은 카메라에서 가려져 있다.",
        "built_space": "파란 원형 식탁 하나와 각 인물이 차지한 파란 의자 두 개의 일부가 보인다. 이현우는 가까운 왼쪽, 쿠마는 식탁 건너편에 앉는다. 원형 상판의 가까운 테두리가 비스듬히 흐르며 대화 축을 형성한다. 주황색 천막과 전구 네 개, 금속 기둥, 수산물 판매대, 상자, 젖은 부두와 어선이 참조 장소를 유지한다. 쿠마의 상반신을 중심으로 한 미디엄 숏이고 의자는 몸 주변의 일부만 드러난다.",
        "entities": "이현우는 뒤쪽에서 보이는 젊은 남성으로 헝클어진 검은 머리와 때 묻은 회색 후드가 참조에 부합한다. 얼굴은 대부분 가려져 정확한 얼굴 및 상처 상태는 판단하기 어렵다. 쿠마는 참조와 유사한 젊은 성인 남성 얼굴과 장식 가죽 조끼, 다색 무늬 셔츠를 유지하지만 검은 모자가 없다. 민족적 배경 자체는 외모만으로 확정할 수 없다. 식탁에는 참조의 젓가락통, 양념병, 작은 그릇이 있으며 생선 쟁반은 없다. 명확히 읽히는 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 하체와 좌판 접촉점은 식탁과 프레임에 가려졌지만 각자의 등 뒤로 의자 등받이가 보이고 상체 높이와 자세가 착석에 부합한다. 쿠마의 숙인 머리와 처진 어깨는 자연스러운 정지 동작이다. 젓가락은 통에 담겨 있고 통, 병, 그릇은 상판에 안정적으로 놓여 있다. 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.905
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.655
   },
   "violations": {
    "B": [
     "[gpt-high] 쟁반 도착 전 장면인데 두 사람 사이 식탁 위에 큰 금속 쟁반이 이미 놓여 있다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 655
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 카메라 구도(테이블 너머로 고개를 숙인 쿠마)와 장소 레퍼런스(파란색 원형 테이블)를 매우 정확하게 구현했으나, 쿠마의 모자가 누락되었습니다."
   },
   {
    "label": "B",
    "score": 655,
    "verdict_ko": "카메라 구도 지시를 어기고 인물들을 측면으로 배치했으며, 장소 레퍼런스에 없는 직사각형 테이블과 의자, 빈 식판을 임의로 등장시켜 감점되었습니다.  ★위반: [gpt-high] 쟁반 도착 전 장면인데 두 사람 사이 식탁 위에 큰 금속 쟁반이 이미 놓여 있다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_harbor_street_diner_053abf.png",
    "asset_id": "b6a644ba-eb88-4ecd-8ecd-14b83870dbc0",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:904970>",
    "asset_id": "afafdeae-9f05-477d-8a3f-6cedcc1096d3",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace86-03a6-7b61-8990-6b7382be0414",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S74sh3__bgfirst_bg.png",
   "bg_asset_id": "4ed7e42b-8f2d-47f4-8075-89a8d0fd5e18",
   "bg_record_key": "S74sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "harbor_street_diner",
   "groupbg_asset_id": "b6a644ba-eb88-4ecd-8ecd-14b83870dbc0"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S74sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:31:56.805148+00:00",
  "fingerprint": "ad2bf8868c7702bb3f7ccf6fac43aba5d2d93cd63e9a97a265d108c6e5879875",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S74sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S74sh3_sel.png",
  "source_sha256": "453d332e24c14f8608464317ed8c6224831862da85ce6689344db1c899d7ff9d",
  "file": "S74sh3_cine.png",
  "staged_sha256": "8857b13f263e642c406bcd24f57a4a5a7871fc6c7e0ca6001943ed835bf48726",
  "latency_ms": 17247
 },
 "S74sh6::signage": {
  "fp": "49ebaf74500dfdf1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S74sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:32:59.492828+00:00",
  "fingerprint": "ca51a8f10a9c55c5f3c0cf21dfd250a2b09cf7d94269d6e0fa1f4c643f280775",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S74sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S74sh6_sel.png",
  "source_sha256": "d37c4e5d7c8fec65f414a3d04ad8395b740411243169682305897deeee689985",
  "file": "S74sh6_cine.png",
  "staged_sha256": "9e301927db1e08e923500e93f5364505084edf212b3f2410cb505f3afa9961d1",
  "latency_ms": 10127
 },
 "S75sh4::signage": {
  "fp": "43254fdef864e49f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S75sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:35:02.802765+00:00",
  "fingerprint": "3f6728f7f26e3a54799361c74e154a5b79fd985be69c7c3260a18618db03412c",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S75sh4_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S75sh4_sel.png",
  "source_sha256": "5f8acb8cd2c4a92c27e87b67850a700cec1b13c5b135817d18a3d3238b04f0af",
  "file": "S75sh4_cine.png",
  "staged_sha256": "b87f9e614f7c2e7be5e60d8b18991a4a9f418c7ae13183fa881f50b1b0342398",
  "latency_ms": 12336
 },
 "S75sh8::signage": {
  "fp": "df4b3a56325daba2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::9f50c0c873dea96b": {
  "subjects": [],
  "subject_text": "목포공항 격납고 내부\n거대한 창고문 안쪽으로 넓은 실내가 펼쳐지는 낡은 격납고. 오래된 구조물과 실내 전구가 보이는 대형 창고형 공간.",
  "identity": "canonical",
  "scope_id": "L119",
  "scope_role": "location_interior",
  "scope_sha": "17ca1d5681102162"
 },
 "S75sh8::bgfirst_bg": {
  "input_fingerprint": "7612e861c58de656",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 입을 벌린 채 이현우와 쿠마가 있는 창고 문 앞을 향해 맹렬히 뛰어오는 도중 한 발이 지면에서 막 떨어진 신부의 mid-action 찰나.\n\nLOCATION (lock): On the outdoor approach to the open hangar entrance at the abandoned airfield at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open warehouse entrance, destination of the priest's left-to-right approach in the middle-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Warehouse entrance (Open, with 이현우 and 쿠마 at its edge) — Seen obliquely at screen right from just outside; used as The priest's fixed destination and the spatial anchor for the following pan; Airport approach ground (The visible running route before the doorway) — Extends beneath the priest toward the entrance; used as Makes the lifted foot and remaining distance to the men readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient light keeps the priest's open mouth and lifted foot legible without relocating the exterior campfire into the doorway.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 입을 벌린 채 이현우와 쿠마가 있는 창고 문 앞을 향해 맹렬히 뛰어오는 도중 한 발이 지면에서 막 떨어진 신부의 mid-action 찰나.\n\nLOCATION (lock): On the outdoor approach to the open hangar entrance at the abandoned airfield at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open warehouse entrance, destination of the priest's left-to-right approach in the middle-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Warehouse entrance (Open, with 이현우 and 쿠마 at its edge) — Seen obliquely at screen right from just outside; used as The priest's fixed destination and the spatial anchor for the following pan; Airport approach ground (The visible running route before the doorway) — Extends beneath the priest toward the entrance; used as Makes the lifted foot and remaining distance to the men readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient light keeps the priest's open mouth and lifted foot legible without relocating the exterior campfire into the doorway.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S75sh8__bgfirst_bg.png",
  "asset_id": "961f7105-7d6a-46c5-9e65-9e35853740a1",
  "input_asset_ids": [
   "7a02bf8d-e147-40f5-b761-d37c4a6591eb",
   "031a67fa-32c2-421a-8088-c2d5cf845e26"
  ]
 },
 "S75sh8": {
  "input_fingerprint": "55175e16bd14b5d3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 입을 벌린 채 이현우와 쿠마가 있는 창고 문 앞을 향해 맹렬히 뛰어오는 도중 한 발이 지면에서 막 떨어진 신부의 mid-action 찰나.\n\nLOCATION (lock): On the outdoor approach to the open hangar entrance at the abandoned airfield at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open warehouse entrance, destination of the priest's left-to-right approach in the middle-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Warehouse entrance (Open, with 이현우 and 쿠마 at its edge) — Seen obliquely at screen right from just outside; used as The priest's fixed destination and the spatial anchor for the following pan; Airport approach ground (The visible running route before the doorway) — Extends beneath the priest toward the entrance; used as Makes the lifted foot and remaining distance to the men readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient light keeps the priest's open mouth and lifted foot legible without relocating the exterior campfire into the doorway.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hangar remains open with the inoperative light aircraft inside. The campfire continues burning in the derelict airport grounds. 신부: He has arrived at the airport urgently.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 입을 벌린 채 이현우와 쿠마가 있는 창고 문 앞을 향해 맹렬히 뛰어오는 도중 한 발이 지면에서 막 떨어진 신부의 mid-action 찰나.\n\nLOCATION (lock): On the outdoor approach to the open hangar entrance at the abandoned airfield at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open warehouse entrance, destination of the priest's left-to-right approach in the middle-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Warehouse entrance (Open, with 이현우 and 쿠마 at its edge) — Seen obliquely at screen right from just outside; used as The priest's fixed destination and the spatial anchor for the following pan; Airport approach ground (The visible running route before the doorway) — Extends beneath the priest toward the entrance; used as Makes the lifted foot and remaining distance to the men readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient light keeps the priest's open mouth and lifted foot legible without relocating the exterior campfire into the doorway.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hangar remains open with the inoperative light aircraft inside. The campfire continues burning in the derelict airport grounds. 신부: He has arrived at the airport urgently.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 입을 벌린 채 이현우와 쿠마가 있는 창고 문 앞을 향해 맹렬히 뛰어오는 도중 한 발이 지면에서 막 떨어진 신부의 mid-action 찰나.\n\nLOCATION (lock): On the outdoor approach to the open hangar entrance at the abandoned airfield at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open warehouse entrance, destination of the priest's left-to-right approach in the middle-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Warehouse entrance (Open, with 이현우 and 쿠마 at its edge) — Seen obliquely at screen right from just outside; used as The priest's fixed destination and the spatial anchor for the following pan; Airport approach ground (The visible running route before the doorway) — Extends beneath the priest toward the entrance; used as Makes the lifted foot and remaining distance to the men readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient light keeps the priest's open mouth and lifted foot legible without relocating the exterior campfire into the doorway.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hangar remains open with the inoperative light aircraft inside. The campfire continues burning in the derelict airport grounds. 신부: He has arrived at the airport urgently.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (60대 남성, 노년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S75sh8__bgfirst_bg.png",
     "asset_id": "961f7105-7d6a-46c5-9e65-9e35853740a1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S75sh8.png",
     "asset_id": "7a02bf8d-e147-40f5-b761-d37c4a6591eb",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:672084>",
     "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L119B03.png",
     "asset_id": "031a67fa-32c2-421a-8088-c2d5cf845e26",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:672084>",
     "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "우측의 남성들을 향해 달리고 있으며, 입을 벌린 시선 방향이 명확함.",
    "built_space": "격납고 배경이 보이나 원본에 없는 물웅덩이가 바닥에 추가됨. 두 남성이 입구가 아닌 전경에 지나치게 가깝게 배치됨.",
    "entities": "신부의 외모는 레퍼런스와 일치하나, 우측 두 남성이 실사가 아닌 2D 애니메이션 스타일로 렌더링됨.",
    "hard_violations": [
     "[gemini-pro] 2D 일러스트 인물이 합성된 형태 (collage/panels 위반)"
    ],
    "physics": "오른발이 바닥에 닿아 체중을 지탱하고 있으며 물을 튀기고 있음."
   },
   {
    "label": "B",
    "direction": "격납고 입구에 서 있는 두 남성을 향해 달려가고 있으며, 입을 크게 벌리고 있음.",
    "built_space": "레퍼런스와 일치하는 격납고와 바닥 재질. 지시된 대로 입구 경계에 두 남성이 서 있고 배경 좌측에 모닥불이 위치함.",
    "entities": "신부의 외모 및 복장이 레퍼런스와 정확히 일치하며, 모든 인물과 배경이 실사로 렌더링됨.",
    "hard_violations": [
     "[gpt-high] 출입구 왼쪽 문기둥의 숫자 ‘8’이 판독 가능하여, 읽을 수 있는 문자를 어디에도 두지 말라는 조건을 위반한다."
    ],
    "physics": "달리는 도중 양발이 공중에 뜬 도약 상태이며, 모멘텀을 가진 자연스러운 자세를 보여줌."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "우측 인물들이 2D 일러스트로 렌더링되어 실사 촬영 조건을 심각하게 위반함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스의 배경과 인물을 정확히 반영하고, 지시된 달리기 액션과 구도를 사실적으로 완벽히 구현함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "우측의 남성들을 향해 달리고 있으며, 입을 벌린 시선 방향이 명확함.",
        "built_space": "격납고 배경이 보이나 원본에 없는 물웅덩이가 바닥에 추가됨. 두 남성이 입구가 아닌 전경에 지나치게 가깝게 배치됨.",
        "entities": "신부의 외모는 레퍼런스와 일치하나, 우측 두 남성이 실사가 아닌 2D 애니메이션 스타일로 렌더링됨.",
        "hard_violations": [
         "2D 일러스트 인물이 합성된 형태 (collage/panels 위반)"
        ],
        "physics": "오른발이 바닥에 닿아 체중을 지탱하고 있으며 물을 튀기고 있음."
       },
       {
        "label": "B",
        "direction": "격납고 입구에 서 있는 두 남성을 향해 달려가고 있으며, 입을 크게 벌리고 있음.",
        "built_space": "레퍼런스와 일치하는 격납고와 바닥 재질. 지시된 대로 입구 경계에 두 남성이 서 있고 배경 좌측에 모닥불이 위치함.",
        "entities": "신부의 외모 및 복장이 레퍼런스와 정확히 일치하며, 모든 인물과 배경이 실사로 렌더링됨.",
        "hard_violations": [],
        "physics": "달리는 도중 양발이 공중에 뜬 도약 상태이며, 모멘텀을 가진 자연스러운 자세를 보여줌."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "우측 인물들이 2D 일러스트로 렌더링되어 실사 촬영 조건을 심각하게 위반함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스의 배경과 인물을 정확히 반영하고, 지시된 달리기 액션과 구도를 사실적으로 완벽히 구현함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "우측의 남성들을 향해 달리고 있으며, 입을 벌린 시선 방향이 명확함.",
        "built_space": "격납고 배경이 보이나 원본에 없는 물웅덩이가 바닥에 추가됨. 두 남성이 입구가 아닌 전경에 지나치게 가깝게 배치됨.",
        "entities": "신부의 외모는 레퍼런스와 일치하나, 우측 두 남성이 실사가 아닌 2D 애니메이션 스타일로 렌더링됨.",
        "hard_violations": [
         "2D 일러스트 인물이 합성된 형태 (collage/panels 위반)"
        ],
        "physics": "오른발이 바닥에 닿아 체중을 지탱하고 있으며 물을 튀기고 있음."
       },
       {
        "label": "B",
        "direction": "격납고 입구에 서 있는 두 남성을 향해 달려가고 있으며, 입을 크게 벌리고 있음.",
        "built_space": "레퍼런스와 일치하는 격납고와 바닥 재질. 지시된 대로 입구 경계에 두 남성이 서 있고 배경 좌측에 모닥불이 위치함.",
        "entities": "신부의 외모 및 복장이 레퍼런스와 정확히 일치하며, 모든 인물과 배경이 실사로 렌더링됨.",
        "hard_violations": [],
        "physics": "달리는 도중 양발이 공중에 뜬 도약 상태이며, 모멘텀을 가진 자연스러운 자세를 보여줌."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "왼쪽에서 오른쪽 출입구로 달리는 와이드 구도와 신부의 외형은 더 충실하지만, 문기둥의 숫자 ‘8’이 선명하게 읽혀 무문자 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "입을 벌리고 두 남자를 향해 달리는 동작과 이륙 직후의 발을 설득력 있게 보여주지만, 전경 남자들이 커진 어깨너머 구도와 신부의 얼굴은 지정 구도·참조에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "신부의 얼굴, 시선과 몸의 진행 방향이 화면 오른쪽 열린 격납고 가장자리의 두 남자를 향한다. 두 남자도 왼쪽 신부를 바라본다. 왼쪽에서 오른쪽으로 접근하는 동선과 목적지가 명확하다.",
        "built_space": "오른쪽에 큰 개방 출입구 하나가 있고, 두 남자는 그 문턱 안쪽 가장자리에 서 있다. 녹슨 밝은 골판금 외벽, 철제 지붕 보, 외벽등 두 개와 내부 매달린 등 하나가 보인다. 내부에는 경비행기 한 대와 뒤쪽의 노출된 기체 골조가 보이며, 야외 모닥불 하나는 왼쪽 먼 지면에 남아 있다. 콘크리트 접근로가 신부 아래에서 문턱까지 이어져 거리와 이동 방향이 잘 읽힌다.",
        "entities": "회색 짧은 머리와 밝은 피부의 노년 남성 신부는 참조의 얼굴·체형에 비교적 가깝고, 긴 검은 수단과 흰 로만 칼라, 검은 구두도 맞는다. 입이 열려 있다. 문가에는 갈색 재킷과 회색 재킷을 입은 젊은 성인 남성 두 명이 있으며, 이현우와 쿠마의 위치에 해당하지만 개별 신원은 확인할 참조가 없다. 정지한 낡은 경비행기와 외부 모닥불은 요구된 대상이다. 문기둥에는 숫자 ‘8’이 뚜렷하게 읽힌다.",
        "hard_violations": [
         "출입구 왼쪽 문기둥의 숫자 ‘8’이 판독 가능하여, 읽을 수 있는 문자를 어디에도 두지 말라는 조건을 위반한다."
        ],
        "physics": "신부의 두 구두는 모두 지면에서 떨어져 보인다. 뒤로 뻗은 다리, 앞으로 나온 반대 다리와 팔의 교차 동작은 달리기의 짧은 공중 국면으로 가능하며, 앞발이 내려설 콘크리트도 바로 아래에 있다. 따라서 무근거한 부유는 아니지만, 한 발이 막 떨어지는 순간보다는 이미 보폭 중간에 들어선 모습이다. 두 남자는 발로 바닥을 딛고 있고 비행기는 착륙장치로 지지된다."
       },
       {
        "label": "B",
        "direction": "신부는 화면 왼쪽에서 오른쪽 전경의 두 남자를 향해 비스듬히 달려오며, 열린 입과 시선도 그들에게 향한다. 두 남자는 고개와 상체를 신부 쪽으로 돌리고 있다. 목적지는 맞지만 이동이 순수한 좌우 횡단보다는 카메라 쪽으로 다가오는 대각선으로 강조된다.",
        "built_space": "오른쪽에 열린 격납고 출입구 하나가 있으며, 두 남자는 가까운 문 가장자리 전경에 서 있다. 녹슨 골판금 외벽, 철제 지붕 보, 외벽등 두 개, 내부 천장등 두 개, 선반 여러 개가 보인다. 내부에는 앞뒤로 경비행기 두 대가 있고, 모닥불 하나는 왼쪽 먼 야외에 있다. 접근로와 문턱은 연결되지만 두 남자의 큰 상반신이 오른쪽을 차지해, 출입구 자체를 공간 기준점으로 삼는 와이드 구도보다 어깨너머 구도에 가까워졌다. 젖은 바닥의 조명 반사는 가능한 방향이다.",
        "entities": "신부는 짧은 회색 머리의 노년 남성이며 검은 수단, 흰 로만 칼라와 검은 구두를 착용하고 입을 크게 벌린다. 다만 얼굴의 윤곽과 이목구비는 참조의 신부와 차이가 나고 동아시아계 인상으로 보인다. 전경에는 검은 머리의 남성과 삭발에 가까운 남성, 총 두 명이 있으며 둘 다 어두운 옷을 입었다. 이현우와 쿠마의 자리에는 맞지만 얼굴이 충분히 보이지 않아 개별 정체성은 검증하기 어렵다. 경비행기와 야외 모닥불이 존재하고 판독 가능한 문자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "신부의 뒤쪽 다리는 무릎이 굽혀져 들려 있고 앞쪽 구두도 젖은 지면 바로 위에 있다. 구두 아래 물 튀김과 달리기 자세가 직전 지면 접촉 및 도약을 뒷받침하며, 착지할 접근로도 확보되어 있다. 한 발이 지면을 막 벗어난 달리기 국면으로 읽을 수 있다. 두 남자의 발은 화면 밖이지만 몸이 공중에 떠 있다는 징후는 없으며, 비행기들은 바퀴로 바닥에 지지된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "왼쪽에서 오른쪽 출입구로 달리는 와이드 구도와 신부의 외형은 더 충실하지만, 문기둥의 숫자 ‘8’이 선명하게 읽혀 무문자 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "입을 벌리고 두 남자를 향해 달리는 동작과 이륙 직후의 발을 설득력 있게 보여주지만, 전경 남자들이 커진 어깨너머 구도와 신부의 얼굴은 지정 구도·참조에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "신부의 얼굴, 시선과 몸의 진행 방향이 화면 오른쪽 열린 격납고 가장자리의 두 남자를 향한다. 두 남자도 왼쪽 신부를 바라본다. 왼쪽에서 오른쪽으로 접근하는 동선과 목적지가 명확하다.",
        "built_space": "오른쪽에 큰 개방 출입구 하나가 있고, 두 남자는 그 문턱 안쪽 가장자리에 서 있다. 녹슨 밝은 골판금 외벽, 철제 지붕 보, 외벽등 두 개와 내부 매달린 등 하나가 보인다. 내부에는 경비행기 한 대와 뒤쪽의 노출된 기체 골조가 보이며, 야외 모닥불 하나는 왼쪽 먼 지면에 남아 있다. 콘크리트 접근로가 신부 아래에서 문턱까지 이어져 거리와 이동 방향이 잘 읽힌다.",
        "entities": "회색 짧은 머리와 밝은 피부의 노년 남성 신부는 참조의 얼굴·체형에 비교적 가깝고, 긴 검은 수단과 흰 로만 칼라, 검은 구두도 맞는다. 입이 열려 있다. 문가에는 갈색 재킷과 회색 재킷을 입은 젊은 성인 남성 두 명이 있으며, 이현우와 쿠마의 위치에 해당하지만 개별 신원은 확인할 참조가 없다. 정지한 낡은 경비행기와 외부 모닥불은 요구된 대상이다. 문기둥에는 숫자 ‘8’이 뚜렷하게 읽힌다.",
        "hard_violations": [
         "출입구 왼쪽 문기둥의 숫자 ‘8’이 판독 가능하여, 읽을 수 있는 문자를 어디에도 두지 말라는 조건을 위반한다."
        ],
        "physics": "신부의 두 구두는 모두 지면에서 떨어져 보인다. 뒤로 뻗은 다리, 앞으로 나온 반대 다리와 팔의 교차 동작은 달리기의 짧은 공중 국면으로 가능하며, 앞발이 내려설 콘크리트도 바로 아래에 있다. 따라서 무근거한 부유는 아니지만, 한 발이 막 떨어지는 순간보다는 이미 보폭 중간에 들어선 모습이다. 두 남자는 발로 바닥을 딛고 있고 비행기는 착륙장치로 지지된다."
       },
       {
        "label": "A",
        "direction": "신부는 화면 왼쪽에서 오른쪽 전경의 두 남자를 향해 비스듬히 달려오며, 열린 입과 시선도 그들에게 향한다. 두 남자는 고개와 상체를 신부 쪽으로 돌리고 있다. 목적지는 맞지만 이동이 순수한 좌우 횡단보다는 카메라 쪽으로 다가오는 대각선으로 강조된다.",
        "built_space": "오른쪽에 열린 격납고 출입구 하나가 있으며, 두 남자는 가까운 문 가장자리 전경에 서 있다. 녹슨 골판금 외벽, 철제 지붕 보, 외벽등 두 개, 내부 천장등 두 개, 선반 여러 개가 보인다. 내부에는 앞뒤로 경비행기 두 대가 있고, 모닥불 하나는 왼쪽 먼 야외에 있다. 접근로와 문턱은 연결되지만 두 남자의 큰 상반신이 오른쪽을 차지해, 출입구 자체를 공간 기준점으로 삼는 와이드 구도보다 어깨너머 구도에 가까워졌다. 젖은 바닥의 조명 반사는 가능한 방향이다.",
        "entities": "신부는 짧은 회색 머리의 노년 남성이며 검은 수단, 흰 로만 칼라와 검은 구두를 착용하고 입을 크게 벌린다. 다만 얼굴의 윤곽과 이목구비는 참조의 신부와 차이가 나고 동아시아계 인상으로 보인다. 전경에는 검은 머리의 남성과 삭발에 가까운 남성, 총 두 명이 있으며 둘 다 어두운 옷을 입었다. 이현우와 쿠마의 자리에는 맞지만 얼굴이 충분히 보이지 않아 개별 정체성은 검증하기 어렵다. 경비행기와 야외 모닥불이 존재하고 판독 가능한 문자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "신부의 뒤쪽 다리는 무릎이 굽혀져 들려 있고 앞쪽 구두도 젖은 지면 바로 위에 있다. 구두 아래 물 튀김과 달리기 자세가 직전 지면 접촉 및 도약을 뒷받침하며, 착지할 접근로도 확보되어 있다. 한 발이 지면을 막 벗어난 달리기 국면으로 읽을 수 있다. 두 남자의 발은 화면 밖이지만 몸이 공중에 떠 있다는 징후는 없으며, 비행기들은 바퀴로 바닥에 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.321
   },
   "violations": {
    "A": [
     "[gemini-pro] 2D 일러스트 인물이 합성된 형태 (collage/panels 위반)"
    ],
    "B": [
     "[gpt-high] 출입구 왼쪽 문기둥의 숫자 ‘8’이 판독 가능하여, 읽을 수 있는 문자를 어디에도 두지 말라는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "A": 1179,
   "B": 1321
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "우측 인물들이 2D 일러스트로 렌더링되어 실사 촬영 조건을 심각하게 위반함.  ★위반: [gemini-pro] 2D 일러스트 인물이 합성된 형태 (collage/panels 위반)"
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "레퍼런스의 배경과 인물을 정확히 반영하고, 지시된 달리기 액션과 구도를 사실적으로 완벽히 구현함.  ★위반: [gpt-high] 출입구 왼쪽 문기둥의 숫자 ‘8’이 판독 가능하여, 읽을 수 있는 문자를 어디에도 두지 말라는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L119B03.png",
    "asset_id": "031a67fa-32c2-421a-8088-c2d5cf845e26",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:672084>",
    "asset_id": "3e7519d8-1bbf-4a2f-aefd-8ec09bfa4a67",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aace9b-37f2-7bfd-a0e2-1ef3268ac6e1",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S75sh8__bgfirst_bg.png",
   "bg_asset_id": "961f7105-7d6a-46c5-9e65-9e35853740a1",
   "bg_record_key": "S75sh8::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S75sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:36:59.458649+00:00",
  "fingerprint": "b044930b1e4ca78b56dff3987a89e698e6cd337ffad2d77ee6180e3bcd2e1b76",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S75sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S75sh8_sel.png",
  "source_sha256": "887dbfa704a94d960e2436a9da99c2fcd666730b8064f1eccc90c1cfaeba87ef",
  "file": "S75sh8_cine.png",
  "staged_sha256": "61c77261d9eb7367e76f0f3db830998ddaf2d714425f54225027cb1814c0c5f2",
  "latency_ms": 31584
 },
 "S76sh4::signage": {
  "fp": "6eed1913eaba95e9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S76sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:38:26.317875+00:00",
  "fingerprint": "be5545996d65b5d4b9268f5714c8dee8856e0670846e3b8ba4735c28e4e18e7c",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S76sh4_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S76sh4_sel.png",
  "source_sha256": "470ba46adf280c98a3e58f1edfbc8d6b6b0f10dba3adfc789588913302f2cbfd",
  "file": "S76sh4_cine.png",
  "staged_sha256": "8d24e2840adec1831b55ff0b13354520535581402c4bf280aca5a3742cdeb634",
  "latency_ms": 10609
 },
 "S76sh9::signage": {
  "fp": "cc8a772b9928363f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S76sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:39:40.573728+00:00",
  "fingerprint": "c868f67b6971e22443d588a1b2b44d77de77d4ac757e73ad4832c5ca7da8cd01",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S76sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S76sh9_sel.png",
  "source_sha256": "c4bcd1e0fa29485cd36a53a2bd3357122cbfa8423c90529c9ba008b73ba55ea2",
  "file": "S76sh9_cine.png",
  "staged_sha256": "6e32860186dedd7785ab6c79c3b40a93ed68a64d796f9367e794d59c36aece1b",
  "latency_ms": 14576
 },
 "S77sh18::signage": {
  "fp": "14c0678a5cb32524",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S77sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:40:41.823366+00:00",
  "fingerprint": "ec3cba9253b34b5ca05016e0c1ffedeb0dd3019b95e581e5de61df9d9e3d57e6",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S77sh18_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S77sh18_sel.png",
  "source_sha256": "d6b176d251b476557ce42d44642553ac3019198a2945ebb6b972bca159e63288",
  "file": "S77sh18_cine.png",
  "staged_sha256": "84317c6ddb6e65c5e5d8a869d3ebd39071d08e3ad783db38f2e7bd57345ce770",
  "latency_ms": 10504
 },
 "S77sh43::signage": {
  "fp": "d9631091faed2fe9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S77sh43::bgfirst_bg": {
  "input_fingerprint": "201734018ab4b455",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 비행하는 헬기 안, 앰버가 조심스레 펼친 두 손바닥 위에 보라색 제비꽃 한 송이가 놓여있는 클로즈업.\n\nLOCATION (lock): Inside the passenger cabin of the departing helicopter, at the girl's seat with daylight entering through the windows.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Single violet (Resting on 앰버's open palms) — The blossom is visible obliquely from above; used as Small central emotional anchor, scaled against both hands; Helicopter interior (Occupied during flight) — Only a peripheral portion of the cabin is visible; used as Soft contextual margin around the hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft daytime ambient light preserves the delicate violet color and the gentle tonal separation of her hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 비행하는 헬기 안, 앰버가 조심스레 펼친 두 손바닥 위에 보라색 제비꽃 한 송이가 놓여있는 클로즈업.\n\nLOCATION (lock): Inside the passenger cabin of the departing helicopter, at the girl's seat with daylight entering through the windows.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Single violet (Resting on 앰버's open palms) — The blossom is visible obliquely from above; used as Small central emotional anchor, scaled against both hands; Helicopter interior (Occupied during flight) — Only a peripheral portion of the cabin is visible; used as Soft contextual margin around the hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft daytime ambient light preserves the delicate violet color and the gentle tonal separation of her hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S77sh43__bgfirst_bg.png",
  "asset_id": "144b05c8-c592-4762-b6d3-0f958e2255dc",
  "input_asset_ids": [
   "18f1a580-0086-4aa6-8980-51c965050cbf",
   "d623b34c-4206-41a2-83ad-a7729f0c1edc"
  ]
 },
 "S77sh43": {
  "input_fingerprint": "c82589c308e2b174",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 비행하는 헬기 안, 앰버가 조심스레 펼친 두 손바닥 위에 보라색 제비꽃 한 송이가 놓여있는 클로즈업.\n\nLOCATION (lock): Inside the passenger cabin of the departing helicopter, at the girl's seat with daylight entering through the windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Single violet (Resting on 앰버's open palms) — The blossom is visible obliquely from above; used as Small central emotional anchor, scaled against both hands; Helicopter interior (Occupied during flight) — Only a peripheral portion of the cabin is visible; used as Soft contextual margin around the hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft daytime ambient light preserves the delicate violet color and the gentle tonal separation of her hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The helicopter is airborne in daylight, departing the airport. 앰버: She is aboard the helicopter, holding and examining a violet flower. She remains seriously ill after emergency treatment and has not yet received the required surgery.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버 right now, so 앰버's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 비행하는 헬기 안, 앰버가 조심스레 펼친 두 손바닥 위에 보라색 제비꽃 한 송이가 놓여있는 클로즈업.\n\nLOCATION (lock): Inside the passenger cabin of the departing helicopter, at the girl's seat with daylight entering through the windows. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Single violet (Resting on 앰버's open palms) — The blossom is visible obliquely from above; used as Small central emotional anchor, scaled against both hands; Helicopter interior (Occupied during flight) — Only a peripheral portion of the cabin is visible; used as Soft contextual margin around the hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft daytime ambient light preserves the delicate violet color and the gentle tonal separation of her hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The helicopter is airborne in daylight, departing the airport. 앰버: She is aboard the helicopter, holding and examining a violet flower. She remains seriously ill after emergency treatment and has not yet received the required surgery.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버 right now, so 앰버's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 비행하는 헬기 안, 앰버가 조심스레 펼친 두 손바닥 위에 보라색 제비꽃 한 송이가 놓여있는 클로즈업.\n\nLOCATION (lock): Inside the passenger cabin of the departing helicopter, at the girl's seat with daylight entering through the windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Single violet (Resting on 앰버's open palms) — The blossom is visible obliquely from above; used as Small central emotional anchor, scaled against both hands; Helicopter interior (Occupied during flight) — Only a peripheral portion of the cabin is visible; used as Soft contextual margin around the hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft daytime ambient light preserves the delicate violet color and the gentle tonal separation of her hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The helicopter is airborne in daylight, departing the airport. 앰버: She is aboard the helicopter, holding and examining a violet flower. She remains seriously ill after emergency treatment and has not yet received the required surgery.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버 right now, so 앰버's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S77sh43__bgfirst_bg.png",
     "asset_id": "144b05c8-c592-4762-b6d3-0f958e2255dc",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S77sh43.png",
     "asset_id": "18f1a580-0086-4aa6-8980-51c965050cbf",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L119B01.png",
     "asset_id": "d623b34c-4206-41a2-83ad-a7729f0c1edc",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:844202>",
     "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라 시점은 1인칭(POV)으로 펼친 두 손바닥과 제비꽃을 직접 향함.",
    "built_space": "헬기 내부. 좌측에 안전벨트가 있는 빈 좌석, 우측 창문 밖으로 야간 도시 불빛이 보임.",
    "entities": "초록색 재킷 소매를 입은 두 손, 보라색 제비꽃 한 송이.",
    "hard_violations": [],
    "physics": "제비꽃이 두 손바닥 위에 안정적으로 놓여 있으며, 손은 프레임 밖의 팔에 의해 지지됨."
   },
   {
    "label": "B",
    "direction": "인물의 시선은 손바닥 위 제비꽃을 향하며, 카메라는 측면에서 손과 인물을 담음.",
    "built_space": "헬기 내부. 좌측에 좌석, 우측 창문 밖으로 야간 도시 불빛이 보임.",
    "entities": "금발 머리와 마스크를 쓴 앰버, 초록색 재킷, 제비꽃, 레퍼런스에 없는 회색 손목 밴드.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스와 프롬프트에 없는 임의의 사물(왼팔의 회색 손목 밴드) 생성"
    ],
    "physics": "제비꽃이 손바닥에 놓여 있고 팔이 양손을 지지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 손과 제비꽃의 클로즈업 구도를 정확히 구현하였고, 레퍼런스의 의상 특징과 야간 배경을 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "클로즈업 지시를 어기고 인물의 측면 상반신을 넓게 포함시켰으며, 레퍼런스에 없는 손목 밴드가 임의로 생성되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라 시점은 1인칭(POV)으로 펼친 두 손바닥과 제비꽃을 직접 향함.",
        "built_space": "헬기 내부. 좌측에 안전벨트가 있는 빈 좌석, 우측 창문 밖으로 야간 도시 불빛이 보임.",
        "entities": "초록색 재킷 소매를 입은 두 손, 보라색 제비꽃 한 송이.",
        "hard_violations": [],
        "physics": "제비꽃이 두 손바닥 위에 안정적으로 놓여 있으며, 손은 프레임 밖의 팔에 의해 지지됨."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 손바닥 위 제비꽃을 향하며, 카메라는 측면에서 손과 인물을 담음.",
        "built_space": "헬기 내부. 좌측에 좌석, 우측 창문 밖으로 야간 도시 불빛이 보임.",
        "entities": "금발 머리와 마스크를 쓴 앰버, 초록색 재킷, 제비꽃, 레퍼런스에 없는 회색 손목 밴드.",
        "hard_violations": [
         "레퍼런스와 프롬프트에 없는 임의의 사물(왼팔의 회색 손목 밴드) 생성"
        ],
        "physics": "제비꽃이 손바닥에 놓여 있고 팔이 양손을 지지함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 손과 제비꽃의 클로즈업 구도를 정확히 구현하였고, 레퍼런스의 의상 특징과 야간 배경을 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "클로즈업 지시를 어기고 인물의 측면 상반신을 넓게 포함시켰으며, 레퍼런스에 없는 손목 밴드가 임의로 생성되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라 시점은 1인칭(POV)으로 펼친 두 손바닥과 제비꽃을 직접 향함.",
        "built_space": "헬기 내부. 좌측에 안전벨트가 있는 빈 좌석, 우측 창문 밖으로 야간 도시 불빛이 보임.",
        "entities": "초록색 재킷 소매를 입은 두 손, 보라색 제비꽃 한 송이.",
        "hard_violations": [],
        "physics": "제비꽃이 두 손바닥 위에 안정적으로 놓여 있으며, 손은 프레임 밖의 팔에 의해 지지됨."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 손바닥 위 제비꽃을 향하며, 카메라는 측면에서 손과 인물을 담음.",
        "built_space": "헬기 내부. 좌측에 좌석, 우측 창문 밖으로 야간 도시 불빛이 보임.",
        "entities": "금발 머리와 마스크를 쓴 앰버, 초록색 재킷, 제비꽃, 레퍼런스에 없는 회색 손목 밴드.",
        "hard_violations": [
         "레퍼런스와 프롬프트에 없는 임의의 사물(왼팔의 회색 손목 밴드) 생성"
        ],
        "physics": "제비꽃이 손바닥에 놓여 있고 팔이 양손을 지지함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "두 손바닥에 놓인 제비꽃과 내려다보는 행동은 맞지만, 얼굴·상체와 넓은 객실이 들어와 손 중심 클로즈업 및 주변부로 제한한 배경 지시에서 멀어진다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 손과 그 위의 제비꽃을 크게 담아 지정된 클로즈업에 더 충실하고 객실 재질도 잘 맞지만, 배경 비중이 여전히 크고 손의 아동성은 뚜렷하지 않다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 손바닥이 위로 펼쳐져 중앙의 꽃을 받치고, 꽃의 앞면은 카메라 쪽 위를 향해 비스듬히 보인다. 오른쪽 위에 보이는 아이의 얼굴은 손과 꽃을 향해 숙여져 있으나 눈동자의 정확한 방향은 흐려 확인하기 어렵다. 창밖 불빛만으로 헬기의 진행 방향은 알 수 없다.",
        "built_space": "왼쪽에 연속된 천 벤치 한 줄과 안전벨트 버클 세 개, 오른쪽에 큰 창 하나와 위쪽에 잘린 작은 창 일부, 리벳이 있는 금속 벽체가 보인다. 고정 설비의 배치와 재질은 장소 참조에 부합한다. 아이는 오른쪽 전경에 있으나 실제 착석 접점은 화면 밖이다. 불가능한 반사나 중복 설비는 보이지 않는다. 다만 벤치와 창이 넓고 비교적 선명하게 보여 객실을 부드러운 주변부로만 쓰라는 지시보다 강조된다.",
        "entities": "인물은 한 명이며 금발, 흰 마스크, 녹색 계열 재킷이 참조와 대체로 맞는다. 얼굴이 부분적으로 흐리게 보여 정확한 나이와 한국계·백인 혼혈 정체성은 확인하기 어렵지만 어린 여자아이로 읽힌다. 양손 위에는 보라색 제비꽃 한 송이가 있다. 왼쪽 손목의 회색 밴드는 참조에 없는 복장 차이다. 창밖은 야간으로 보이며, 손의 조명은 요청된 부드러운 주간광보다 어둡고 차갑다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "꽃은 서로 모은 손바닥에 접촉해 받쳐져 있다. 손목과 팔은 소매를 통해 인물의 몸으로 이어지고, 손을 살짝 오므려 작은 꽃을 조심스럽게 받치는 동작은 가능하다. 몸통의 좌석 접촉은 잘려 있어 확인할 수 없지만 공중에 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "양손이 위로 열린 채 서로 맞닿아 꽃을 받친다. 꽃의 앞면과 짧은 줄기가 위쪽의 카메라에 비스듬히 드러나 요청된 관찰 방향에 맞는다. 얼굴은 프레임 밖이므로 시선은 확인할 수 없으며, 비행 방향 역시 특정할 수 없다.",
        "built_space": "왼쪽에 회청색 천 벤치 한 줄과 버클 세 개, 오른쪽에 큰 창 하나, 상단에 작은 창 일부가 보인다. 낡은 금속 패널과 리벳, 창틀 및 좌석의 관계는 장소 참조와 잘 맞으며 중복 설비나 불가능한 반사는 없다. 손은 전경 중앙을 크게 차지하고 인물의 착석 접점은 프레임 밖이다. A보다 손 중심 구도에 가깝지만 객실은 여전히 주변 여백 이상으로 넓고 선명하다.",
        "entities": "보이는 인체는 한 사람의 두 손과 손목뿐이며, 연한 녹색 재킷 소매는 인물 참조와 잘 맞는다. 피부와 짧은 손톱은 자연스럽지만 손바닥의 두께와 주름 때문에 10세 아이의 손이라는 특징은 확실하지 않다. 얼굴·머리·눈이 없는 것은 지정된 클로즈업에서 결점이 아니다. 중앙에는 보라색 제비꽃 한 송이가 있고 추가 소품이나 읽을 수 있는 글자는 없다. 창밖 야경은 밤 조건에 맞으며 손에는 주간광처럼 부드럽고 밝은 빛이 들어온다.",
        "hard_violations": [],
        "physics": "꽃과 줄기는 양손이 만드는 오목한 면에 놓여 지지된다. 양손은 손목과 소매로 자연스럽게 이어지고 손가락의 펼침도 가능한 자세다. 얼굴과 몸통을 제외한 크롭 때문에 착석 상태 자체는 검증할 수 없지만, 보이는 범위에 지지 없이 떠 있는 물체나 불가능한 신체 자세는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "두 손바닥에 놓인 제비꽃과 내려다보는 행동은 맞지만, 얼굴·상체와 넓은 객실이 들어와 손 중심 클로즈업 및 주변부로 제한한 배경 지시에서 멀어진다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 손과 그 위의 제비꽃을 크게 담아 지정된 클로즈업에 더 충실하고 객실 재질도 잘 맞지만, 배경 비중이 여전히 크고 손의 아동성은 뚜렷하지 않다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 손바닥이 위로 펼쳐져 중앙의 꽃을 받치고, 꽃의 앞면은 카메라 쪽 위를 향해 비스듬히 보인다. 오른쪽 위에 보이는 아이의 얼굴은 손과 꽃을 향해 숙여져 있으나 눈동자의 정확한 방향은 흐려 확인하기 어렵다. 창밖 불빛만으로 헬기의 진행 방향은 알 수 없다.",
        "built_space": "왼쪽에 연속된 천 벤치 한 줄과 안전벨트 버클 세 개, 오른쪽에 큰 창 하나와 위쪽에 잘린 작은 창 일부, 리벳이 있는 금속 벽체가 보인다. 고정 설비의 배치와 재질은 장소 참조에 부합한다. 아이는 오른쪽 전경에 있으나 실제 착석 접점은 화면 밖이다. 불가능한 반사나 중복 설비는 보이지 않는다. 다만 벤치와 창이 넓고 비교적 선명하게 보여 객실을 부드러운 주변부로만 쓰라는 지시보다 강조된다.",
        "entities": "인물은 한 명이며 금발, 흰 마스크, 녹색 계열 재킷이 참조와 대체로 맞는다. 얼굴이 부분적으로 흐리게 보여 정확한 나이와 한국계·백인 혼혈 정체성은 확인하기 어렵지만 어린 여자아이로 읽힌다. 양손 위에는 보라색 제비꽃 한 송이가 있다. 왼쪽 손목의 회색 밴드는 참조에 없는 복장 차이다. 창밖은 야간으로 보이며, 손의 조명은 요청된 부드러운 주간광보다 어둡고 차갑다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "꽃은 서로 모은 손바닥에 접촉해 받쳐져 있다. 손목과 팔은 소매를 통해 인물의 몸으로 이어지고, 손을 살짝 오므려 작은 꽃을 조심스럽게 받치는 동작은 가능하다. 몸통의 좌석 접촉은 잘려 있어 확인할 수 없지만 공중에 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "양손이 위로 열린 채 서로 맞닿아 꽃을 받친다. 꽃의 앞면과 짧은 줄기가 위쪽의 카메라에 비스듬히 드러나 요청된 관찰 방향에 맞는다. 얼굴은 프레임 밖이므로 시선은 확인할 수 없으며, 비행 방향 역시 특정할 수 없다.",
        "built_space": "왼쪽에 회청색 천 벤치 한 줄과 버클 세 개, 오른쪽에 큰 창 하나, 상단에 작은 창 일부가 보인다. 낡은 금속 패널과 리벳, 창틀 및 좌석의 관계는 장소 참조와 잘 맞으며 중복 설비나 불가능한 반사는 없다. 손은 전경 중앙을 크게 차지하고 인물의 착석 접점은 프레임 밖이다. A보다 손 중심 구도에 가깝지만 객실은 여전히 주변 여백 이상으로 넓고 선명하다.",
        "entities": "보이는 인체는 한 사람의 두 손과 손목뿐이며, 연한 녹색 재킷 소매는 인물 참조와 잘 맞는다. 피부와 짧은 손톱은 자연스럽지만 손바닥의 두께와 주름 때문에 10세 아이의 손이라는 특징은 확실하지 않다. 얼굴·머리·눈이 없는 것은 지정된 클로즈업에서 결점이 아니다. 중앙에는 보라색 제비꽃 한 송이가 있고 추가 소품이나 읽을 수 있는 글자는 없다. 창밖 야경은 밤 조건에 맞으며 손에는 주간광처럼 부드럽고 밝은 빛이 들어온다.",
        "hard_violations": [],
        "physics": "꽃과 줄기는 양손이 만드는 오목한 면에 놓여 지지된다. 양손은 손목과 소매로 자연스럽게 이어지고 손가락의 펼침도 가능한 자세다. 얼굴과 몸통을 제외한 크롭 때문에 착석 상태 자체는 검증할 수 없지만, 보이는 범위에 지지 없이 떠 있는 물체나 불가능한 신체 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.321
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.071
   },
   "violations": {
    "B": [
     "[gemini-pro] 레퍼런스와 프롬프트에 없는 임의의 사물(왼팔의 회색 손목 밴드) 생성"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1071
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 손과 제비꽃의 클로즈업 구도를 정확히 구현하였고, 레퍼런스의 의상 특징과 야간 배경을 충실히 반영했습니다."
   },
   {
    "label": "B",
    "score": 1071,
    "verdict_ko": "클로즈업 지시를 어기고 인물의 측면 상반신을 넓게 포함시켰으며, 레퍼런스에 없는 손목 밴드가 임의로 생성되었습니다.  ★위반: [gemini-pro] 레퍼런스와 프롬프트에 없는 임의의 사물(왼팔의 회색 손목 밴드) 생성"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L119B01.png",
    "asset_id": "d623b34c-4206-41a2-83ad-a7729f0c1edc",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:844202>",
    "asset_id": "3e5ee313-49e4-46e2-85a1-8442296e281d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aaceb0-4b03-7569-98a0-c98636e6524b",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S77sh43__bgfirst_bg.png",
   "bg_asset_id": "144b05c8-c592-4762-b6d3-0f958e2255dc",
   "bg_record_key": "S77sh43::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S77sh43::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:42:51.663196+00:00",
  "fingerprint": "121b9fcc8545642adc157add961e1698f7ab0590ce8909b069fd096efd86041b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S77sh43_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S77sh43_sel.png",
  "source_sha256": "3dcf17850ea9486e8d2c85db11021c5716308bc38419fd30aee771b46eee03f9",
  "file": "S77sh43_cine.png",
  "staged_sha256": "4afda77d4e289f2be537517474d6513eaf40053dc4f8eabec0a949b23872b786",
  "latency_ms": 14150
 },
 "S77sh46::signage": {
  "fp": "c48d6d1c6bf4657e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S77sh46::bgfirst_bg": {
  "input_fingerprint": "ac0628cd538ee059",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 낡은 배낭을 멘 이현우와 찰리가 텅 빈 활주로를 따라 나란히 걷는 mid-stride로 한 발을 내딛은 전신 구도.\n\nLOCATION (lock): On the empty runway of the abandoned airfield in daylight, after the helicopter has departed.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Empty runway (Unoccupied along the pair's route) — Recedes diagonally from the lower frame into the space ahead; used as Carries the walking direction and separates both feet from the frame edge; Worn backpacks (Worn by 이현우 and 찰리) — Rear and side surfaces are visible from the trailing camera; used as Links the two departing silhouettes without obscuring their different strides.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light maintains an even, restrained tonal range without changing the established departure lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 낡은 배낭을 멘 이현우와 찰리가 텅 빈 활주로를 따라 나란히 걷는 mid-stride로 한 발을 내딛은 전신 구도.\n\nLOCATION (lock): On the empty runway of the abandoned airfield in daylight, after the helicopter has departed.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Empty runway (Unoccupied along the pair's route) — Recedes diagonally from the lower frame into the space ahead; used as Carries the walking direction and separates both feet from the frame edge; Worn backpacks (Worn by 이현우 and 찰리) — Rear and side surfaces are visible from the trailing camera; used as Links the two departing silhouettes without obscuring their different strides.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light maintains an even, restrained tonal range without changing the established departure lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S77sh46__bgfirst_bg.png",
  "asset_id": "92bd179a-1409-45cb-b33d-c522fa779c22",
  "input_asset_ids": [
   "98cb0775-4ba7-4723-ac48-4e4ef20d6876",
   "8270cddc-18bf-40cf-8d42-b93a59c6de05"
  ]
 },
 "S77sh46": {
  "input_fingerprint": "1ac9a93300ae26ca",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 낡은 배낭을 멘 이현우와 찰리가 텅 빈 활주로를 따라 나란히 걷는 mid-stride로 한 발을 내딛은 전신 구도.\n\nLOCATION (lock): On the empty runway of the abandoned airfield in daylight, after the helicopter has departed. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Empty runway (Unoccupied along the pair's route) — Recedes diagonally from the lower frame into the space ahead; used as Carries the walking direction and separates both feet from the frame edge; Worn backpacks (Worn by 이현우 and 찰리) — Rear and side surfaces are visible from the trailing camera; used as Links the two departing silhouettes without obscuring their different strides.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light maintains an even, restrained tonal range without changing the established departure lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The helicopter has disappeared into the distance. The old airport remains in daylight, with its derelict grounds and hangars. 이현우: He walks away wearing a backpack. His skin and clothing remain grimy, and his accumulated injuries remain. 찰리: He walks away wearing a backpack. His unrepaired battle damage, exposed chest opening and salvaged chest component remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 낡은 배낭을 멘 이현우와 찰리가 텅 빈 활주로를 따라 나란히 걷는 mid-stride로 한 발을 내딛은 전신 구도.\n\nLOCATION (lock): On the empty runway of the abandoned airfield in daylight, after the helicopter has departed. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Empty runway (Unoccupied along the pair's route) — Recedes diagonally from the lower frame into the space ahead; used as Carries the walking direction and separates both feet from the frame edge; Worn backpacks (Worn by 이현우 and 찰리) — Rear and side surfaces are visible from the trailing camera; used as Links the two departing silhouettes without obscuring their different strides.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light maintains an even, restrained tonal range without changing the established departure lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The helicopter has disappeared into the distance. The old airport remains in daylight, with its derelict grounds and hangars. 이현우: He walks away wearing a backpack. His skin and clothing remain grimy, and his accumulated injuries remain. 찰리: He walks away wearing a backpack. His unrepaired battle damage, exposed chest opening and salvaged chest component remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 낡은 배낭을 멘 이현우와 찰리가 텅 빈 활주로를 따라 나란히 걷는 mid-stride로 한 발을 내딛은 전신 구도.\n\nLOCATION (lock): On the empty runway of the abandoned airfield in daylight, after the helicopter has departed. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Empty runway (Unoccupied along the pair's route) — Recedes diagonally from the lower frame into the space ahead; used as Carries the walking direction and separates both feet from the frame edge; Worn backpacks (Worn by 이현우 and 찰리) — Rear and side surfaces are visible from the trailing camera; used as Links the two departing silhouettes without obscuring their different strides.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light maintains an even, restrained tonal range without changing the established departure lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The helicopter has disappeared into the distance. The old airport remains in daylight, with its derelict grounds and hangars. 이현우: He walks away wearing a backpack. His skin and clothing remain grimy, and his accumulated injuries remain. 찰리: He walks away wearing a backpack. His unrepaired battle damage, exposed chest opening and salvaged chest component remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S77sh46__bgfirst_bg.png",
     "asset_id": "92bd179a-1409-45cb-b33d-c522fa779c22",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S77sh46.png",
     "asset_id": "98cb0775-4ba7-4723-ac48-4e4ef20d6876",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:885644>",
     "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L119B02.png",
     "asset_id": "8270cddc-18bf-40cf-8d42-b93a59c6de05",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:885644>",
     "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물이 활주로를 따라 카메라 반대편으로 멀어지며 걷고 있음",
    "built_space": "갈라진 틈과 잡초가 있는 텅 빈 활주로가 배경과 잘 일치함",
    "entities": "이현우는 후드티와 청바지, 배낭을 착용해 레퍼런스와 일치함; 찰리는 고릴라 형태의 기계 머리를 가졌으나, 레퍼런스에 있는 모든 의상(모자, 옷, 바지, 신발)이 누락된 채 금속 골조만 노출되어 있음",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 발을 딛고 자연스러운 보행 동작(mid-stride)을 보여줌"
   },
   {
    "label": "B",
    "direction": "두 인물이 활주로를 따라 카메라 반대편으로 나란히 멀어지며 걷고 있음",
    "built_space": "갈라진 틈과 잡초가 있는 텅 빈 활주로가 레퍼런스와 잘 일치함",
    "entities": "이현우는 후드티와 청바지를 입고 배낭을 메고 있음; 찰리는 둥근 형태의 기계 뒷머리를 보이며, 흉부 파손 설정을 위해 상의는 탈의했으나 레퍼런스의 초록색 바지와 노란 장화는 정확히 착용하고 배낭을 멘 상태임",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 발을 안정적으로 딛고 나란히 걷는 보행 동작(mid-stride)을 잘 보여줌"
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "캐릭터 레퍼런스의 의상(초록색 바지, 노란 장화)을 잘 유지했으며 나란히 걷는 자연스러운 포즈를 구현하여 지시문을 더 충실히 따랐습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "찰리의 의상 레퍼런스를 완전히 무시하고 전신을 기계로만 표현했으며, 부자연스럽게 상체를 틀고 제스처를 취하고 있어 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 활주로를 따라 카메라 반대편으로 멀어지며 걷고 있음",
        "built_space": "갈라진 틈과 잡초가 있는 텅 빈 활주로가 배경과 잘 일치함",
        "entities": "이현우는 후드티와 청바지, 배낭을 착용해 레퍼런스와 일치함; 찰리는 고릴라 형태의 기계 머리를 가졌으나, 레퍼런스에 있는 모든 의상(모자, 옷, 바지, 신발)이 누락된 채 금속 골조만 노출되어 있음",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 발을 딛고 자연스러운 보행 동작(mid-stride)을 보여줌"
       },
       {
        "label": "B",
        "direction": "두 인물이 활주로를 따라 카메라 반대편으로 나란히 멀어지며 걷고 있음",
        "built_space": "갈라진 틈과 잡초가 있는 텅 빈 활주로가 레퍼런스와 잘 일치함",
        "entities": "이현우는 후드티와 청바지를 입고 배낭을 메고 있음; 찰리는 둥근 형태의 기계 뒷머리를 보이며, 흉부 파손 설정을 위해 상의는 탈의했으나 레퍼런스의 초록색 바지와 노란 장화는 정확히 착용하고 배낭을 멘 상태임",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 발을 안정적으로 딛고 나란히 걷는 보행 동작(mid-stride)을 잘 보여줌"
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "캐릭터 레퍼런스의 의상(초록색 바지, 노란 장화)을 잘 유지했으며 나란히 걷는 자연스러운 포즈를 구현하여 지시문을 더 충실히 따랐습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "찰리의 의상 레퍼런스를 완전히 무시하고 전신을 기계로만 표현했으며, 부자연스럽게 상체를 틀고 제스처를 취하고 있어 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 활주로를 따라 카메라 반대편으로 멀어지며 걷고 있음",
        "built_space": "갈라진 틈과 잡초가 있는 텅 빈 활주로가 배경과 잘 일치함",
        "entities": "이현우는 후드티와 청바지, 배낭을 착용해 레퍼런스와 일치함; 찰리는 고릴라 형태의 기계 머리를 가졌으나, 레퍼런스에 있는 모든 의상(모자, 옷, 바지, 신발)이 누락된 채 금속 골조만 노출되어 있음",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 발을 딛고 자연스러운 보행 동작(mid-stride)을 보여줌"
       },
       {
        "label": "B",
        "direction": "두 인물이 활주로를 따라 카메라 반대편으로 나란히 멀어지며 걷고 있음",
        "built_space": "갈라진 틈과 잡초가 있는 텅 빈 활주로가 레퍼런스와 잘 일치함",
        "entities": "이현우는 후드티와 청바지를 입고 배낭을 메고 있음; 찰리는 둥근 형태의 기계 뒷머리를 보이며, 흉부 파손 설정을 위해 상의는 탈의했으나 레퍼런스의 초록색 바지와 노란 장화는 정확히 착용하고 배낭을 멘 상태임",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 발을 안정적으로 딛고 나란히 걷는 보행 동작(mid-stride)을 잘 보여줌"
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "밤의 빈 활주로를 배낭을 메고 나란히 떠나는 전신 후방 구도를 충족하며, 찰리의 기계 머리·초록 바지·노란 장화가 B보다 참조에 가깝지만 상의와 모자가 빠졌다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "배낭을 멘 두 인물의 보행과 후방 전신 구도는 맞지만, 찰리가 흰 기계 얼굴판 대신 사실적인 고릴라 얼굴을 하고 참조 의상도 모두 잃어 정체성 일치도가 낮다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물의 머리와 몸은 화면 중앙 먼 활주로를 향하며 카메라에서 멀어지는 방향으로 나란히 걷는다. 배낭 뒷면과 일부 측면이 보인다. 진행축은 요청한 뚜렷한 사선보다는 중앙 소실점을 향하는 정면 후방 축에 가깝다.",
        "built_space": "갈라진 포장과 잡초, 닳은 흰 활주로 표식, 왼쪽 원경의 관제탑 한 기와 낮은 공항 건물군이 보인다. 인물 둘은 활주로 위에 있고 앞길은 비어 있으며 헬리콥터는 없다. 발 아래로 충분한 포장면이 남는 전신 와이드 구도다. 참조보다 왼쪽 중경의 켜진 조명 기둥 한 기가 두드러진다. 명시된 밤 잠금은 지켰으나 본문에 함께 적힌 주간 조명과는 일치하지 않는다.",
        "entities": "등장 대상은 이현우와 찰리로 읽히는 두 인물, 낡은 배낭 두 개다. 이현우의 헝클어진 검은 머리, 젊고 마른 체형, 때 묻은 어두운 후드와 청바지는 참조와 대체로 맞는다. 얼굴이 뒤로 가려져 한국계 얼굴과 정확한 나이, 얼굴 상처는 확인할 수 없다. 찰리는 금속 머리와 손상된 기계 팔, 초록 바지와 노란 장화를 갖췄지만 참조의 밀짚모자와 다색 외투가 없다. 흰 얼굴판과 가슴 개구부·회수 부품은 후방 구도에서 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우는 왼발을 지면에 두고 오른발을 뒤로 들어 밑창을 드러낸다. 찰리는 오른쪽 장화로 지면을 딛고 왼쪽 장화를 들어 보행 중간 자세를 만든다. 두 인물 모두 디딤발과 지면 쪽으로 이어지는 그림자가 있으며, 공중에 지지 없이 매달린 자세로 단정할 근거는 없다. 배낭은 어깨끈으로 몸에 고정되어 있다."
       },
       {
        "label": "B",
        "direction": "두 몸과 다리는 먼 활주로를 향해 이동하지만 이현우는 오른쪽 찰리 쪽으로, 찰리는 왼쪽 이현우 쪽으로 머리를 돌린다. 찰리의 왼손도 동행자 쪽으로 펼쳐져 대화하는 순간처럼 보인다. 이동 방향은 맞지만 활주로 진행축은 A처럼 중앙 소실점을 향해 있어 사선 배치가 약하다.",
        "built_space": "균열과 잡초가 있는 활주로 한 면, 마모된 흰 표식, 왼쪽 원경의 관제탑 한 기와 낮은 건물군이 보인다. 두 인물 외에 길을 점유한 사람이나 항공기는 없다. 배경 건물은 원경 크기를 유지하고 두 발도 화면 안에 완전히 들어온다. 주변 공항 불빛과 하늘의 달이 보이는 야간이며, 참조의 폐활주로 재질과 지형은 대체로 유지된다.",
        "entities": "검은 머리의 젊은 남성과 손상된 기계 몸을 가진 찰리, 낡은 배낭 두 개가 보인다. 이현우의 후드·청바지와 체형은 참조에 가깝지만 얼굴은 작은 측면만 보여 정확한 얼굴 일치를 판단하기 어렵다. 찰리의 머리는 털과 살 같은 표면의 사실적인 고릴라 얼굴로, 지정된 각진 기계 외장과 흰 얼굴판·점눈·선입과 다르다. 밀짚모자, 다색 외투, 초록 바지와 노란 장화도 없고 기계 다리가 그대로 노출된다. 가슴 개구부와 회수 부품은 뒤쪽 시점 때문에 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우는 앞쪽 발에 체중을 옮기며 뒤쪽 발의 뒤꿈치를 드는 보행 자세이고, 찰리도 앞쪽 다리로 몸을 지탱하면서 뒤쪽 기계 발의 밑창을 드러낸다. 발과 활주로의 접촉 및 그림자가 이어져 지지 관계가 성립한다. 찰리의 들어 올린 왼손은 굽힌 팔에 연결되어 있으며, 배낭 두 개는 각각 어깨끈으로 지지된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "밤의 빈 활주로를 배낭을 메고 나란히 떠나는 전신 후방 구도를 충족하며, 찰리의 기계 머리·초록 바지·노란 장화가 B보다 참조에 가깝지만 상의와 모자가 빠졌다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "배낭을 멘 두 인물의 보행과 후방 전신 구도는 맞지만, 찰리가 흰 기계 얼굴판 대신 사실적인 고릴라 얼굴을 하고 참조 의상도 모두 잃어 정체성 일치도가 낮다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 인물의 머리와 몸은 화면 중앙 먼 활주로를 향하며 카메라에서 멀어지는 방향으로 나란히 걷는다. 배낭 뒷면과 일부 측면이 보인다. 진행축은 요청한 뚜렷한 사선보다는 중앙 소실점을 향하는 정면 후방 축에 가깝다.",
        "built_space": "갈라진 포장과 잡초, 닳은 흰 활주로 표식, 왼쪽 원경의 관제탑 한 기와 낮은 공항 건물군이 보인다. 인물 둘은 활주로 위에 있고 앞길은 비어 있으며 헬리콥터는 없다. 발 아래로 충분한 포장면이 남는 전신 와이드 구도다. 참조보다 왼쪽 중경의 켜진 조명 기둥 한 기가 두드러진다. 명시된 밤 잠금은 지켰으나 본문에 함께 적힌 주간 조명과는 일치하지 않는다.",
        "entities": "등장 대상은 이현우와 찰리로 읽히는 두 인물, 낡은 배낭 두 개다. 이현우의 헝클어진 검은 머리, 젊고 마른 체형, 때 묻은 어두운 후드와 청바지는 참조와 대체로 맞는다. 얼굴이 뒤로 가려져 한국계 얼굴과 정확한 나이, 얼굴 상처는 확인할 수 없다. 찰리는 금속 머리와 손상된 기계 팔, 초록 바지와 노란 장화를 갖췄지만 참조의 밀짚모자와 다색 외투가 없다. 흰 얼굴판과 가슴 개구부·회수 부품은 후방 구도에서 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우는 왼발을 지면에 두고 오른발을 뒤로 들어 밑창을 드러낸다. 찰리는 오른쪽 장화로 지면을 딛고 왼쪽 장화를 들어 보행 중간 자세를 만든다. 두 인물 모두 디딤발과 지면 쪽으로 이어지는 그림자가 있으며, 공중에 지지 없이 매달린 자세로 단정할 근거는 없다. 배낭은 어깨끈으로 몸에 고정되어 있다."
       },
       {
        "label": "A",
        "direction": "두 몸과 다리는 먼 활주로를 향해 이동하지만 이현우는 오른쪽 찰리 쪽으로, 찰리는 왼쪽 이현우 쪽으로 머리를 돌린다. 찰리의 왼손도 동행자 쪽으로 펼쳐져 대화하는 순간처럼 보인다. 이동 방향은 맞지만 활주로 진행축은 A처럼 중앙 소실점을 향해 있어 사선 배치가 약하다.",
        "built_space": "균열과 잡초가 있는 활주로 한 면, 마모된 흰 표식, 왼쪽 원경의 관제탑 한 기와 낮은 건물군이 보인다. 두 인물 외에 길을 점유한 사람이나 항공기는 없다. 배경 건물은 원경 크기를 유지하고 두 발도 화면 안에 완전히 들어온다. 주변 공항 불빛과 하늘의 달이 보이는 야간이며, 참조의 폐활주로 재질과 지형은 대체로 유지된다.",
        "entities": "검은 머리의 젊은 남성과 손상된 기계 몸을 가진 찰리, 낡은 배낭 두 개가 보인다. 이현우의 후드·청바지와 체형은 참조에 가깝지만 얼굴은 작은 측면만 보여 정확한 얼굴 일치를 판단하기 어렵다. 찰리의 머리는 털과 살 같은 표면의 사실적인 고릴라 얼굴로, 지정된 각진 기계 외장과 흰 얼굴판·점눈·선입과 다르다. 밀짚모자, 다색 외투, 초록 바지와 노란 장화도 없고 기계 다리가 그대로 노출된다. 가슴 개구부와 회수 부품은 뒤쪽 시점 때문에 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우는 앞쪽 발에 체중을 옮기며 뒤쪽 발의 뒤꿈치를 드는 보행 자세이고, 찰리도 앞쪽 다리로 몸을 지탱하면서 뒤쪽 기계 발의 밑창을 드러낸다. 발과 활주로의 접촉 및 그림자가 이어져 지지 관계가 성립한다. 찰리의 들어 올린 왼손은 굽힌 팔에 연결되어 있으며, 배낭 두 개는 각각 어깨끈으로 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.429,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1429
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "캐릭터 레퍼런스의 의상(초록색 바지, 노란 장화)을 잘 유지했으며 나란히 걷는 자연스러운 포즈를 구현하여 지시문을 더 충실히 따랐습니다."
   },
   {
    "label": "A",
    "score": 1429,
    "verdict_ko": "찰리의 의상 레퍼런스를 완전히 무시하고 전신을 기계로만 표현했으며, 부자연스럽게 상체를 틀고 제스처를 취하고 있어 아쉽습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L119B02.png",
    "asset_id": "8270cddc-18bf-40cf-8d42-b93a59c6de05",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:885644>",
    "asset_id": "25c5174a-642f-4904-bc56-adfc1b73bf14",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aaceb8-a2e6-7afc-bf70-ace4943457ce",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S77sh46__bgfirst_bg.png",
   "bg_asset_id": "92bd179a-1409-45cb-b33d-c522fa779c22",
   "bg_record_key": "S77sh46::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S77sh46::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:44:40.830980+00:00",
  "fingerprint": "6c99232ec92dcf727046d05977d00776ff06a4e34adb6ca985d7494181f738e4",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S77sh46_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S77sh46_sel.png",
  "source_sha256": "4e0c155340dd2c81794b4237192df3169265cb510e7f024254bea080272de9e7",
  "file": "S77sh46_cine.png",
  "staged_sha256": "05bef20b963023b67d45a9a5652a3755808e7a3fab19c8f06858209bb8f2bf86",
  "latency_ms": 11912
 },
 "S78sh2::signage": {
  "fp": "ad0fc6f9a1ff636c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::60a2115426d15f03": {
  "subjects": [],
  "subject_text": "목포 항구 부두\n출렁이는 바다와 맞닿은 부둣가. 가장자리에 닻줄을 묶는 시설이 있고 생선을 진열한 작은 판매대가 가까이 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L121",
  "scope_role": "location_exterior",
  "scope_sha": "98aca2f15c64ddaf"
 },
 "groupbg::harbor_waiting_pier": {
  "input_fingerprint": "fc57f64bf58b7f19",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "harbor_waiting_pier",
    "tags": [
     "S78sh2"
    ]
   },
   "context_sig": "b0c3c51ce868ff97"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the harbor quay's edge beside an old mooring post, overlooking moored boats in the evening.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구 부두: 출렁이는 검은 바닷물 위로 낡은 밀항선들이 정박해 있고, 두꺼운 밧줄과 닻이 묶인 어스름한 저녁의 선착장. (특징: 배를 고정하는 굵은 정박용 밧줄; 검은 바닷물 위로 흔들리는 소형 선박들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 커다란 모포를 뒤집어쓴 찰리와 현우가 부둣가에 걸터앉아있다.\n\nTIME OF DAY (lock): dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the harbor quay's edge beside an old mooring post, overlooking moored boats in the evening.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구 부두: 출렁이는 검은 바닷물 위로 낡은 밀항선들이 정박해 있고, 두꺼운 밧줄과 닻이 묶인 어스름한 저녁의 선착장. (특징: 배를 고정하는 굵은 정박용 밧줄; 검은 바닷물 위로 흔들리는 소형 선박들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 커다란 모포를 뒤집어쓴 찰리와 현우가 부둣가에 걸터앉아있다.\n\nTIME OF DAY (lock): dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_harbor_waiting_pier_96014f.png",
  "asset_id": "edc5290b-4296-4327-b3bd-31cfc2dee164",
  "input_asset_ids": [
   "d1f7a3c6-44c8-4a99-80c4-9c2586c36402"
  ],
  "origin_tag": "S78sh2",
  "place_text": "At the harbor quay's edge beside an old mooring post, overlooking moored boats in the evening.",
  "origin_inputs": {
   "place_text": "At the harbor quay's edge beside an old mooring post, overlooking moored boats in the evening.",
   "time_of_day_en": "dusk",
   "conti_asset_id": "d1f7a3c6-44c8-4a99-80c4-9c2586c36402"
  }
 },
 "S78sh2::bgfirst_bg": {
  "input_fingerprint": "b627e23983f03dbc",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 커다란 모포를 뒤집어쓴 찰리와 이현우가 부둣가 낡은 나무 기둥 옆에 나란히 걸터앉은 구도.\n\nLOCATION (lock): At the harbor quay's edge beside an old mooring post, overlooking moored boats in the evening.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old wooden quay post (Beside the seated pair) — Seen obliquely at the right frame edge; used as Local spatial anchor, kept narrow and subordinate to the figures; Quay edge (Used as a seat by 찰리 and 이현우) — Runs obliquely beneath the pair; used as Establishes their shared seating plane; Moored boats (Stationary in the harbor) — Partial hull profiles are visible beyond the seated figures; used as Provides the distant destination of their searching attention; Large blanket covering (Draped over the seated pair) — Outer folds face the oblique camera; used as Connects their silhouettes while preserving individual posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim evening ambient light gives the harbor subdued contrast while keeping the blanket-covered figures readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 커다란 모포를 뒤집어쓴 찰리와 이현우가 부둣가 낡은 나무 기둥 옆에 나란히 걸터앉은 구도.\n\nLOCATION (lock): At the harbor quay's edge beside an old mooring post, overlooking moored boats in the evening.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old wooden quay post (Beside the seated pair) — Seen obliquely at the right frame edge; used as Local spatial anchor, kept narrow and subordinate to the figures; Quay edge (Used as a seat by 찰리 and 이현우) — Runs obliquely beneath the pair; used as Establishes their shared seating plane; Moored boats (Stationary in the harbor) — Partial hull profiles are visible beyond the seated figures; used as Provides the distant destination of their searching attention; Large blanket covering (Draped over the seated pair) — Outer folds face the oblique camera; used as Connects their silhouettes while preserving individual posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim evening ambient light gives the harbor subdued contrast while keeping the blanket-covered figures readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S78sh2__bgfirst_bg.png",
  "asset_id": "88ed0dc9-9516-407b-a421-e54223b7fcad",
  "input_asset_ids": [
   "d1f7a3c6-44c8-4a99-80c4-9c2586c36402",
   "edc5290b-4296-4327-b3bd-31cfc2dee164"
  ]
 },
 "S78sh2": {
  "input_fingerprint": "363664901cf390d9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 커다란 모포를 뒤집어쓴 찰리와 이현우가 부둣가 낡은 나무 기둥 옆에 나란히 걸터앉은 구도.\n\nLOCATION (lock): At the harbor quay's edge beside an old mooring post, overlooking moored boats in the evening. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old wooden quay post (Beside the seated pair) — Seen obliquely at the right frame edge; used as Local spatial anchor, kept narrow and subordinate to the figures; Quay edge (Used as a seat by 찰리 and 이현우) — Runs obliquely beneath the pair; used as Establishes their shared seating plane; Moored boats (Stationary in the harbor) — Partial hull profiles are visible beyond the seated figures; used as Provides the distant destination of their searching attention; Large blanket covering (Draped over the seated pair) — Outer folds face the oblique camera; used as Connects their silhouettes while preserving individual posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim evening ambient light gives the harbor subdued contrast while keeping the blanket-covered figures readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boats are moored along the harbor in fading evening light, and fish are displayed for sale beside the dock. The sea moves against the quay. 이현우: He sits on the dock with his backpack. His grimy appearance and accumulated injuries remain. 찰리: He sits on the dock covered by a large blanket, retaining his backpack and salvaged chest component. His unrepaired battle damage and exposed chest opening remain beneath the covering.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 커다란 모포를 뒤집어쓴 찰리와 이현우가 부둣가 낡은 나무 기둥 옆에 나란히 걸터앉은 구도.\n\nLOCATION (lock): At the harbor quay's edge beside an old mooring post, overlooking moored boats in the evening. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old wooden quay post (Beside the seated pair) — Seen obliquely at the right frame edge; used as Local spatial anchor, kept narrow and subordinate to the figures; Quay edge (Used as a seat by 찰리 and 이현우) — Runs obliquely beneath the pair; used as Establishes their shared seating plane; Moored boats (Stationary in the harbor) — Partial hull profiles are visible beyond the seated figures; used as Provides the distant destination of their searching attention; Large blanket covering (Draped over the seated pair) — Outer folds face the oblique camera; used as Connects their silhouettes while preserving individual posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim evening ambient light gives the harbor subdued contrast while keeping the blanket-covered figures readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boats are moored along the harbor in fading evening light, and fish are displayed for sale beside the dock. The sea moves against the quay. 이현우: He sits on the dock with his backpack. His grimy appearance and accumulated injuries remain. 찰리: He sits on the dock covered by a large blanket, retaining his backpack and salvaged chest component. His unrepaired battle damage and exposed chest opening remain beneath the covering.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 커다란 모포를 뒤집어쓴 찰리와 이현우가 부둣가 낡은 나무 기둥 옆에 나란히 걸터앉은 구도.\n\nLOCATION (lock): At the harbor quay's edge beside an old mooring post, overlooking moored boats in the evening. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old wooden quay post (Beside the seated pair) — Seen obliquely at the right frame edge; used as Local spatial anchor, kept narrow and subordinate to the figures; Quay edge (Used as a seat by 찰리 and 이현우) — Runs obliquely beneath the pair; used as Establishes their shared seating plane; Moored boats (Stationary in the harbor) — Partial hull profiles are visible beyond the seated figures; used as Provides the distant destination of their searching attention; Large blanket covering (Draped over the seated pair) — Outer folds face the oblique camera; used as Connects their silhouettes while preserving individual posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim evening ambient light gives the harbor subdued contrast while keeping the blanket-covered figures readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boats are moored along the harbor in fading evening light, and fish are displayed for sale beside the dock. The sea moves against the quay. 이현우: He sits on the dock with his backpack. His grimy appearance and accumulated injuries remain. 찰리: He sits on the dock covered by a large blanket, retaining his backpack and salvaged chest component. His unrepaired battle damage and exposed chest opening remain beneath the covering.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S78sh2__bgfirst_bg.png",
     "asset_id": "88ed0dc9-9516-407b-a421-e54223b7fcad",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S78sh2.png",
     "asset_id": "d1f7a3c6-44c8-4a99-80c4-9c2586c36402",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:948699>",
     "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_harbor_waiting_pier_96014f.png",
     "asset_id": "edc5290b-4296-4327-b3bd-31cfc2dee164",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:948699>",
     "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 바다와 정박된 배들을 향해 시선을 두고 있습니다.",
    "built_space": "참조 이미지와 동일한 부둣가 환경입니다. 화면 우측에 낡은 나무 기둥 1개가 있고, 두 인물은 부둣가 가장자리에 걸터앉아 있습니다.",
    "entities": "기계 머리가 노출된 찰리와 옆에 앉은 검은 머리의 이현우가 확인됩니다. 두 사람을 덮은 모포는 굵은 펜선이 들어간 만화적 질감으로 표현되었습니다.",
    "hard_violations": [
     "[gemini-pro] 모포가 실사가 아닌 2D 만화 그래픽처럼 렌더링되어 재질 사실주의(Material Realism) 지침 위반"
    ],
    "physics": "두 인물 모두 부둣가 바닥에 엉덩이를 대고 안정적으로 앉아 있습니다."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 정박된 배가 있는 바다 쪽을 바라보고 있습니다.",
    "built_space": "참조 이미지의 부둣가를 정확히 반영했습니다. 우측에 나무 기둥 1개가 배치되어 있으며 인물들은 가장자리에 앉아 있습니다.",
    "entities": "모포를 머리까지 뒤집어쓴 찰리(배낭과 기계 몸체 일부 노출)와 배낭을 멘 이현우가 확인됩니다. 실사 재질의 큰 모포가 두 사람을 자연스럽게 덮고 있습니다.",
    "hard_violations": [],
    "physics": "두 인물 모두 부둣가 가장자리에 자연스럽게 걸터앉아 바닥의 지지를 받고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 부둣가 배경에서 두 인물이 모포를 덮고 나란히 앉아있는 구도와 실사 질감을 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "모포가 굵은 외곽선을 가진 2D 그래픽처럼 렌더링되어 재질 사실주의(Photorealistic) 지침을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 바다와 정박된 배들을 향해 시선을 두고 있습니다.",
        "built_space": "참조 이미지와 동일한 부둣가 환경입니다. 화면 우측에 낡은 나무 기둥 1개가 있고, 두 인물은 부둣가 가장자리에 걸터앉아 있습니다.",
        "entities": "기계 머리가 노출된 찰리와 옆에 앉은 검은 머리의 이현우가 확인됩니다. 두 사람을 덮은 모포는 굵은 펜선이 들어간 만화적 질감으로 표현되었습니다.",
        "hard_violations": [
         "모포가 실사가 아닌 2D 만화 그래픽처럼 렌더링되어 재질 사실주의(Material Realism) 지침 위반"
        ],
        "physics": "두 인물 모두 부둣가 바닥에 엉덩이를 대고 안정적으로 앉아 있습니다."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 정박된 배가 있는 바다 쪽을 바라보고 있습니다.",
        "built_space": "참조 이미지의 부둣가를 정확히 반영했습니다. 우측에 나무 기둥 1개가 배치되어 있으며 인물들은 가장자리에 앉아 있습니다.",
        "entities": "모포를 머리까지 뒤집어쓴 찰리(배낭과 기계 몸체 일부 노출)와 배낭을 멘 이현우가 확인됩니다. 실사 재질의 큰 모포가 두 사람을 자연스럽게 덮고 있습니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 부둣가 가장자리에 자연스럽게 걸터앉아 바닥의 지지를 받고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 부둣가 배경에서 두 인물이 모포를 덮고 나란히 앉아있는 구도와 실사 질감을 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "모포가 굵은 외곽선을 가진 2D 그래픽처럼 렌더링되어 재질 사실주의(Photorealistic) 지침을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 바다와 정박된 배들을 향해 시선을 두고 있습니다.",
        "built_space": "참조 이미지와 동일한 부둣가 환경입니다. 화면 우측에 낡은 나무 기둥 1개가 있고, 두 인물은 부둣가 가장자리에 걸터앉아 있습니다.",
        "entities": "기계 머리가 노출된 찰리와 옆에 앉은 검은 머리의 이현우가 확인됩니다. 두 사람을 덮은 모포는 굵은 펜선이 들어간 만화적 질감으로 표현되었습니다.",
        "hard_violations": [
         "모포가 실사가 아닌 2D 만화 그래픽처럼 렌더링되어 재질 사실주의(Material Realism) 지침 위반"
        ],
        "physics": "두 인물 모두 부둣가 바닥에 엉덩이를 대고 안정적으로 앉아 있습니다."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 정박된 배가 있는 바다 쪽을 바라보고 있습니다.",
        "built_space": "참조 이미지의 부둣가를 정확히 반영했습니다. 우측에 나무 기둥 1개가 배치되어 있으며 인물들은 가장자리에 앉아 있습니다.",
        "entities": "모포를 머리까지 뒤집어쓴 찰리(배낭과 기계 몸체 일부 노출)와 배낭을 멘 이현우가 확인됩니다. 실사 재질의 큰 모포가 두 사람을 자연스럽게 덮고 있습니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 부둣가 가장자리에 자연스럽게 걸터앉아 바닥의 지지를 받고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "공유 모포를 뒤집어쓰고 부두 가장자리에 나란히 걸터앉은 자세와 어두운 저녁빛이 더 충실하지만, 시선이 정박선에 모이지 않고 오른쪽 기둥도 지나치게 두드러집니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "장소와 공유 모포는 맞지만, 걸터앉기보다 부두 안쪽에 주저앉은 인상이 강하며 현우의 시선도 정박선 아닌 오른쪽 수면을 향합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물은 카메라를 등지고 항구를 향한다. 찰리는 머리가 대부분 가려져 정확한 눈 방향을 확인할 수 없다. 현우는 고개를 오른쪽 아래로 숙여 가까운 수면을 보는 모습이다. 정박선들은 두 사람의 앞쪽 중에서도 화면 왼쪽에 있으므로, 두 사람이 그 배들을 함께 살피는 시선은 명확히 구현되지 않았다. 무기나 손에 든 지향성 물체는 없다.",
        "built_space": "오른쪽 전경에 밧줄을 감은 큰 목재 계류기둥 하나, 부두를 따라 중경에 뚜렷한 목재 기둥 두 개와 여러 낮은 금속 계류시설이 보인다. 왼쪽 시장 차양과 생선 상자, 뒤쪽 어선들, 방파제와 등대 하나가 장소 참조와 대응한다. 부두 가장자리는 인물 아래에서 대각선으로 이어지며 두 사람의 엉덩이가 같은 가장자리에 놓인다. 다만 오른쪽 기둥은 화면 끝에 좁게 종속된 요소라기보다 폭과 높이가 상당한 전경 물체다.",
        "entities": "인물은 찰리와 현우 두 명이다. 찰리는 모포 밖으로 기계 머리의 일부와 금속 하체가 보이지만, 흰 얼굴판과 점 형태 눈, 선 형태 입은 가려져 확인할 수 없다. 현우는 헝클어진 검은 머리와 낡은 어두운 상의, 청바지를 착용한 젊은 남성으로 보이나 뒷모습이라 정확한 얼굴 일치는 판단하기 어렵다. 큰 모포 하나가 두 사람을 연결하고, 각자의 등에 배낭 하나씩 보인다. 흉부 부품과 손상 부위는 모포에 가려져 있다. 정박 어선, 진열된 생선, 물결이 보이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 골반과 엉덩이는 단단한 부두 면에 지지된다. 현우의 굽힌 다리는 가장자리 바깥으로 내려가며 걸터앉기와 양립한다. 찰리의 금속 하체도 부두 위에 닿아 있다. 배낭은 몸과 부두에 기대고, 모포는 머리와 어깨 및 배낭 위에 걸려 중력 방향으로 처진다. 배들은 수면에 떠 있고 계류줄이 연결되어 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리는 오른쪽으로 약간 돌린 기계 머리의 뒷면을 보이며 항구 쪽을 향하지만 눈과 구체적인 주시점은 보이지 않는다. 현우는 오른쪽 옆얼굴을 드러내며 화면 오른쪽의 열린 수면을 바라본다. 정박선들은 화면 왼쪽과 중앙에 있으므로 현우의 시선은 배들을 향하지 않는다. 무기나 조작 중인 물체는 없다.",
        "built_space": "오른쪽 전경의 밧줄 감긴 목재 기둥 하나와 중경의 뚜렷한 목재 기둥 두 개, 낮은 계류시설들이 보인다. 왼쪽 차양과 생선 상자, 어선들, 방파제와 등대 하나는 장소 참조와 잘 대응한다. 부두 가장자리는 대각선이지만 두 사람의 하체와 접촉점이 넓게 펼쳐진 모포에 가려져, 같은 가장자리에 걸터앉았다기보다 그 안쪽 바닥에 앉은 인상이 강하다. 오른쪽 기둥은 여기서도 좁고 부차적인 배경 기준점치고 크게 강조된다.",
        "entities": "인물은 두 명이며, 찰리의 기계 머리와 현우의 젊은 동아시아계 남성 옆얼굴이 보인다. 현우의 검은 헝클어진 머리와 얼굴의 때는 설정에 맞지만 정확히 18세인지와 참조 얼굴의 세부 일치는 단정하기 어렵다. 찰리의 노출된 머리 외장은 참조보다 둥글고, 얼굴판의 특징은 보이지 않는다. 큰 모포 하나가 두 사람의 어깨와 몸을 연결한다. 현우 오른쪽에 배낭 하나가 놓여 있으며 찰리의 배낭과 흉부 부품은 가림 때문에 확인되지 않는다. 생선 진열과 정박선은 있고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 몸은 부두 바닥에 지지되고, 현우 옆의 배낭도 바닥에 놓여 있다. 모포는 어깨에서 내려와 부두 위에 넓게 쌓이므로 무게와 지지는 자연스럽다. 현우의 무릎처럼 보이는 부분이 위로 굽혀져 있어 바닥에 다리를 접고 앉은 자세로도 읽힌다. 이는 걸터앉기 연출의 불명확함이지 공중부양이나 불가능한 자세는 아니다. 배들은 물의 부력과 계류줄로 지지된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "공유 모포를 뒤집어쓰고 부두 가장자리에 나란히 걸터앉은 자세와 어두운 저녁빛이 더 충실하지만, 시선이 정박선에 모이지 않고 오른쪽 기둥도 지나치게 두드러집니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "장소와 공유 모포는 맞지만, 걸터앉기보다 부두 안쪽에 주저앉은 인상이 강하며 현우의 시선도 정박선 아닌 오른쪽 수면을 향합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "두 인물은 카메라를 등지고 항구를 향한다. 찰리는 머리가 대부분 가려져 정확한 눈 방향을 확인할 수 없다. 현우는 고개를 오른쪽 아래로 숙여 가까운 수면을 보는 모습이다. 정박선들은 두 사람의 앞쪽 중에서도 화면 왼쪽에 있으므로, 두 사람이 그 배들을 함께 살피는 시선은 명확히 구현되지 않았다. 무기나 손에 든 지향성 물체는 없다.",
        "built_space": "오른쪽 전경에 밧줄을 감은 큰 목재 계류기둥 하나, 부두를 따라 중경에 뚜렷한 목재 기둥 두 개와 여러 낮은 금속 계류시설이 보인다. 왼쪽 시장 차양과 생선 상자, 뒤쪽 어선들, 방파제와 등대 하나가 장소 참조와 대응한다. 부두 가장자리는 인물 아래에서 대각선으로 이어지며 두 사람의 엉덩이가 같은 가장자리에 놓인다. 다만 오른쪽 기둥은 화면 끝에 좁게 종속된 요소라기보다 폭과 높이가 상당한 전경 물체다.",
        "entities": "인물은 찰리와 현우 두 명이다. 찰리는 모포 밖으로 기계 머리의 일부와 금속 하체가 보이지만, 흰 얼굴판과 점 형태 눈, 선 형태 입은 가려져 확인할 수 없다. 현우는 헝클어진 검은 머리와 낡은 어두운 상의, 청바지를 착용한 젊은 남성으로 보이나 뒷모습이라 정확한 얼굴 일치는 판단하기 어렵다. 큰 모포 하나가 두 사람을 연결하고, 각자의 등에 배낭 하나씩 보인다. 흉부 부품과 손상 부위는 모포에 가려져 있다. 정박 어선, 진열된 생선, 물결이 보이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 골반과 엉덩이는 단단한 부두 면에 지지된다. 현우의 굽힌 다리는 가장자리 바깥으로 내려가며 걸터앉기와 양립한다. 찰리의 금속 하체도 부두 위에 닿아 있다. 배낭은 몸과 부두에 기대고, 모포는 머리와 어깨 및 배낭 위에 걸려 중력 방향으로 처진다. 배들은 수면에 떠 있고 계류줄이 연결되어 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리는 오른쪽으로 약간 돌린 기계 머리의 뒷면을 보이며 항구 쪽을 향하지만 눈과 구체적인 주시점은 보이지 않는다. 현우는 오른쪽 옆얼굴을 드러내며 화면 오른쪽의 열린 수면을 바라본다. 정박선들은 화면 왼쪽과 중앙에 있으므로 현우의 시선은 배들을 향하지 않는다. 무기나 조작 중인 물체는 없다.",
        "built_space": "오른쪽 전경의 밧줄 감긴 목재 기둥 하나와 중경의 뚜렷한 목재 기둥 두 개, 낮은 계류시설들이 보인다. 왼쪽 차양과 생선 상자, 어선들, 방파제와 등대 하나는 장소 참조와 잘 대응한다. 부두 가장자리는 대각선이지만 두 사람의 하체와 접촉점이 넓게 펼쳐진 모포에 가려져, 같은 가장자리에 걸터앉았다기보다 그 안쪽 바닥에 앉은 인상이 강하다. 오른쪽 기둥은 여기서도 좁고 부차적인 배경 기준점치고 크게 강조된다.",
        "entities": "인물은 두 명이며, 찰리의 기계 머리와 현우의 젊은 동아시아계 남성 옆얼굴이 보인다. 현우의 검은 헝클어진 머리와 얼굴의 때는 설정에 맞지만 정확히 18세인지와 참조 얼굴의 세부 일치는 단정하기 어렵다. 찰리의 노출된 머리 외장은 참조보다 둥글고, 얼굴판의 특징은 보이지 않는다. 큰 모포 하나가 두 사람의 어깨와 몸을 연결한다. 현우 오른쪽에 배낭 하나가 놓여 있으며 찰리의 배낭과 흉부 부품은 가림 때문에 확인되지 않는다. 생선 진열과 정박선은 있고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 몸은 부두 바닥에 지지되고, 현우 옆의 배낭도 바닥에 놓여 있다. 모포는 어깨에서 내려와 부두 위에 넓게 쌓이므로 무게와 지지는 자연스럽다. 현우의 무릎처럼 보이는 부분이 위로 굽혀져 있어 바닥에 다리를 접고 앉은 자세로도 읽힌다. 이는 걸터앉기 연출의 불명확함이지 공중부양이나 불가능한 자세는 아니다. 배들은 물의 부력과 계류줄로 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.286,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.036,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 모포가 실사가 아닌 2D 만화 그래픽처럼 렌더링되어 재질 사실주의(Material Realism) 지침 위반"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1036
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 부둣가 배경에서 두 인물이 모포를 덮고 나란히 앉아있는 구도와 실사 질감을 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1036,
    "verdict_ko": "모포가 굵은 외곽선을 가진 2D 그래픽처럼 렌더링되어 재질 사실주의(Photorealistic) 지침을 심각하게 위반했습니다.  ★위반: [gemini-pro] 모포가 실사가 아닌 2D 만화 그래픽처럼 렌더링되어 재질 사실주의(Material Realism) 지침 위반"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_harbor_waiting_pier_96014f.png",
    "asset_id": "edc5290b-4296-4327-b3bd-31cfc2dee164",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:948699>",
    "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacebf-51c9-7f40-a165-86e7e9f10db6",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S78sh2__bgfirst_bg.png",
   "bg_asset_id": "88ed0dc9-9516-407b-a421-e54223b7fcad",
   "bg_record_key": "S78sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "harbor_waiting_pier",
   "groupbg_asset_id": "edc5290b-4296-4327-b3bd-31cfc2dee164"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S78sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:47:00.966378+00:00",
  "fingerprint": "662e0079161ee53e61dfb941ad01cb6f17a53fae4dd296da3df9ed258c037ba0",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S78sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S78sh2_sel.png",
  "source_sha256": "161c200484dbe0a5df35199868b784800ec7bd5f3b4cec6477aae324e3095101",
  "file": "S78sh2_cine.png",
  "staged_sha256": "4e1ab95bd4ca12049413be001a4f5ddc18e59bb6551e1849d21a3162fd958967",
  "latency_ms": 13461
 },
 "S78sh14::signage": {
  "fp": "8b8ab395e7f22c68",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::boat_boarding_berth": {
  "input_fingerprint": "810d094de9c7d485",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "boat_boarding_berth",
    "tags": [
     "S78sh14"
    ]
   },
   "context_sig": "97c5d05ff4e45879"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the quayside mooring of an old small vessel, where its captain and crew are carrying out repairs at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구 부두: 출렁이는 검은 바닷물 위로 낡은 밀항선들이 정박해 있고, 두꺼운 밧줄과 닻이 묶인 어스름한 저녁의 선착장. (특징: 배를 고정하는 굵은 정박용 밧줄; 검은 바닷물 위로 흔들리는 소형 선박들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 선원이 어디론가로 현우를 인도한다.\n- 낡은 배를 정비하는 40대 초반의 크리스와 몇몇 선원들 보인다.\n\nTIME OF DAY (lock): dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the quayside mooring of an old small vessel, where its captain and crew are carrying out repairs at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구 부두: 출렁이는 검은 바닷물 위로 낡은 밀항선들이 정박해 있고, 두꺼운 밧줄과 닻이 묶인 어스름한 저녁의 선착장. (특징: 배를 고정하는 굵은 정박용 밧줄; 검은 바닷물 위로 흔들리는 소형 선박들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 선원이 어디론가로 현우를 인도한다.\n- 낡은 배를 정비하는 40대 초반의 크리스와 몇몇 선원들 보인다.\n\nTIME OF DAY (lock): dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_boat_boarding_berth_655ef6.png",
  "asset_id": "a29b91e8-e78c-4774-b743-3959ad9e6734",
  "input_asset_ids": [
   "cf55fbf5-2f77-4eae-a6e9-18a111c56c5e"
  ],
  "origin_tag": "S78sh14",
  "place_text": "At the quayside mooring of an old small vessel, where its captain and crew are carrying out repairs at night.",
  "origin_inputs": {
   "place_text": "At the quayside mooring of an old small vessel, where its captain and crew are carrying out repairs at night.",
   "time_of_day_en": "dusk",
   "conti_asset_id": "cf55fbf5-2f77-4eae-a6e9-18a111c56c5e"
  }
 },
 "S78sh14::bgfirst_bg": {
  "input_fingerprint": "dd815ae7aa30fdab",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 크리스가 모포를 쓴 찰리의 거대한 덩치에 시선을 둔 채, 미간을 살짝 찌푸린 의아한 얼굴.\n\nLOCATION (lock): At the quayside mooring of an old small vessel, where its captain and crew are carrying out repairs at night.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리's blanket (Covering his shoulder and concealing his bulk) — Only the outer shoulder fold appears along the left edge; used as Narrow foreground reference establishing the subject of 크리스's scrutiny; Old boat (Moored for maintenance) — A soft partial side view remains behind 크리스; used as Background context for the encounter and subsequent boarding.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the harbor's dim evening ambient light, retaining subtle detail in 크리스's questioning expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 크리스가 모포를 쓴 찰리의 거대한 덩치에 시선을 둔 채, 미간을 살짝 찌푸린 의아한 얼굴.\n\nLOCATION (lock): At the quayside mooring of an old small vessel, where its captain and crew are carrying out repairs at night.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리's blanket (Covering his shoulder and concealing his bulk) — Only the outer shoulder fold appears along the left edge; used as Narrow foreground reference establishing the subject of 크리스's scrutiny; Old boat (Moored for maintenance) — A soft partial side view remains behind 크리스; used as Background context for the encounter and subsequent boarding.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the harbor's dim evening ambient light, retaining subtle detail in 크리스's questioning expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S78sh14__bgfirst_bg.png",
  "asset_id": "a24895dc-4b0f-4532-9a5c-dc697c5d2c93",
  "input_asset_ids": [
   "cf55fbf5-2f77-4eae-a6e9-18a111c56c5e",
   "a29b91e8-e78c-4774-b743-3959ad9e6734"
  ]
 },
 "S78sh14": {
  "input_fingerprint": "067a26d270fd76b5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 크리스가 모포를 쓴 찰리의 거대한 덩치에 시선을 둔 채, 미간을 살짝 찌푸린 의아한 얼굴.\n\nLOCATION (lock): At the quayside mooring of an old small vessel, where its captain and crew are carrying out repairs at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리's blanket (Covering his shoulder and concealing his bulk) — Only the outer shoulder fold appears along the left edge; used as Narrow foreground reference establishing the subject of 크리스's scrutiny; Old boat (Moored for maintenance) — A soft partial side view remains behind 크리스; used as Background context for the encounter and subsequent boarding.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the harbor's dim evening ambient light, retaining subtle detail in 크리스's questioning expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An old boat is moored at the evening harbor for maintenance and boarding. 크리스: He is beside the old boat at the maintenance and boarding point.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 크리스 (40대 초반 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 바닷물과 잦은 비에 노출되어 얼룩진 두꺼운 노란색 계열의 어부용 방수 점퍼. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 크리스가 모포를 쓴 찰리의 거대한 덩치에 시선을 둔 채, 미간을 살짝 찌푸린 의아한 얼굴.\n\nLOCATION (lock): At the quayside mooring of an old small vessel, where its captain and crew are carrying out repairs at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리's blanket (Covering his shoulder and concealing his bulk) — Only the outer shoulder fold appears along the left edge; used as Narrow foreground reference establishing the subject of 크리스's scrutiny; Old boat (Moored for maintenance) — A soft partial side view remains behind 크리스; used as Background context for the encounter and subsequent boarding.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the harbor's dim evening ambient light, retaining subtle detail in 크리스's questioning expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An old boat is moored at the evening harbor for maintenance and boarding. 크리스: He is beside the old boat at the maintenance and boarding point.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 크리스 (40대 초반 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 바닷물과 잦은 비에 노출되어 얼룩진 두꺼운 노란색 계열의 어부용 방수 점퍼. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 크리스가 모포를 쓴 찰리의 거대한 덩치에 시선을 둔 채, 미간을 살짝 찌푸린 의아한 얼굴.\n\nLOCATION (lock): At the quayside mooring of an old small vessel, where its captain and crew are carrying out repairs at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 찰리's blanket (Covering his shoulder and concealing his bulk) — Only the outer shoulder fold appears along the left edge; used as Narrow foreground reference establishing the subject of 크리스's scrutiny; Old boat (Moored for maintenance) — A soft partial side view remains behind 크리스; used as Background context for the encounter and subsequent boarding.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the harbor's dim evening ambient light, retaining subtle detail in 크리스's questioning expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An old boat is moored at the evening harbor for maintenance and boarding. 크리스: He is beside the old boat at the maintenance and boarding point.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 크리스 (40대 초반 남성, 중년의 얼굴, 머리 모양 미상) — wearing: 바닷물과 잦은 비에 노출되어 얼룩진 두꺼운 노란색 계열의 어부용 방수 점퍼. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S78sh14__bgfirst_bg.png",
     "asset_id": "a24895dc-4b0f-4532-9a5c-dc697c5d2c93",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S78sh14.png",
     "asset_id": "cf55fbf5-2f77-4eae-a6e9-18a111c56c5e",
     "role": "conti_light"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_boat_boarding_berth_655ef6.png",
     "asset_id": "a29b91e8-e78c-4774-b743-3959ad9e6734",
     "role": "bgfirst_group_bg"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "크리스의 시선이 화면 왼쪽에 자리 잡은 인물의 형태를 향하고 있음.",
    "built_space": "레퍼런스와 일치하는 노을 진 항구, 등대, 정박된 어선이 배경에 올바르게 배치됨.",
    "entities": "노란 방수 점퍼를 입은 40대 남성(크리스)과 화면 왼쪽을 넓게 차지하는 후드 쓴 뒷모습(찰리), 배경의 어선이 존재함. 점퍼의 얼룩 표현은 다소 약함.",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "크리스의 시선이 화면 왼쪽 가장자리의 모포 쪽을 정확히 향하고 있음.",
    "built_space": "어두워지는 항구를 배경으로 낡은 어선이 인물 뒤에 배치되어 있음.",
    "entities": "비와 바닷물에 심하게 얼룩진 노란 방수 점퍼를 입은 40대 남성(크리스), 왼쪽 가장자리에 좁게 걸친 모포(찰리의 어깨), 배경의 낡은 어선이 지시대로 존재함.",
    "hard_violations": [],
    "physics": "인물이 지면에 안정적으로 서 있으며 옷깃과 자세가 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "왼쪽 가장자리에 모포의 어깨선만 좁게 배치하라는 엄격한 프레이밍 지시를 정확히 따랐으며, 얼룩진 노란색 방수 점퍼의 묘사도 지시문과 매우 잘 부합합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "배경과 조명은 레퍼런스를 잘 반영했으나, 왼쪽 가장자리에 모포만 좁게 보여야 한다는 프레이밍 지시를 어기고 후드를 쓴 인물의 머리와 뒷모습을 너무 넓게 노출했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "크리스의 시선이 화면 왼쪽 가장자리의 모포 쪽을 정확히 향하고 있음.",
        "built_space": "어두워지는 항구를 배경으로 낡은 어선이 인물 뒤에 배치되어 있음.",
        "entities": "비와 바닷물에 심하게 얼룩진 노란 방수 점퍼를 입은 40대 남성(크리스), 왼쪽 가장자리에 좁게 걸친 모포(찰리의 어깨), 배경의 낡은 어선이 지시대로 존재함.",
        "hard_violations": [],
        "physics": "인물이 지면에 안정적으로 서 있으며 옷깃과 자세가 자연스러움."
       },
       {
        "label": "A",
        "direction": "크리스의 시선이 화면 왼쪽에 자리 잡은 인물의 형태를 향하고 있음.",
        "built_space": "레퍼런스와 일치하는 노을 진 항구, 등대, 정박된 어선이 배경에 올바르게 배치됨.",
        "entities": "노란 방수 점퍼를 입은 40대 남성(크리스)과 화면 왼쪽을 넓게 차지하는 후드 쓴 뒷모습(찰리), 배경의 어선이 존재함. 점퍼의 얼룩 표현은 다소 약함.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "왼쪽 가장자리에 모포의 어깨선만 좁게 배치하라는 엄격한 프레이밍 지시를 정확히 따랐으며, 얼룩진 노란색 방수 점퍼의 묘사도 지시문과 매우 잘 부합합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "배경과 조명은 레퍼런스를 잘 반영했으나, 왼쪽 가장자리에 모포만 좁게 보여야 한다는 프레이밍 지시를 어기고 후드를 쓴 인물의 머리와 뒷모습을 너무 넓게 노출했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "크리스의 시선이 화면 왼쪽 가장자리의 모포 쪽을 정확히 향하고 있음.",
        "built_space": "어두워지는 항구를 배경으로 낡은 어선이 인물 뒤에 배치되어 있음.",
        "entities": "비와 바닷물에 심하게 얼룩진 노란 방수 점퍼를 입은 40대 남성(크리스), 왼쪽 가장자리에 좁게 걸친 모포(찰리의 어깨), 배경의 낡은 어선이 지시대로 존재함.",
        "hard_violations": [],
        "physics": "인물이 지면에 안정적으로 서 있으며 옷깃과 자세가 자연스러움."
       },
       {
        "label": "A",
        "direction": "크리스의 시선이 화면 왼쪽에 자리 잡은 인물의 형태를 향하고 있음.",
        "built_space": "레퍼런스와 일치하는 노을 진 항구, 등대, 정박된 어선이 배경에 올바르게 배치됨.",
        "entities": "노란 방수 점퍼를 입은 40대 남성(크리스)과 화면 왼쪽을 넓게 차지하는 후드 쓴 뒷모습(찰리), 배경의 어선이 존재함. 점퍼의 얼룩 표현은 다소 약함.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 모포를 의아하게 바라보는 크리스의 클로즈업, 얼룩진 노란 방수 점퍼와 흐릿한 배경 선박이 잘 맞지만 모포가 요구된 좁은 가장자리보다 넓다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "시선과 의아한 표정, 황혼의 장소는 맞지만 찰리의 머리 뒤와 넓은 등까지 전경에 넣어 ‘왼쪽 가장자리의 어깨 모포 주름만’이라는 핵심 구도를 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "크리스의 얼굴과 두 눈은 화면 왼쪽의 모포 쪽으로 향하며, 약간 위를 올려다본다. 시선의 대상은 카메라가 아니라 모포에 가려진 찰리다. 미간의 주름과 살짝 벌어진 입이 의아함을 드러낸다. 무기나 이동 동작은 없다.",
        "built_space": "크리스 뒤로 낡은 배 한 척의 선체 측면과 선실 일부가 보인다. 선실에는 정면 창 두 개와 측면 창 하나, 따뜻한 조명 두 곳이 식별되며, 오른쪽에 검은 배기관 하나와 대각선으로 걸린 굵은 밧줄이 있다. 오른쪽 아래에는 부두 가장자리 일부가 보인다. 참고 장소의 낡은 금속 선체·선실·계류줄과 어울리며, 배가 흐릿한 부분 배경으로 남는다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "크리스는 짧은 검은 머리의 동아시아계로 보이는 중년 남성으로, 40대 초반 설정과 대체로 맞는다. 두꺼운 노란 후드형 방수 점퍼에는 검은 얼룩과 마모가 뚜렷하다. 왼쪽에는 찰리를 가리는 회갈색 모포 일부만 보이며, 신체가 가려져 나이·성별·실제 체격은 확인할 수 없다. 모포는 세로로 긴 띠를 이루지만 요구된 좁은 어깨 주름보다 면적이 크다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "크리스의 머리와 상체는 자연스럽게 이어지며, 발과 지면 접촉은 클로즈업 밖이다. 왼쪽 모포는 화면 밖으로 이어지는 찰리의 몸 위에서 아래로 늘어진 천으로 읽히고, 공중에 독립적으로 떠 있는 물체는 없다. 배는 수면에 떠 있고 계류줄이 부두 방향으로 이어진다. 도약이나 충격처럼 별도 지지 설명이 필요한 동작은 없다."
       },
       {
        "label": "B",
        "direction": "크리스는 고개와 눈을 왼쪽으로 돌려 가까운 찰리의 모포 덮인 어깨와 상체를 바라본다. 미간을 찌푸리고 입을 조금 벌린 표정은 의아함에 부합한다. 카메라를 응시하지 않으며, 무기나 이동 동작은 없다.",
        "built_space": "왼쪽 뒤에 낡은 선박 한 척의 선실과 돛대가, 오른쪽에 부두와 바다가 보인다. 선실 정면 창 두 개와 측면 창 하나, 돛대 조명 하나와 선실 외벽 조명 하나, 검은 배기관 하나가 식별된다. 오른쪽 부두에는 계선주 세 개가 거리 순서대로 보이고, 먼 방파제에는 붉은 등대 하나가 있다. 참고 장소의 배치와 황혼 풍경은 잘 유지되지만 선실·부두·등대가 비교적 또렷하여 요구된 부드러운 선박 부분 배경보다 존재감이 강하다.",
        "entities": "크리스는 검은 머리의 동아시아계로 보이는 중년 남성이며, 노란 방수 외투와 검은 안쪽 상의를 입었다. 외투에 얼룩은 있으나 A보다 훨씬 깨끗해 보인다. 찰리는 왼쪽 전경에서 머리 뒤쪽의 덮개, 어깨와 넓은 등을 드러내며 모포가 체격을 감싼다. 이는 어깨 바깥 주름만 보이라는 지시보다 노출 범위가 크다. 추가 인물과 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "찰리의 모포는 머리·어깨·등의 윤곽에 걸쳐 지지되며 아래로 접혀 내려온다. 크리스의 목과 상체 자세도 자연스럽다. 두 사람의 발은 프레임 밖이므로 직접적인 지면 접촉은 확인할 수 없지만, 부유나 불가능한 자세의 증거는 없다. 계선주는 부두에 고정되어 있고 밧줄은 계선주에 감겨 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 모포를 의아하게 바라보는 크리스의 클로즈업, 얼룩진 노란 방수 점퍼와 흐릿한 배경 선박이 잘 맞지만 모포가 요구된 좁은 가장자리보다 넓다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "시선과 의아한 표정, 황혼의 장소는 맞지만 찰리의 머리 뒤와 넓은 등까지 전경에 넣어 ‘왼쪽 가장자리의 어깨 모포 주름만’이라는 핵심 구도를 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "크리스의 얼굴과 두 눈은 화면 왼쪽의 모포 쪽으로 향하며, 약간 위를 올려다본다. 시선의 대상은 카메라가 아니라 모포에 가려진 찰리다. 미간의 주름과 살짝 벌어진 입이 의아함을 드러낸다. 무기나 이동 동작은 없다.",
        "built_space": "크리스 뒤로 낡은 배 한 척의 선체 측면과 선실 일부가 보인다. 선실에는 정면 창 두 개와 측면 창 하나, 따뜻한 조명 두 곳이 식별되며, 오른쪽에 검은 배기관 하나와 대각선으로 걸린 굵은 밧줄이 있다. 오른쪽 아래에는 부두 가장자리 일부가 보인다. 참고 장소의 낡은 금속 선체·선실·계류줄과 어울리며, 배가 흐릿한 부분 배경으로 남는다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "크리스는 짧은 검은 머리의 동아시아계로 보이는 중년 남성으로, 40대 초반 설정과 대체로 맞는다. 두꺼운 노란 후드형 방수 점퍼에는 검은 얼룩과 마모가 뚜렷하다. 왼쪽에는 찰리를 가리는 회갈색 모포 일부만 보이며, 신체가 가려져 나이·성별·실제 체격은 확인할 수 없다. 모포는 세로로 긴 띠를 이루지만 요구된 좁은 어깨 주름보다 면적이 크다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "크리스의 머리와 상체는 자연스럽게 이어지며, 발과 지면 접촉은 클로즈업 밖이다. 왼쪽 모포는 화면 밖으로 이어지는 찰리의 몸 위에서 아래로 늘어진 천으로 읽히고, 공중에 독립적으로 떠 있는 물체는 없다. 배는 수면에 떠 있고 계류줄이 부두 방향으로 이어진다. 도약이나 충격처럼 별도 지지 설명이 필요한 동작은 없다."
       },
       {
        "label": "A",
        "direction": "크리스는 고개와 눈을 왼쪽으로 돌려 가까운 찰리의 모포 덮인 어깨와 상체를 바라본다. 미간을 찌푸리고 입을 조금 벌린 표정은 의아함에 부합한다. 카메라를 응시하지 않으며, 무기나 이동 동작은 없다.",
        "built_space": "왼쪽 뒤에 낡은 선박 한 척의 선실과 돛대가, 오른쪽에 부두와 바다가 보인다. 선실 정면 창 두 개와 측면 창 하나, 돛대 조명 하나와 선실 외벽 조명 하나, 검은 배기관 하나가 식별된다. 오른쪽 부두에는 계선주 세 개가 거리 순서대로 보이고, 먼 방파제에는 붉은 등대 하나가 있다. 참고 장소의 배치와 황혼 풍경은 잘 유지되지만 선실·부두·등대가 비교적 또렷하여 요구된 부드러운 선박 부분 배경보다 존재감이 강하다.",
        "entities": "크리스는 검은 머리의 동아시아계로 보이는 중년 남성이며, 노란 방수 외투와 검은 안쪽 상의를 입었다. 외투에 얼룩은 있으나 A보다 훨씬 깨끗해 보인다. 찰리는 왼쪽 전경에서 머리 뒤쪽의 덮개, 어깨와 넓은 등을 드러내며 모포가 체격을 감싼다. 이는 어깨 바깥 주름만 보이라는 지시보다 노출 범위가 크다. 추가 인물과 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "찰리의 모포는 머리·어깨·등의 윤곽에 걸쳐 지지되며 아래로 접혀 내려온다. 크리스의 목과 상체 자세도 자연스럽다. 두 사람의 발은 프레임 밖이므로 직접적인 지면 접촉은 확인할 수 없지만, 부유나 불가능한 자세의 증거는 없다. 계선주는 부두에 고정되어 있고 밧줄은 계선주에 감겨 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.464,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.464,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1464
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "왼쪽 가장자리에 모포의 어깨선만 좁게 배치하라는 엄격한 프레이밍 지시를 정확히 따랐으며, 얼룩진 노란색 방수 점퍼의 묘사도 지시문과 매우 잘 부합합니다."
   },
   {
    "label": "A",
    "score": 1464,
    "verdict_ko": "배경과 조명은 레퍼런스를 잘 반영했으나, 왼쪽 가장자리에 모포만 좁게 보여야 한다는 프레이밍 지시를 어기고 후드를 쓴 인물의 머리와 뒷모습을 너무 넓게 노출했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_boat_boarding_berth_655ef6.png",
    "asset_id": "a29b91e8-e78c-4774-b743-3959ad9e6734",
    "role": "bgfirst_group_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacec8-2c9f-73d7-b7fa-23b7a91c82ca",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S78sh14__bgfirst_bg.png",
   "bg_asset_id": "a24895dc-4b0f-4532-9a5c-dc697c5d2c93",
   "bg_record_key": "S78sh14::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "boat_boarding_berth",
   "groupbg_asset_id": "a29b91e8-e78c-4774-b743-3959ad9e6734"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S78sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:48:59.183499+00:00",
  "fingerprint": "1296348439b9c1f8879673048805faf618b987716d7d671bca6919349db601d4",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S78sh14_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S78sh14_sel.png",
  "source_sha256": "11726be05f7b832da6365c13aa1106ae1a872166786afff6d2840737f03b0f8e",
  "file": "S78sh14_cine.png",
  "staged_sha256": "ddc87f120a3185e8687b55be92b3fc8181122d1f2483ba941b9b67e98a90a5f9",
  "latency_ms": 12407
 },
 "S78sh19::signage": {
  "fp": "703410cf4b063ee2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S78sh19::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:50:05.907518+00:00",
  "fingerprint": "3989ff8e6ff19523e554c98c07c0314373aa08f963b22d2bd9de39f5ea559b59",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S78sh19_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S78sh19_sel.png",
  "source_sha256": "3ef0a4d5069da479b7ab8c05e838ac5a5e2bbc3443face61261e67561f7e4783",
  "file": "S78sh19_cine.png",
  "staged_sha256": "2e92f4d8e8cf4b1e8afb4be444b5524acc47844ad76c5fe434060cd232cfbe3f",
  "latency_ms": 11758
 },
 "S79sh6::signage": {
  "fp": "515d5a863b5d62a2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::ffaee7876c805584": {
  "subjects": [],
  "subject_text": "크리스의 선박 선원실\n두 명 정도가 쉴 수 있는 작은 선원실. 벽 가까이에 소파가 놓여 있고 출입문과 좌석 사이의 여유 공간이 좁다.",
  "identity": "canonical",
  "scope_id": "L123",
  "scope_role": "location_interior",
  "scope_sha": "dcf9fcb3ead632f6"
 },
 "S79sh6::bgfirst_bg": {
  "input_fingerprint": "ea5a5e8a8bc3bf64",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우 옆에서 찰리가 등받이에 기댄 채 눈가 금속 부품을 아래로 떨군 묘한 표정.\n\nLOCATION (lock): On the sofa inside a small two-person crew cabin aboard the vessel, under subdued cabin lighting at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sofa (Supporting both reclining figures) — Seat and backrest are viewed obliquely from beside 이현우; used as Shared support plane connecting the foreground listener and 찰리; Small cabin wall (Enclosing the seated figures) — Visible behind the sofa; used as Restrains background depth and reinforces the intimate shared space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient cabin illumination preserves facial and mechanical detail with quiet contrast and no perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우 옆에서 찰리가 등받이에 기댄 채 눈가 금속 부품을 아래로 떨군 묘한 표정.\n\nLOCATION (lock): On the sofa inside a small two-person crew cabin aboard the vessel, under subdued cabin lighting at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sofa (Supporting both reclining figures) — Seat and backrest are viewed obliquely from beside 이현우; used as Shared support plane connecting the foreground listener and 찰리; Small cabin wall (Enclosing the seated figures) — Visible behind the sofa; used as Restrains background depth and reinforces the intimate shared space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient cabin illumination preserves facial and mechanical detail with quiet contrast and no perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S79sh6__bgfirst_bg.png",
  "asset_id": "bc30f250-b6d4-44ad-ab87-8a1b5db90cc7",
  "input_asset_ids": [
   "9fa928a6-95e5-4ed7-8867-e60207229ca5",
   "7833fa7b-2804-404e-9e53-0cffdffa29c7"
  ]
 },
 "S79sh6": {
  "input_fingerprint": "3ad2349fc44ca3fe",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우 옆에서 찰리가 등받이에 기댄 채 눈가 금속 부품을 아래로 떨군 묘한 표정.\n\nLOCATION (lock): On the sofa inside a small two-person crew cabin aboard the vessel, under subdued cabin lighting at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sofa (Supporting both reclining figures) — Seat and backrest are viewed obliquely from beside 이현우; used as Shared support plane connecting the foreground listener and 찰리; Small cabin wall (Enclosing the seated figures) — Visible behind the sofa; used as Restrains background depth and reinforces the intimate shared space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient cabin illumination preserves facial and mechanical detail with quiet contrast and no perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small two-person cabin contains a sofa, and its door is closed. The boat remains underway at night. 이현우: He reclines on the sofa with his legs stretched out, retaining his backpack. His grime and accumulated injuries remain. 찰리: He reclines against the sofa back, still covered by the blanket and retaining his backpack and salvaged chest component. His unrepaired body damage remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우 옆에서 찰리가 등받이에 기댄 채 눈가 금속 부품을 아래로 떨군 묘한 표정.\n\nLOCATION (lock): On the sofa inside a small two-person crew cabin aboard the vessel, under subdued cabin lighting at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sofa (Supporting both reclining figures) — Seat and backrest are viewed obliquely from beside 이현우; used as Shared support plane connecting the foreground listener and 찰리; Small cabin wall (Enclosing the seated figures) — Visible behind the sofa; used as Restrains background depth and reinforces the intimate shared space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient cabin illumination preserves facial and mechanical detail with quiet contrast and no perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small two-person cabin contains a sofa, and its door is closed. The boat remains underway at night. 이현우: He reclines on the sofa with his legs stretched out, retaining his backpack. His grime and accumulated injuries remain. 찰리: He reclines against the sofa back, still covered by the blanket and retaining his backpack and salvaged chest component. His unrepaired body damage remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우 옆에서 찰리가 등받이에 기댄 채 눈가 금속 부품을 아래로 떨군 묘한 표정.\n\nLOCATION (lock): On the sofa inside a small two-person crew cabin aboard the vessel, under subdued cabin lighting at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sofa (Supporting both reclining figures) — Seat and backrest are viewed obliquely from beside 이현우; used as Shared support plane connecting the foreground listener and 찰리; Small cabin wall (Enclosing the seated figures) — Visible behind the sofa; used as Restrains background depth and reinforces the intimate shared space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient cabin illumination preserves facial and mechanical detail with quiet contrast and no perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small two-person cabin contains a sofa, and its door is closed. The boat remains underway at night. 이현우: He reclines on the sofa with his legs stretched out, retaining his backpack. His grime and accumulated injuries remain. 찰리: He reclines against the sofa back, still covered by the blanket and retaining his backpack and salvaged chest component. His unrepaired body damage remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S79sh6__bgfirst_bg.png",
     "asset_id": "bc30f250-b6d4-44ad-ab87-8a1b5db90cc7",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S79sh6.png",
     "asset_id": "9fa928a6-95e5-4ed7-8867-e60207229ca5",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:948699>",
     "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L123B02.png",
     "asset_id": "7833fa7b-2804-404e-9e53-0cffdffa29c7",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:948699>",
     "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "전경의 이현우가 찰리를 보고, 찰리는 정면 아래를 향함.",
    "built_space": "카메라가 이현우 옆에 위치함. 문 옆 벽면에 기준에 없는 붉은 소화함이 추가됨.",
    "entities": "이현우 반팔 착용(후드티 누락). 찰리의 망토와 가슴 부품 누락. 소화함에 'FIRE BLANKET' 텍스트 노출.",
    "hard_violations": [
     "[gemini-pro] 기준 장소에 없는 벽면 소화함 임의 생성",
     "[gemini-pro] 읽을 수 있는 영문 텍스트(FIRE BLANKET) 노출",
     "[gpt-high] 참고 장소에 없는 붉은 소화 담요 주머니를 추가했고, 그 표면의 영문 표제가 읽혀 명시적인 문자 금지 조건을 위반한다."
    ],
    "physics": "두 인물 모두 소파에 앉아 안정적으로 지지됨."
   },
   {
    "label": "B",
    "direction": "찰리는 앞을 향하고, 배경의 이현우는 찰리를 바라봄.",
    "built_space": "선실 내부 구조와 일치. 단 카메라는 이현우가 아닌 찰리 옆에 배치되어 프레이밍 지시를 어김.",
    "entities": "찰리의 담요 망토, 백팩, 떨어진 금속 부품 구현됨. 이현우의 복장과 부상 상태 기준과 일치함.",
    "hard_violations": [],
    "physics": "두 인물 모두 소파에 정상적으로 기대어 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지시된 전경 배치는 지켰으나, 금지된 텍스트 노출과 임의의 구조물 생성(하드 위반), 심각한 복장 오류로 탈락."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "카메라 앵글(전후경 배치)이 지시와 반대지만, 찰리의 안면 부품 이탈과 두 인물의 복장 상태를 충실히 구현함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경의 이현우가 찰리를 보고, 찰리는 정면 아래를 향함.",
        "built_space": "카메라가 이현우 옆에 위치함. 문 옆 벽면에 기준에 없는 붉은 소화함이 추가됨.",
        "entities": "이현우 반팔 착용(후드티 누락). 찰리의 망토와 가슴 부품 누락. 소화함에 'FIRE BLANKET' 텍스트 노출.",
        "hard_violations": [
         "기준 장소에 없는 벽면 소화함 임의 생성",
         "읽을 수 있는 영문 텍스트(FIRE BLANKET) 노출"
        ],
        "physics": "두 인물 모두 소파에 앉아 안정적으로 지지됨."
       },
       {
        "label": "B",
        "direction": "찰리는 앞을 향하고, 배경의 이현우는 찰리를 바라봄.",
        "built_space": "선실 내부 구조와 일치. 단 카메라는 이현우가 아닌 찰리 옆에 배치되어 프레이밍 지시를 어김.",
        "entities": "찰리의 담요 망토, 백팩, 떨어진 금속 부품 구현됨. 이현우의 복장과 부상 상태 기준과 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 소파에 정상적으로 기대어 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지시된 전경 배치는 지켰으나, 금지된 텍스트 노출과 임의의 구조물 생성(하드 위반), 심각한 복장 오류로 탈락."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "카메라 앵글(전후경 배치)이 지시와 반대지만, 찰리의 안면 부품 이탈과 두 인물의 복장 상태를 충실히 구현함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "전경의 이현우가 찰리를 보고, 찰리는 정면 아래를 향함.",
        "built_space": "카메라가 이현우 옆에 위치함. 문 옆 벽면에 기준에 없는 붉은 소화함이 추가됨.",
        "entities": "이현우 반팔 착용(후드티 누락). 찰리의 망토와 가슴 부품 누락. 소화함에 'FIRE BLANKET' 텍스트 노출.",
        "hard_violations": [
         "기준 장소에 없는 벽면 소화함 임의 생성",
         "읽을 수 있는 영문 텍스트(FIRE BLANKET) 노출"
        ],
        "physics": "두 인물 모두 소파에 앉아 안정적으로 지지됨."
       },
       {
        "label": "B",
        "direction": "찰리는 앞을 향하고, 배경의 이현우는 찰리를 바라봄.",
        "built_space": "선실 내부 구조와 일치. 단 카메라는 이현우가 아닌 찰리 옆에 배치되어 프레이밍 지시를 어김.",
        "entities": "찰리의 담요 망토, 백팩, 떨어진 금속 부품 구현됨. 이현우의 복장과 부상 상태 기준과 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 소파에 정상적으로 기대어 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 인물의 기대는 자세와 담요·배낭·야간 선실은 잘 구현했지만, 이현우 대신 찰리를 전경에 두어 지정된 관찰 위치와 어긋난다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이현우 옆에서 찰리를 보는 배치는 더 정확하지만, 배경에 새로 추가한 소화 담요 주머니의 읽히는 영문이 문자 금지 조건을 명백히 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 이현우는 왼쪽의 찰리 얼굴을 바라본다. 찰리는 고개를 약간 뒤로 기댄 채 앞쪽 아래를 향하며, 눈 주변 판이 내려와 눈이 가늘게 보인다. 특정 물체를 응시하는지는 불명확하지만 카메라를 정면으로 바라보지는 않는다.",
        "built_space": "회색 천 소파 하나에 찰리가 왼쪽 전경, 이현우가 오른쪽 후경으로 나란히 기대 있다. 뒤에는 목재 상부 수납장과 책이 있는 개방 선반, 벽 사진들, 수납장 아래 조명 하나, 오른쪽의 닫힌 금속문 하나가 보인다. 참고 장소의 주요 재료와 배치는 유지한다. 다만 소파를 비스듬히 보는 카메라가 이현우 옆이 아니라 찰리 옆에 있어 전경 청자와 후경 찰리라는 구도를 뒤집었다.",
        "entities": "등장인물은 둘뿐이다. 찰리는 흰 고릴라형 얼굴판, 금속 머리 외장, 선형 입, 손상과 마모가 있는 기계 몸체를 갖추고 회색 담요와 배낭을 유지한다. 머리는 참고보다 둥글고 눈은 점보다는 좁은 틈에 가깝다. 이현우는 젊은 한국계 남성으로 읽히며 헝클어진 검은 머리, 더러운 후드 상의, 찢어진 청바지와 배낭이 참고에 부합한다. 찰리 무릎에는 여러 작은 금속 조각이 있지만 지정된 회수 흉부 부품 하나로 명확하게 식별되지는 않는다. 읽을 수 있는 문자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 엉덩이는 소파 좌면에, 등과 배낭은 등받이에 받쳐져 있다. 이현우는 한쪽 무릎을 세우고 다른 다리를 소파 위로 뻗고 있어 완전히 두 다리를 뻗은 상태는 아니다. 찰리의 손과 금속 조각들은 허벅지 위 담요에 놓여 있고 담요도 몸과 좌면에 지지된다. 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 전경의 이현우는 오른쪽 찰리를 바라본다. 찰리는 턱과 얼굴을 아래로 숙여 무릎 쪽을 향하고, 눈 옆의 기계 부품이 아래로 늘어져 있다. 청자와 찰리의 방향 관계 및 눈가 부품을 떨군 순간은 비교적 명확하다.",
        "built_space": "회색 소파 하나의 왼쪽에 이현우, 오른쪽에 찰리가 앉고 가운데 좌석 일부가 비어 있다. 이현우 옆에서 좌면과 등받이를 비스듬히 보는 구도는 지시에 맞지만, 무릎 아래 담요와 바닥까지 담아 미디엄 숏보다 넓다. 목재 상부장, 개방 책 선반, 벽 사진들, 하부 조명 하나와 오른쪽 닫힌 금속문 하나가 보인다. 참고에는 없는 붉은 소화 담요 주머니가 문 왼쪽 벽에 추가되어 있다.",
        "entities": "등장인물은 이현우와 찰리 둘이다. 이현우는 헝클어진 검은 머리와 상처·때가 있는 젊은 한국계 남성으로 보이지만, 참고의 후드 상의 대신 검은 반소매 티셔츠를 입었다. 배낭 끈은 보인다. 찰리는 흰 얼굴판과 어두운 기계 몸체를 유지하나 머리가 참고보다 둥글다. 담요는 하체만 덮으며 배낭과 별도 회수 흉부 부품은 명확히 식별되지 않는다. 붉은 벽 주머니의 영문 표제는 읽을 수 있다.",
        "hard_violations": [
         "참고 장소에 없는 붉은 소화 담요 주머니를 추가했고, 그 표면의 영문 표제가 읽혀 명시적인 문자 금지 조건을 위반한다."
        ],
        "physics": "두 사람 모두 소파 좌면에 앉아 등받이에 등을 기대며, 손은 담요로 덮인 허벅지 위에 놓여 있다. 찰리의 눈 옆 부품은 머리 측면에 연결된 채 아래로 매달려 보인다. 담요는 무릎과 좌면에 지지되고 찰리 쪽 끝자락은 바닥으로 내려온다. 다리는 담요에 가려 뻗은 정도를 확인할 수 없으며, 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "두 인물의 기대는 자세와 담요·배낭·야간 선실은 잘 구현했지만, 이현우 대신 찰리를 전경에 두어 지정된 관찰 위치와 어긋난다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이현우 옆에서 찰리를 보는 배치는 더 정확하지만, 배경에 새로 추가한 소화 담요 주머니의 읽히는 영문이 문자 금지 조건을 명백히 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 이현우는 왼쪽의 찰리 얼굴을 바라본다. 찰리는 고개를 약간 뒤로 기댄 채 앞쪽 아래를 향하며, 눈 주변 판이 내려와 눈이 가늘게 보인다. 특정 물체를 응시하는지는 불명확하지만 카메라를 정면으로 바라보지는 않는다.",
        "built_space": "회색 천 소파 하나에 찰리가 왼쪽 전경, 이현우가 오른쪽 후경으로 나란히 기대 있다. 뒤에는 목재 상부 수납장과 책이 있는 개방 선반, 벽 사진들, 수납장 아래 조명 하나, 오른쪽의 닫힌 금속문 하나가 보인다. 참고 장소의 주요 재료와 배치는 유지한다. 다만 소파를 비스듬히 보는 카메라가 이현우 옆이 아니라 찰리 옆에 있어 전경 청자와 후경 찰리라는 구도를 뒤집었다.",
        "entities": "등장인물은 둘뿐이다. 찰리는 흰 고릴라형 얼굴판, 금속 머리 외장, 선형 입, 손상과 마모가 있는 기계 몸체를 갖추고 회색 담요와 배낭을 유지한다. 머리는 참고보다 둥글고 눈은 점보다는 좁은 틈에 가깝다. 이현우는 젊은 한국계 남성으로 읽히며 헝클어진 검은 머리, 더러운 후드 상의, 찢어진 청바지와 배낭이 참고에 부합한다. 찰리 무릎에는 여러 작은 금속 조각이 있지만 지정된 회수 흉부 부품 하나로 명확하게 식별되지는 않는다. 읽을 수 있는 문자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 엉덩이는 소파 좌면에, 등과 배낭은 등받이에 받쳐져 있다. 이현우는 한쪽 무릎을 세우고 다른 다리를 소파 위로 뻗고 있어 완전히 두 다리를 뻗은 상태는 아니다. 찰리의 손과 금속 조각들은 허벅지 위 담요에 놓여 있고 담요도 몸과 좌면에 지지된다. 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 전경의 이현우는 오른쪽 찰리를 바라본다. 찰리는 턱과 얼굴을 아래로 숙여 무릎 쪽을 향하고, 눈 옆의 기계 부품이 아래로 늘어져 있다. 청자와 찰리의 방향 관계 및 눈가 부품을 떨군 순간은 비교적 명확하다.",
        "built_space": "회색 소파 하나의 왼쪽에 이현우, 오른쪽에 찰리가 앉고 가운데 좌석 일부가 비어 있다. 이현우 옆에서 좌면과 등받이를 비스듬히 보는 구도는 지시에 맞지만, 무릎 아래 담요와 바닥까지 담아 미디엄 숏보다 넓다. 목재 상부장, 개방 책 선반, 벽 사진들, 하부 조명 하나와 오른쪽 닫힌 금속문 하나가 보인다. 참고에는 없는 붉은 소화 담요 주머니가 문 왼쪽 벽에 추가되어 있다.",
        "entities": "등장인물은 이현우와 찰리 둘이다. 이현우는 헝클어진 검은 머리와 상처·때가 있는 젊은 한국계 남성으로 보이지만, 참고의 후드 상의 대신 검은 반소매 티셔츠를 입었다. 배낭 끈은 보인다. 찰리는 흰 얼굴판과 어두운 기계 몸체를 유지하나 머리가 참고보다 둥글다. 담요는 하체만 덮으며 배낭과 별도 회수 흉부 부품은 명확히 식별되지 않는다. 붉은 벽 주머니의 영문 표제는 읽을 수 있다.",
        "hard_violations": [
         "참고 장소에 없는 붉은 소화 담요 주머니를 추가했고, 그 표면의 영문 표제가 읽혀 명시적인 문자 금지 조건을 위반한다."
        ],
        "physics": "두 사람 모두 소파 좌면에 앉아 등받이에 등을 기대며, 손은 담요로 덮인 허벅지 위에 놓여 있다. 찰리의 눈 옆 부품은 머리 측면에 연결된 채 아래로 매달려 보인다. 담요는 무릎과 좌면에 지지되고 찰리 쪽 끝자락은 바닥으로 내려온다. 다리는 담요에 가려 뻗은 정도를 확인할 수 없으며, 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.929,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.679,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 기준 장소에 없는 벽면 소화함 임의 생성",
     "[gemini-pro] 읽을 수 있는 영문 텍스트(FIRE BLANKET) 노출",
     "[gpt-high] 참고 장소에 없는 붉은 소화 담요 주머니를 추가했고, 그 표면의 영문 표제가 읽혀 명시적인 문자 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 679,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 679,
    "verdict_ko": "지시된 전경 배치는 지켰으나, 금지된 텍스트 노출과 임의의 구조물 생성(하드 위반), 심각한 복장 오류로 탈락.  ★위반: [gemini-pro] 기준 장소에 없는 벽면 소화함 임의 생성 / [gemini-pro] 읽을 수 있는 영문 텍스트(FIRE BLANKET) 노출 / [gpt-high] 참고 장소에 없는 붉은 소화 담요 주머니를 추가했고, 그 표면의 영문 표제가 읽혀 명시적인 문자 금지 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "카메라 앵글(전후경 배치)이 지시와 반대지만, 찰리의 안면 부품 이탈과 두 인물의 복장 상태를 충실히 구현함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L123B02.png",
    "asset_id": "7833fa7b-2804-404e-9e53-0cffdffa29c7",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:948699>",
    "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aaced3-a08a-7d46-aa14-4db26039ebc6",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S79sh6__bgfirst_bg.png",
   "bg_asset_id": "bc30f250-b6d4-44ad-ab87-8a1b5db90cc7",
   "bg_record_key": "S79sh6::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S79sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:52:39.022207+00:00",
  "fingerprint": "519dbda7c3bbe33b4096b8d982d5f3be31c8bbb8d2214bb31ca4354334f2b016",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S79sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S79sh6_sel.png",
  "source_sha256": "ca4b42a4b2a4f987e88c41064c4012770a57fe16a6d738f1cbb409a1e54f2388",
  "file": "S79sh6_cine.png",
  "staged_sha256": "728211f78458678f1f892f304249ed2f372966148591bbbfc4df5e0c1dbfb5af",
  "latency_ms": 13737
 },
 "S79sh8::signage": {
  "fp": "7eeb28e66170fe39",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S79sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:54:07.454907+00:00",
  "fingerprint": "d0ab31a4631de2f17e52943f13d2fd421d3a935802867b8b1f89eeca39fbde65",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S79sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S79sh8_sel.png",
  "source_sha256": "1834f650b4c6ca15ac53d026088e87a70476b4ff36f5196fafe65e252914f69b",
  "file": "S79sh8_cine.png",
  "staged_sha256": "ff918e2b8b1c89a168e3183ca5a566cfc50bb02caa3867123f55e5db00c1b065",
  "latency_ms": 12178
 },
 "S79sh9::signage": {
  "fp": "3f560a4efe6277f4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S79sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:55:18.768525+00:00",
  "fingerprint": "39cde076720cda1f17b6de38a28803d11dc64e1d3735a17b1f301e1aa6b108ec",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S79sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S79sh9_sel.png",
  "source_sha256": "83d35f4ec91f6f69d1e2f2968154cf0e7bce3969bc19e58f0949baef3178c8e4",
  "file": "S79sh9_cine.png",
  "staged_sha256": "208b9b48003f2597f4037424cf02e02183308648bad10f6de36fa113ca43d93e",
  "latency_ms": 11359
 },
 "S80sh6::signage": {
  "fp": "7f519594782007d7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::6c3851ac6a9e965d": {
  "subjects": [],
  "subject_text": "크리스의 선박 갑판과 주변 해상\n낡은 소형 선박의 좁고 개방된 갑판. 붙잡을 수 있는 기둥과 가장자리 구조물이 있으며 주변으로 바다가 펼쳐진다.",
  "identity": "canonical",
  "scope_id": "L122",
  "scope_role": "location_exterior",
  "scope_sha": "2e6ecf2a61b6c645"
 },
 "S80sh6::bgfirst_bg": {
  "input_fingerprint": "993167ce503ae408",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 거대한 금속 손이 허공에 뜬 이현우의 손목을 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): At the exposed edge of the vessel's storm-battered deck, beside the post used to brace against the waves.\n\nTIME OF DAY (lock): night, typhoon and torrential rain.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ship's edge (Exposed to torrential rain and waves) — Runs obliquely behind the connected arms; used as Separates 찰리's support side from 이현우's suspended position; Pillar (Gripped by 찰리 with his other hand) — A partial side is visible at the upper-left margin; used as Peripheral evidence of the rescue's physical support.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime illumination through torrential rain gives the wet skin and metal readable highlights without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 거대한 금속 손이 허공에 뜬 이현우의 손목을 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): At the exposed edge of the vessel's storm-battered deck, beside the post used to brace against the waves.\n\nTIME OF DAY (lock): night, typhoon and torrential rain.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ship's edge (Exposed to torrential rain and waves) — Runs obliquely behind the connected arms; used as Separates 찰리's support side from 이현우's suspended position; Pillar (Gripped by 찰리 with his other hand) — A partial side is visible at the upper-left margin; used as Peripheral evidence of the rescue's physical support.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime illumination through torrential rain gives the wet skin and metal readable highlights without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S80sh6__bgfirst_bg.png",
  "asset_id": "1d7f2f31-f0ce-4f62-9b96-d470b5cf7891",
  "input_asset_ids": [
   "875ca88c-cf91-4447-a5cd-cd2e8f199f32",
   "6abdf41d-41d4-42ad-934a-d5ad17b629b4"
  ]
 },
 "S80sh6": {
  "input_fingerprint": "1ff3fdebd93e692b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, typhoon and torrential rain.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 손이 허공에 뜬 이현우의 손목을 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): At the exposed edge of the vessel's storm-battered deck, beside the post used to brace against the waves. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ship's edge (Exposed to torrential rain and waves) — Runs obliquely behind the connected arms; used as Separates 찰리's support side from 이현우's suspended position; Pillar (Gripped by 찰리 with his other hand) — A partial side is visible at the upper-left margin; used as Peripheral evidence of the rescue's physical support.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime illumination through torrential rain gives the wet skin and metal readable highlights without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Torrential rain and typhoon-driven waves batter the boat at night, severely limiting visibility. The deck and its upright post are drenched. 이현우: He is soaked and suspended at the edge of the boat, with one arm stretched upward. His backpack and earlier injuries remain. 찰리: He braces on deck with one hand gripping a post and the other extended outward. He is soaked, with his earlier body damage, blanket, backpack and salvaged chest component retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, typhoon and torrential rain.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 손이 허공에 뜬 이현우의 손목을 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): At the exposed edge of the vessel's storm-battered deck, beside the post used to brace against the waves. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ship's edge (Exposed to torrential rain and waves) — Runs obliquely behind the connected arms; used as Separates 찰리's support side from 이현우's suspended position; Pillar (Gripped by 찰리 with his other hand) — A partial side is visible at the upper-left margin; used as Peripheral evidence of the rescue's physical support.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime illumination through torrential rain gives the wet skin and metal readable highlights without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Torrential rain and typhoon-driven waves batter the boat at night, severely limiting visibility. The deck and its upright post are drenched. 이현우: He is soaked and suspended at the edge of the boat, with one arm stretched upward. His backpack and earlier injuries remain. 찰리: He braces on deck with one hand gripping a post and the other extended outward. He is soaked, with his earlier body damage, blanket, backpack and salvaged chest component retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, typhoon and torrential rain.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 손이 허공에 뜬 이현우의 손목을 빈틈없이 꽉 틀어쥔 타격감 있는 클로즈업.\n\nLOCATION (lock): At the exposed edge of the vessel's storm-battered deck, beside the post used to brace against the waves. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ship's edge (Exposed to torrential rain and waves) — Runs obliquely behind the connected arms; used as Separates 찰리's support side from 이현우's suspended position; Pillar (Gripped by 찰리 with his other hand) — A partial side is visible at the upper-left margin; used as Peripheral evidence of the rescue's physical support.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime illumination through torrential rain gives the wet skin and metal readable highlights without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Torrential rain and typhoon-driven waves batter the boat at night, severely limiting visibility. The deck and its upright post are drenched. 이현우: He is soaked and suspended at the edge of the boat, with one arm stretched upward. His backpack and earlier injuries remain. 찰리: He braces on deck with one hand gripping a post and the other extended outward. He is soaked, with his earlier body damage, blanket, backpack and salvaged chest component retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S80sh6__bgfirst_bg.png",
     "asset_id": "1d7f2f31-f0ce-4f62-9b96-d470b5cf7891",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S80sh6.png",
     "asset_id": "875ca88c-cf91-4447-a5cd-cd2e8f199f32",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:948699>",
     "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L122B01.png",
     "asset_id": "6abdf41d-41d4-42ad-934a-d5ad17b629b4",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:948699>",
     "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917742>",
     "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 금속 팔이 아래를 향해 뻗어 이현우의 오른팔 목을 정확히 겨냥해 쥐고 있으며, 이현우의 시선은 위쪽의 찰리를 향함.",
    "built_space": "선박 갑판 가장자리. 좌측 가장자리에 기둥이 보이고 난간이 대각선으로 프레임을 가로지름. 찰리는 난간 안쪽, 이현우는 바다 쪽 바깥에 위치함.",
    "entities": "찰리는 금속 팔을 가졌으나 복장이 레퍼런스의 기계 몸체와 다름. 이현우는 얼굴은 일치하나 긴소매 대신 짧은 소매의 상의를 입고 있음.",
    "hard_violations": [
     "[gpt-high] 장소 참조의 통로에 없는 별도의 내부 금속 난간을 추가하여 갑판의 고정 구조와 구조 동작의 경계를 바꾸었다.",
     "[gpt-high] 선외에 매달려 있어야 할 이현우를 바깥 선측 벽 안쪽의 갑판 영역에 배치했다."
    ],
    "physics": "이현우는 난간 밖 허공에 매달려 있으며, 찰리의 강한 그립에 의해 체중이 전적으로 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "찰리는 고개를 숙여 아래를 내려다보고, 이현우는 위를 올려다봄. 찰리의 팔이 난간 너머로 뻗어 이현우의 손을 잡음.",
    "built_space": "선박 갑판. 좌측에 기둥이 있고 난간이 있음. 찰리는 갑판 위에서 한 손을 난간에 기대고 있고, 이현우는 선체 외벽에 매달려 있음.",
    "entities": "찰리의 로봇 머리, 망토, 배낭이 레퍼런스와 정확히 일치함. 이현우 역시 젖은 후드티와 배낭 등 지정된 외형을 잘 반영함.",
    "hard_violations": [],
    "physics": "이현우는 한 손을 찰리에게 잡힌 채 다른 손과 발로 선체 외벽을 디디고 버티고 있어, 완전히 허공에 뜬 상태의 하중을 보여주지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "클로즈업 프레임 척도와 손목을 강하게 틀어쥔 액션, 허공에 매달린 물리적 상태 등 우선순위가 높은 샷 연출 지시를 잘 따랐으나 인물들의 복장 일치도는 다소 떨어집니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "캐릭터 외형과 복장의 일치도는 훌륭하지만, 클로즈업 지시를 무시하고 프레임을 넓게 잡았으며 지정된 손목 그립 대신 손을 맞잡는 자세를 취해 핵심 연출에서 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 금속 팔이 아래를 향해 뻗어 이현우의 오른팔 목을 정확히 겨냥해 쥐고 있으며, 이현우의 시선은 위쪽의 찰리를 향함.",
        "built_space": "선박 갑판 가장자리. 좌측 가장자리에 기둥이 보이고 난간이 대각선으로 프레임을 가로지름. 찰리는 난간 안쪽, 이현우는 바다 쪽 바깥에 위치함.",
        "entities": "찰리는 금속 팔을 가졌으나 복장이 레퍼런스의 기계 몸체와 다름. 이현우는 얼굴은 일치하나 긴소매 대신 짧은 소매의 상의를 입고 있음.",
        "hard_violations": [],
        "physics": "이현우는 난간 밖 허공에 매달려 있으며, 찰리의 강한 그립에 의해 체중이 전적으로 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "찰리는 고개를 숙여 아래를 내려다보고, 이현우는 위를 올려다봄. 찰리의 팔이 난간 너머로 뻗어 이현우의 손을 잡음.",
        "built_space": "선박 갑판. 좌측에 기둥이 있고 난간이 있음. 찰리는 갑판 위에서 한 손을 난간에 기대고 있고, 이현우는 선체 외벽에 매달려 있음.",
        "entities": "찰리의 로봇 머리, 망토, 배낭이 레퍼런스와 정확히 일치함. 이현우 역시 젖은 후드티와 배낭 등 지정된 외형을 잘 반영함.",
        "hard_violations": [],
        "physics": "이현우는 한 손을 찰리에게 잡힌 채 다른 손과 발로 선체 외벽을 디디고 버티고 있어, 완전히 허공에 뜬 상태의 하중을 보여주지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "클로즈업 프레임 척도와 손목을 강하게 틀어쥔 액션, 허공에 매달린 물리적 상태 등 우선순위가 높은 샷 연출 지시를 잘 따랐으나 인물들의 복장 일치도는 다소 떨어집니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "캐릭터 외형과 복장의 일치도는 훌륭하지만, 클로즈업 지시를 무시하고 프레임을 넓게 잡았으며 지정된 손목 그립 대신 손을 맞잡는 자세를 취해 핵심 연출에서 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 금속 팔이 아래를 향해 뻗어 이현우의 오른팔 목을 정확히 겨냥해 쥐고 있으며, 이현우의 시선은 위쪽의 찰리를 향함.",
        "built_space": "선박 갑판 가장자리. 좌측 가장자리에 기둥이 보이고 난간이 대각선으로 프레임을 가로지름. 찰리는 난간 안쪽, 이현우는 바다 쪽 바깥에 위치함.",
        "entities": "찰리는 금속 팔을 가졌으나 복장이 레퍼런스의 기계 몸체와 다름. 이현우는 얼굴은 일치하나 긴소매 대신 짧은 소매의 상의를 입고 있음.",
        "hard_violations": [],
        "physics": "이현우는 난간 밖 허공에 매달려 있으며, 찰리의 강한 그립에 의해 체중이 전적으로 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "찰리는 고개를 숙여 아래를 내려다보고, 이현우는 위를 올려다봄. 찰리의 팔이 난간 너머로 뻗어 이현우의 손을 잡음.",
        "built_space": "선박 갑판. 좌측에 기둥이 있고 난간이 있음. 찰리는 갑판 위에서 한 손을 난간에 기대고 있고, 이현우는 선체 외벽에 매달려 있음.",
        "entities": "찰리의 로봇 머리, 망토, 배낭이 레퍼런스와 정확히 일치함. 이현우 역시 젖은 후드티와 배낭 등 지정된 외형을 잘 반영함.",
        "hard_violations": [],
        "physics": "이현우는 한 손을 찰리에게 잡힌 채 다른 손과 발로 선체 외벽을 디디고 버티고 있어, 완전히 허공에 뜬 상태의 하중을 보여주지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "손목 클로즈업보다 훨씬 넓고 찰리가 기둥 대신 선측을 짚지만, 갑판 안의 찰리가 선외에 매달린 이현우를 붙드는 핵심 공간 관계는 맞는다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "거대한 금속 손의 손목 압박은 더 선명하지만, 참조에 없는 내부 난간을 추가하고 이현우를 선측 안쪽에 배치해 구조 장면의 공간 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 금속 팔은 갑판 안쪽에서 오른쪽 선외의 이현우에게 뻗어 올라간 손목과 손의 경계를 붙든다. 이현우는 붙잡힌 손과 찰리를 올려다보고, 찰리도 결합된 손 쪽으로 고개를 숙인다. 이현우의 다른 손은 선체 벽을 향한다.",
        "built_space": "대각선으로 이어지는 선측 벽 하나가 왼쪽 갑판의 찰리와 오른쪽 바다 위 이현우를 분리한다. 찰리 바로 뒤와 더 먼 배경에 흰 수직 기둥이 각각 보이며, 아래쪽에는 검은 계류 기둥 하나가 일부 보인다. 젖은 갑판, 녹슨 흰 선측, 화물과 선실은 장소 참조에 부합한다. 다만 지지 기둥은 좌상단 가장자리의 작은 단면이 아니라 찰리 뒤에 보이고, 찰리의 다른 손은 기둥이 아닌 선측 상단을 짚는다.",
        "entities": "찰리는 흰 얼굴판을 가진 고릴라형 기계 머리, 금속 팔, 젖은 회색 담요와 배낭으로 표현된다. 이현우는 젖고 헝클어진 검은 머리의 젊은 동아시아계 남성이며, 어두운 후드 상의와 배낭이 보인다. 얼굴의 정확한 동일성은 측면과 빗물 때문에 제한적으로만 확인된다. 두 인물 외의 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우의 체중을 받는 지지점은 찰리에게 붙잡힌 위쪽 손목이며, 아래쪽 신발도 선체에 닿아 버티는 모습이다. 찰리는 갑판 안에서 몸을 기울이고 다른 금속 손을 선측 상단에 걸쳐 지지한다. 발은 담요와 프레임에 가려졌지만 인물이 공중에 떠 있는 모습은 아니다. 구조 동작은 물리적으로 성립하나, 지정된 기둥 파지는 구현되지 않았다."
       },
       {
        "label": "B",
        "direction": "찰리의 금속 팔은 좌상단에서 우하단으로 뻗어 이현우의 맨손목을 감싼다. 이현우의 팔은 그 손을 향해 위로 뻗고, 시선도 찰리와 붙잡힌 손 쪽을 향한다. 손목을 꽉 붙드는 표적과 동작 자체는 명확하다.",
        "built_space": "왼쪽 가장자리에 녹슨 흰 기둥 하나, 뒤쪽에 가느다란 수직 기둥 하나가 보인다. 오른쪽에는 바다와 맞닿은 흰 선측 벽이 있고, 그 안쪽 갑판에 별도의 금속 난간이 길게 추가되어 있다. 이현우의 머리와 몸통은 바깥 선측 벽보다 카메라 쪽, 즉 내부 난간과 선측 사이의 갑판 영역에 놓인다. 따라서 선측이 찰리의 지지 쪽과 이현우의 선외 매달림을 나누지 못한다.",
        "entities": "큰 금속 손과 마모된 팔 외장은 찰리의 기계 신체에 부합하며, 왼쪽에 회색 천과 신체 일부가 보인다. 찰리의 얼굴은 프레임 밖이므로 머리 형태는 평가할 수 없다. 이현우는 검은 젖은 머리와 어두운 후드 상의를 입은 동아시아계 남성으로 보이지만, 참조의 18세 얼굴보다 성숙하고 노출된 팔도 더 굵게 표현된다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장소 참조의 통로에 없는 별도의 내부 금속 난간을 추가하여 갑판의 고정 구조와 구조 동작의 경계를 바꾸었다.",
         "선외에 매달려 있어야 할 이현우를 바깥 선측 벽 안쪽의 갑판 영역에 배치했다."
        ],
        "physics": "금속 손가락이 이현우의 손목을 감싸고 있어 팔을 끌어올리거나 몸을 지탱할 접촉점은 분명하다. 찰리의 팔은 왼쪽 신체로 이어지며, 이현우의 몸은 아래 프레임 밖으로 계속된다. 지지 없는 부유로 볼 근거는 없다. 다만 찰리의 다른 손이 기둥을 잡는 접촉은 보이지 않으며, 보이는 공간에서는 이현우가 선외에 매달린 상태가 성립하지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "손목 클로즈업보다 훨씬 넓고 찰리가 기둥 대신 선측을 짚지만, 갑판 안의 찰리가 선외에 매달린 이현우를 붙드는 핵심 공간 관계는 맞는다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "거대한 금속 손의 손목 압박은 더 선명하지만, 참조에 없는 내부 난간을 추가하고 이현우를 선측 안쪽에 배치해 구조 장면의 공간 조건을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 금속 팔은 갑판 안쪽에서 오른쪽 선외의 이현우에게 뻗어 올라간 손목과 손의 경계를 붙든다. 이현우는 붙잡힌 손과 찰리를 올려다보고, 찰리도 결합된 손 쪽으로 고개를 숙인다. 이현우의 다른 손은 선체 벽을 향한다.",
        "built_space": "대각선으로 이어지는 선측 벽 하나가 왼쪽 갑판의 찰리와 오른쪽 바다 위 이현우를 분리한다. 찰리 바로 뒤와 더 먼 배경에 흰 수직 기둥이 각각 보이며, 아래쪽에는 검은 계류 기둥 하나가 일부 보인다. 젖은 갑판, 녹슨 흰 선측, 화물과 선실은 장소 참조에 부합한다. 다만 지지 기둥은 좌상단 가장자리의 작은 단면이 아니라 찰리 뒤에 보이고, 찰리의 다른 손은 기둥이 아닌 선측 상단을 짚는다.",
        "entities": "찰리는 흰 얼굴판을 가진 고릴라형 기계 머리, 금속 팔, 젖은 회색 담요와 배낭으로 표현된다. 이현우는 젖고 헝클어진 검은 머리의 젊은 동아시아계 남성이며, 어두운 후드 상의와 배낭이 보인다. 얼굴의 정확한 동일성은 측면과 빗물 때문에 제한적으로만 확인된다. 두 인물 외의 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "이현우의 체중을 받는 지지점은 찰리에게 붙잡힌 위쪽 손목이며, 아래쪽 신발도 선체에 닿아 버티는 모습이다. 찰리는 갑판 안에서 몸을 기울이고 다른 금속 손을 선측 상단에 걸쳐 지지한다. 발은 담요와 프레임에 가려졌지만 인물이 공중에 떠 있는 모습은 아니다. 구조 동작은 물리적으로 성립하나, 지정된 기둥 파지는 구현되지 않았다."
       },
       {
        "label": "A",
        "direction": "찰리의 금속 팔은 좌상단에서 우하단으로 뻗어 이현우의 맨손목을 감싼다. 이현우의 팔은 그 손을 향해 위로 뻗고, 시선도 찰리와 붙잡힌 손 쪽을 향한다. 손목을 꽉 붙드는 표적과 동작 자체는 명확하다.",
        "built_space": "왼쪽 가장자리에 녹슨 흰 기둥 하나, 뒤쪽에 가느다란 수직 기둥 하나가 보인다. 오른쪽에는 바다와 맞닿은 흰 선측 벽이 있고, 그 안쪽 갑판에 별도의 금속 난간이 길게 추가되어 있다. 이현우의 머리와 몸통은 바깥 선측 벽보다 카메라 쪽, 즉 내부 난간과 선측 사이의 갑판 영역에 놓인다. 따라서 선측이 찰리의 지지 쪽과 이현우의 선외 매달림을 나누지 못한다.",
        "entities": "큰 금속 손과 마모된 팔 외장은 찰리의 기계 신체에 부합하며, 왼쪽에 회색 천과 신체 일부가 보인다. 찰리의 얼굴은 프레임 밖이므로 머리 형태는 평가할 수 없다. 이현우는 검은 젖은 머리와 어두운 후드 상의를 입은 동아시아계 남성으로 보이지만, 참조의 18세 얼굴보다 성숙하고 노출된 팔도 더 굵게 표현된다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장소 참조의 통로에 없는 별도의 내부 금속 난간을 추가하여 갑판의 고정 구조와 구조 동작의 경계를 바꾸었다.",
         "선외에 매달려 있어야 할 이현우를 바깥 선측 벽 안쪽의 갑판 영역에 배치했다."
        ],
        "physics": "금속 손가락이 이현우의 손목을 감싸고 있어 팔을 끌어올리거나 몸을 지탱할 접촉점은 분명하다. 찰리의 팔은 왼쪽 신체로 이어지며, 이현우의 몸은 아래 프레임 밖으로 계속된다. 지지 없는 부유로 볼 근거는 없다. 다만 찰리의 다른 손이 기둥을 잡는 접촉은 보이지 않으며, 보이는 공간에서는 이현우가 선외에 매달린 상태가 성립하지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.35,
    "B": 1.667
   },
   "violations": {
    "A": [
     "[gpt-high] 장소 참조의 통로에 없는 별도의 내부 금속 난간을 추가하여 갑판의 고정 구조와 구조 동작의 경계를 바꾸었다.",
     "[gpt-high] 선외에 매달려 있어야 할 이현우를 바깥 선측 벽 안쪽의 갑판 영역에 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1350,
   "B": 1667
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "클로즈업 프레임 척도와 손목을 강하게 틀어쥔 액션, 허공에 매달린 물리적 상태 등 우선순위가 높은 샷 연출 지시를 잘 따랐으나 인물들의 복장 일치도는 다소 떨어집니다.  ★위반: [gpt-high] 장소 참조의 통로에 없는 별도의 내부 금속 난간을 추가하여 갑판의 고정 구조와 구조 동작의 경계를 바꾸었다. / [gpt-high] 선외에 매달려 있어야 할 이현우를 바깥 선측 벽 안쪽의 갑판 영역에 배치했다."
   },
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "캐릭터 외형과 복장의 일치도는 훌륭하지만, 클로즈업 지시를 무시하고 프레임을 넓게 잡았으며 지정된 손목 그립 대신 손을 맞잡는 자세를 취해 핵심 연출에서 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L122B01.png",
    "asset_id": "6abdf41d-41d4-42ad-934a-d5ad17b629b4",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:948699>",
    "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917742>",
    "asset_id": "42b5a726-7afc-44b7-82a3-8943670a5fe2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacee7-27df-7c5c-9cad-d5b71116c545",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S80sh6__bgfirst_bg.png",
   "bg_asset_id": "1d7f2f31-f0ce-4f62-9b96-d470b5cf7891",
   "bg_record_key": "S80sh6::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S80sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:57:22.381628+00:00",
  "fingerprint": "b667edb0676c77aaf7b5a40e4f94864e634c0ee03fb582ccbdccbd84ce8df2af",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S80sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S80sh6_sel.png",
  "source_sha256": "e7adca0c1503405dbf74dba38dcb112b5e30ca76af4e174dda2d47d220b8459c",
  "file": "S80sh6_cine.png",
  "staged_sha256": "233a0947f639887812d86487a83b311aa56880060f4e0c2221bc616bee18ed84",
  "latency_ms": 8736
 },
 "S80sh12::signage": {
  "fp": "e9a2907895dc0c4c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S80sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T07:59:59.089530+00:00",
  "fingerprint": "2e39983b81eadd17bd2687122a9fef58f397b57ac170eb5b26d1c8ea3154e2d4",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S80sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S80sh12_sel.png",
  "source_sha256": "5119f44bd2b6600519c113ad826c81e41484bd5a9447c70bc1c7acfbb5183811",
  "file": "S80sh12_cine.png",
  "staged_sha256": "f3f2162bdb2c612b5dc38233a414477c030899b1b99dddd576d9e0d3b748a5e4",
  "latency_ms": 9407
 },
 "S80sh17::signage": {
  "fp": "2280aaec1df0c203",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S80sh17::bgfirst_bg": {
  "input_fingerprint": "69436b72d28c7968",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거대한 선박의 짐짝들과 함께 찰리와 이현우의 몸이 어두운 심해로 끝없이 가라앉고 있는 상태의 수중 mid-action 전경.\n\nLOCATION (lock): Deep in the open sea beneath the storm-struck vessel, among sinking cargo in increasingly dark water.\n\nTIME OF DAY (lock): night, typhoon and torrential rain.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Heavy ship cargo (Submerged and sinking alongside the two bodies) — Separate pieces present different oblique faces along the downward path; used as Establish depth and scale around the small figures, with no individual piece exceeding a minor portion of the frame; Deep water (Surrounding the descending bodies and cargo); used as Provide unbroken negative space around the final retreating composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The surroundings darken toward the scripted fade-out, allowing the bodies and cargo to lose separation gradually without adding visible underwater light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거대한 선박의 짐짝들과 함께 찰리와 이현우의 몸이 어두운 심해로 끝없이 가라앉고 있는 상태의 수중 mid-action 전경.\n\nLOCATION (lock): Deep in the open sea beneath the storm-struck vessel, among sinking cargo in increasingly dark water.\n\nTIME OF DAY (lock): night, typhoon and torrential rain.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Heavy ship cargo (Submerged and sinking alongside the two bodies) — Separate pieces present different oblique faces along the downward path; used as Establish depth and scale around the small figures, with no individual piece exceeding a minor portion of the frame; Deep water (Surrounding the descending bodies and cargo); used as Provide unbroken negative space around the final retreating composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The surroundings darken toward the scripted fade-out, allowing the bodies and cargo to lose separation gradually without adding visible underwater light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S80sh17__bgfirst_bg.png",
  "asset_id": "c18fc6e1-38f6-47c1-ab2f-f8763351e721",
  "input_asset_ids": [
   "9248eed0-e73d-485c-a3de-296b590984b4",
   "0df69e86-e8d9-4180-8c5c-4ea567fa6cc2"
  ]
 },
 "S80sh17": {
  "input_fingerprint": "e372edbe89c4ee72",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, typhoon and torrential rain.\n\nSHOT TEXT (authoritative, Korean): 거대한 선박의 짐짝들과 함께 찰리와 이현우의 몸이 어두운 심해로 끝없이 가라앉고 있는 상태의 수중 mid-action 전경.\n\nLOCATION (lock): Deep in the open sea beneath the storm-struck vessel, among sinking cargo in increasingly dark water. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Heavy ship cargo (Submerged and sinking alongside the two bodies) — Separate pieces present different oblique faces along the downward path; used as Establish depth and scale around the small figures, with no individual piece exceeding a minor portion of the frame; Deep water (Surrounding the descending bodies and cargo); used as Provide unbroken negative space around the final retreating composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The surroundings darken toward the scripted fade-out, allowing the bodies and cargo to lose separation gradually without adding visible underwater light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is unconscious and sinking freely beneath the sea among the ship's heavy cargo, with no supporting surface specified. His exact head angle, torso orientation, and arm and leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy cargo from the vessel sinks into increasingly dark seawater. Charlie is submerged and descending, with his already damaged body and chest opening still unrepaired. 이현우: He is unconscious and sinking deeper underwater, with his clothing soaked through.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, typhoon and torrential rain.\n\nSHOT TEXT (authoritative, Korean): 거대한 선박의 짐짝들과 함께 찰리와 이현우의 몸이 어두운 심해로 끝없이 가라앉고 있는 상태의 수중 mid-action 전경.\n\nLOCATION (lock): Deep in the open sea beneath the storm-struck vessel, among sinking cargo in increasingly dark water. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Heavy ship cargo (Submerged and sinking alongside the two bodies) — Separate pieces present different oblique faces along the downward path; used as Establish depth and scale around the small figures, with no individual piece exceeding a minor portion of the frame; Deep water (Surrounding the descending bodies and cargo); used as Provide unbroken negative space around the final retreating composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The surroundings darken toward the scripted fade-out, allowing the bodies and cargo to lose separation gradually without adding visible underwater light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is unconscious and sinking freely beneath the sea among the ship's heavy cargo, with no supporting surface specified. His exact head angle, torso orientation, and arm and leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy cargo from the vessel sinks into increasingly dark seawater. Charlie is submerged and descending, with his already damaged body and chest opening still unrepaired. 이현우: He is unconscious and sinking deeper underwater, with his clothing soaked through.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night, typhoon and torrential rain.\n\nSHOT TEXT (authoritative, Korean): 거대한 선박의 짐짝들과 함께 찰리와 이현우의 몸이 어두운 심해로 끝없이 가라앉고 있는 상태의 수중 mid-action 전경.\n\nLOCATION (lock): Deep in the open sea beneath the storm-struck vessel, among sinking cargo in increasingly dark water. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Heavy ship cargo (Submerged and sinking alongside the two bodies) — Separate pieces present different oblique faces along the downward path; used as Establish depth and scale around the small figures, with no individual piece exceeding a minor portion of the frame; Deep water (Surrounding the descending bodies and cargo); used as Provide unbroken negative space around the final retreating composition.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The surroundings darken toward the scripted fade-out, allowing the bodies and cargo to lose separation gradually without adding visible underwater light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is unconscious and sinking freely beneath the sea among the ship's heavy cargo, with no supporting surface specified. His exact head angle, torso orientation, and arm and leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy cargo from the vessel sinks into increasingly dark seawater. Charlie is submerged and descending, with his already damaged body and chest opening still unrepaired. 이현우: He is unconscious and sinking deeper underwater, with his clothing soaked through.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 머리, 각진 머리 외장, 흰 얼굴판, 점 형태의 눈과 선 형태의 입); 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S80sh17__bgfirst_bg.png",
     "asset_id": "c18fc6e1-38f6-47c1-ab2f-f8763351e721",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S80sh17.png",
     "asset_id": "9248eed0-e73d-485c-a3de-296b590984b4",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:948699>",
     "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1218943>",
     "asset_id": "717460f5-31cc-4f17-a49b-efdad9614ee1",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L122B02.png",
     "asset_id": "0df69e86-e8d9-4180-8c5c-4ea567fa6cc2",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:948699>",
     "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1218943>",
     "asset_id": "717460f5-31cc-4f17-a49b-efdad9614ee1",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 깊은 바다 속에서 아래를 향해 가라앉는 인물들과 화물들을 비스듬히 바라보며, 빛은 수면 위쪽에서 들어옵니다.",
    "built_space": "어두운 심해 배경이며, 여러 개의 철창과 상자 형태의 화물들이 물속에 떠다니며 가라앉고 있습니다.",
    "entities": "찰리(고릴라형 기계 얼굴, 손상된 가슴)와 이현우(어두운 머리, 후드티, 청바지, 의식 잃음)가 프롬프트의 지시대로 화면 중앙에 작게 배치되어 있습니다.",
    "hard_violations": [],
    "physics": "두 인물과 화물들 모두 물속에서 부력과 중력의 영향을 받아 자연스럽게 부유하며 가라앉고 있고, 이현우는 의식을 잃고 축 늘어진 상태를 유지합니다."
   },
   {
    "label": "B",
    "direction": "카메라는 바다 속에서 가라앉는 인물들과 화물들을 수평에 가까운 각도에서 바라봅니다.",
    "built_space": "심해 배경 속에 밧줄로 묶인 거대한 화물 상자들과 그물들이 주변에 크게 떠 있습니다.",
    "entities": "찰리(고릴라형 기계 얼굴, 가슴의 뚜렷한 손상)와 이현우(어두운 머리, 후드티, 청바지)가 묘사되었으나, 화면에 너무 큰 비율로 꽉 차게 배치되었습니다.",
    "hard_violations": [
     "[gpt-high] 왼쪽 아래 컨테이너에 식별 가능한 흰 숫자 표기가 남아 있어, 읽을 수 있는 글자를 전면 금지한 조건을 위반합니다."
    ],
    "physics": "인물들과 화물이 물속에 가라앉고 있으며, 이현우는 머리가 아래로 향한 채 떠 있고 찰리는 다소 꼿꼿한 자세로 부유합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 강력하게 요구한 '와이드 샷', '작은 인물들', '퇴각하는 구도를 감싸는 넓은 여백'의 화면 구성을 매우 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물들이 화면에 너무 크게 배치되었고 전경의 화물이 화면의 큰 부분을 차지하여 '작은 인물들'과 '와이드 샷' 지시를 명백히 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 깊은 바다 속에서 아래를 향해 가라앉는 인물들과 화물들을 비스듬히 바라보며, 빛은 수면 위쪽에서 들어옵니다.",
        "built_space": "어두운 심해 배경이며, 여러 개의 철창과 상자 형태의 화물들이 물속에 떠다니며 가라앉고 있습니다.",
        "entities": "찰리(고릴라형 기계 얼굴, 손상된 가슴)와 이현우(어두운 머리, 후드티, 청바지, 의식 잃음)가 프롬프트의 지시대로 화면 중앙에 작게 배치되어 있습니다.",
        "hard_violations": [],
        "physics": "두 인물과 화물들 모두 물속에서 부력과 중력의 영향을 받아 자연스럽게 부유하며 가라앉고 있고, 이현우는 의식을 잃고 축 늘어진 상태를 유지합니다."
       },
       {
        "label": "B",
        "direction": "카메라는 바다 속에서 가라앉는 인물들과 화물들을 수평에 가까운 각도에서 바라봅니다.",
        "built_space": "심해 배경 속에 밧줄로 묶인 거대한 화물 상자들과 그물들이 주변에 크게 떠 있습니다.",
        "entities": "찰리(고릴라형 기계 얼굴, 가슴의 뚜렷한 손상)와 이현우(어두운 머리, 후드티, 청바지)가 묘사되었으나, 화면에 너무 큰 비율로 꽉 차게 배치되었습니다.",
        "hard_violations": [],
        "physics": "인물들과 화물이 물속에 가라앉고 있으며, 이현우는 머리가 아래로 향한 채 떠 있고 찰리는 다소 꼿꼿한 자세로 부유합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 강력하게 요구한 '와이드 샷', '작은 인물들', '퇴각하는 구도를 감싸는 넓은 여백'의 화면 구성을 매우 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물들이 화면에 너무 크게 배치되었고 전경의 화물이 화면의 큰 부분을 차지하여 '작은 인물들'과 '와이드 샷' 지시를 명백히 어겼습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 깊은 바다 속에서 아래를 향해 가라앉는 인물들과 화물들을 비스듬히 바라보며, 빛은 수면 위쪽에서 들어옵니다.",
        "built_space": "어두운 심해 배경이며, 여러 개의 철창과 상자 형태의 화물들이 물속에 떠다니며 가라앉고 있습니다.",
        "entities": "찰리(고릴라형 기계 얼굴, 손상된 가슴)와 이현우(어두운 머리, 후드티, 청바지, 의식 잃음)가 프롬프트의 지시대로 화면 중앙에 작게 배치되어 있습니다.",
        "hard_violations": [],
        "physics": "두 인물과 화물들 모두 물속에서 부력과 중력의 영향을 받아 자연스럽게 부유하며 가라앉고 있고, 이현우는 의식을 잃고 축 늘어진 상태를 유지합니다."
       },
       {
        "label": "B",
        "direction": "카메라는 바다 속에서 가라앉는 인물들과 화물들을 수평에 가까운 각도에서 바라봅니다.",
        "built_space": "심해 배경 속에 밧줄로 묶인 거대한 화물 상자들과 그물들이 주변에 크게 떠 있습니다.",
        "entities": "찰리(고릴라형 기계 얼굴, 가슴의 뚜렷한 손상)와 이현우(어두운 머리, 후드티, 청바지)가 묘사되었으나, 화면에 너무 큰 비율로 꽉 차게 배치되었습니다.",
        "hard_violations": [],
        "physics": "인물들과 화물이 물속에 가라앉고 있으며, 이현우는 머리가 아래로 향한 채 떠 있고 찰리는 다소 꼿꼿한 자세로 부유합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "작은 두 인물과 넓은 중앙 여백으로 침강 전경에 더 가깝지만, 거대한 전경 화물과 선명한 수면은 심해의 최종 후퇴 구도를 훼손합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물과 전경 화물이 지나치게 크고, 찰리의 외형·가슴 개구부 재현이 부족하며 컨테이너의 식별 가능한 숫자도 무문자 조건을 위반합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 찰리는 얼굴을 오른쪽 위로 들고 발을 아래로 늘어뜨렸으며, 특정 상대를 바라보는지는 불명확합니다. 오른쪽 현우는 머리가 왼쪽 아래, 발이 오른쪽 위를 향하고 얼굴은 아래로 떨어져 있습니다. 두 몸의 방향은 서로 다르지만 화물 사이로 가라앉는 상태로 해석할 수 있습니다. 기포는 위로 이어지며 무기나 조준 대상은 없습니다.",
        "built_space": "고정 건축물이나 바닥 없이 어두운 바닷물이 이어집니다. 왼쪽 위 그물 화물, 왼쪽 가장자리 대형 금속 화물, 왼쪽 아래 컨테이너, 오른쪽 위·중간의 상자들, 오른쪽 아래 대형 그물 화물과 그 뒤 금속 상자 등 약 아홉 덩어리가 구분됩니다. 재질과 배치는 장소 참고와 유사하지만, 양쪽 전경 화물이 화면의 큰 부분을 차지하고 찰리도 화면 높이의 절반 정도여서 작은 인물 주변에 소규모 화물을 분산하라는 구도와 다릅니다.",
        "entities": "등장 주체는 기계형 찰리와 젊은 검은 머리 남성 현우, 두 명뿐입니다. 현우의 어두운 후드와 청바지는 참고에 부합하며 얼굴은 작고 아래로 향해 정확한 나이와 동일인 여부를 확정하기 어렵습니다. 찰리는 금속 몸과 고릴라형 머리는 있으나 참고의 흰 얼굴판·점눈·선입보다 조각된 고릴라 얼굴에 가깝고 회색 망토가 없습니다. 가슴에는 손상이 보이지만 열린 흉부는 뚜렷하지 않습니다. 금속·그물 화물은 실물 질감으로 보이나 왼쪽 아래 컨테이너에 흰 숫자 표기가 보입니다.",
        "hard_violations": [
         "왼쪽 아래 컨테이너에 식별 가능한 흰 숫자 표기가 남아 있어, 읽을 수 있는 글자를 전면 금지한 조건을 위반합니다."
        ],
        "physics": "인물과 화물은 모두 물속에 잠겨 있으며 바닥이나 손잡이에는 의지하지 않습니다. 부력과 물의 저항을 받는 침강 장면이므로 공중에 근거 없이 떠 있는 배치는 아닙니다. 현우의 한 팔은 아래로 처지고 다른 팔과 다리는 몸 뒤쪽에 남아 있어 침강 중 수중 저항으로 가능한 자세입니다. 찰리의 손발에도 별도 지지물은 없으며, 기계 몸이 자유 침강하는 것으로 읽힙니다."
       },
       {
        "label": "B",
        "direction": "찰리는 화면 아래 중앙 왼쪽에서 얼굴과 몸 앞면을 카메라 쪽으로 비스듬히 향하고 한 팔을 오른쪽 위로 뻗고 있습니다. 현우는 그 오른쪽에서 얼굴을 위로 젖힌 채 머리가 오른쪽 위, 다리가 왼쪽 아래로 향합니다. 서로를 응시하거나 붙잡지는 않습니다. 화물은 여러 각도로 기울고 기포는 위쪽 수면으로 이어져, 두 인물이 그 아래로 멀어지는 관계가 보입니다.",
        "built_space": "상단에 물결치는 수면이 있고 그 아래 넓은 중앙 수역이 펼쳐집니다. 왼쪽 위 겹친 화물군, 왼쪽 가장자리의 대형 화물군, 중앙과 오른쪽의 작은 상자들, 하단 중앙과 오른쪽 아래의 화물군 등 십여 개 이상의 상자와 짐이 여러 깊이에 놓여 있습니다. 고정 시설이나 반사는 없습니다. 두 인물을 작게 배치한 점은 요구에 가깝지만, 양쪽 가장자리와 오른쪽 아래 화물은 여전히 과대하며 가까운 수면은 깊은 심해라는 장소감을 약화합니다.",
        "entities": "찰리와 현우 두 주체만 보입니다. 찰리는 밝은 얼굴판과 각진 기계 머리, 금속 몸을 갖추고 가슴에 어두운 개구부가 보여 A보다 지정된 정체성과 손상 상태에 가깝습니다. 다만 회색 망토는 보이지 않습니다. 현우는 검은 머리의 젊은 남성으로 어두운 상의와 청바지를 입었으며, 이 크기에서는 정확한 얼굴 일치와 18세 여부를 확인하기 어렵습니다. 중량 화물은 금속 테두리 상자와 그물 포장 짐으로 표현되어 장소 참고의 재질 계열에 맞습니다. 읽을 수 있는 문자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "두 몸과 상자는 물속에서 부력과 저항을 받으며 자유롭게 침강하는 상태입니다. 현우의 팔은 옆으로 느슨하게 벌어지고 다리도 비대칭으로 풀려 있어 수영 동작보다는 의식 없는 몸의 수중 자세로 읽힙니다. 찰리의 위로 뻗은 팔은 다소 능동적으로 보이지만 찰리에게 현우와 같은 무의식 자세가 명시되지는 않았습니다. 두 인물 모두 화물에 기대거나 매달리지 않으며, 물속 침강이라는 물리적 맥락이 있어 지지 없는 공중 부유에 해당하지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "작은 두 인물과 넓은 중앙 여백으로 침강 전경에 더 가깝지만, 거대한 전경 화물과 선명한 수면은 심해의 최종 후퇴 구도를 훼손합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물과 전경 화물이 지나치게 크고, 찰리의 외형·가슴 개구부 재현이 부족하며 컨테이너의 식별 가능한 숫자도 무문자 조건을 위반합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 찰리는 얼굴을 오른쪽 위로 들고 발을 아래로 늘어뜨렸으며, 특정 상대를 바라보는지는 불명확합니다. 오른쪽 현우는 머리가 왼쪽 아래, 발이 오른쪽 위를 향하고 얼굴은 아래로 떨어져 있습니다. 두 몸의 방향은 서로 다르지만 화물 사이로 가라앉는 상태로 해석할 수 있습니다. 기포는 위로 이어지며 무기나 조준 대상은 없습니다.",
        "built_space": "고정 건축물이나 바닥 없이 어두운 바닷물이 이어집니다. 왼쪽 위 그물 화물, 왼쪽 가장자리 대형 금속 화물, 왼쪽 아래 컨테이너, 오른쪽 위·중간의 상자들, 오른쪽 아래 대형 그물 화물과 그 뒤 금속 상자 등 약 아홉 덩어리가 구분됩니다. 재질과 배치는 장소 참고와 유사하지만, 양쪽 전경 화물이 화면의 큰 부분을 차지하고 찰리도 화면 높이의 절반 정도여서 작은 인물 주변에 소규모 화물을 분산하라는 구도와 다릅니다.",
        "entities": "등장 주체는 기계형 찰리와 젊은 검은 머리 남성 현우, 두 명뿐입니다. 현우의 어두운 후드와 청바지는 참고에 부합하며 얼굴은 작고 아래로 향해 정확한 나이와 동일인 여부를 확정하기 어렵습니다. 찰리는 금속 몸과 고릴라형 머리는 있으나 참고의 흰 얼굴판·점눈·선입보다 조각된 고릴라 얼굴에 가깝고 회색 망토가 없습니다. 가슴에는 손상이 보이지만 열린 흉부는 뚜렷하지 않습니다. 금속·그물 화물은 실물 질감으로 보이나 왼쪽 아래 컨테이너에 흰 숫자 표기가 보입니다.",
        "hard_violations": [
         "왼쪽 아래 컨테이너에 식별 가능한 흰 숫자 표기가 남아 있어, 읽을 수 있는 글자를 전면 금지한 조건을 위반합니다."
        ],
        "physics": "인물과 화물은 모두 물속에 잠겨 있으며 바닥이나 손잡이에는 의지하지 않습니다. 부력과 물의 저항을 받는 침강 장면이므로 공중에 근거 없이 떠 있는 배치는 아닙니다. 현우의 한 팔은 아래로 처지고 다른 팔과 다리는 몸 뒤쪽에 남아 있어 침강 중 수중 저항으로 가능한 자세입니다. 찰리의 손발에도 별도 지지물은 없으며, 기계 몸이 자유 침강하는 것으로 읽힙니다."
       },
       {
        "label": "A",
        "direction": "찰리는 화면 아래 중앙 왼쪽에서 얼굴과 몸 앞면을 카메라 쪽으로 비스듬히 향하고 한 팔을 오른쪽 위로 뻗고 있습니다. 현우는 그 오른쪽에서 얼굴을 위로 젖힌 채 머리가 오른쪽 위, 다리가 왼쪽 아래로 향합니다. 서로를 응시하거나 붙잡지는 않습니다. 화물은 여러 각도로 기울고 기포는 위쪽 수면으로 이어져, 두 인물이 그 아래로 멀어지는 관계가 보입니다.",
        "built_space": "상단에 물결치는 수면이 있고 그 아래 넓은 중앙 수역이 펼쳐집니다. 왼쪽 위 겹친 화물군, 왼쪽 가장자리의 대형 화물군, 중앙과 오른쪽의 작은 상자들, 하단 중앙과 오른쪽 아래의 화물군 등 십여 개 이상의 상자와 짐이 여러 깊이에 놓여 있습니다. 고정 시설이나 반사는 없습니다. 두 인물을 작게 배치한 점은 요구에 가깝지만, 양쪽 가장자리와 오른쪽 아래 화물은 여전히 과대하며 가까운 수면은 깊은 심해라는 장소감을 약화합니다.",
        "entities": "찰리와 현우 두 주체만 보입니다. 찰리는 밝은 얼굴판과 각진 기계 머리, 금속 몸을 갖추고 가슴에 어두운 개구부가 보여 A보다 지정된 정체성과 손상 상태에 가깝습니다. 다만 회색 망토는 보이지 않습니다. 현우는 검은 머리의 젊은 남성으로 어두운 상의와 청바지를 입었으며, 이 크기에서는 정확한 얼굴 일치와 18세 여부를 확인하기 어렵습니다. 중량 화물은 금속 테두리 상자와 그물 포장 짐으로 표현되어 장소 참고의 재질 계열에 맞습니다. 읽을 수 있는 문자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "두 몸과 상자는 물속에서 부력과 저항을 받으며 자유롭게 침강하는 상태입니다. 현우의 팔은 옆으로 느슨하게 벌어지고 다리도 비대칭으로 풀려 있어 수영 동작보다는 의식 없는 몸의 수중 자세로 읽힙니다. 찰리의 위로 뻗은 팔은 다소 능동적으로 보이지만 찰리에게 현우와 같은 무의식 자세가 명시되지는 않았습니다. 두 인물 모두 화물에 기대거나 매달리지 않으며, 물속 침강이라는 물리적 맥락이 있어 지지 없는 공중 부유에 해당하지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.071
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.821
   },
   "violations": {
    "B": [
     "[gpt-high] 왼쪽 아래 컨테이너에 식별 가능한 흰 숫자 표기가 남아 있어, 읽을 수 있는 글자를 전면 금지한 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 821
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 강력하게 요구한 '와이드 샷', '작은 인물들', '퇴각하는 구도를 감싸는 넓은 여백'의 화면 구성을 매우 충실하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 821,
    "verdict_ko": "인물들이 화면에 너무 크게 배치되었고 전경의 화물이 화면의 큰 부분을 차지하여 '작은 인물들'과 '와이드 샷' 지시를 명백히 어겼습니다.  ★위반: [gpt-high] 왼쪽 아래 컨테이너에 식별 가능한 흰 숫자 표기가 남아 있어, 읽을 수 있는 글자를 전면 금지한 조건을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L122B02.png",
    "asset_id": "0df69e86-e8d9-4180-8c5c-4ea567fa6cc2",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:948699>",
    "asset_id": "0bc5704b-429f-4d3a-8af0-75423c952778",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1218943>",
    "asset_id": "717460f5-31cc-4f17-a49b-efdad9614ee1",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacef8-8de7-7525-82d2-3bc5200e9237",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S80sh17__bgfirst_bg.png",
   "bg_asset_id": "c18fc6e1-38f6-47c1-ab2f-f8763351e721",
   "bg_record_key": "S80sh17::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S80sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:01:44.567305+00:00",
  "fingerprint": "c9472be4e069975f58e16dd02383fa30efe978ad48ff2b825f71b7783a346412",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S80sh17_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S80sh17_sel.png",
  "source_sha256": "b3948e5c2cd3a4b616c54ffe7f3394f385463cbfa254359ed2f3130b947c3177",
  "file": "S80sh17_cine.png",
  "staged_sha256": "4aa6e2f093e25c588ed5d731953e6501d95bec5e05a05b96cc1b775bcfd17790",
  "latency_ms": 9146
 },
 "S81sh1::signage": {
  "fp": "288b75a240ed5007",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::ff9db9d9f1a953f4": {
  "subjects": [],
  "subject_text": "제주도 임시보호소 숙소 방\n침대가 놓인 단출한 숙소 방. 옷을 갈아입을 자리와 샤워 공간이 마련되어 있고 화상통화가 가능한 전화 장치가 있다.",
  "identity": "canonical",
  "scope_id": "L127",
  "scope_role": "location_interior",
  "scope_sha": "0c71d8ba84732400"
 },
 "S81sh1::bgfirst_bg": {
  "input_fingerprint": "8f5c15fe01faa20f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 눈부신 형광등 불빛 아래 새하얀 침대에서 이현우가 두 눈을 번쩍 뜬 얼굴 클로즈업.\n\nLOCATION (lock): On a white bed inside the island research institute's temporary shelter room, beneath bright fluorescent lights with a turquoise sea visible outside the window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: White bed (Occupied by 이현우 as he awakens) — Only the upper surface immediately around his head and shoulders is visible; used as Border the face without introducing unrelated room details.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Intense fluorescent illumination strikes his newly opened eyes against the white bed, giving the awakening an abrupt, exposing brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 눈부신 형광등 불빛 아래 새하얀 침대에서 이현우가 두 눈을 번쩍 뜬 얼굴 클로즈업.\n\nLOCATION (lock): On a white bed inside the island research institute's temporary shelter room, beneath bright fluorescent lights with a turquoise sea visible outside the window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: White bed (Occupied by 이현우 as he awakens) — Only the upper surface immediately around his head and shoulders is visible; used as Border the face without introducing unrelated room details.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Intense fluorescent illumination strikes his newly opened eyes against the white bed, giving the awakening an abrupt, exposing brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S81sh1__bgfirst_bg.png",
  "asset_id": "21fc7969-a438-4859-becf-903656b8b3b4",
  "input_asset_ids": [
   "05182809-a86d-4e44-857b-350db4b17ce5",
   "7921e778-417f-49f1-9784-6734e0e6fcd3"
  ]
 },
 "S81sh1": {
  "input_fingerprint": "044c2c5d3ecba337",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈부신 형광등 불빛 아래 새하얀 침대에서 이현우가 두 눈을 번쩍 뜬 얼굴 클로즈업.\n\nLOCATION (lock): On a white bed inside the island research institute's temporary shelter room, beneath bright fluorescent lights with a turquoise sea visible outside the window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: White bed (Occupied by 이현우 as he awakens) — Only the upper surface immediately around his head and shoulders is visible; used as Border the face without introducing unrelated room details.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Intense fluorescent illumination strikes his newly opened eyes against the white bed, giving the awakening an abrupt, exposing brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is reclining on a white bed in the temporary shelter, with his body still supported by the bed as his eyes open. His exact head support, torso orientation, and arm and leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A white bed stands beneath strong fluorescent lighting in the temporary shelter. The emerald-blue sea of Jeju is visible through the window. 이현우: He is waking on the white bed, dressed in fresh white clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈부신 형광등 불빛 아래 새하얀 침대에서 이현우가 두 눈을 번쩍 뜬 얼굴 클로즈업.\n\nLOCATION (lock): On a white bed inside the island research institute's temporary shelter room, beneath bright fluorescent lights with a turquoise sea visible outside the window. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: White bed (Occupied by 이현우 as he awakens) — Only the upper surface immediately around his head and shoulders is visible; used as Border the face without introducing unrelated room details.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Intense fluorescent illumination strikes his newly opened eyes against the white bed, giving the awakening an abrupt, exposing brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is reclining on a white bed in the temporary shelter, with his body still supported by the bed as his eyes open. His exact head support, torso orientation, and arm and leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A white bed stands beneath strong fluorescent lighting in the temporary shelter. The emerald-blue sea of Jeju is visible through the window. 이현우: He is waking on the white bed, dressed in fresh white clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈부신 형광등 불빛 아래 새하얀 침대에서 이현우가 두 눈을 번쩍 뜬 얼굴 클로즈업.\n\nLOCATION (lock): On a white bed inside the island research institute's temporary shelter room, beneath bright fluorescent lights with a turquoise sea visible outside the window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: White bed (Occupied by 이현우 as he awakens) — Only the upper surface immediately around his head and shoulders is visible; used as Border the face without introducing unrelated room details.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Intense fluorescent illumination strikes his newly opened eyes against the white bed, giving the awakening an abrupt, exposing brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is reclining on a white bed in the temporary shelter, with his body still supported by the bed as his eyes open. His exact head support, torso orientation, and arm and leg positions are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A white bed stands beneath strong fluorescent lighting in the temporary shelter. The emerald-blue sea of Jeju is visible through the window. 이현우: He is waking on the white bed, dressed in fresh white clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S81sh1__bgfirst_bg.png",
     "asset_id": "21fc7969-a438-4859-becf-903656b8b3b4",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S81sh1.png",
     "asset_id": "05182809-a86d-4e44-857b-350db4b17ce5",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L127B01.png",
     "asset_id": "7921e778-417f-49f1-9784-6734e0e6fcd3",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 우측 상단 허공을 향해 있음.",
    "built_space": "나무 헤드보드 뒤쪽으로 창문과 바다 풍경이 배치됨.",
    "entities": "이현우 레퍼런스와 유사한 젊은 남성, 헝클어진 검은 머리, 흰옷 착용.",
    "hard_violations": [
     "[gpt-high] 참조에서 막힌 왼쪽 벽에 별도의 창을 추가해, 한 벽면의 창을 두 벽면의 코너 창 구조로 바꿨습니다."
    ],
    "physics": "머리와 상체가 침대와 베개에 안정적으로 지지되어 누워있음."
   },
   {
    "label": "B",
    "direction": "눈을 크게 뜨고 위쪽 정면을 똑바로 응시함.",
    "built_space": "베개 바로 뒤로 창틀과 바다가 좁은 면적으로 보임.",
    "entities": "이현우 레퍼런스와 이목구비가 매우 일치하는 남성, 헝클어진 검은 머리, 흰옷 착용.",
    "hard_violations": [],
    "physics": "머리와 상체가 침대와 베개에 안정적으로 지지되어 누워있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "눈을 번쩍 뜬 표정과 인물의 외모가 레퍼런스와 완벽히 일치하며, 주변을 배제한 타이트한 클로즈업 프레이밍 지시를 훌륭히 따랐습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "시선이 옆을 향해 갑작스럽게 깨어난 느낌이 부족하며, 자연광 그림자가 강해 눈부신 형광등 조명이라는 지시와 다소 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 우측 상단 허공을 향해 있음.",
        "built_space": "나무 헤드보드 뒤쪽으로 창문과 바다 풍경이 배치됨.",
        "entities": "이현우 레퍼런스와 유사한 젊은 남성, 헝클어진 검은 머리, 흰옷 착용.",
        "hard_violations": [],
        "physics": "머리와 상체가 침대와 베개에 안정적으로 지지되어 누워있음."
       },
       {
        "label": "B",
        "direction": "눈을 크게 뜨고 위쪽 정면을 똑바로 응시함.",
        "built_space": "베개 바로 뒤로 창틀과 바다가 좁은 면적으로 보임.",
        "entities": "이현우 레퍼런스와 이목구비가 매우 일치하는 남성, 헝클어진 검은 머리, 흰옷 착용.",
        "hard_violations": [],
        "physics": "머리와 상체가 침대와 베개에 안정적으로 지지되어 누워있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "눈을 번쩍 뜬 표정과 인물의 외모가 레퍼런스와 완벽히 일치하며, 주변을 배제한 타이트한 클로즈업 프레이밍 지시를 훌륭히 따랐습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "시선이 옆을 향해 갑작스럽게 깨어난 느낌이 부족하며, 자연광 그림자가 강해 눈부신 형광등 조명이라는 지시와 다소 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 우측 상단 허공을 향해 있음.",
        "built_space": "나무 헤드보드 뒤쪽으로 창문과 바다 풍경이 배치됨.",
        "entities": "이현우 레퍼런스와 유사한 젊은 남성, 헝클어진 검은 머리, 흰옷 착용.",
        "hard_violations": [],
        "physics": "머리와 상체가 침대와 베개에 안정적으로 지지되어 누워있음."
       },
       {
        "label": "B",
        "direction": "눈을 크게 뜨고 위쪽 정면을 똑바로 응시함.",
        "built_space": "베개 바로 뒤로 창틀과 바다가 좁은 면적으로 보임.",
        "entities": "이현우 레퍼런스와 이목구비가 매우 일치하는 남성, 헝클어진 검은 머리, 흰옷 착용.",
        "hard_violations": [],
        "physics": "머리와 상체가 침대와 베개에 안정적으로 지지되어 누워있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "흰 침구에 지지된 얼굴과 번쩍 뜬 두 눈을 중심으로 잡아 더 충실하지만, 창 배경이 끼어들고 눈부신 형광등보다는 자연광처럼 보입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "누운 자세와 각성 표정은 맞지만, 참조에 없는 왼쪽 벽 창을 추가하고 창·벽·헤드보드까지 드러내 얼굴 주변 침구만 보이라는 구도를 벗어납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남성은 두 눈을 크게 뜨고 얼굴 위 카메라 부근을 바라봅니다. 시선이 향하는 별도의 물체는 보이지 않으며, 지정된 응시 대상이 없는 각성 장면과 맞습니다.",
        "built_space": "흰 침대 하나의 머리 부분, 베개 하나, 이불과 상의가 보입니다. 화면 위에는 바다가 보이는 창 일부와 수평 창턱이 들어옵니다. 얼굴을 침구가 둘러싸는 구성은 대체로 맞지만, 침구만 배경으로 제한하라는 지시에서 벗어납니다. 형광등 기구는 프레임 밖이며, 얼굴의 밝고 어두운 면은 강한 자연광에 더 가깝습니다. 나머지 실내 설비는 잘려 있어 확인할 수 없습니다.",
        "entities": "참조 인물과 유사한 얼굴형과 헝클어진 검은 머리의 젊은 동아시아계 남성 한 명입니다. 한국계 18세라는 설정과 외관상 크게 충돌하지 않으며 흰 상의를 입었습니다. 흰 침구와 청록색 바다가 보이고, 눈의 홍채·동공·흰자는 정상적인 형태입니다. 읽을 수 있는 글자나 다른 사람은 없습니다.",
        "hard_violations": [],
        "physics": "뒤통수와 목 주변은 눌린 베개가 받치며, 어깨와 몸통은 침대 위에 누워 있습니다. 이불은 가슴 위에 자연스럽게 얹혀 있습니다. 공중에 떠 있거나 지지 없이 들린 신체 부위와 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "남성은 두 눈을 크게 뜨고 얼굴 위쪽의 카메라 부근을 바라봅니다. 머리는 약간 비스듬하지만 특정한 실내 물체를 주시한다고 볼 근거는 없습니다. 깨어나는 순간의 시선으로는 자연스럽습니다.",
        "built_space": "흰 침대 하나, 베개 하나, 나무 헤드보드 하나가 보이며 뒤쪽과 왼쪽의 서로 직각인 두 벽면에 창이 있습니다. 참조는 뒤쪽 벽의 창 한 곳과 막힌 왼쪽 벽이므로 왼쪽 창은 추가된 고정 설비입니다. 침구 이외의 창·벽·헤드보드가 상당한 면적을 차지합니다. 형광등은 보이지 않고 벽과 얼굴에 생긴 방향성 있는 그림자는 주간 자연광처럼 읽힙니다.",
        "entities": "젊은 동아시아계 남성 한 명이며, 얼굴과 헝클어진 검은 머리는 참조 인물에 대체로 부합합니다. 흰 상의, 흰 침구, 나무 헤드보드와 청록색 바다가 보입니다. 눈은 정상적인 인간의 형태이고, 다른 인물이나 읽을 수 있는 문자는 없습니다.",
        "hard_violations": [
         "참조에서 막힌 왼쪽 벽에 별도의 창을 추가해, 한 벽면의 창을 두 벽면의 코너 창 구조로 바꿨습니다."
        ],
        "physics": "머리는 베개에 기대어 있고 목과 어깨도 침구와 침대에 지지됩니다. 몸통 위 이불은 무게에 따라 주름져 있으며, 눈을 뜨는 동작에 불필요한 팔다리 들림이나 지지 없는 부유는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "흰 침구에 지지된 얼굴과 번쩍 뜬 두 눈을 중심으로 잡아 더 충실하지만, 창 배경이 끼어들고 눈부신 형광등보다는 자연광처럼 보입니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "누운 자세와 각성 표정은 맞지만, 참조에 없는 왼쪽 벽 창을 추가하고 창·벽·헤드보드까지 드러내 얼굴 주변 침구만 보이라는 구도를 벗어납니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "남성은 두 눈을 크게 뜨고 얼굴 위 카메라 부근을 바라봅니다. 시선이 향하는 별도의 물체는 보이지 않으며, 지정된 응시 대상이 없는 각성 장면과 맞습니다.",
        "built_space": "흰 침대 하나의 머리 부분, 베개 하나, 이불과 상의가 보입니다. 화면 위에는 바다가 보이는 창 일부와 수평 창턱이 들어옵니다. 얼굴을 침구가 둘러싸는 구성은 대체로 맞지만, 침구만 배경으로 제한하라는 지시에서 벗어납니다. 형광등 기구는 프레임 밖이며, 얼굴의 밝고 어두운 면은 강한 자연광에 더 가깝습니다. 나머지 실내 설비는 잘려 있어 확인할 수 없습니다.",
        "entities": "참조 인물과 유사한 얼굴형과 헝클어진 검은 머리의 젊은 동아시아계 남성 한 명입니다. 한국계 18세라는 설정과 외관상 크게 충돌하지 않으며 흰 상의를 입었습니다. 흰 침구와 청록색 바다가 보이고, 눈의 홍채·동공·흰자는 정상적인 형태입니다. 읽을 수 있는 글자나 다른 사람은 없습니다.",
        "hard_violations": [],
        "physics": "뒤통수와 목 주변은 눌린 베개가 받치며, 어깨와 몸통은 침대 위에 누워 있습니다. 이불은 가슴 위에 자연스럽게 얹혀 있습니다. 공중에 떠 있거나 지지 없이 들린 신체 부위와 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "남성은 두 눈을 크게 뜨고 얼굴 위쪽의 카메라 부근을 바라봅니다. 머리는 약간 비스듬하지만 특정한 실내 물체를 주시한다고 볼 근거는 없습니다. 깨어나는 순간의 시선으로는 자연스럽습니다.",
        "built_space": "흰 침대 하나, 베개 하나, 나무 헤드보드 하나가 보이며 뒤쪽과 왼쪽의 서로 직각인 두 벽면에 창이 있습니다. 참조는 뒤쪽 벽의 창 한 곳과 막힌 왼쪽 벽이므로 왼쪽 창은 추가된 고정 설비입니다. 침구 이외의 창·벽·헤드보드가 상당한 면적을 차지합니다. 형광등은 보이지 않고 벽과 얼굴에 생긴 방향성 있는 그림자는 주간 자연광처럼 읽힙니다.",
        "entities": "젊은 동아시아계 남성 한 명이며, 얼굴과 헝클어진 검은 머리는 참조 인물에 대체로 부합합니다. 흰 상의, 흰 침구, 나무 헤드보드와 청록색 바다가 보입니다. 눈은 정상적인 인간의 형태이고, 다른 인물이나 읽을 수 있는 문자는 없습니다.",
        "hard_violations": [
         "참조에서 막힌 왼쪽 벽에 별도의 창을 추가해, 한 벽면의 창을 두 벽면의 코너 창 구조로 바꿨습니다."
        ],
        "physics": "머리는 베개에 기대어 있고 목과 어깨도 침구와 침대에 지지됩니다. 몸통 위 이불은 무게에 따라 주름져 있으며, 눈을 뜨는 동작에 불필요한 팔다리 들림이나 지지 없는 부유는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.75,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gpt-high] 참조에서 막힌 왼쪽 벽에 별도의 창을 추가해, 한 벽면의 창을 두 벽면의 코너 창 구조로 바꿨습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "눈을 번쩍 뜬 표정과 인물의 외모가 레퍼런스와 완벽히 일치하며, 주변을 배제한 타이트한 클로즈업 프레이밍 지시를 훌륭히 따랐습니다."
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "시선이 옆을 향해 갑작스럽게 깨어난 느낌이 부족하며, 자연광 그림자가 강해 눈부신 형광등 조명이라는 지시와 다소 어긋납니다.  ★위반: [gpt-high] 참조에서 막힌 왼쪽 벽에 별도의 창을 추가해, 한 벽면의 창을 두 벽면의 코너 창 구조로 바꿨습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L127B01.png",
    "asset_id": "7921e778-417f-49f1-9784-6734e0e6fcd3",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aaceff-2133-7ebb-bdda-c5d9d15090e3",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S81sh1__bgfirst_bg.png",
   "bg_asset_id": "21fc7969-a438-4859-becf-903656b8b3b4",
   "bg_record_key": "S81sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S81sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:03:32.970802+00:00",
  "fingerprint": "68697d5e0a63c41a1e95887797a774c84cb4caff0f22c73ff08ad86ff06d6bf9",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S81sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S81sh1_sel.png",
  "source_sha256": "aa32f294f8d2b08acb365379ea2ff889fe7e7caae6078905f39a64f2511994a8",
  "file": "S81sh1_cine.png",
  "staged_sha256": "ef5c11f99bae04372aef7938373e056ef6020d8e052e0d23772091fdf73ddd99",
  "latency_ms": 9232
 },
 "S81sh5::signage": {
  "fp": "5a8bb4cdb05becf4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S81sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:04:40.579957+00:00",
  "fingerprint": "da7e0a6198fb1dd93a051c5d529ddfeb58d3d9cbc05bc5ff24015d02b22f3367",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S81sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S81sh5_sel.png",
  "source_sha256": "43d3b1b7109ee01c11d4f5b4eec6578477c652182da09144cbedb1275d541caf",
  "file": "S81sh5_cine.png",
  "staged_sha256": "79a56a8c4851299eeb0e973f2ffdee82cf507ba375f135d58a28baf936be9e0f",
  "latency_ms": 9671
 },
 "S81sh8::signage": {
  "fp": "30015c3c40a7ce50",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S81sh8::bgfirst_bg": {
  "input_fingerprint": "617defef1aae7595",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 새하얀 연구소 복도, 서지민의 뒤쪽에서 발걸음을 우뚝 멈추고 선 이현우의 전신.\n\nLOCATION (lock): In the white interior corridor of the island research institute, on the route toward a card-controlled door under institutional lighting.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: White research corridor (Occupied by 서지민 ahead and the halted 이현우 behind) — The side wall and floor recede diagonally behind the nearer figure; used as Keep a visible strip of floor between the figures so their interrupted following relationship reads clearly.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient illumination appropriate to the white corridor, preserving facial readability without assigning the room's fluorescent source to this separate space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 새하얀 연구소 복도, 서지민의 뒤쪽에서 발걸음을 우뚝 멈추고 선 이현우의 전신.\n\nLOCATION (lock): In the white interior corridor of the island research institute, on the route toward a card-controlled door under institutional lighting.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: White research corridor (Occupied by 서지민 ahead and the halted 이현우 behind) — The side wall and floor recede diagonally behind the nearer figure; used as Keep a visible strip of floor between the figures so their interrupted following relationship reads clearly.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient illumination appropriate to the white corridor, preserving facial readability without assigning the room's fluorescent source to this separate space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S81sh8__bgfirst_bg.png",
  "asset_id": "f5cc858a-e5e2-400f-8657-0fa6f1771b25",
  "input_asset_ids": [
   "6f8a7b3e-abe3-4305-8014-48420ac98eb5",
   "7921e778-417f-49f1-9784-6734e0e6fcd3"
  ]
 },
 "S81sh8": {
  "input_fingerprint": "d481745cb68d6f1c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 새하얀 연구소 복도, 서지민의 뒤쪽에서 발걸음을 우뚝 멈추고 선 이현우의 전신.\n\nLOCATION (lock): In the white interior corridor of the island research institute, on the route toward a card-controlled door under institutional lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: White research corridor (Occupied by 서지민 ahead and the halted 이현우 behind) — The side wall and floor recede diagonally behind the nearer figure; used as Keep a visible strip of floor between the figures so their interrupted following relationship reads clearly.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient illumination appropriate to the white corridor, preserving facial readability without assigning the room's fluorescent source to this separate space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 이현우: Still dressed in white, he has stopped in the research facility corridor. 서지민: She is in the research facility corridor wearing her research coat and carrying her researcher access card.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 서지민 (20세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 새하얀 연구소 복도, 서지민의 뒤쪽에서 발걸음을 우뚝 멈추고 선 이현우의 전신.\n\nLOCATION (lock): In the white interior corridor of the island research institute, on the route toward a card-controlled door under institutional lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: White research corridor (Occupied by 서지민 ahead and the halted 이현우 behind) — The side wall and floor recede diagonally behind the nearer figure; used as Keep a visible strip of floor between the figures so their interrupted following relationship reads clearly.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient illumination appropriate to the white corridor, preserving facial readability without assigning the room's fluorescent source to this separate space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 이현우: Still dressed in white, he has stopped in the research facility corridor. 서지민: She is in the research facility corridor wearing her research coat and carrying her researcher access card.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 서지민 (20세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 새하얀 연구소 복도, 서지민의 뒤쪽에서 발걸음을 우뚝 멈추고 선 이현우의 전신.\n\nLOCATION (lock): In the white interior corridor of the island research institute, on the route toward a card-controlled door under institutional lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: White research corridor (Occupied by 서지민 ahead and the halted 이현우 behind) — The side wall and floor recede diagonally behind the nearer figure; used as Keep a visible strip of floor between the figures so their interrupted following relationship reads clearly.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient illumination appropriate to the white corridor, preserving facial readability without assigning the room's fluorescent source to this separate space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 이현우: Still dressed in white, he has stopped in the research facility corridor. 서지민: She is in the research facility corridor wearing her research coat and carrying her researcher access card.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리); 서지민 (20세 여성, 젊은 성인의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S81sh8__bgfirst_bg.png",
     "asset_id": "f5cc858a-e5e2-400f-8657-0fa6f1771b25",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S81sh8.png",
     "asset_id": "6f8a7b3e-abe3-4305-8014-48420ac98eb5",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 서지민: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:746082>",
     "asset_id": "a9a13f64-38f2-4a18-91db-eab511b31c43",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L127B01.png",
     "asset_id": "7921e778-417f-49f1-9784-6734e0e6fcd3",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 서지민: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:746082>",
     "asset_id": "a9a13f64-38f2-4a18-91db-eab511b31c43",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "서지민은 화면 오른쪽 밖을 바라보고 있으며, 이현우는 정면을 응시하고 있음.",
    "built_space": "우측에 유리 벽면과 좌측 벽에 핸드레일이 있는 복도로, 참조 이미지의 회색 바닥과 일반 문틀 구조와 완전히 다름.",
    "entities": "서지민과 이현우 모두 참조된 외모와 의상을 따르고 있으며, 서지민은 신분증을 들고 있음. 두 인물 모두 옷과 카드에 읽을 수 있는 글씨가 남아있음.",
    "hard_violations": [],
    "physics": "두 인물 모두 복도 바닥에 올바르게 서 있음."
   },
   {
    "label": "B",
    "direction": "이현우는 정면에서 약간 왼쪽을 바라보고, 서지민은 복도 안쪽을 향해 걸어가고 있음.",
    "built_space": "흰색 벽, 회색 바닥, 회색 문틀 등 참조 사진의 장소적 특징을 잘 반영함. 우측에 지시된 카드 리더기 문이 있음.",
    "entities": "이현우는 전신이 노출되며 참조 이미지와 일치하나 셔츠에 글씨가 있음. 서지민은 뒷모습이며 가운을 입고 오른손에 카드를 들고 있음.",
    "hard_violations": [
     "[gpt-high] 이현우의 상의 가슴에 판독 가능한 영문 인쇄가 있어, 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반합니다."
    ],
    "physics": "이현우는 바닥에 서 있고, 서지민은 걷는 자세로 바닥을 딛고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "참조된 장소의 구조와 '앞서가는 서지민, 그 뒤에 멈춰 선 이현우'의 원근 구도를 지시문대로 정확히 구현했으나, 의상에 글씨가 노출된 점이 감점 요소입니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지정된 위치(일반 복도)와 다른 유리 벽 복도를 생성하여 장소 일치 조건(Location Lock)을 위반했으며, 두 인물의 공간적 관계 연출이 지시문보다 덜 효과적입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "서지민은 화면 오른쪽 밖을 바라보고 있으며, 이현우는 정면을 응시하고 있음.",
        "built_space": "우측에 유리 벽면과 좌측 벽에 핸드레일이 있는 복도로, 참조 이미지의 회색 바닥과 일반 문틀 구조와 완전히 다름.",
        "entities": "서지민과 이현우 모두 참조된 외모와 의상을 따르고 있으며, 서지민은 신분증을 들고 있음. 두 인물 모두 옷과 카드에 읽을 수 있는 글씨가 남아있음.",
        "hard_violations": [],
        "physics": "두 인물 모두 복도 바닥에 올바르게 서 있음."
       },
       {
        "label": "B",
        "direction": "이현우는 정면에서 약간 왼쪽을 바라보고, 서지민은 복도 안쪽을 향해 걸어가고 있음.",
        "built_space": "흰색 벽, 회색 바닥, 회색 문틀 등 참조 사진의 장소적 특징을 잘 반영함. 우측에 지시된 카드 리더기 문이 있음.",
        "entities": "이현우는 전신이 노출되며 참조 이미지와 일치하나 셔츠에 글씨가 있음. 서지민은 뒷모습이며 가운을 입고 오른손에 카드를 들고 있음.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 서 있고, 서지민은 걷는 자세로 바닥을 딛고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "참조된 장소의 구조와 '앞서가는 서지민, 그 뒤에 멈춰 선 이현우'의 원근 구도를 지시문대로 정확히 구현했으나, 의상에 글씨가 노출된 점이 감점 요소입니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지정된 위치(일반 복도)와 다른 유리 벽 복도를 생성하여 장소 일치 조건(Location Lock)을 위반했으며, 두 인물의 공간적 관계 연출이 지시문보다 덜 효과적입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "서지민은 화면 오른쪽 밖을 바라보고 있으며, 이현우는 정면을 응시하고 있음.",
        "built_space": "우측에 유리 벽면과 좌측 벽에 핸드레일이 있는 복도로, 참조 이미지의 회색 바닥과 일반 문틀 구조와 완전히 다름.",
        "entities": "서지민과 이현우 모두 참조된 외모와 의상을 따르고 있으며, 서지민은 신분증을 들고 있음. 두 인물 모두 옷과 카드에 읽을 수 있는 글씨가 남아있음.",
        "hard_violations": [],
        "physics": "두 인물 모두 복도 바닥에 올바르게 서 있음."
       },
       {
        "label": "B",
        "direction": "이현우는 정면에서 약간 왼쪽을 바라보고, 서지민은 복도 안쪽을 향해 걸어가고 있음.",
        "built_space": "흰색 벽, 회색 바닥, 회색 문틀 등 참조 사진의 장소적 특징을 잘 반영함. 우측에 지시된 카드 리더기 문이 있음.",
        "entities": "이현우는 전신이 노출되며 참조 이미지와 일치하나 셔츠에 글씨가 있음. 서지민은 뒷모습이며 가운을 입고 오른손에 카드를 들고 있음.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 서 있고, 서지민은 걷는 자세로 바닥을 딛고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이현우의 전신과 앞서가는 서지민 사이의 바닥 간격은 잘 보이지만, 가슴의 판독 가능한 영문이 무문자 조건을 위반하고 지정 장소의 일치도도 부족합니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "앞선 서지민과 뒤에 멈춘 이현우의 전신·시선 관계가 명확하지만, 유리벽 중심의 복도는 장소 참조와 다르고 이현우의 자세는 급정지보다 정자세에 가깝습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "서지민은 등을 보인 채 복도 끝 닫힌 문으로 걸어갑니다. 가까운 오른쪽의 이현우는 고개를 왼쪽 복도 안쪽으로 돌렸으며 서지민 쪽을 바라보는 것으로 읽힙니다. 서지민이 앞서고 이현우가 뒤에 남은 이동 관계는 성립합니다.",
        "built_space": "좁은 흰 복도에 회색 바닥과 걸레받이가 있고, 왼쪽 벽과 바닥이 끝문 쪽으로 대각선 후퇴합니다. 정면 끝문 하나, 가까운 오른쪽 유리창 달린 문 하나, 측면의 추가 문틀들이 보입니다. 끝문과 가까운 오른쪽 문에 각각 검은 잠금장치가 있고 천장에는 직사각 조명 하나가 뚜렷합니다. 두 사람 사이에는 빈 바닥이 충분합니다. 다만 참조에서 확인되는 방 출입구와 옆 복도의 연결 구조를 재현했다고 확인할 특징은 부족합니다. 참조의 침대 등은 방 내부 물건이므로 복도 프레임에 없다는 것 자체는 결함이 아닙니다.",
        "entities": "인물은 두 명뿐입니다. 이현우는 젊은 한국계 남성의 외모, 검은 머리, 마른 체형, 흰 긴팔 상하의와 흰 신발을 갖췄으며 참조와 대체로 맞습니다. 서지민은 짧은 검은 머리의 여성으로 연구복과 짙은 바지, 어두운 신발을 착용했고 오른손에 작은 카드를 들고 있습니다. 뒷모습이므로 얼굴 일치는 판단할 수 없습니다. 이현우의 가슴에는 읽을 수 있는 영문 인쇄가 남아 있습니다.",
        "hard_violations": [
         "이현우의 상의 가슴에 판독 가능한 영문 인쇄가 있어, 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반합니다."
        ],
        "physics": "이현우의 두 신발은 바닥에 닿아 몸을 지지합니다. 다만 발이 거의 나란하고 팔이 아래로 내려와 있어 걷다가 우뚝 멈춘 비대칭 체중 이동은 약합니다. 서지민은 한 발로 체중을 받으며 다른 발을 내딛는 보행 자세이고, 카드는 오른손으로 잡고 있습니다. 지지 없이 떠 있는 신체나 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "서지민은 화면 앞쪽으로 이동하면서 고개와 눈을 오른쪽 뒤의 이현우 쪽으로 돌립니다. 이현우는 화면 왼쪽 앞에 있는 서지민을 바라봅니다. 선행하던 서지민과 뒤에서 멈춘 이현우의 관계가 시선과 거리로 드러납니다.",
        "built_space": "왼쪽에는 흰 벽과 금속 문틀의 문 두 개가 뚜렷하고, 오른쪽에는 여러 칸의 유리벽과 가까운 유리문 하나가 이어집니다. 오른쪽 가까운 문에는 손잡이와 검은 카드 판독기 하나가 보이며, 복도 끝에는 바깥 풍경이 보이는 유리문이 있습니다. 천장 직사각 조명은 원근을 따라 다섯 개 보입니다. 두 사람 사이의 바닥은 넓게 드러나고 벽·바닥이 대각선으로 후퇴합니다. 다만 참조의 불투명 벽과 짧은 옆 복도 대신 길고 유리 칸막이가 많은 공간이어서 정확한 장소 잠금에는 미달합니다. 유리와 바닥의 약한 반사는 광학적으로 무리하지 않습니다.",
        "entities": "인물은 서지민과 이현우 두 명뿐입니다. 이현우의 젊은 한국계 남성 얼굴, 헝클어진 검은 머리, 체형과 흰 옷·신발은 참조에 가깝습니다. 서지민은 젊은 여성으로 피부색과 얼굴 인상이 참조에 대체로 가깝지만, 머리를 뒤로 묶었고 참조의 셔츠·짙은 넥타이 대신 밝은 둥근 목 상의를 입었습니다. 연구복과 짙은 바지는 유지하며 목걸이형 출입카드를 손으로 잡고 있습니다. 카드와 상의의 작은 표시는 있으나 문구를 확실히 판독하기 어렵습니다.",
        "hard_violations": [],
        "physics": "이현우는 두 발로 바닥에 안정적으로 서 있습니다. 발 간격과 처진 양팔이 거의 대칭이어서 급히 멈춘 순간보다는 서서 기다리는 자세에 가깝습니다. 서지민의 발은 프레임 밖이지만 몸통과 다리는 자연스러운 보행으로 이어지며, 잘렸다는 이유로 지지 불능이라고 볼 근거는 없습니다. 출입카드는 목줄과 손이 지지합니다. 공중에 떠 있는 물체나 불가능한 관절은 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이현우의 전신과 앞서가는 서지민 사이의 바닥 간격은 잘 보이지만, 가슴의 판독 가능한 영문이 무문자 조건을 위반하고 지정 장소의 일치도도 부족합니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "앞선 서지민과 뒤에 멈춘 이현우의 전신·시선 관계가 명확하지만, 유리벽 중심의 복도는 장소 참조와 다르고 이현우의 자세는 급정지보다 정자세에 가깝습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "서지민은 등을 보인 채 복도 끝 닫힌 문으로 걸어갑니다. 가까운 오른쪽의 이현우는 고개를 왼쪽 복도 안쪽으로 돌렸으며 서지민 쪽을 바라보는 것으로 읽힙니다. 서지민이 앞서고 이현우가 뒤에 남은 이동 관계는 성립합니다.",
        "built_space": "좁은 흰 복도에 회색 바닥과 걸레받이가 있고, 왼쪽 벽과 바닥이 끝문 쪽으로 대각선 후퇴합니다. 정면 끝문 하나, 가까운 오른쪽 유리창 달린 문 하나, 측면의 추가 문틀들이 보입니다. 끝문과 가까운 오른쪽 문에 각각 검은 잠금장치가 있고 천장에는 직사각 조명 하나가 뚜렷합니다. 두 사람 사이에는 빈 바닥이 충분합니다. 다만 참조에서 확인되는 방 출입구와 옆 복도의 연결 구조를 재현했다고 확인할 특징은 부족합니다. 참조의 침대 등은 방 내부 물건이므로 복도 프레임에 없다는 것 자체는 결함이 아닙니다.",
        "entities": "인물은 두 명뿐입니다. 이현우는 젊은 한국계 남성의 외모, 검은 머리, 마른 체형, 흰 긴팔 상하의와 흰 신발을 갖췄으며 참조와 대체로 맞습니다. 서지민은 짧은 검은 머리의 여성으로 연구복과 짙은 바지, 어두운 신발을 착용했고 오른손에 작은 카드를 들고 있습니다. 뒷모습이므로 얼굴 일치는 판단할 수 없습니다. 이현우의 가슴에는 읽을 수 있는 영문 인쇄가 남아 있습니다.",
        "hard_violations": [
         "이현우의 상의 가슴에 판독 가능한 영문 인쇄가 있어, 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반합니다."
        ],
        "physics": "이현우의 두 신발은 바닥에 닿아 몸을 지지합니다. 다만 발이 거의 나란하고 팔이 아래로 내려와 있어 걷다가 우뚝 멈춘 비대칭 체중 이동은 약합니다. 서지민은 한 발로 체중을 받으며 다른 발을 내딛는 보행 자세이고, 카드는 오른손으로 잡고 있습니다. 지지 없이 떠 있는 신체나 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "서지민은 화면 앞쪽으로 이동하면서 고개와 눈을 오른쪽 뒤의 이현우 쪽으로 돌립니다. 이현우는 화면 왼쪽 앞에 있는 서지민을 바라봅니다. 선행하던 서지민과 뒤에서 멈춘 이현우의 관계가 시선과 거리로 드러납니다.",
        "built_space": "왼쪽에는 흰 벽과 금속 문틀의 문 두 개가 뚜렷하고, 오른쪽에는 여러 칸의 유리벽과 가까운 유리문 하나가 이어집니다. 오른쪽 가까운 문에는 손잡이와 검은 카드 판독기 하나가 보이며, 복도 끝에는 바깥 풍경이 보이는 유리문이 있습니다. 천장 직사각 조명은 원근을 따라 다섯 개 보입니다. 두 사람 사이의 바닥은 넓게 드러나고 벽·바닥이 대각선으로 후퇴합니다. 다만 참조의 불투명 벽과 짧은 옆 복도 대신 길고 유리 칸막이가 많은 공간이어서 정확한 장소 잠금에는 미달합니다. 유리와 바닥의 약한 반사는 광학적으로 무리하지 않습니다.",
        "entities": "인물은 서지민과 이현우 두 명뿐입니다. 이현우의 젊은 한국계 남성 얼굴, 헝클어진 검은 머리, 체형과 흰 옷·신발은 참조에 가깝습니다. 서지민은 젊은 여성으로 피부색과 얼굴 인상이 참조에 대체로 가깝지만, 머리를 뒤로 묶었고 참조의 셔츠·짙은 넥타이 대신 밝은 둥근 목 상의를 입었습니다. 연구복과 짙은 바지는 유지하며 목걸이형 출입카드를 손으로 잡고 있습니다. 카드와 상의의 작은 표시는 있으나 문구를 확실히 판독하기 어렵습니다.",
        "hard_violations": [],
        "physics": "이현우는 두 발로 바닥에 안정적으로 서 있습니다. 발 간격과 처진 양팔이 거의 대칭이어서 급히 멈춘 순간보다는 서서 기다리는 자세에 가깝습니다. 서지민의 발은 프레임 밖이지만 몸통과 다리는 자연스러운 보행으로 이어지며, 잘렸다는 이유로 지지 불능이라고 볼 근거는 없습니다. 출입카드는 목줄과 손이 지지합니다. 공중에 떠 있는 물체나 불가능한 관절은 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gpt-high] 이현우의 상의 가슴에 판독 가능한 영문 인쇄가 있어, 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1250,
   "A": 1571
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "참조된 장소의 구조와 '앞서가는 서지민, 그 뒤에 멈춰 선 이현우'의 원근 구도를 지시문대로 정확히 구현했으나, 의상에 글씨가 노출된 점이 감점 요소입니다.  ★위반: [gpt-high] 이현우의 상의 가슴에 판독 가능한 영문 인쇄가 있어, 어디에도 읽을 수 있는 글자가 없어야 한다는 조건을 위반합니다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "지정된 위치(일반 복도)와 다른 유리 벽 복도를 생성하여 장소 일치 조건(Location Lock)을 위반했으며, 두 인물의 공간적 관계 연출이 지시문보다 덜 효과적입니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L127B01.png",
    "asset_id": "7921e778-417f-49f1-9784-6734e0e6fcd3",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 서지민: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:746082>",
    "asset_id": "a9a13f64-38f2-4a18-91db-eab511b31c43",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf0a-29c0-7abf-bd75-9407fb9534ab",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S81sh8__bgfirst_bg.png",
   "bg_asset_id": "f5cc858a-e5e2-400f-8657-0fa6f1771b25",
   "bg_record_key": "S81sh8::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S81sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:06:22.787877+00:00",
  "fingerprint": "24222990e61752471f71f89f3f3ad6fb1d5a212c5695eee17cb25860826c674d",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S81sh8_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S81sh8_sel.png",
  "source_sha256": "e50da6ffeeac861257454a4c658aa68387a541b5ea45d100bde5c04077893a03",
  "file": "S81sh8_cine.png",
  "staged_sha256": "87c6abb37e8ce1b5b1699a3ec1d978845f5aa7fbb6264d1a76fe47460a5893d9",
  "latency_ms": 9643
 },
 "S82sh5::signage": {
  "fp": "6ae4a5e209152a39",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::917fdaa6dcf743be": {
  "subjects": [],
  "subject_text": "제주도 연구소 메인 연구센터 공용 공간\n발전소처럼 거대한 규모의 연구 공간. 각종 컴퓨터와 연구 장치 사이로 열대 식물과 꽃이 풍성하게 배치되어 있다.",
  "identity": "canonical",
  "scope_id": "L129",
  "scope_role": "location_interior",
  "scope_sha": "dc6f106b95737a2d"
 },
 "S82sh5::bgfirst_bg": {
  "input_fingerprint": "93b7ad9ec219112d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 세련된 흰색 연구복 차림의 50대 지소영이 두 사람을 향해 한쪽 발을 앞으로 내디디며 걷고 있는 위압적인 전신.\n\nLOCATION (lock): In the expansive main hall of the island energy research center, among tropical plants and research equipment under ambient research-hall lighting.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Research equipment (Distributed throughout the main research center) — Oblique sides recede around the open approach path; used as Establish the facility's scale while keeping 지소영's full silhouette unobstructed; Tropical plants and flowers (Numerous within the main research space); used as Break up the equipment-filled background without placing a large foreground obstruction over the approaching figure; Main research-center interior (A vast space containing computers, equipment, and planting) — Viewed diagonally past the two foreground observers; used as Preserve expansive depth above and beside the full-length figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient light appropriate to the research center, balancing the white coat against the surrounding equipment without inventing a distinctive source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 세련된 흰색 연구복 차림의 50대 지소영이 두 사람을 향해 한쪽 발을 앞으로 내디디며 걷고 있는 위압적인 전신.\n\nLOCATION (lock): In the expansive main hall of the island energy research center, among tropical plants and research equipment under ambient research-hall lighting.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Research equipment (Distributed throughout the main research center) — Oblique sides recede around the open approach path; used as Establish the facility's scale while keeping 지소영's full silhouette unobstructed; Tropical plants and flowers (Numerous within the main research space); used as Break up the equipment-filled background without placing a large foreground obstruction over the approaching figure; Main research-center interior (A vast space containing computers, equipment, and planting) — Viewed diagonally past the two foreground observers; used as Preserve expansive depth above and beside the full-length figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient light appropriate to the research center, balancing the white coat against the surrounding equipment without inventing a distinctive source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S82sh5__bgfirst_bg.png",
  "asset_id": "44e2cc69-8214-4697-a78d-543ddf925b63",
  "input_asset_ids": [
   "8ce9deaa-5e60-4b66-a9fd-d69c2650e24d",
   "4a6aa631-eaf4-4be4-a449-2ca4a933e26a"
  ]
 },
 "S82sh5": {
  "input_fingerprint": "d0a06ec97c1d8207",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 세련된 흰색 연구복 차림의 50대 지소영이 두 사람을 향해 한쪽 발을 앞으로 내디디며 걷고 있는 위압적인 전신.\n\nLOCATION (lock): In the expansive main hall of the island energy research center, among tropical plants and research equipment under ambient research-hall lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Research equipment (Distributed throughout the main research center) — Oblique sides recede around the open approach path; used as Establish the facility's scale while keeping 지소영's full silhouette unobstructed; Tropical plants and flowers (Numerous within the main research space); used as Break up the equipment-filled background without placing a large foreground obstruction over the approaching figure; Main research-center interior (A vast space containing computers, equipment, and planting) — Viewed diagonally past the two foreground observers; used as Preserve expansive depth above and beside the full-length figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient light appropriate to the research center, balancing the white coat against the surrounding equipment without inventing a distinctive source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main research center contains numerous computers and research devices among tropical plants and flowers. 지소영: She wears a clean research coat and is approaching through the main center.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 세련된 흰색 연구복 차림의 50대 지소영이 두 사람을 향해 한쪽 발을 앞으로 내디디며 걷고 있는 위압적인 전신.\n\nLOCATION (lock): In the expansive main hall of the island energy research center, among tropical plants and research equipment under ambient research-hall lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Research equipment (Distributed throughout the main research center) — Oblique sides recede around the open approach path; used as Establish the facility's scale while keeping 지소영's full silhouette unobstructed; Tropical plants and flowers (Numerous within the main research space); used as Break up the equipment-filled background without placing a large foreground obstruction over the approaching figure; Main research-center interior (A vast space containing computers, equipment, and planting) — Viewed diagonally past the two foreground observers; used as Preserve expansive depth above and beside the full-length figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient light appropriate to the research center, balancing the white coat against the surrounding equipment without inventing a distinctive source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main research center contains numerous computers and research devices among tropical plants and flowers. 지소영: She wears a clean research coat and is approaching through the main center.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 세련된 흰색 연구복 차림의 50대 지소영이 두 사람을 향해 한쪽 발을 앞으로 내디디며 걷고 있는 위압적인 전신.\n\nLOCATION (lock): In the expansive main hall of the island energy research center, among tropical plants and research equipment under ambient research-hall lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Research equipment (Distributed throughout the main research center) — Oblique sides recede around the open approach path; used as Establish the facility's scale while keeping 지소영's full silhouette unobstructed; Tropical plants and flowers (Numerous within the main research space); used as Break up the equipment-filled background without placing a large foreground obstruction over the approaching figure; Main research-center interior (A vast space containing computers, equipment, and planting) — Viewed diagonally past the two foreground observers; used as Preserve expansive depth above and beside the full-length figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use ambient light appropriate to the research center, balancing the white coat against the surrounding equipment without inventing a distinctive source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main research center contains numerous computers and research devices among tropical plants and flowers. 지소영: She wears a clean research coat and is approaching through the main center.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (50대 중반 여성, 중년의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S82sh5__bgfirst_bg.png",
     "asset_id": "44e2cc69-8214-4697-a78d-543ddf925b63",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S82sh5.png",
     "asset_id": "8ce9deaa-5e60-4b66-a9fd-d69c2650e24d",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:808347>",
     "asset_id": "f15ee474-495b-45e2-8cb1-804018e9010e",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L129B01.png",
     "asset_id": "4a6aa631-eaf4-4be4-a449-2ca4a933e26a",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:808347>",
     "asset_id": "f15ee474-495b-45e2-8cb1-804018e9010e",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "지소영이 전경에 배치된 두 사람을 향해 걷고 있음.",
    "built_space": "레퍼런스와 일치하는 넓은 연구소 공간으로, 좌측의 거대 장비와 우측의 유리실이 올바른 비례로 배치됨.",
    "entities": "지소영의 외모와 흰색 연구복이 레퍼런스와 일치하며, 지문에 요구된 두 명의 관찰자가 전경 양측에 존재함.",
    "hard_violations": [],
    "physics": "내디딘 발과 체중 이동이 바닥에 안정적으로 지지되어 걷는 동작을 자연스럽게 보여줌."
   },
   {
    "label": "B",
    "direction": "지소영이 카메라 쪽을 향해 걷고 있음.",
    "built_space": "연구소의 기본 요소는 있으나, 우측 전경에 거대한 화단이 배치되어 공간감이 달라짐.",
    "entities": "지소영의 묘사는 일치하나, 지정된 인물 외에 데스크에 앉아 있는 다수의 연구원들이 배경에 나타남.",
    "hard_violations": [
     "[gemini-pro] 지문에 명시되지 않은 다수의 인물(배경에서 작업 중인 연구원들)이 발명되어 추가됨",
     "[gpt-high] 지소영과 두 관찰자 외에 작업 중인 연구원 최소 일곱 명을 추가하여, 샷에 지정되지 않은 인물들이 등장한다."
    ],
    "physics": "걷는 포즈와 발의 접지가 바닥면에 올바르게 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지문에 명시된 인물들만 정확하게 배치하고, 전경의 두 사람을 걸쳐 찍는 앵글과 연구소 배경을 충실히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지문에 없는 인물들이 배경에 다수 추가되어 '추가 인물 금지' 규칙을 심각하게 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지소영이 전경에 배치된 두 사람을 향해 걷고 있음.",
        "built_space": "레퍼런스와 일치하는 넓은 연구소 공간으로, 좌측의 거대 장비와 우측의 유리실이 올바른 비례로 배치됨.",
        "entities": "지소영의 외모와 흰색 연구복이 레퍼런스와 일치하며, 지문에 요구된 두 명의 관찰자가 전경 양측에 존재함.",
        "hard_violations": [],
        "physics": "내디딘 발과 체중 이동이 바닥에 안정적으로 지지되어 걷는 동작을 자연스럽게 보여줌."
       },
       {
        "label": "B",
        "direction": "지소영이 카메라 쪽을 향해 걷고 있음.",
        "built_space": "연구소의 기본 요소는 있으나, 우측 전경에 거대한 화단이 배치되어 공간감이 달라짐.",
        "entities": "지소영의 묘사는 일치하나, 지정된 인물 외에 데스크에 앉아 있는 다수의 연구원들이 배경에 나타남.",
        "hard_violations": [
         "지문에 명시되지 않은 다수의 인물(배경에서 작업 중인 연구원들)이 발명되어 추가됨"
        ],
        "physics": "걷는 포즈와 발의 접지가 바닥면에 올바르게 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지문에 명시된 인물들만 정확하게 배치하고, 전경의 두 사람을 걸쳐 찍는 앵글과 연구소 배경을 충실히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지문에 없는 인물들이 배경에 다수 추가되어 '추가 인물 금지' 규칙을 심각하게 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "지소영이 전경에 배치된 두 사람을 향해 걷고 있음.",
        "built_space": "레퍼런스와 일치하는 넓은 연구소 공간으로, 좌측의 거대 장비와 우측의 유리실이 올바른 비례로 배치됨.",
        "entities": "지소영의 외모와 흰색 연구복이 레퍼런스와 일치하며, 지문에 요구된 두 명의 관찰자가 전경 양측에 존재함.",
        "hard_violations": [],
        "physics": "내디딘 발과 체중 이동이 바닥에 안정적으로 지지되어 걷는 동작을 자연스럽게 보여줌."
       },
       {
        "label": "B",
        "direction": "지소영이 카메라 쪽을 향해 걷고 있음.",
        "built_space": "연구소의 기본 요소는 있으나, 우측 전경에 거대한 화단이 배치되어 공간감이 달라짐.",
        "entities": "지소영의 묘사는 일치하나, 지정된 인물 외에 데스크에 앉아 있는 다수의 연구원들이 배경에 나타남.",
        "hard_violations": [
         "지문에 명시되지 않은 다수의 인물(배경에서 작업 중인 연구원들)이 발명되어 추가됨"
        ],
        "physics": "걷는 포즈와 발의 접지가 바닥면에 올바르게 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "전신 보행과 넓은 연구홀은 구현했지만, 지정된 세 사람 외에 여러 연구원을 추가한 치명적 위반이 있으며 지소영의 접근 방향도 오른쪽 관찰자들과 덜 맞는다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "두 전경 관찰자 사이로 다가오는 지소영의 위압적인 전신, 한 발을 내딛는 순간, 참조 연구홀의 주요 구조와 인물 외형을 가장 충실하게 구현했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지소영의 얼굴과 몸, 앞으로 내민 신발은 거의 카메라 정면을 향한다. 두 관찰자는 화면 오른쪽 전경에서 지소영 쪽으로 고개를 돌리고 있으나, 지소영의 시선과 보행 축은 이들보다 화면 중앙의 카메라 쪽에 가깝다. 배경 연구원들은 책상과 모니터 쪽을 보고 있다.",
        "built_space": "왼쪽에 대형 원통형 연구장비 두 대와 여러 컴퓨터 작업대가 있고, 중앙 뒤와 오른쪽 전경에는 금속 식재함이 보인다. 왼쪽 높은 창, 선형 천장등, 광택 콘크리트 바닥은 참조와 유사하다. 다만 오른쪽은 긴 유리벽과 상부 창 구획으로 바뀌어, 참조의 원형 천장 설비가 있는 유리 실험실보다 다른 홀 구획처럼 보인다. 중앙 통로는 열려 있고 지소영의 전신을 가리지 않는다.",
        "entities": "중앙에는 짧은 회갈색 머리의 중년 여성이 흰 연구복, 흰 셔츠, 짙은 바지와 검은 신발을 착용해 참조 지소영의 외형과 대체로 맞는다. 오른쪽 전경에는 검은 머리의 성인 관찰자 두 명이 있다. 이들 외에도 왼쪽과 뒤쪽 작업대에 최소 일곱 명의 연구원이 추가로 보인다. 컴퓨터, 연구장비, 무성한 열대 식물과 일부 꽃은 있으며, 뚜렷하게 읽히는 문구는 확인되지 않는다.",
        "hard_violations": [
         "지소영과 두 관찰자 외에 작업 중인 연구원 최소 일곱 명을 추가하여, 샷에 지정되지 않은 인물들이 등장한다."
        ],
        "physics": "지소영은 앞쪽 검은 신발로 바닥을 디디고 다른 다리를 뒤에 두어 보행 중 체중 이동으로 해석할 수 있다. 배경의 앉은 인물들은 의자 좌판에 지지되고 책상과 자연스럽게 연결된다. 전경 관찰자들은 하체가 프레임 밖이므로 발 접촉은 확인되지 않지만 상체 자세에 부유 징후는 없다. 장비와 식재함은 바닥에 놓여 있으며 바닥 반사도 가능한 방향이다."
       },
       {
        "label": "B",
        "direction": "지소영의 얼굴과 몸은 양쪽 전경 관찰자 사이를 향하고, 한쪽 발을 그쪽으로 내딛는다. 두 관찰자는 각각 화면 안쪽의 지소영을 향해 고개를 돌리고 있어 접근 대상과 상호 시선 관계가 명확하다. 오른쪽 유리실의 로봇 팔은 실험실 내부를 향하며 사람을 겨냥하지 않는다.",
        "built_space": "왼쪽에는 대형 원통형 장비 두 대, 컴퓨터 작업대와 의자들이 후퇴하며 배치되어 있다. 오른쪽에는 유리로 구획된 실험실 한 곳과 보이는 로봇 팔 한 대가 있다. 천장에는 겹쳐 보이는 원형 발광 윤곽들이 있어 참조의 단일 원형 설비와 세부 형태는 다르다. 높은 왼쪽 창, 선형 천장등, 금속 식재함, 광택 바닥과 중앙의 빈 접근 통로는 참조 장소에 가깝다. 관찰자들은 좌우 가장자리에 있어 지소영의 전신을 가리지 않는다.",
        "entities": "중앙 여성은 참조와 가까운 중년 얼굴, 짧은 회갈색 머리와 체형을 보이며, 흰 연구복과 단추 셔츠, 짙은 바지, 검은 신발도 일치한다. 양쪽 전경에는 검은 머리의 성인 관찰자 두 명만 있고 추가 인물은 보이지 않는다. 컴퓨터와 연구장비, 열대 잎 식물 및 일부 꽃이 배치되어 있다. 얼굴만으로 국적이나 민족을 확정할 수는 없으며, 지소영의 시각적 정체성은 참조와 잘 맞는다. 판독 가능한 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "지소영은 화면 왼쪽의 앞발을 바닥에 디뎌 체중을 받치고, 반대쪽 무릎을 굽힌 채 신발을 들어 다음 보폭으로 옮기는 중이다. 들린 발은 지지 다리가 있는 정상적인 보행 동작으로 설명된다. 두 관찰자의 발은 프레임 밖이지만 상체 배치는 자연스럽다. 연구장비와 식재함은 바닥에, 모니터는 작업대에 지지되어 있으며 유리와 바닥의 반사도 명백한 광학적 모순이 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "전신 보행과 넓은 연구홀은 구현했지만, 지정된 세 사람 외에 여러 연구원을 추가한 치명적 위반이 있으며 지소영의 접근 방향도 오른쪽 관찰자들과 덜 맞는다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "두 전경 관찰자 사이로 다가오는 지소영의 위압적인 전신, 한 발을 내딛는 순간, 참조 연구홀의 주요 구조와 인물 외형을 가장 충실하게 구현했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "지소영의 얼굴과 몸, 앞으로 내민 신발은 거의 카메라 정면을 향한다. 두 관찰자는 화면 오른쪽 전경에서 지소영 쪽으로 고개를 돌리고 있으나, 지소영의 시선과 보행 축은 이들보다 화면 중앙의 카메라 쪽에 가깝다. 배경 연구원들은 책상과 모니터 쪽을 보고 있다.",
        "built_space": "왼쪽에 대형 원통형 연구장비 두 대와 여러 컴퓨터 작업대가 있고, 중앙 뒤와 오른쪽 전경에는 금속 식재함이 보인다. 왼쪽 높은 창, 선형 천장등, 광택 콘크리트 바닥은 참조와 유사하다. 다만 오른쪽은 긴 유리벽과 상부 창 구획으로 바뀌어, 참조의 원형 천장 설비가 있는 유리 실험실보다 다른 홀 구획처럼 보인다. 중앙 통로는 열려 있고 지소영의 전신을 가리지 않는다.",
        "entities": "중앙에는 짧은 회갈색 머리의 중년 여성이 흰 연구복, 흰 셔츠, 짙은 바지와 검은 신발을 착용해 참조 지소영의 외형과 대체로 맞는다. 오른쪽 전경에는 검은 머리의 성인 관찰자 두 명이 있다. 이들 외에도 왼쪽과 뒤쪽 작업대에 최소 일곱 명의 연구원이 추가로 보인다. 컴퓨터, 연구장비, 무성한 열대 식물과 일부 꽃은 있으며, 뚜렷하게 읽히는 문구는 확인되지 않는다.",
        "hard_violations": [
         "지소영과 두 관찰자 외에 작업 중인 연구원 최소 일곱 명을 추가하여, 샷에 지정되지 않은 인물들이 등장한다."
        ],
        "physics": "지소영은 앞쪽 검은 신발로 바닥을 디디고 다른 다리를 뒤에 두어 보행 중 체중 이동으로 해석할 수 있다. 배경의 앉은 인물들은 의자 좌판에 지지되고 책상과 자연스럽게 연결된다. 전경 관찰자들은 하체가 프레임 밖이므로 발 접촉은 확인되지 않지만 상체 자세에 부유 징후는 없다. 장비와 식재함은 바닥에 놓여 있으며 바닥 반사도 가능한 방향이다."
       },
       {
        "label": "A",
        "direction": "지소영의 얼굴과 몸은 양쪽 전경 관찰자 사이를 향하고, 한쪽 발을 그쪽으로 내딛는다. 두 관찰자는 각각 화면 안쪽의 지소영을 향해 고개를 돌리고 있어 접근 대상과 상호 시선 관계가 명확하다. 오른쪽 유리실의 로봇 팔은 실험실 내부를 향하며 사람을 겨냥하지 않는다.",
        "built_space": "왼쪽에는 대형 원통형 장비 두 대, 컴퓨터 작업대와 의자들이 후퇴하며 배치되어 있다. 오른쪽에는 유리로 구획된 실험실 한 곳과 보이는 로봇 팔 한 대가 있다. 천장에는 겹쳐 보이는 원형 발광 윤곽들이 있어 참조의 단일 원형 설비와 세부 형태는 다르다. 높은 왼쪽 창, 선형 천장등, 금속 식재함, 광택 바닥과 중앙의 빈 접근 통로는 참조 장소에 가깝다. 관찰자들은 좌우 가장자리에 있어 지소영의 전신을 가리지 않는다.",
        "entities": "중앙 여성은 참조와 가까운 중년 얼굴, 짧은 회갈색 머리와 체형을 보이며, 흰 연구복과 단추 셔츠, 짙은 바지, 검은 신발도 일치한다. 양쪽 전경에는 검은 머리의 성인 관찰자 두 명만 있고 추가 인물은 보이지 않는다. 컴퓨터와 연구장비, 열대 잎 식물 및 일부 꽃이 배치되어 있다. 얼굴만으로 국적이나 민족을 확정할 수는 없으며, 지소영의 시각적 정체성은 참조와 잘 맞는다. 판독 가능한 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "지소영은 화면 왼쪽의 앞발을 바닥에 디뎌 체중을 받치고, 반대쪽 무릎을 굽힌 채 신발을 들어 다음 보폭으로 옮기는 중이다. 들린 발은 지지 다리가 있는 정상적인 보행 동작으로 설명된다. 두 관찰자의 발은 프레임 밖이지만 상체 배치는 자연스럽다. 연구장비와 식재함은 바닥에, 모니터는 작업대에 지지되어 있으며 유리와 바닥의 반사도 명백한 광학적 모순이 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.762
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.512
   },
   "violations": {
    "B": [
     "[gemini-pro] 지문에 명시되지 않은 다수의 인물(배경에서 작업 중인 연구원들)이 발명되어 추가됨",
     "[gpt-high] 지소영과 두 관찰자 외에 작업 중인 연구원 최소 일곱 명을 추가하여, 샷에 지정되지 않은 인물들이 등장한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 512
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지문에 명시된 인물들만 정확하게 배치하고, 전경의 두 사람을 걸쳐 찍는 앵글과 연구소 배경을 충실히 구현함."
   },
   {
    "label": "B",
    "score": 512,
    "verdict_ko": "지문에 없는 인물들이 배경에 다수 추가되어 '추가 인물 금지' 규칙을 심각하게 위반함.  ★위반: [gemini-pro] 지문에 명시되지 않은 다수의 인물(배경에서 작업 중인 연구원들)이 발명되어 추가됨 / [gpt-high] 지소영과 두 관찰자 외에 작업 중인 연구원 최소 일곱 명을 추가하여, 샷에 지정되지 않은 인물들이 등장한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L129B01.png",
    "asset_id": "4a6aa631-eaf4-4be4-a449-2ca4a933e26a",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:808347>",
    "asset_id": "f15ee474-495b-45e2-8cb1-804018e9010e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf10-8cd3-75fc-8a97-5a39c8e30505",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S82sh5__bgfirst_bg.png",
   "bg_asset_id": "44e2cc69-8214-4697-a78d-543ddf925b63",
   "bg_record_key": "S82sh5::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S82sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:08:13.413863+00:00",
  "fingerprint": "afffdba37d054194f12a012399ec5f74829a1ecb54b190f29a2cf38e48f4b438",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S82sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S82sh5_sel.png",
  "source_sha256": "50ffd554044d0bf991c3fcdd0e91b0448500a410163cf376ed4d3b1b7366fd9b",
  "file": "S82sh5_cine.png",
  "staged_sha256": "d6bd2ee57a962c52142131c79c36e9ec87ddc6beb86a589771a64c00454abe6f",
  "latency_ms": 9590
 },
 "S82sh11::signage": {
  "fp": "001dede2354e498e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S82sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:09:11.732666+00:00",
  "fingerprint": "7c07b939979b29117ff5b2632e2476581b14e2a2e7e26c9777342f1fcabdeaa8",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S82sh11_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S82sh11_sel.png",
  "source_sha256": "685ca09c41759995e8b590d4c8e7862305b623e6a6f7b661e0839008da4fe761",
  "file": "S82sh11_cine.png",
  "staged_sha256": "698764625fb49b7ff5159fe543f8c1309561ddce529920720a3e9479c490b84c",
  "latency_ms": 9734
 },
 "S82sh16::signage": {
  "fp": "4b676251d633f414",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S82sh16::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:10:33.934200+00:00",
  "fingerprint": "cb3691627a2c7b2e96bcfb24ac36ca873d6d23cc3c1d050fa2cb75a633838164",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S82sh16_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S82sh16_sel.png",
  "source_sha256": "ed2c8b9dafd35c141ac54b87cddea0d0e3b630907d8c1119ba8b913722f0a995",
  "file": "S82sh16_cine.png",
  "staged_sha256": "2351bfe373abd856c59d1e7964e8b08ed3a46bddbe9aae1d19d550319223815f",
  "latency_ms": 9678
 },
 "S83sh9::signage": {
  "fp": "392171e9b4b9472a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S83sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:11:32.685478+00:00",
  "fingerprint": "fbb7a5a596be5d1eaca7e55911bf48ea7422480263796985a5ac0bf8bba22307",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S83sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S83sh9_sel.png",
  "source_sha256": "3b9e70a6c480b5c27a0c412e612d8f03a29b56462f8b52dcb815337115a4271a",
  "file": "S83sh9_cine.png",
  "staged_sha256": "746d09ccd8cc4b9ae7618e8062d07cdc82c79c911661240bb6a634a04ed61df6",
  "latency_ms": 9604
 },
 "S83sh11::signage": {
  "fp": "02306bd7afc6786c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S83sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:12:30.444368+00:00",
  "fingerprint": "f0b90089d62c7911a1a44c63e1c4726c221d2279306b2ac1af94d49f3197030a",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S83sh11_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S83sh11_sel.png",
  "source_sha256": "dd4b14a2a21c26228573a4522bad02d71118e793aa88d91f9cbd8a791dd0da55",
  "file": "S83sh11_cine.png",
  "staged_sha256": "65b6ecbc5f3a3cbde509c5d3969e30d4d26b5bf12a0dd247749cf1923218cc0c",
  "latency_ms": 8962
 },
 "S83sh17::signage": {
  "fp": "17b6a1ad53781657",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S83sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:13:24.434440+00:00",
  "fingerprint": "f0a7064a728879f42685820c3ef899961eb34a19df8d938679e692f00aa9c5a5",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S83sh17_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S83sh17_sel.png",
  "source_sha256": "1feb777225ae36aead9eb9d9f3ba323ee33056835907103a5d955f6f72438cc4",
  "file": "S83sh17_cine.png",
  "staged_sha256": "89392fd1d1fe2a0878b7ed33bae327eab639cf275bf5024853c151eb3750334d",
  "latency_ms": 10585
 },
 "S84sh1::signage": {
  "fp": "0230c3624ba33967",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S84sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:14:29.910613+00:00",
  "fingerprint": "0aae58f4a76a5ff5422344e2d955673c5d545d85c124d83c5d7c9968c99d227f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S84sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S84sh1_sel.png",
  "source_sha256": "d59f75e0370539c81a01460ec628fc28e5a0e47c8d885591d18f26b7d9dccc8f",
  "file": "S84sh1_cine.png",
  "staged_sha256": "d72f7f89358ce64d9355441b9cdfed3d7e09e12aa36077b6d4907d6f115aee70",
  "latency_ms": 9566
 },
 "S84sh18::signage": {
  "fp": "d48987330e60e1cd",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "얼굴 데이터"
   }
  ],
  "dropped": []
 },
 "S84sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:15:42.768342+00:00",
  "fingerprint": "73874cdcbd93333d5c95657e16e7849d81e47b06d438aa7c43bcdfeb0a9f4205",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S84sh18_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S84sh18_sel.png",
  "source_sha256": "35f3d2bb5ea4c537e017976c0ab9b6c1bc0298b469ae0acd332c4f6fd172d944",
  "file": "S84sh18_cine.png",
  "staged_sha256": "b47d22d6a8d72226e5aaf0d09608c4c6d4a5788cb9c4134efb82a46e66d8e67b",
  "latency_ms": 10803
 },
 "S84sh21::signage": {
  "fp": "e071daa1d872e26f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S84sh21::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:17:24.553908+00:00",
  "fingerprint": "a6ece183dfb5fff029cb620634e4fbb1a005699757778679a2df1b43d0769e93",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S84sh21_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S84sh21_sel.png",
  "source_sha256": "9e769e22ee0a6f57e063dd088a2b47a3ea252680d5e2933fa6fa7733d769ad97",
  "file": "S84sh21_cine.png",
  "staged_sha256": "61649d9d18892b8d63c1facf3ecf2a8ceb50c81a6d0aacaf5c49bd0b5470694b",
  "latency_ms": 10081
 },
 "S85sh1::signage": {
  "fp": "ff4aac72eb98a3de",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S85sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:19:00.843166+00:00",
  "fingerprint": "ec2d0aa9b20ef72f00e3c2b1542b5e078cb94545e8ad99b595a06137d60895fe",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S85sh1_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S85sh1_sel.png",
  "source_sha256": "483cf819cefc34f01a00ab3de018ef8384e10fa76ef5c5ee0b84812df02251c2",
  "file": "S85sh1_cine.png",
  "staged_sha256": "5238cde79dbad56c724507aa833b5666515e949bc101865daf0951281c573c02",
  "latency_ms": 9041
 },
 "S85sh12::signage": {
  "fp": "1641688cc29abc10",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S85sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:20:15.317006+00:00",
  "fingerprint": "282f6b8c7efaa30992614f0aa6843f1da409a217c5e0bbf2266601118235d725",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S85sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S85sh12_sel.png",
  "source_sha256": "320e5eb4af14f5e7cbe71e7c330be7ab4bd2596691827650f1500f59da67657c",
  "file": "S85sh12_cine.png",
  "staged_sha256": "8ab64a30cc58ce2d848f25bb805c96633bea481fa9e2fa7c8f77f71484ca412b",
  "latency_ms": 8953
 },
 "S85sh22::signage": {
  "fp": "38b2331ae7ce53f1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::7be52edca62f73f8": {
  "subjects": [],
  "subject_text": "제주도 연구소 메인 실험실\n중앙의 원형 유리벽 안에 스테인리스 침대와 기계 팔이 설치된 실험실. 주변에는 컴퓨터 작업대와 여러 전선이 배치되어 있다.",
  "identity": "canonical",
  "scope_id": "L130",
  "scope_role": "location_interior",
  "scope_sha": "3d6449847bb036fa"
 },
 "S85sh22::bgfirst_bg": {
  "input_fingerprint": "40913a4a3f0fbe31",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 테이블을 두 손으로 짚고 상체를 반쯤 일으킨 채 입을 크게 벌려 고함치는 역동적인 찰나.\n\nLOCATION (lock): At a dining table inside the research institute's communal cafeteria, under daytime dining-room lighting.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Tabletop retaining both planted hand contacts in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Dining table (Supporting both of 이현우's hands and his meal) — The top recedes diagonally from the lower frame toward his torso; used as Keeps the two hand contacts visible in the lower quarter without exaggerated foreground scale; Bread and food (Set before 이현우 and not yet eaten); used as Remain as small contextual shapes between his hands, contrasting the interrupted meal with his reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve neutral dining-room ambient light and clear facial contrast, allowing the abrupt posture and expression to carry the shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 테이블을 두 손으로 짚고 상체를 반쯤 일으킨 채 입을 크게 벌려 고함치는 역동적인 찰나.\n\nLOCATION (lock): At a dining table inside the research institute's communal cafeteria, under daytime dining-room lighting.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Tabletop retaining both planted hand contacts in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Dining table (Supporting both of 이현우's hands and his meal) — The top recedes diagonally from the lower frame toward his torso; used as Keeps the two hand contacts visible in the lower quarter without exaggerated foreground scale; Bread and food (Set before 이현우 and not yet eaten); used as Remain as small contextual shapes between his hands, contrasting the interrupted meal with his reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve neutral dining-room ambient light and clear facial contrast, allowing the abrupt posture and expression to carry the shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S85sh22__bgfirst_bg.png",
  "asset_id": "3dcc6b13-3596-4042-8e11-fc4da9b052ce",
  "input_asset_ids": [
   "2bba71ea-c759-45ae-b398-a51f11c2d202",
   "908318fe-739e-4b05-9e68-94bf66189db3"
  ]
 },
 "S85sh22": {
  "input_fingerprint": "6de73af44cf673fd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 테이블을 두 손으로 짚고 상체를 반쯤 일으킨 채 입을 크게 벌려 고함치는 역동적인 찰나.\n\nLOCATION (lock): At a dining table inside the research institute's communal cafeteria, under daytime dining-room lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Tabletop retaining both planted hand contacts in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Dining table (Supporting both of 이현우's hands and his meal) — The top recedes diagonally from the lower frame toward his torso; used as Keeps the two hand contacts visible in the lower quarter without exaggerated foreground scale; Bread and food (Set before 이현우 and not yet eaten); used as Remain as small contextual shapes between his hands, contrasting the interrupted meal with his reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve neutral dining-room ambient light and clear facial contrast, allowing the abrupt posture and expression to carry the shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Appetizing bread and other food remain on the dining table, still uneaten. 이현우: He is at the dining table in his white clothing, with an abruptly shocked and angry expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 테이블을 두 손으로 짚고 상체를 반쯤 일으킨 채 입을 크게 벌려 고함치는 역동적인 찰나.\n\nLOCATION (lock): At a dining table inside the research institute's communal cafeteria, under daytime dining-room lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Tabletop retaining both planted hand contacts in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Dining table (Supporting both of 이현우's hands and his meal) — The top recedes diagonally from the lower frame toward his torso; used as Keeps the two hand contacts visible in the lower quarter without exaggerated foreground scale; Bread and food (Set before 이현우 and not yet eaten); used as Remain as small contextual shapes between his hands, contrasting the interrupted meal with his reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve neutral dining-room ambient light and clear facial contrast, allowing the abrupt posture and expression to carry the shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Appetizing bread and other food remain on the dining table, still uneaten. 이현우: He is at the dining table in his white clothing, with an abruptly shocked and angry expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 테이블을 두 손으로 짚고 상체를 반쯤 일으킨 채 입을 크게 벌려 고함치는 역동적인 찰나.\n\nLOCATION (lock): At a dining table inside the research institute's communal cafeteria, under daytime dining-room lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Tabletop retaining both planted hand contacts in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Dining table (Supporting both of 이현우's hands and his meal) — The top recedes diagonally from the lower frame toward his torso; used as Keeps the two hand contacts visible in the lower quarter without exaggerated foreground scale; Bread and food (Set before 이현우 and not yet eaten); used as Remain as small contextual shapes between his hands, contrasting the interrupted meal with his reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve neutral dining-room ambient light and clear facial contrast, allowing the abrupt posture and expression to carry the shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Appetizing bread and other food remain on the dining table, still uneaten. 이현우: He is at the dining table in his white clothing, with an abruptly shocked and angry expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S85sh22__bgfirst_bg.png",
     "asset_id": "3dcc6b13-3596-4042-8e11-fc4da9b052ce",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S85sh22.png",
     "asset_id": "2bba71ea-c759-45ae-b398-a51f11c2d202",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L130B01.png",
     "asset_id": "908318fe-739e-4b05-9e68-94bf66189db3",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 전경에 있는 인물을 향해 강하게 고함치고 있음.",
    "built_space": "식당 내부, 유리벽 너머의 공간이 보이나 레퍼런스의 특정 위치(수술대, 로봇 팔 등)와는 다소 차이가 있으며 식탁들이 늘어서 있음.",
    "entities": "이현우(흰 옷, 헝클어진 머리), 테이블 위의 음식, 그리고 프롬프트에 없는 뒷모습의 여성 인물이 추가됨.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 없는 추가 인물 등장 (전경 왼쪽)",
     "[gpt-high] 지문이 허용한 이현우 외에 맞은편에 앉은 인물의 머리·상체·손을 추가했다."
    ],
    "physics": "두 손으로 테이블을 단단히 짚고 상체를 세운 자세가 올바르게 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "이현우는 카메라 우측 바깥을 향해 입을 크게 벌려 고함치고 있음.",
    "built_space": "식당 내부 테이블에 위치, 배경에는 레퍼런스 이미지와 일치하는 수술대, 로봇 팔, 모니터가 있는 유리벽 내부 및 외부 사무 공간이 정확히 배치됨.",
    "entities": "이현우(레퍼런스와 일치하는 로고가 박힌 흰 옷, 헝클어진 머리), 두 손 사이에 놓인 빵과 음식들. 명시되지 않은 다른 인물은 없음.",
    "hard_violations": [],
    "physics": "상체를 반쯤 일으키고 두 손으로 테이블 테두리를 강하게 짚고 있는 자세가 테이블에 의해 물리적으로 자연스럽게 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지시된 위치의 배경을 정확히 구현하고, 프롬프트에 명시된 인물(이현우)과 그의 행동(고함치기, 두 손으로 테이블 짚기), 미디엄 샷 프레이밍을 완벽하게 따랐습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트에 언급되지 않은 인물이 화면 전경에 추가되어 치명적인 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 전경에 있는 인물을 향해 강하게 고함치고 있음.",
        "built_space": "식당 내부, 유리벽 너머의 공간이 보이나 레퍼런스의 특정 위치(수술대, 로봇 팔 등)와는 다소 차이가 있으며 식탁들이 늘어서 있음.",
        "entities": "이현우(흰 옷, 헝클어진 머리), 테이블 위의 음식, 그리고 프롬프트에 없는 뒷모습의 여성 인물이 추가됨.",
        "hard_violations": [
         "프롬프트에 없는 추가 인물 등장 (전경 왼쪽)"
        ],
        "physics": "두 손으로 테이블을 단단히 짚고 상체를 세운 자세가 올바르게 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "이현우는 카메라 우측 바깥을 향해 입을 크게 벌려 고함치고 있음.",
        "built_space": "식당 내부 테이블에 위치, 배경에는 레퍼런스 이미지와 일치하는 수술대, 로봇 팔, 모니터가 있는 유리벽 내부 및 외부 사무 공간이 정확히 배치됨.",
        "entities": "이현우(레퍼런스와 일치하는 로고가 박힌 흰 옷, 헝클어진 머리), 두 손 사이에 놓인 빵과 음식들. 명시되지 않은 다른 인물은 없음.",
        "hard_violations": [],
        "physics": "상체를 반쯤 일으키고 두 손으로 테이블 테두리를 강하게 짚고 있는 자세가 테이블에 의해 물리적으로 자연스럽게 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지시된 위치의 배경을 정확히 구현하고, 프롬프트에 명시된 인물(이현우)과 그의 행동(고함치기, 두 손으로 테이블 짚기), 미디엄 샷 프레이밍을 완벽하게 따랐습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트에 언급되지 않은 인물이 화면 전경에 추가되어 치명적인 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 전경에 있는 인물을 향해 강하게 고함치고 있음.",
        "built_space": "식당 내부, 유리벽 너머의 공간이 보이나 레퍼런스의 특정 위치(수술대, 로봇 팔 등)와는 다소 차이가 있으며 식탁들이 늘어서 있음.",
        "entities": "이현우(흰 옷, 헝클어진 머리), 테이블 위의 음식, 그리고 프롬프트에 없는 뒷모습의 여성 인물이 추가됨.",
        "hard_violations": [
         "프롬프트에 없는 추가 인물 등장 (전경 왼쪽)"
        ],
        "physics": "두 손으로 테이블을 단단히 짚고 상체를 세운 자세가 올바르게 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "이현우는 카메라 우측 바깥을 향해 입을 크게 벌려 고함치고 있음.",
        "built_space": "식당 내부 테이블에 위치, 배경에는 레퍼런스 이미지와 일치하는 수술대, 로봇 팔, 모니터가 있는 유리벽 내부 및 외부 사무 공간이 정확히 배치됨.",
        "entities": "이현우(레퍼런스와 일치하는 로고가 박힌 흰 옷, 헝클어진 머리), 두 손 사이에 놓인 빵과 음식들. 명시되지 않은 다른 인물은 없음.",
        "hard_violations": [],
        "physics": "상체를 반쯤 일으키고 두 손으로 테이블 테두리를 강하게 짚고 있는 자세가 테이블에 의해 물리적으로 자연스럽게 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 손으로 식탁을 짚고 반쯤 일어나 고함치는 단독 인물의 순간은 충실하지만, 전경 의자가 식탁을 가리고 식당의 공간 구획도 참조와 다르다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "손의 접촉과 격한 고함은 명확하지만, 허용되지 않은 맞은편 인물을 추가하여 어깨너머 구도로 바꾼 것이 결정적인 위반이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 얼굴과 시선은 화면 오른쪽 앞의 프레임 밖을 향하고, 입을 벌려 그 방향으로 고함친다. 지문은 고함의 특정 상대를 지정하지 않으므로 방향상 모순은 없다. 두 팔은 아래로 뻗어 식탁의 서로 떨어진 지점을 짚는다.",
        "built_space": "전경 식탁 하나와 중앙의 큰 의자 등받이 하나, 오른쪽에 잘린 등받이 하나가 보이고 뒤에는 여러 식탁과 의자가 있다. 왼쪽에는 곡면 유리 연구실과 천장 부착 관절 장치 하나가 보이며, 오른쪽에는 모니터들이 놓인 연구용 작업대가 있다. 유리와 금속, 밝은 천장은 참조의 연구소와 유사하지만, 참조에서 별도 출입구 오른쪽에 있던 식당의 구획은 재현되지 않는다. 두 손은 하단 사분면에 들어오지만 중앙 의자가 식탁 전경을 크게 가리고, 상판의 대각선 후퇴보다 수평 가장자리가 강조된다.",
        "entities": "보이는 사람은 이현우 한 명이다. 젊은 한국계 남성으로 읽히는 얼굴, 헝클어진 검은 머리, 마른 체격과 흰 긴팔 상의·흰 바지가 참조와 대체로 맞는다. 표정에는 분노와 충격이 드러나며 입이 크게 열려 있다. 두 손 사이 식탁에는 빵과 다른 음식이 놓여 있고 먹는 동작은 없다. 상의 가슴에는 작은 인쇄 흔적이 남아 있어 무문자 요구에는 덜 충실하다.",
        "hard_violations": [],
        "physics": "두 손바닥과 손가락이 식탁 위에 닿아 상체를 지지한다. 몸통은 앞으로 기울고 팔을 통해 체중을 받는 형태여서 자리에서 급히 일어나는 동작으로 가능하다. 하체는 화면 아래로 이어지며 발은 프레임 밖이므로 발 접촉 여부는 판단할 수 없다. 접시와 음식은 식탁에 놓여 있고 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "이현우는 화면 왼쪽 전경에 앉은 추가 인물의 얼굴 쪽을 똑바로 보며 고함친다. 전경 인물도 이현우를 향해 있다. 방향 관계 자체는 자연스럽지만, 지문에 없는 상대를 만들어 고함의 대상으로 지정했다. 두 손은 식사 양옆의 식탁을 향해 내려가 있다.",
        "built_space": "전경 식탁 하나에 양쪽 사람의 식사가 놓여 있으며 이현우 뒤에는 의자 등받이 하나가 보인다. 뒤쪽에는 여러 식탁과 의자, 오른쪽에는 급식용 작업대와 금속 음료 용기가 있다. 왼쪽 유리 연구실 안에는 천장 관절 장치 두 개가 보여 참조의 장치 수와 부합하지만, 참조의 곡면 유리실과 분리된 식당 대신 긴 직선 유리벽에 접한 식당으로 바뀌었다. 손 두 개와 식사는 잘 보이나 추가 인물의 머리와 어깨가 왼쪽 전경을 차지해 단독 중간 숏이 어깨너머 구도가 되었다.",
        "entities": "중앙 인물은 젊은 한국계 남성으로 읽히며 검은 헝클어진 머리와 흰옷, 참조와 유사한 체격을 갖췄다. 입을 크게 벌리고 눈썹을 찡그려 격한 분노를 표현한다. 그 앞에는 손대지 않은 빵과 음식 그릇, 컵이 있다. 그러나 왼쪽에 별도의 검은 머리 인물과 그 사람의 식사가 추가되어 인물 제한을 어긴다. 뚜렷하게 읽히는 글자는 보이지 않는다.",
        "hard_violations": [
         "지문이 허용한 이현우 외에 맞은편에 앉은 인물의 머리·상체·손을 추가했다."
        ],
        "physics": "이현우의 두 손바닥과 펼친 손가락이 식탁에 명확히 닿고, 팔과 앞으로 기운 몸통이 일어서는 체중 이동을 자연스럽게 보여 준다. 골반 뒤로 의자가 보이며 발은 프레임 밖이다. 추가 인물은 식탁 앞에 앉은 자세로 보이지만 좌면 접촉은 가려져 있다. 식판, 컵, 음식은 모두 식탁이나 식기에 받쳐져 있고 공중에 뜬 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "두 손으로 식탁을 짚고 반쯤 일어나 고함치는 단독 인물의 순간은 충실하지만, 전경 의자가 식탁을 가리고 식당의 공간 구획도 참조와 다르다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "손의 접촉과 격한 고함은 명확하지만, 허용되지 않은 맞은편 인물을 추가하여 어깨너머 구도로 바꾼 것이 결정적인 위반이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 얼굴과 시선은 화면 오른쪽 앞의 프레임 밖을 향하고, 입을 벌려 그 방향으로 고함친다. 지문은 고함의 특정 상대를 지정하지 않으므로 방향상 모순은 없다. 두 팔은 아래로 뻗어 식탁의 서로 떨어진 지점을 짚는다.",
        "built_space": "전경 식탁 하나와 중앙의 큰 의자 등받이 하나, 오른쪽에 잘린 등받이 하나가 보이고 뒤에는 여러 식탁과 의자가 있다. 왼쪽에는 곡면 유리 연구실과 천장 부착 관절 장치 하나가 보이며, 오른쪽에는 모니터들이 놓인 연구용 작업대가 있다. 유리와 금속, 밝은 천장은 참조의 연구소와 유사하지만, 참조에서 별도 출입구 오른쪽에 있던 식당의 구획은 재현되지 않는다. 두 손은 하단 사분면에 들어오지만 중앙 의자가 식탁 전경을 크게 가리고, 상판의 대각선 후퇴보다 수평 가장자리가 강조된다.",
        "entities": "보이는 사람은 이현우 한 명이다. 젊은 한국계 남성으로 읽히는 얼굴, 헝클어진 검은 머리, 마른 체격과 흰 긴팔 상의·흰 바지가 참조와 대체로 맞는다. 표정에는 분노와 충격이 드러나며 입이 크게 열려 있다. 두 손 사이 식탁에는 빵과 다른 음식이 놓여 있고 먹는 동작은 없다. 상의 가슴에는 작은 인쇄 흔적이 남아 있어 무문자 요구에는 덜 충실하다.",
        "hard_violations": [],
        "physics": "두 손바닥과 손가락이 식탁 위에 닿아 상체를 지지한다. 몸통은 앞으로 기울고 팔을 통해 체중을 받는 형태여서 자리에서 급히 일어나는 동작으로 가능하다. 하체는 화면 아래로 이어지며 발은 프레임 밖이므로 발 접촉 여부는 판단할 수 없다. 접시와 음식은 식탁에 놓여 있고 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "이현우는 화면 왼쪽 전경에 앉은 추가 인물의 얼굴 쪽을 똑바로 보며 고함친다. 전경 인물도 이현우를 향해 있다. 방향 관계 자체는 자연스럽지만, 지문에 없는 상대를 만들어 고함의 대상으로 지정했다. 두 손은 식사 양옆의 식탁을 향해 내려가 있다.",
        "built_space": "전경 식탁 하나에 양쪽 사람의 식사가 놓여 있으며 이현우 뒤에는 의자 등받이 하나가 보인다. 뒤쪽에는 여러 식탁과 의자, 오른쪽에는 급식용 작업대와 금속 음료 용기가 있다. 왼쪽 유리 연구실 안에는 천장 관절 장치 두 개가 보여 참조의 장치 수와 부합하지만, 참조의 곡면 유리실과 분리된 식당 대신 긴 직선 유리벽에 접한 식당으로 바뀌었다. 손 두 개와 식사는 잘 보이나 추가 인물의 머리와 어깨가 왼쪽 전경을 차지해 단독 중간 숏이 어깨너머 구도가 되었다.",
        "entities": "중앙 인물은 젊은 한국계 남성으로 읽히며 검은 헝클어진 머리와 흰옷, 참조와 유사한 체격을 갖췄다. 입을 크게 벌리고 눈썹을 찡그려 격한 분노를 표현한다. 그 앞에는 손대지 않은 빵과 음식 그릇, 컵이 있다. 그러나 왼쪽에 별도의 검은 머리 인물과 그 사람의 식사가 추가되어 인물 제한을 어긴다. 뚜렷하게 읽히는 글자는 보이지 않는다.",
        "hard_violations": [
         "지문이 허용한 이현우 외에 맞은편에 앉은 인물의 머리·상체·손을 추가했다."
        ],
        "physics": "이현우의 두 손바닥과 펼친 손가락이 식탁에 명확히 닿고, 팔과 앞으로 기운 몸통이 일어서는 체중 이동을 자연스럽게 보여 준다. 골반 뒤로 의자가 보이며 발은 프레임 밖이다. 추가 인물은 식탁 앞에 앉은 자세로 보이지만 좌면 접촉은 가려져 있다. 식판, 컵, 음식은 모두 식탁이나 식기에 받쳐져 있고 공중에 뜬 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.661,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.411,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 없는 추가 인물 등장 (전경 왼쪽)",
     "[gpt-high] 지문이 허용한 이현우 외에 맞은편에 앉은 인물의 머리·상체·손을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 411
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 위치의 배경을 정확히 구현하고, 프롬프트에 명시된 인물(이현우)과 그의 행동(고함치기, 두 손으로 테이블 짚기), 미디엄 샷 프레이밍을 완벽하게 따랐습니다."
   },
   {
    "label": "A",
    "score": 411,
    "verdict_ko": "프롬프트에 언급되지 않은 인물이 화면 전경에 추가되어 치명적인 오류가 발생했습니다.  ★위반: [gemini-pro] 프롬프트에 없는 추가 인물 등장 (전경 왼쪽) / [gpt-high] 지문이 허용한 이현우 외에 맞은편에 앉은 인물의 머리·상체·손을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L130B01.png",
    "asset_id": "908318fe-739e-4b05-9e68-94bf66189db3",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf44-8a21-7294-9914-4e03e09ce373",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S85sh22__bgfirst_bg.png",
   "bg_asset_id": "3dcc6b13-3596-4042-8e11-fc4da9b052ce",
   "bg_record_key": "S85sh22::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S85sh22::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:21:57.422609+00:00",
  "fingerprint": "2cb470da80d3bd84d31d07eeb9472fe386f0e826a8bc2c494bdc14d8988fbafd",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S85sh22_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S85sh22_sel.png",
  "source_sha256": "684bae51eeb1da1379ee471b0400b74833eb3493521febcee9b97a387d39bd20",
  "file": "S85sh22_cine.png",
  "staged_sha256": "42869116458f842c3b69038eaf3a2d0186b4dc27d94c135db63528046d5bb2e9",
  "latency_ms": 9021
 },
 "S86sh2::signage": {
  "fp": "8ee7108981e73f6d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::d16bd2a8299abc18": {
  "subjects": [],
  "subject_text": "제주도 연구소 소장실\n깔끔하고 모던한 인테리어의 넓은 집무 공간. 중앙 홀에 낡은 피아노가 있고 실내 기둥과 책상이 배치되어 있다.",
  "identity": "canonical",
  "scope_id": "L131",
  "scope_role": "location_interior",
  "scope_sha": "ca219010cf106ac5"
 },
 "S86sh2::bgfirst_bg": {
  "input_fingerprint": "f2dc955d6ea43e1d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 지소영의 책상 앞에 선 채 두 주먹을 꽉 쥐고 쏘아붙이는 분노의 표정.\n\nLOCATION (lock): Directly before the desk inside the research center director's office, under nighttime office lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지소영's desk (Between the confronting positions) — A corner and adjoining top edges are seen diagonally from the camera's lateral position; used as Establishes the confrontation boundary along the lower right without obscuring either fist.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the office, preserving readable tension in his face and hands without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 지소영의 책상 앞에 선 채 두 주먹을 꽉 쥐고 쏘아붙이는 분노의 표정.\n\nLOCATION (lock): Directly before the desk inside the research center director's office, under nighttime office lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지소영's desk (Between the confronting positions) — A corner and adjoining top edges are seen diagonally from the camera's lateral position; used as Establishes the confrontation boundary along the lower right without obscuring either fist.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the office, preserving readable tension in his face and hands without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S86sh2__bgfirst_bg.png",
  "asset_id": "67fd87e8-aaa9-4568-a435-f7c5c8bf4ea2",
  "input_asset_ids": [
   "e36081f1-674f-4ca6-9d76-773c0fc75ced",
   "f8206cb3-a7f1-4e62-8d2a-fe2973dcefd7"
  ]
 },
 "S86sh2": {
  "input_fingerprint": "ec39b0b432262a34",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 지소영의 책상 앞에 선 채 두 주먹을 꽉 쥐고 쏘아붙이는 분노의 표정.\n\nLOCATION (lock): Directly before the desk inside the research center director's office, under nighttime office lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지소영's desk (Between the confronting positions) — A corner and adjoining top edges are seen diagonally from the camera's lateral position; used as Establishes the confrontation boundary along the lower right without obscuring either fist.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the office, preserving readable tension in his face and hands without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office retains its clean, modern furnishings and the old piano in the central hall. 이현우: He has entered the director's office in his white clothing and stands confronting the room in anger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 지소영의 책상 앞에 선 채 두 주먹을 꽉 쥐고 쏘아붙이는 분노의 표정.\n\nLOCATION (lock): Directly before the desk inside the research center director's office, under nighttime office lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지소영's desk (Between the confronting positions) — A corner and adjoining top edges are seen diagonally from the camera's lateral position; used as Establishes the confrontation boundary along the lower right without obscuring either fist.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the office, preserving readable tension in his face and hands without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office retains its clean, modern furnishings and the old piano in the central hall. 이현우: He has entered the director's office in his white clothing and stands confronting the room in anger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 지소영의 책상 앞에 선 채 두 주먹을 꽉 쥐고 쏘아붙이는 분노의 표정.\n\nLOCATION (lock): Directly before the desk inside the research center director's office, under nighttime office lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지소영's desk (Between the confronting positions) — A corner and adjoining top edges are seen diagonally from the camera's lateral position; used as Establishes the confrontation boundary along the lower right without obscuring either fist.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the office, preserving readable tension in his face and hands without introducing a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office retains its clean, modern furnishings and the old piano in the central hall. 이현우: He has entered the director's office in his white clothing and stands confronting the room in anger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S86sh2__bgfirst_bg.png",
     "asset_id": "67fd87e8-aaa9-4568-a435-f7c5c8bf4ea2",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S86sh2.png",
     "asset_id": "e36081f1-674f-4ca6-9d76-773c0fc75ced",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L131B01.png",
     "asset_id": "f8206cb3-a7f1-4e62-8d2a-fe2973dcefd7",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 날카로운 시선이 우측 전경에 앉아 있는 여성(지소영으로 추정)을 정확히 향하고 있음.",
    "built_space": "기준 사진과 매우 일치하는 구조. 좌측에 콘크리트 기둥, 화분, 피아노가 있고 우측에 책상이 있음. 카메라가 여성 뒤에서 책상 모서리를 우측 하단에 걸치도록 정확히 프레이밍함.",
    "entities": "이현우는 18세 한국계 남성, 헝클어진 검은 머리, 흰옷 등 참조 이미지를 잘 반영함. 전경의 인물은 금발 여성으로 묘사됨.",
    "hard_violations": [
     "[gpt-high] 이현우만 등장할 수 있는데 오른쪽 전경에 별도 인물의 머리와 상체가 크게 등장한다."
    ],
    "physics": "이현우는 바닥에 안정적으로 서 있으며, 두 주먹을 꽉 쥔 자세가 자연스러움. 노트북과 펜꽂이는 책상 위에 정상적으로 놓여 있음."
   },
   {
    "label": "B",
    "direction": "이현우가 분노에 찬 표정으로 우측 전경의 흑발 여성을 쏘아보고 있음.",
    "built_space": "공간 묘사에서 기준 사진의 좌측 콘크리트 기둥이 누락됨. 책상이 우측 하단 모서리에만 배치되지 않고 화면 하단 전체를 가로지르고 있어 지정된 구도를 벗어남.",
    "entities": "이현우의 외모와 의상은 참조 이미지와 잘 일치함. 전경에는 흑발 여성이 배치됨.",
    "hard_violations": [
     "[gpt-high] 이현우만 등장할 수 있는데 오른쪽 책상 뒤에 별도 인물의 머리·상체·손이 등장한다."
    ],
    "physics": "인물이 바닥에 서서 체중을 지탱하고 있으며 주먹 쥔 자세가 현실적임. 탁자 위 물건들도 표면에 제대로 지지되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 프레이밍(우측 하단 책상 모서리)과 배경(좌측 콘크리트 기둥, 피아노)을 정확히 재현하며 요구된 분노의 표정을 자연스럽게 담아냈습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "기준 사진의 주요 배경 구조(좌측 기둥)를 누락했으며, 책상이 하단 전체를 가로지르게 배치되어 지시된 화면 구성을 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 날카로운 시선이 우측 전경에 앉아 있는 여성(지소영으로 추정)을 정확히 향하고 있음.",
        "built_space": "기준 사진과 매우 일치하는 구조. 좌측에 콘크리트 기둥, 화분, 피아노가 있고 우측에 책상이 있음. 카메라가 여성 뒤에서 책상 모서리를 우측 하단에 걸치도록 정확히 프레이밍함.",
        "entities": "이현우는 18세 한국계 남성, 헝클어진 검은 머리, 흰옷 등 참조 이미지를 잘 반영함. 전경의 인물은 금발 여성으로 묘사됨.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 안정적으로 서 있으며, 두 주먹을 꽉 쥔 자세가 자연스러움. 노트북과 펜꽂이는 책상 위에 정상적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "이현우가 분노에 찬 표정으로 우측 전경의 흑발 여성을 쏘아보고 있음.",
        "built_space": "공간 묘사에서 기준 사진의 좌측 콘크리트 기둥이 누락됨. 책상이 우측 하단 모서리에만 배치되지 않고 화면 하단 전체를 가로지르고 있어 지정된 구도를 벗어남.",
        "entities": "이현우의 외모와 의상은 참조 이미지와 잘 일치함. 전경에는 흑발 여성이 배치됨.",
        "hard_violations": [],
        "physics": "인물이 바닥에 서서 체중을 지탱하고 있으며 주먹 쥔 자세가 현실적임. 탁자 위 물건들도 표면에 제대로 지지되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 프레이밍(우측 하단 책상 모서리)과 배경(좌측 콘크리트 기둥, 피아노)을 정확히 재현하며 요구된 분노의 표정을 자연스럽게 담아냈습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "기준 사진의 주요 배경 구조(좌측 기둥)를 누락했으며, 책상이 하단 전체를 가로지르게 배치되어 지시된 화면 구성을 어겼습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 날카로운 시선이 우측 전경에 앉아 있는 여성(지소영으로 추정)을 정확히 향하고 있음.",
        "built_space": "기준 사진과 매우 일치하는 구조. 좌측에 콘크리트 기둥, 화분, 피아노가 있고 우측에 책상이 있음. 카메라가 여성 뒤에서 책상 모서리를 우측 하단에 걸치도록 정확히 프레이밍함.",
        "entities": "이현우는 18세 한국계 남성, 헝클어진 검은 머리, 흰옷 등 참조 이미지를 잘 반영함. 전경의 인물은 금발 여성으로 묘사됨.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 안정적으로 서 있으며, 두 주먹을 꽉 쥔 자세가 자연스러움. 노트북과 펜꽂이는 책상 위에 정상적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "이현우가 분노에 찬 표정으로 우측 전경의 흑발 여성을 쏘아보고 있음.",
        "built_space": "공간 묘사에서 기준 사진의 좌측 콘크리트 기둥이 누락됨. 책상이 우측 하단 모서리에만 배치되지 않고 화면 하단 전체를 가로지르고 있어 지정된 구도를 벗어남.",
        "entities": "이현우의 외모와 의상은 참조 이미지와 잘 일치함. 전경에는 흑발 여성이 배치됨.",
        "hard_violations": [],
        "physics": "인물이 바닥에 서서 체중을 지탱하고 있으며 주먹 쥔 자세가 현실적임. 탁자 위 물건들도 표면에 제대로 지지되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "금지된 추가 인물 때문에 부적합하지만, B보다 미디엄 샷에 가깝고 두 주먹을 쥔 채 쏘아붙이는 분노의 순간을 더 정확히 구현했다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "금지된 추가 인물이 크게 등장하며, 다리까지 넓힌 구도와 입을 다문 표정이 지정된 미디엄 샷 및 쏘아붙이는 순간에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 얼굴과 시선은 오른쪽 전경의 책상 뒤 인물을 향하고 있어 대치 방향은 분명하다. 상대 인물도 이현우 쪽으로 고개를 돌렸다. 두 주먹은 몸 양옆에서 아래로 향하며 상대를 때리려고 뻗은 자세는 아니다. 다만 시선의 대상인 추가 인물 자체가 등장 허용 조건을 위반한다.",
        "built_space": "오른쪽 아래에 목재 책상 한 개의 상판과 비스듬한 가장자리가 보이고, 이현우는 책상 앞에 서 있다. 두 주먹은 책상에 가리지 않는다. 왼쪽에는 검은 업라이트 피아노 한 대, 뒤에는 수납장 한 개와 회색 패널 벽이 보인다. 책상에는 모니터 한 대, 작은 탁상 액자 또는 기기 한 개, 필기구와 서류가 있다. 참조의 주요 재료와 설비는 이어지지만, 책상 뒤 회색 벽을 배경으로 둔 구도 때문에 참조의 대치 공간과 카메라 위치의 연결은 덜 명확하다. 인물은 허벅지 부근까지 보여 지정된 미디엄 샷보다 약간 넓다.",
        "entities": "검은 머리의 젊은 한국계 남성으로 보이는 인물 한 명이 흰 긴소매 상의와 흰 바지를 입고 있다. 얼굴과 체격은 참조에 대체로 가깝지만 머리는 참조보다 정돈되어 있다. 눈썹을 찌푸리고 이를 드러내며 입을 열어 쏘아붙이는 분노를 표현하고, 양손 모두 주먹이다. 오른쪽에는 검은 단발머리와 어두운 옷을 입은 별도 인물의 머리·어깨·손 일부가 보인다. 허용된 인물은 이현우뿐이므로 이 추가 인물은 조건에 맞지 않는다. 판독 가능한 글자는 보이지 않는다.",
        "hard_violations": [
         "이현우만 등장할 수 있는데 오른쪽 책상 뒤에 별도 인물의 머리·상체·손이 등장한다."
        ],
        "physics": "이현우는 몸통과 다리가 자연스럽게 이어진 직립 자세이고, 두 주먹은 각각 손목과 팔에 연결되어 있다. 발과 바닥의 접촉은 프레임 밖이므로 직접 확인할 수 없지만 공중에 뜬 징후는 없다. 책상 위 물건은 상판이나 자체 받침에 지지되어 있다. 오른쪽 인물의 좌면은 보이지 않아 착석 접촉은 확인할 수 없으나 불가능한 부유 자세로 보이지는 않는다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 책상 뒤에 앉은 여성으로 보이는 인물을 노려본다. 상대도 왼쪽의 이현우를 향한다. 두 주먹은 허벅지 옆에서 아래로 향하고 몸은 상대 쪽으로 조금 기울어져 있다. 대치 방향은 읽히지만, 대상 인물의 가시적 등장은 허용되지 않았다.",
        "built_space": "오른쪽 아래에 목재 책상 한 개의 모서리와 인접 상판 가장자리가 뚜렷하게 보이며 두 주먹을 가리지 않는다. 책상 뒤에는 검은 사무용 의자 한 개와 추가 인물이 있다. 왼쪽에는 콘크리트 기둥 한 개, 업라이트 피아노 한 대와 벤치 한 개, 큰 화분 한 개가 보이고, 뒤에는 수납장 한 개와 회색 패널 벽이 있다. 천장에는 카세트형 냉난방기 한 개와 선형 조명이 보여 참조 장소의 설비와 재료를 많이 유지한다. 다만 인물의 종아리 부근까지 포함한 넓은 구도로, 요구된 미디엄 샷보다 확연히 넓다.",
        "entities": "이현우는 헝클어진 검은 머리, 젊은 한국계 남성으로 보이는 얼굴, 흰 긴소매 상의와 흰 바지로 표현되어 참조의 주요 특징을 따른다. 두 손은 주먹이지만 입을 굳게 다물고 있어 쏘아붙이기보다는 화를 참으며 노려보는 순간에 가깝다. 오른쪽에는 긴 갈색 머리와 밝은 상의를 입은 별도 여성으로 보이는 인물이 크게 등장한다. 책상에는 펼친 노트북 한 대, 전화기 한 대, 펜꽂이 한 개와 서류철이 있다. 참조의 데스크톱 모니터 대신 노트북이 놓여 있다. 창밖은 어둡고 판독 가능한 글자는 보이지 않는다.",
        "hard_violations": [
         "이현우만 등장할 수 있는데 오른쪽 전경에 별도 인물의 머리와 상체가 크게 등장한다."
        ],
        "physics": "이현우의 몸은 서 있는 다리 위에서 상대 쪽으로 약간 기울어져 있고 팔과 주먹의 연결도 자연스럽다. 발은 프레임 밖이라 지면 접촉을 직접 확인할 수 없지만 부유를 시사하는 모습은 없다. 추가 인물의 등 뒤에는 의자 등받이가 있으며 착석 자세로 읽힌다. 노트북, 전화기, 펜꽂이와 서류철은 책상 상판에 놓여 있어 지지가 명확하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "금지된 추가 인물 때문에 부적합하지만, B보다 미디엄 샷에 가깝고 두 주먹을 쥔 채 쏘아붙이는 분노의 순간을 더 정확히 구현했다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "금지된 추가 인물이 크게 등장하며, 다리까지 넓힌 구도와 입을 다문 표정이 지정된 미디엄 샷 및 쏘아붙이는 순간에서 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 얼굴과 시선은 오른쪽 전경의 책상 뒤 인물을 향하고 있어 대치 방향은 분명하다. 상대 인물도 이현우 쪽으로 고개를 돌렸다. 두 주먹은 몸 양옆에서 아래로 향하며 상대를 때리려고 뻗은 자세는 아니다. 다만 시선의 대상인 추가 인물 자체가 등장 허용 조건을 위반한다.",
        "built_space": "오른쪽 아래에 목재 책상 한 개의 상판과 비스듬한 가장자리가 보이고, 이현우는 책상 앞에 서 있다. 두 주먹은 책상에 가리지 않는다. 왼쪽에는 검은 업라이트 피아노 한 대, 뒤에는 수납장 한 개와 회색 패널 벽이 보인다. 책상에는 모니터 한 대, 작은 탁상 액자 또는 기기 한 개, 필기구와 서류가 있다. 참조의 주요 재료와 설비는 이어지지만, 책상 뒤 회색 벽을 배경으로 둔 구도 때문에 참조의 대치 공간과 카메라 위치의 연결은 덜 명확하다. 인물은 허벅지 부근까지 보여 지정된 미디엄 샷보다 약간 넓다.",
        "entities": "검은 머리의 젊은 한국계 남성으로 보이는 인물 한 명이 흰 긴소매 상의와 흰 바지를 입고 있다. 얼굴과 체격은 참조에 대체로 가깝지만 머리는 참조보다 정돈되어 있다. 눈썹을 찌푸리고 이를 드러내며 입을 열어 쏘아붙이는 분노를 표현하고, 양손 모두 주먹이다. 오른쪽에는 검은 단발머리와 어두운 옷을 입은 별도 인물의 머리·어깨·손 일부가 보인다. 허용된 인물은 이현우뿐이므로 이 추가 인물은 조건에 맞지 않는다. 판독 가능한 글자는 보이지 않는다.",
        "hard_violations": [
         "이현우만 등장할 수 있는데 오른쪽 책상 뒤에 별도 인물의 머리·상체·손이 등장한다."
        ],
        "physics": "이현우는 몸통과 다리가 자연스럽게 이어진 직립 자세이고, 두 주먹은 각각 손목과 팔에 연결되어 있다. 발과 바닥의 접촉은 프레임 밖이므로 직접 확인할 수 없지만 공중에 뜬 징후는 없다. 책상 위 물건은 상판이나 자체 받침에 지지되어 있다. 오른쪽 인물의 좌면은 보이지 않아 착석 접촉은 확인할 수 없으나 불가능한 부유 자세로 보이지는 않는다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 책상 뒤에 앉은 여성으로 보이는 인물을 노려본다. 상대도 왼쪽의 이현우를 향한다. 두 주먹은 허벅지 옆에서 아래로 향하고 몸은 상대 쪽으로 조금 기울어져 있다. 대치 방향은 읽히지만, 대상 인물의 가시적 등장은 허용되지 않았다.",
        "built_space": "오른쪽 아래에 목재 책상 한 개의 모서리와 인접 상판 가장자리가 뚜렷하게 보이며 두 주먹을 가리지 않는다. 책상 뒤에는 검은 사무용 의자 한 개와 추가 인물이 있다. 왼쪽에는 콘크리트 기둥 한 개, 업라이트 피아노 한 대와 벤치 한 개, 큰 화분 한 개가 보이고, 뒤에는 수납장 한 개와 회색 패널 벽이 있다. 천장에는 카세트형 냉난방기 한 개와 선형 조명이 보여 참조 장소의 설비와 재료를 많이 유지한다. 다만 인물의 종아리 부근까지 포함한 넓은 구도로, 요구된 미디엄 샷보다 확연히 넓다.",
        "entities": "이현우는 헝클어진 검은 머리, 젊은 한국계 남성으로 보이는 얼굴, 흰 긴소매 상의와 흰 바지로 표현되어 참조의 주요 특징을 따른다. 두 손은 주먹이지만 입을 굳게 다물고 있어 쏘아붙이기보다는 화를 참으며 노려보는 순간에 가깝다. 오른쪽에는 긴 갈색 머리와 밝은 상의를 입은 별도 여성으로 보이는 인물이 크게 등장한다. 책상에는 펼친 노트북 한 대, 전화기 한 대, 펜꽂이 한 개와 서류철이 있다. 참조의 데스크톱 모니터 대신 노트북이 놓여 있다. 창밖은 어둡고 판독 가능한 글자는 보이지 않는다.",
        "hard_violations": [
         "이현우만 등장할 수 있는데 오른쪽 전경에 별도 인물의 머리와 상체가 크게 등장한다."
        ],
        "physics": "이현우의 몸은 서 있는 다리 위에서 상대 쪽으로 약간 기울어져 있고 팔과 주먹의 연결도 자연스럽다. 발은 프레임 밖이라 지면 접촉을 직접 확인할 수 없지만 부유를 시사하는 모습은 없다. 추가 인물의 등 뒤에는 의자 등받이가 있으며 착석 자세로 읽힌다. 노트북, 전화기, 펜꽂이와 서류철은 책상 상판에 놓여 있어 지지가 명확하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.321
   },
   "violations": {
    "B": [
     "[gpt-high] 이현우만 등장할 수 있는데 오른쪽 책상 뒤에 별도 인물의 머리·상체·손이 등장한다."
    ],
    "A": [
     "[gpt-high] 이현우만 등장할 수 있는데 오른쪽 전경에 별도 인물의 머리와 상체가 크게 등장한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1321
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "지정된 프레이밍(우측 하단 책상 모서리)과 배경(좌측 콘크리트 기둥, 피아노)을 정확히 재현하며 요구된 분노의 표정을 자연스럽게 담아냈습니다.  ★위반: [gpt-high] 이현우만 등장할 수 있는데 오른쪽 전경에 별도 인물의 머리와 상체가 크게 등장한다."
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "기준 사진의 주요 배경 구조(좌측 기둥)를 누락했으며, 책상이 하단 전체를 가로지르게 배치되어 지시된 화면 구성을 어겼습니다.  ★위반: [gpt-high] 이현우만 등장할 수 있는데 오른쪽 책상 뒤에 별도 인물의 머리·상체·손이 등장한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L131B01.png",
    "asset_id": "f8206cb3-a7f1-4e62-8d2a-fe2973dcefd7",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf4a-ecf6-7501-941b-f7a2c9fa01bb",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S86sh2__bgfirst_bg.png",
   "bg_asset_id": "67fd87e8-aaa9-4568-a435-f7c5c8bf4ea2",
   "bg_record_key": "S86sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S86sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:24:08.808265+00:00",
  "fingerprint": "847b73bffc6e25d9ccd2593c5f1f8a144ce2d6e40c23c922d7dc7bed681f1944",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S86sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S86sh2_sel.png",
  "source_sha256": "40e40e87201435790f0e578f3b311ced06693f71e6f8735d60119a8ffbcda608",
  "file": "S86sh2_cine.png",
  "staged_sha256": "0eb037a5e508d65e3776a454e276693afcb97cdc69d4e43a4f08e201691172f7",
  "latency_ms": 9632
 },
 "S86sh6::signage": {
  "fp": "fde8acc85eee5028",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S86sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:25:05.517049+00:00",
  "fingerprint": "d695384b18e6dc42ed7298744296f57ac1455de06fb0f6e04d41b3d61ef91a33",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S86sh6_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S86sh6_sel.png",
  "source_sha256": "b0d535793a86609b4f34ce5c36137eaa3adab3df5bd8a70e5aa2258f86b56089",
  "file": "S86sh6_cine.png",
  "staged_sha256": "dabbc6c982c236017c5f4992b75f52c0f6ba9059921bf0f5256842d1228530e4",
  "latency_ms": 10693
 },
 "S87sh5::signage": {
  "fp": "56eb72ad3fe2790e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::faa38306b054fd1a": {
  "subjects": [],
  "subject_text": "특임대 이강준의 사무실\n책상과 문서가 배치된 군 지휘부의 실내 집무 공간. 출입문에서 책상 앞으로 이어지는 동선이 확보되어 있다.",
  "identity": "canonical",
  "scope_id": "L134",
  "scope_role": "location_interior",
  "scope_sha": "5da433bed32be053"
 },
 "groupbg::special_unit_office": {
  "input_fingerprint": "2f92d4315e6b7216",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "special_unit_office",
    "tags": [
     "S87sh5"
    ]
   },
   "context_sig": "b62ae868231a9e9b"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the command desk inside the special operations colonel's office, under nighttime office lighting.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n특임대 이강준의 사무실: 감시 카메라 영상이 분석되는 대형 스크린과 작전 지시가 이루어지는 딱딱한 군 간부 집무실. (특징: 나비드론 영상이 띄워진 분석용 모니터들; 작전 문건이 놓인 각진 책상)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 틸업하면 특임대장 사무실 대령에게 문건을 내민다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: At the command desk inside the special operations colonel's office, under nighttime office lighting.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n특임대 이강준의 사무실: 감시 카메라 영상이 분석되는 대형 스크린과 작전 지시가 이루어지는 딱딱한 군 간부 집무실. (특징: 나비드론 영상이 띄워진 분석용 모니터들; 작전 문건이 놓인 각진 책상)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 틸업하면 특임대장 사무실 대령에게 문건을 내민다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_special_unit_office_2f9882.png",
  "asset_id": "65af9d29-2020-4be5-bc9d-d619c049f765",
  "input_asset_ids": [
   "37dfc33d-7dc4-4664-aee7-7e7e37eb0f46"
  ],
  "origin_tag": "S87sh5",
  "place_text": "At the command desk inside the special operations colonel's office, under nighttime office lighting.",
  "origin_inputs": {
   "place_text": "At the command desk inside the special operations colonel's office, under nighttime office lighting.",
   "time_of_day_en": "night",
   "conti_asset_id": "37dfc33d-7dc4-4664-aee7-7e7e37eb0f46"
  }
 },
 "S87sh5::bgfirst_bg": {
  "input_fingerprint": "e6bcf9aeba2e9a30",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이강준이 두 주먹을 책상 위에 꽉 쥔 채 앞으로 몸을 쑥 내민 결연한 구도.\n\nLOCATION (lock): At the command desk inside the special operations colonel's office, under nighttime office lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Commander’s desk (Supporting 이강준's clenched fists) — Its top is seen at a shallow angle from beside the soldier's position; used as Forms a restrained lower-frame plane connecting both fists and anchoring his forward lean.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled neutral ambient contrast gives his face and clenched hands firm definition without adding an unestablished practical light or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이강준이 두 주먹을 책상 위에 꽉 쥔 채 앞으로 몸을 쑥 내민 결연한 구도.\n\nLOCATION (lock): At the command desk inside the special operations colonel's office, under nighttime office lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Commander’s desk (Supporting 이강준's clenched fists) — Its top is seen at a shallow angle from beside the soldier's position; used as Forms a restrained lower-frame plane connecting both fists and anchoring his forward lean.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled neutral ambient contrast gives his face and clenched hands firm definition without adding an unestablished practical light or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S87sh5__bgfirst_bg.png",
  "asset_id": "76690bf0-a73d-4e13-bec9-5c5bf00d2661",
  "input_asset_ids": [
   "37dfc33d-7dc4-4664-aee7-7e7e37eb0f46",
   "65af9d29-2020-4be5-bc9d-d619c049f765"
  ]
 },
 "S87sh5": {
  "input_fingerprint": "2f9246601d0a6d73",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이강준이 두 주먹을 책상 위에 꽉 쥔 채 앞으로 몸을 쑥 내민 결연한 구도.\n\nLOCATION (lock): At the command desk inside the special operations colonel's office, under nighttime office lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Commander’s desk (Supporting 이강준's clenched fists) — Its top is seen at a shallow angle from beside the soldier's position; used as Forms a restrained lower-frame plane connecting both fists and anchoring his forward lean.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled neutral ambient contrast gives his face and clenched hands firm definition without adding an unestablished practical light or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 이강준: He is in the special-task-force commander's office receiving the report.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이강준 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — wearing: 검은색과 회색 위장 패턴이 섞인 첨단 소재의 특임대 전술 전투복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이강준이 두 주먹을 책상 위에 꽉 쥔 채 앞으로 몸을 쑥 내민 결연한 구도.\n\nLOCATION (lock): At the command desk inside the special operations colonel's office, under nighttime office lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Commander’s desk (Supporting 이강준's clenched fists) — Its top is seen at a shallow angle from beside the soldier's position; used as Forms a restrained lower-frame plane connecting both fists and anchoring his forward lean.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled neutral ambient contrast gives his face and clenched hands firm definition without adding an unestablished practical light or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 이강준: He is in the special-task-force commander's office receiving the report.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이강준 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — wearing: 검은색과 회색 위장 패턴이 섞인 첨단 소재의 특임대 전술 전투복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이강준이 두 주먹을 책상 위에 꽉 쥔 채 앞으로 몸을 쑥 내민 결연한 구도.\n\nLOCATION (lock): At the command desk inside the special operations colonel's office, under nighttime office lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Commander’s desk (Supporting 이강준's clenched fists) — Its top is seen at a shallow angle from beside the soldier's position; used as Forms a restrained lower-frame plane connecting both fists and anchoring his forward lean.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled neutral ambient contrast gives his face and clenched hands firm definition without adding an unestablished practical light or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 이강준: He is in the special-task-force commander's office receiving the report.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이강준 (성인 남성 얼굴, 얼굴형 미상, 머리 모양 미상) — wearing: 검은색과 회색 위장 패턴이 섞인 첨단 소재의 특임대 전술 전투복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S87sh5__bgfirst_bg.png",
     "asset_id": "76690bf0-a73d-4e13-bec9-5c5bf00d2661",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S87sh5.png",
     "asset_id": "37dfc33d-7dc4-4664-aee7-7e7e37eb0f46",
     "role": "conti_light"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_special_unit_office_2f9882.png",
     "asset_id": "65af9d29-2020-4be5-bc9d-d619c049f765",
     "role": "bgfirst_group_bg"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "중앙 인물이 화면 왼쪽 앞을 응시함.",
    "built_space": "레퍼런스와 일치하는 지휘관실 배경(데스크, 모니터, 국기 등).",
    "entities": "이강준 및 프롬프트에 명시되지 않은 인물 1명이 좌측에 등장함.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 없는 인물(왼측 전경의 인물) 추가",
     "[gpt-high] 지문에 없고 등장도 금지된 두 번째 군인을 왼쪽 전경에 추가했다."
    ],
    "physics": "주먹을 책상에 짚고 있으나, 하반신이 책상에 파묻힌 것처럼 보임."
   },
   {
    "label": "B",
    "direction": "인물이 정면 아래쪽을 강렬하게 응시함.",
    "built_space": "레퍼런스와 일치하는 지휘관실. 데스크와 배경 요소가 올바르게 배치됨.",
    "entities": "이강준 1명만 명시된 위장복을 입고 등장함.",
    "hard_violations": [],
    "physics": "두 주먹을 책상에 꽉 쥐고 상체를 앞으로 기울인 자세가 자연스럽게 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 단일 인물, 포즈, 카메라 구도 및 배경을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트에 없는 추가 인물이 등장하여 지침을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙 인물이 화면 왼쪽 앞을 응시함.",
        "built_space": "레퍼런스와 일치하는 지휘관실 배경(데스크, 모니터, 국기 등).",
        "entities": "이강준 및 프롬프트에 명시되지 않은 인물 1명이 좌측에 등장함.",
        "hard_violations": [
         "프롬프트에 없는 인물(왼측 전경의 인물) 추가"
        ],
        "physics": "주먹을 책상에 짚고 있으나, 하반신이 책상에 파묻힌 것처럼 보임."
       },
       {
        "label": "B",
        "direction": "인물이 정면 아래쪽을 강렬하게 응시함.",
        "built_space": "레퍼런스와 일치하는 지휘관실. 데스크와 배경 요소가 올바르게 배치됨.",
        "entities": "이강준 1명만 명시된 위장복을 입고 등장함.",
        "hard_violations": [],
        "physics": "두 주먹을 책상에 꽉 쥐고 상체를 앞으로 기울인 자세가 자연스럽게 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 단일 인물, 포즈, 카메라 구도 및 배경을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트에 없는 추가 인물이 등장하여 지침을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "중앙 인물이 화면 왼쪽 앞을 응시함.",
        "built_space": "레퍼런스와 일치하는 지휘관실 배경(데스크, 모니터, 국기 등).",
        "entities": "이강준 및 프롬프트에 명시되지 않은 인물 1명이 좌측에 등장함.",
        "hard_violations": [
         "프롬프트에 없는 인물(왼측 전경의 인물) 추가"
        ],
        "physics": "주먹을 책상에 짚고 있으나, 하반신이 책상에 파묻힌 것처럼 보임."
       },
       {
        "label": "B",
        "direction": "인물이 정면 아래쪽을 강렬하게 응시함.",
        "built_space": "레퍼런스와 일치하는 지휘관실. 데스크와 배경 요소가 올바르게 배치됨.",
        "entities": "이강준 1명만 명시된 위장복을 입고 등장함.",
        "hard_violations": [],
        "physics": "두 주먹을 책상에 꽉 쥐고 상체를 앞으로 기울인 자세가 자연스럽게 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "이강준 한 명의 미디엄 숏으로, 책상에 꽉 쥔 두 주먹을 지지하고 상체를 강하게 내민 결연한 순간과 야간 집무실을 충실히 구현했다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "책상에 기대 앞으로 내민 동작과 장소는 맞지만, 금지된 두 번째 군인을 전경에 추가해 단독 미디엄 숏을 어깨너머 구도로 바꾼 결정적 위반이 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이강준은 상체와 얼굴을 책상 건너편으로 내밀고 화면 왼쪽의 프레임 밖 상대 위치를 응시한다. 실제 상대는 보이지 않으며 시선 대상의 명시는 없는 지문과 충돌하지 않는다. 두 주먹은 서로 떨어진 채 책상 상판을 향해 내려놓여 있다.",
        "built_space": "전경의 지휘관 책상 하나, 인물 뒤 높은 검은 의자 하나, 오른쪽 석재 벽의 금색 문장 하나가 보인다. 왼쪽 통유리창과 하부 수납장, 중앙 왼쪽의 상하 배치 작전 화면, 인물에 가려진 태극기와 오른쪽 부대기가 참조 공간의 배치를 따른다. 상단 대형 화면 하나와 하단 화면 일부가 보이며 가려진 화면 수는 확인할 수 없다. 왼쪽 스탠드 하나와 항공기 모형 하나도 참조에 대응한다. 인물은 의자 앞, 책상 뒤에서 몸을 기울이며, 상판이 낮은 각도의 하단 평면으로 두 주먹을 연결한다.",
        "entities": "보이는 사람은 짧은 검은 머리의 동아시아계로 보이는 성인 남성 한 명뿐이다. 별도 얼굴 참조는 없으며 이강준의 성인 남성 조건에 부합한다. 검정·회색 위장무늬와 기능성 절개가 있는 전술복을 입었다. 두 손 모두 주먹으로 꽉 쥐어져 있고 얼굴에는 집중한 결의가 보인다. 책상, 가죽 패드, 서류철과 필기구는 실물 재질로 표현되며 읽을 수 있는 글자는 보이지 않는다. 창밖은 밤이다.",
        "hard_violations": [],
        "physics": "두 주먹의 아랫면이 책상 패드에 닿아 있고, 굽힌 팔을 통해 앞으로 기울인 상체를 지지한다. 하체와 발은 책상에 가려져 있으나 상체의 접촉과 하중 전달은 자연스럽다. 서류철과 책상용품은 상판에 놓여 있으며 공중에 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이강준은 책상 너머 왼쪽 전경에 추가된 군인의 얼굴 쪽을 똑바로 바라본다. 전경 군인도 머리를 이강준 쪽으로 돌리고 있으나 눈은 보이지 않는다. 이강준의 상체는 그 군인을 향해 기울어 있다. 시선 관계 자체는 성립하지만 지문에 없는 상대를 화면 안에 만들었다.",
        "built_space": "책상 하나와 뒤쪽 높은 검은 의자 하나, 오른쪽 벽 문장 하나, 왼쪽 야간 통유리창과 수납장, 중앙 왼쪽 상하 작전 화면이 보인다. 태극기는 인물 뒤에 일부 가려지고 부대기는 오른쪽 가장자리에 있다. 참조의 목재·석재 마감과 천장 간접조명도 유지된다. 이강준은 책상 뒤 의자 앞에 있지만, 추가 군인이 왼쪽 전경을 크게 차지하면서 단독 미디엄 숏이 어깨너머 숏으로 바뀐다.",
        "entities": "이강준으로 보이는 짧은 검은 머리의 동아시아계 성인 남성은 검정·회색 위장 전술복을 입고 있다. 그러나 왼쪽 전경에 뒤통수·귀·어깨·상체가 보이는 별도의 성인 남성 군인이 추가되어 유일한 허용 인물 조건을 어긴다. 화면 왼쪽 손은 분명한 주먹이고 오른쪽 손은 책상 가장자리에서 상대적으로 느슨하게 말려 있다. 지휘관 책상과 사무용품은 사실적이며 읽을 수 있는 글자는 보이지 않는다. 창밖 야경은 시간 조건에 맞는다.",
        "hard_violations": [
         "지문에 없고 등장도 금지된 두 번째 군인을 왼쪽 전경에 추가했다."
        ],
        "physics": "이강준의 양손은 책상 패드와 앞 가장자리에 접촉하며 펴진 팔을 통해 기울어진 상체를 받친다. 오른손의 쥔 모양은 덜 분명하지만 지지 자체가 불가능하지는 않다. 두 인물의 발은 프레임 밖이며 떠 있다는 징후는 없다. 서류철과 필기구통은 책상 위에 안정적으로 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "이강준 한 명의 미디엄 숏으로, 책상에 꽉 쥔 두 주먹을 지지하고 상체를 강하게 내민 결연한 순간과 야간 집무실을 충실히 구현했다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "책상에 기대 앞으로 내민 동작과 장소는 맞지만, 금지된 두 번째 군인을 전경에 추가해 단독 미디엄 숏을 어깨너머 구도로 바꾼 결정적 위반이 있다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이강준은 상체와 얼굴을 책상 건너편으로 내밀고 화면 왼쪽의 프레임 밖 상대 위치를 응시한다. 실제 상대는 보이지 않으며 시선 대상의 명시는 없는 지문과 충돌하지 않는다. 두 주먹은 서로 떨어진 채 책상 상판을 향해 내려놓여 있다.",
        "built_space": "전경의 지휘관 책상 하나, 인물 뒤 높은 검은 의자 하나, 오른쪽 석재 벽의 금색 문장 하나가 보인다. 왼쪽 통유리창과 하부 수납장, 중앙 왼쪽의 상하 배치 작전 화면, 인물에 가려진 태극기와 오른쪽 부대기가 참조 공간의 배치를 따른다. 상단 대형 화면 하나와 하단 화면 일부가 보이며 가려진 화면 수는 확인할 수 없다. 왼쪽 스탠드 하나와 항공기 모형 하나도 참조에 대응한다. 인물은 의자 앞, 책상 뒤에서 몸을 기울이며, 상판이 낮은 각도의 하단 평면으로 두 주먹을 연결한다.",
        "entities": "보이는 사람은 짧은 검은 머리의 동아시아계로 보이는 성인 남성 한 명뿐이다. 별도 얼굴 참조는 없으며 이강준의 성인 남성 조건에 부합한다. 검정·회색 위장무늬와 기능성 절개가 있는 전술복을 입었다. 두 손 모두 주먹으로 꽉 쥐어져 있고 얼굴에는 집중한 결의가 보인다. 책상, 가죽 패드, 서류철과 필기구는 실물 재질로 표현되며 읽을 수 있는 글자는 보이지 않는다. 창밖은 밤이다.",
        "hard_violations": [],
        "physics": "두 주먹의 아랫면이 책상 패드에 닿아 있고, 굽힌 팔을 통해 앞으로 기울인 상체를 지지한다. 하체와 발은 책상에 가려져 있으나 상체의 접촉과 하중 전달은 자연스럽다. 서류철과 책상용품은 상판에 놓여 있으며 공중에 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이강준은 책상 너머 왼쪽 전경에 추가된 군인의 얼굴 쪽을 똑바로 바라본다. 전경 군인도 머리를 이강준 쪽으로 돌리고 있으나 눈은 보이지 않는다. 이강준의 상체는 그 군인을 향해 기울어 있다. 시선 관계 자체는 성립하지만 지문에 없는 상대를 화면 안에 만들었다.",
        "built_space": "책상 하나와 뒤쪽 높은 검은 의자 하나, 오른쪽 벽 문장 하나, 왼쪽 야간 통유리창과 수납장, 중앙 왼쪽 상하 작전 화면이 보인다. 태극기는 인물 뒤에 일부 가려지고 부대기는 오른쪽 가장자리에 있다. 참조의 목재·석재 마감과 천장 간접조명도 유지된다. 이강준은 책상 뒤 의자 앞에 있지만, 추가 군인이 왼쪽 전경을 크게 차지하면서 단독 미디엄 숏이 어깨너머 숏으로 바뀐다.",
        "entities": "이강준으로 보이는 짧은 검은 머리의 동아시아계 성인 남성은 검정·회색 위장 전술복을 입고 있다. 그러나 왼쪽 전경에 뒤통수·귀·어깨·상체가 보이는 별도의 성인 남성 군인이 추가되어 유일한 허용 인물 조건을 어긴다. 화면 왼쪽 손은 분명한 주먹이고 오른쪽 손은 책상 가장자리에서 상대적으로 느슨하게 말려 있다. 지휘관 책상과 사무용품은 사실적이며 읽을 수 있는 글자는 보이지 않는다. 창밖 야경은 시간 조건에 맞는다.",
        "hard_violations": [
         "지문에 없고 등장도 금지된 두 번째 군인을 왼쪽 전경에 추가했다."
        ],
        "physics": "이강준의 양손은 책상 패드와 앞 가장자리에 접촉하며 펴진 팔을 통해 기울어진 상체를 받친다. 오른손의 쥔 모양은 덜 분명하지만 지지 자체가 불가능하지는 않다. 두 인물의 발은 프레임 밖이며 떠 있다는 징후는 없다. 서류철과 필기구통은 책상 위에 안정적으로 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.651,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.401,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 없는 인물(왼측 전경의 인물) 추가",
     "[gpt-high] 지문에 없고 등장도 금지된 두 번째 군인을 왼쪽 전경에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 401
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 단일 인물, 포즈, 카메라 구도 및 배경을 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 401,
    "verdict_ko": "프롬프트에 없는 추가 인물이 등장하여 지침을 위반했습니다.  ★위반: [gemini-pro] 프롬프트에 없는 인물(왼측 전경의 인물) 추가 / [gpt-high] 지문에 없고 등장도 금지된 두 번째 군인을 왼쪽 전경에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_special_unit_office_2f9882.png",
    "asset_id": "65af9d29-2020-4be5-bc9d-d619c049f765",
    "role": "bgfirst_group_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf56-c95d-7d7e-b2c4-323ce5ba289b",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S87sh5__bgfirst_bg.png",
   "bg_asset_id": "76690bf0-a73d-4e13-bec9-5c5bf00d2661",
   "bg_record_key": "S87sh5::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "special_unit_office",
   "groupbg_asset_id": "65af9d29-2020-4be5-bc9d-d619c049f765"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S87sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:27:04.767069+00:00",
  "fingerprint": "844c3bb7921374db50cc0c1a020af37c5c4f900a2a5739e6a579a0126515bc57",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S87sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S87sh5_sel.png",
  "source_sha256": "19d63271a2f65ac2ad560f09c7f8bd7343f30cb66311780c5d784a1ded1e0660",
  "file": "S87sh5_cine.png",
  "staged_sha256": "e33087388d4d870ecf854f4e64eecd2061065c8ace4ebba7f177d0d693d089c4",
  "latency_ms": 9684
 },
 "S87sh10::signage": {
  "fp": "e2ee4512189fd2bd",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::corporate_chair_office": {
  "input_fingerprint": "6fcb6cf8e4767ad2",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "corporate_chair_office",
    "tags": [
     "S87sh10"
    ]
   },
   "context_sig": "8293e02608e1f0de"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: Inside the corporate chairman's private office at night, near the equipment used to monitor military communications.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n특임대 이강준의 사무실: 감시 카메라 영상이 분석되는 대형 스크린과 작전 지시가 이루어지는 딱딱한 군 간부 집무실. (특징: 나비드론 영상이 띄워진 분석용 모니터들; 작전 문건이 놓인 각진 책상)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /유빅사 윤성찬 회장실 –N\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: Inside the corporate chairman's private office at night, near the equipment used to monitor military communications.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n특임대 이강준의 사무실: 감시 카메라 영상이 분석되는 대형 스크린과 작전 지시가 이루어지는 딱딱한 군 간부 집무실. (특징: 나비드론 영상이 띄워진 분석용 모니터들; 작전 문건이 놓인 각진 책상)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /유빅사 윤성찬 회장실 –N\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_corporate_chair_office_dd85e5.png",
  "asset_id": "733d9612-d340-46cb-904d-b1a9de0b11ea",
  "input_asset_ids": [
   "29cf54f1-42ca-43fa-be8c-00dac6ede326"
  ],
  "origin_tag": "S87sh10",
  "place_text": "Inside the corporate chairman's private office at night, near the equipment used to monitor military communications.",
  "origin_inputs": {
   "place_text": "Inside the corporate chairman's private office at night, near the equipment used to monitor military communications.",
   "time_of_day_en": "night",
   "conti_asset_id": "29cf54f1-42ca-43fa-be8c-00dac6ede326"
  }
 },
 "S87sh10::bgfirst_bg": {
  "input_fingerprint": "d33c91c09c5774fc",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 윤성찬이 검지손가락을 세워 자신의 입술에 댄 채 거만한 눈빛을 뿜어내는 찰나.\n\nLOCATION (lock): Inside the corporate chairman's private office at night, near the equipment used to monitor military communications.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Chairman's office interior (Surrounding 윤성찬 in the separate office location); used as Remains softly unresolved around the close facial composition without adding unsupported furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination preserves detail in his eyes, lips, and raised finger, sustaining a controlled rather than theatrical menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 윤성찬이 검지손가락을 세워 자신의 입술에 댄 채 거만한 눈빛을 뿜어내는 찰나.\n\nLOCATION (lock): Inside the corporate chairman's private office at night, near the equipment used to monitor military communications.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Chairman's office interior (Surrounding 윤성찬 in the separate office location); used as Remains softly unresolved around the close facial composition without adding unsupported furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination preserves detail in his eyes, lips, and raised finger, sustaining a controlled rather than theatrical menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S87sh10__bgfirst_bg.png",
  "asset_id": "5a1e4322-e553-4695-b5fe-44f8654ef515",
  "input_asset_ids": [
   "29cf54f1-42ca-43fa-be8c-00dac6ede326",
   "733d9612-d340-46cb-904d-b1a9de0b11ea"
  ]
 },
 "S87sh10": {
  "input_fingerprint": "7c7adc857fb00245",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 윤성찬이 검지손가락을 세워 자신의 입술에 댄 채 거만한 눈빛을 뿜어내는 찰나.\n\nLOCATION (lock): Inside the corporate chairman's private office at night, near the equipment used to monitor military communications. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Chairman's office interior (Surrounding 윤성찬 in the separate office location); used as Remains softly unresolved around the close facial composition without adding unsupported furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination preserves detail in his eyes, lips, and raised finger, sustaining a controlled rather than theatrical menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬: He is in his Ubik chairman's office during the intercepted military communication.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 윤성찬이 검지손가락을 세워 자신의 입술에 댄 채 거만한 눈빛을 뿜어내는 찰나.\n\nLOCATION (lock): Inside the corporate chairman's private office at night, near the equipment used to monitor military communications. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Chairman's office interior (Surrounding 윤성찬 in the separate office location); used as Remains softly unresolved around the close facial composition without adding unsupported furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination preserves detail in his eyes, lips, and raised finger, sustaining a controlled rather than theatrical menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬: He is in his Ubik chairman's office during the intercepted military communication.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 윤성찬이 검지손가락을 세워 자신의 입술에 댄 채 거만한 눈빛을 뿜어내는 찰나.\n\nLOCATION (lock): Inside the corporate chairman's private office at night, near the equipment used to monitor military communications. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Chairman's office interior (Surrounding 윤성찬 in the separate office location); used as Remains softly unresolved around the close facial composition without adding unsupported furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination preserves detail in his eyes, lips, and raised finger, sustaining a controlled rather than theatrical menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬: He is in his Ubik chairman's office during the intercepted military communication.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (77세 남성, 고령의 얼굴, 머리 모양 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S87sh10__bgfirst_bg.png",
     "asset_id": "5a1e4322-e553-4695-b5fe-44f8654ef515",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S87sh10.png",
     "asset_id": "29cf54f1-42ca-43fa-be8c-00dac6ede326",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:781789>",
     "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_corporate_chair_office_dd85e5.png",
     "asset_id": "733d9612-d340-46cb-904d-b1a9de0b11ea",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:781789>",
     "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선이 우측 전경에 있는 다른 인물을 향함. 검지손가락은 입술에 닿아 있음.",
    "built_space": "참조된 회장실 공간(야경, 모니터)이 배경에 올바르게 배치됨.",
    "entities": "주 인물은 윤성찬의 외형과 일치하나, 우측 전경에 지문에 없는 인물의 뒷모습이 등장함.",
    "hard_violations": [
     "[gemini-pro] 지문에 없는 인물 임의 추가",
     "[gpt-high] 윤성찬만 등장해야 하는 장면에 별도 남성을 오른쪽 전경에 추가했다."
    ],
    "physics": "손가락이 입술에 닿아있고 팔의 자세가 자연스러움."
   },
   {
    "label": "B",
    "direction": "시선은 정면 렌즈를 향하고 있으며, 검지손가락이 입술에 닿아 있음.",
    "built_space": "회장실 내부의 책상, 모니터, 좌측 장식품 등 참조 공간이 정확히 반영됨.",
    "entities": "윤성찬이 단독으로 등장하며 참조 이미지의 외형 및 복장과 정확히 일치함.",
    "hard_violations": [],
    "physics": "손가락을 입술에 댄 자세가 물리적으로 안정적이며 자연스럽게 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지문에서 요구한 단독 등장 조건, 클로즈업 구도, 입술에 손가락을 댄 거만한 표정 및 배경까지 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지문에 명시되지 않은 제3의 인물을 화면에 추가하여 결정적인 제약 사항을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선이 우측 전경에 있는 다른 인물을 향함. 검지손가락은 입술에 닿아 있음.",
        "built_space": "참조된 회장실 공간(야경, 모니터)이 배경에 올바르게 배치됨.",
        "entities": "주 인물은 윤성찬의 외형과 일치하나, 우측 전경에 지문에 없는 인물의 뒷모습이 등장함.",
        "hard_violations": [
         "지문에 없는 인물 임의 추가"
        ],
        "physics": "손가락이 입술에 닿아있고 팔의 자세가 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선은 정면 렌즈를 향하고 있으며, 검지손가락이 입술에 닿아 있음.",
        "built_space": "회장실 내부의 책상, 모니터, 좌측 장식품 등 참조 공간이 정확히 반영됨.",
        "entities": "윤성찬이 단독으로 등장하며 참조 이미지의 외형 및 복장과 정확히 일치함.",
        "hard_violations": [],
        "physics": "손가락을 입술에 댄 자세가 물리적으로 안정적이며 자연스럽게 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지문에서 요구한 단독 등장 조건, 클로즈업 구도, 입술에 손가락을 댄 거만한 표정 및 배경까지 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지문에 명시되지 않은 제3의 인물을 화면에 추가하여 결정적인 제약 사항을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선이 우측 전경에 있는 다른 인물을 향함. 검지손가락은 입술에 닿아 있음.",
        "built_space": "참조된 회장실 공간(야경, 모니터)이 배경에 올바르게 배치됨.",
        "entities": "주 인물은 윤성찬의 외형과 일치하나, 우측 전경에 지문에 없는 인물의 뒷모습이 등장함.",
        "hard_violations": [
         "지문에 없는 인물 임의 추가"
        ],
        "physics": "손가락이 입술에 닿아있고 팔의 자세가 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선은 정면 렌즈를 향하고 있으며, 검지손가락이 입술에 닿아 있음.",
        "built_space": "회장실 내부의 책상, 모니터, 좌측 장식품 등 참조 공간이 정확히 반영됨.",
        "entities": "윤성찬이 단독으로 등장하며 참조 이미지의 외형 및 복장과 정확히 일치함.",
        "hard_violations": [],
        "physics": "손가락을 입술에 댄 자세가 물리적으로 안정적이며 자연스럽게 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "윤성찬 한 명의 얼굴과 입술에 댄 검지를 중심으로 밤의 회장실을 담아 핵심 지시를 충족하지만, 눈빛은 거만함보다 엄격함에 가깝다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "입술에 댄 검지와 회장실은 맞지만, 허용되지 않은 상대 남성을 전경에 추가하고 어깨 너머 구도로 바꾼 것이 결정적 위반이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬은 거의 정면의 카메라를 응시한다. 세운 검지가 자신의 입술 중앙에 실제로 닿으며, 손끝은 코 아래까지 올라온다. 시선의 상대는 화면에 없고, 눈빛은 차갑고 단호하지만 뚜렷하게 거만한 표정까지는 아니다.",
        "built_space": "왼쪽에 세로 장식벽과 조각상 하나, 그 옆에 짙은 대리석 기둥 하나가 보인다. 뒤로 야경 창문과 오른쪽 통신 감시 설비가 있으며, 큰 지도 화면 하나, 오른쪽에 위아래로 놓인 화면 두 개와 하단의 작은 화면 일부가 보인다. 하부 장비 랙과 후방 책상도 참조 공간의 재질과 배치에 부합한다. 인물은 설비 앞에 있고 배경은 얼굴보다 부드럽게 흐려져 있다. 중복된 고정 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "보이는 사람은 고령 남성 윤성찬 한 명뿐이다. 짧게 뒤로 넘긴 회색 머리, 깊은 얼굴 주름, 얼굴 윤곽과 보이는 어깨 체격이 인물 참조에 가깝다. 민족 정체성을 외모만으로 확정할 수는 없지만 참조 인물과 뚜렷한 외형 충돌은 없다. 짙은 남색 정장, 흰 셔츠, 청색 계열 넥타이와 흰 포켓스퀘어도 맞는다. 통신 감시 화면과 밤의 도시 불빛이 보이며 읽을 수 있는 글자나 로고는 없다.",
        "hard_violations": [],
        "physics": "올린 손은 셔츠 소매에서 이어지는 손목과 팔에 연결되어 있고, 검지는 펴고 나머지 손가락은 접어 입술에 대는 동작이 자연스럽다. 하체와 좌석은 구도 밖이므로 지지 방식을 확정할 수 없지만, 상체가 떠 있다는 증거는 없다. 조각상은 받침 위에, 장비와 책상 물품은 각 설치면 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "윤성찬의 얼굴과 눈은 화면 오른쪽 전경의 남성을 향한다. 검지는 자신의 입술 중앙에 닿고 위로 서 있어 핵심 손동작은 맞는다. 상대를 제지하듯 평가하는 눈빛은 보이지만, 그 시선의 대상인 별도 남성은 지시되지 않은 인물이다.",
        "built_space": "윤성찬 뒤에는 검은 가죽 의자 등받이 하나가 있고, 왼쪽에는 도시 사진 액자 하나, 사각 갓 조명 하나와 지구본 하나가 보인다. 왼쪽 아래 책상 조명 하나, 뒤쪽 야경 창문, 오른쪽 지도 화면 하나와 하부 통신 장비 랙도 보인다. 공간의 주요 재질과 설비는 참조에 가깝지만, 전경 남성의 어깨를 넣은 구도로 얼굴 외의 상체와 사무실을 더 넓게 보여준다. 불가능한 반사나 명백히 중복된 설비는 없다.",
        "entities": "윤성찬은 회색 머리와 주름진 얼굴을 가진 고령 남성으로, 참조의 얼굴과 남색 정장·조끼·흰 셔츠·청색 넥타이·흰 포켓스퀘어를 대체로 따른다. 그러나 오른쪽 전경에 검은 머리의 별도 성인 남성이 귀, 뒤통수, 목과 정장 어깨를 드러내고 있다. 이 사람은 허용된 윤성찬 한 명 외의 추가 인물이다. 밤의 회장실과 통신 감시 장비는 보이며 판독 가능한 글자는 없다.",
        "hard_violations": [
         "윤성찬만 등장해야 하는 장면에 별도 남성을 오른쪽 전경에 추가했다."
        ],
        "physics": "윤성찬의 등 뒤에 의자 등받이가 있어 앉은 자세와 부합하며, 실제 좌면 접촉은 프레임 밖이다. 올린 손과 손목은 팔에 자연스럽게 이어지고 검지가 입술에 닿는다. 전경 남성의 머리와 어깨 연결도 물리적으로 가능하다. 조명과 지구본, 책상 물품은 가구 위에 지지되어 있으며 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "윤성찬 한 명의 얼굴과 입술에 댄 검지를 중심으로 밤의 회장실을 담아 핵심 지시를 충족하지만, 눈빛은 거만함보다 엄격함에 가깝다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "입술에 댄 검지와 회장실은 맞지만, 허용되지 않은 상대 남성을 전경에 추가하고 어깨 너머 구도로 바꾼 것이 결정적 위반이다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "윤성찬은 거의 정면의 카메라를 응시한다. 세운 검지가 자신의 입술 중앙에 실제로 닿으며, 손끝은 코 아래까지 올라온다. 시선의 상대는 화면에 없고, 눈빛은 차갑고 단호하지만 뚜렷하게 거만한 표정까지는 아니다.",
        "built_space": "왼쪽에 세로 장식벽과 조각상 하나, 그 옆에 짙은 대리석 기둥 하나가 보인다. 뒤로 야경 창문과 오른쪽 통신 감시 설비가 있으며, 큰 지도 화면 하나, 오른쪽에 위아래로 놓인 화면 두 개와 하단의 작은 화면 일부가 보인다. 하부 장비 랙과 후방 책상도 참조 공간의 재질과 배치에 부합한다. 인물은 설비 앞에 있고 배경은 얼굴보다 부드럽게 흐려져 있다. 중복된 고정 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "보이는 사람은 고령 남성 윤성찬 한 명뿐이다. 짧게 뒤로 넘긴 회색 머리, 깊은 얼굴 주름, 얼굴 윤곽과 보이는 어깨 체격이 인물 참조에 가깝다. 민족 정체성을 외모만으로 확정할 수는 없지만 참조 인물과 뚜렷한 외형 충돌은 없다. 짙은 남색 정장, 흰 셔츠, 청색 계열 넥타이와 흰 포켓스퀘어도 맞는다. 통신 감시 화면과 밤의 도시 불빛이 보이며 읽을 수 있는 글자나 로고는 없다.",
        "hard_violations": [],
        "physics": "올린 손은 셔츠 소매에서 이어지는 손목과 팔에 연결되어 있고, 검지는 펴고 나머지 손가락은 접어 입술에 대는 동작이 자연스럽다. 하체와 좌석은 구도 밖이므로 지지 방식을 확정할 수 없지만, 상체가 떠 있다는 증거는 없다. 조각상은 받침 위에, 장비와 책상 물품은 각 설치면 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "윤성찬의 얼굴과 눈은 화면 오른쪽 전경의 남성을 향한다. 검지는 자신의 입술 중앙에 닿고 위로 서 있어 핵심 손동작은 맞는다. 상대를 제지하듯 평가하는 눈빛은 보이지만, 그 시선의 대상인 별도 남성은 지시되지 않은 인물이다.",
        "built_space": "윤성찬 뒤에는 검은 가죽 의자 등받이 하나가 있고, 왼쪽에는 도시 사진 액자 하나, 사각 갓 조명 하나와 지구본 하나가 보인다. 왼쪽 아래 책상 조명 하나, 뒤쪽 야경 창문, 오른쪽 지도 화면 하나와 하부 통신 장비 랙도 보인다. 공간의 주요 재질과 설비는 참조에 가깝지만, 전경 남성의 어깨를 넣은 구도로 얼굴 외의 상체와 사무실을 더 넓게 보여준다. 불가능한 반사나 명백히 중복된 설비는 없다.",
        "entities": "윤성찬은 회색 머리와 주름진 얼굴을 가진 고령 남성으로, 참조의 얼굴과 남색 정장·조끼·흰 셔츠·청색 넥타이·흰 포켓스퀘어를 대체로 따른다. 그러나 오른쪽 전경에 검은 머리의 별도 성인 남성이 귀, 뒤통수, 목과 정장 어깨를 드러내고 있다. 이 사람은 허용된 윤성찬 한 명 외의 추가 인물이다. 밤의 회장실과 통신 감시 장비는 보이며 판독 가능한 글자는 없다.",
        "hard_violations": [
         "윤성찬만 등장해야 하는 장면에 별도 남성을 오른쪽 전경에 추가했다."
        ],
        "physics": "윤성찬의 등 뒤에 의자 등받이가 있어 앉은 자세와 부합하며, 실제 좌면 접촉은 프레임 밖이다. 올린 손과 손목은 팔에 자연스럽게 이어지고 검지가 입술에 닿는다. 전경 남성의 머리와 어깨 연결도 물리적으로 가능하다. 조명과 지구본, 책상 물품은 가구 위에 지지되어 있으며 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.651,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.401,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지문에 없는 인물 임의 추가",
     "[gpt-high] 윤성찬만 등장해야 하는 장면에 별도 남성을 오른쪽 전경에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 401
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지문에서 요구한 단독 등장 조건, 클로즈업 구도, 입술에 손가락을 댄 거만한 표정 및 배경까지 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 401,
    "verdict_ko": "지문에 명시되지 않은 제3의 인물을 화면에 추가하여 결정적인 제약 사항을 위반했습니다.  ★위반: [gemini-pro] 지문에 없는 인물 임의 추가 / [gpt-high] 윤성찬만 등장해야 하는 장면에 별도 남성을 오른쪽 전경에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_corporate_chair_office_dd85e5.png",
    "asset_id": "733d9612-d340-46cb-904d-b1a9de0b11ea",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:781789>",
    "asset_id": "58c2b9de-61fe-46f6-8bb9-bb5f28bc277c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf5e-2d4e-7970-8305-26987f6f4d5d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S87sh10__bgfirst_bg.png",
   "bg_asset_id": "5a1e4322-e553-4695-b5fe-44f8654ef515",
   "bg_record_key": "S87sh10::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "corporate_chair_office",
   "groupbg_asset_id": "733d9612-d340-46cb-904d-b1a9de0b11ea"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S87sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:29:18.505616+00:00",
  "fingerprint": "234541e8f2cc46093d9f67b132a2200cd324249fb87da8794da99e5aa61e4135",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S87sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S87sh10_sel.png",
  "source_sha256": "ecb2b00a0ca4d75d9aafcc43d1a416715ebaff1109042b3110d439fb31a61e89",
  "file": "S87sh10_cine.png",
  "staged_sha256": "7c90196cef633e9499b4fdef10eec069b1bbb9ff050ddf4a5f1cc5d30af5b10d",
  "latency_ms": 10765
 },
 "S88sh7::signage": {
  "fp": "10631f40c9086329",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S88sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:30:39.940148+00:00",
  "fingerprint": "5e29f8bd8df307790e7863d0dd76c0e36768b181b3b5f892e1f899e676e719cf",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S88sh7_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S88sh7_sel.png",
  "source_sha256": "ea1921803d86de3ce810766364f9f32249b90ddb7045134f5ab2ceb048d9709b",
  "file": "S88sh7_cine.png",
  "staged_sha256": "3f304ba801443b013cb515b9f4457558056c7f5692c07e8db5f4592eac26bd37",
  "latency_ms": 8953
 },
 "S88sh37::signage": {
  "fp": "64babab1131ac871",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S88sh37::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:32:13.172814+00:00",
  "fingerprint": "2e17f8e2ecfbeacba5ffd104d9b5fcb047ca6fe0627409678e6379bb4850a3b0",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S88sh37_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S88sh37_sel.png",
  "source_sha256": "8bc6b94afdf6f13fca53d6dfce6bf8895b84503fcb468ffd5a20bd7f9859ae7c",
  "file": "S88sh37_cine.png",
  "staged_sha256": "f9c7c0048c4e9ac5a2fb9600ba4d6052118245fdae73b92a7396b1aa9815bba0",
  "latency_ms": 10842
 },
 "S88sh41::signage": {
  "fp": "d60740e93c71513a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S88sh41::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:33:18.327194+00:00",
  "fingerprint": "7ff0448fdbadca63649193590bb0a1b3c988e84bf3aac5d169f5671c9397ff2a",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S88sh41_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S88sh41_sel.png",
  "source_sha256": "068aa174f1bed117f41e3f43371e5e41dfab1043d48b56f94275ad61d5b70610",
  "file": "S88sh41_cine.png",
  "staged_sha256": "6c50c6852d304c2c99ba8dde51142dfa7e3724af526e35ab4ef28ecd33b88cb9",
  "latency_ms": 9605
 },
 "S89sh38::signage": {
  "fp": "d216c687293561b9",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "유빅사 로고"
   }
  ],
  "dropped": []
 },
 "S89sh38::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:34:34.336589+00:00",
  "fingerprint": "2fac3419b19ccb871682f645e647b564c119793b47ad535843c5f561cd91fd98",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S89sh38_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S89sh38_sel.png",
  "source_sha256": "0bef95cad2b5583585f83d06e2abef7e68aaf725fa87e8c3c1470835ab3ed706",
  "file": "S89sh38_cine.png",
  "staged_sha256": "d7db3dc94aa3a17b8844722b6731bf42cf1839d7733450fba93e7bda0fdbb636",
  "latency_ms": 9032
 },
 "S89sh58::signage": {
  "fp": "7df903bcc4ef445c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S89sh58::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:35:55.257586+00:00",
  "fingerprint": "17426aa0c0ddf5b535567ad5932abcf353d12de977a36c1ca38b4ac633ad6c4b",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S89sh58_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S89sh58_sel.png",
  "source_sha256": "c29196f53796b7d79e533a2fe980c662755d62752b9475e1428e8c3fa11ea2da",
  "file": "S89sh58_cine.png",
  "staged_sha256": "adc4a9d4a19b1c4dc32a8b740d47c65e79424a5108499ba197f95ceb51543ef3",
  "latency_ms": 9063
 },
 "S89sh73::signage": {
  "fp": "e92116e02f8ff50d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::9a7100a8da3f5060": {
  "subjects": [],
  "subject_text": "제주도 연구소 앞 해변과 상륙 지점\n연구소 앞쪽 바다와 육지가 맞닿는 개방된 해변. 물가에서 해안 안쪽으로 이어지는 공간과 넓은 해상 시야가 보인다.",
  "identity": "canonical",
  "scope_id": "L138",
  "scope_role": "location_exterior",
  "scope_sha": "5102e7af337461a1"
 },
 "S89sh73::bgfirst_bg": {
  "input_fingerprint": "0044fb7a44701574",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 찰리의 차가운 금속 잔해 위로 상체를 엎드린 채 오열하는 슬픈 뒷모습.\n\nLOCATION (lock): At the devastated outdoor aftermath site beside the island research institute, where the robot's remaining fragments lie at night.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ground outside the institute (Exposed around 이현우 and 찰리's remains) — Seen obliquely downward from behind 이현우; used as Creates negative space around the compact embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination holds the bowed back and metal remains legible without retaining any glow from the vanished beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 찰리의 차가운 금속 잔해 위로 상체를 엎드린 채 오열하는 슬픈 뒷모습.\n\nLOCATION (lock): At the devastated outdoor aftermath site beside the island research institute, where the robot's remaining fragments lie at night.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ground outside the institute (Exposed around 이현우 and 찰리's remains) — Seen obliquely downward from behind 이현우; used as Creates negative space around the compact embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination holds the bowed back and metal remains legible without retaining any glow from the vanished beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S89sh73__bgfirst_bg.png",
  "asset_id": "1ce33cdc-bf13-4ac9-9425-cf756e8b577e",
  "input_asset_ids": [
   "38594bd0-bd30-4888-bdd8-80137721ab5d",
   "c3cd0288-8fbb-4409-8af7-a3a514a3a34d"
  ]
 },
 "S89sh73": {
  "input_fingerprint": "e3db6bb659dd41ab",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 차가운 금속 잔해 위로 상체를 엎드린 채 오열하는 슬픈 뒷모습.\n\nLOCATION (lock): At the devastated outdoor aftermath site beside the island research institute, where the robot's remaining fragments lie at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ground outside the institute (Exposed around 이현우 and 찰리's remains) — Seen obliquely downward from behind 이현우; used as Creates negative space around the compact embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination holds the bowed back and metal remains legible without retaining any glow from the vanished beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's powered-off remains consist of only part of his burned metal torso, held against Hyunwoo as Hyunwoo folds his upper body over them in grief. No intact limb arrangement or precise orientation of the surviving torso is specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Night has fallen outside the devastated research center, and the white beam is gone. Only part of Charlie's burned-away torso remains; its remaining power has now shut off. 이현우: He is outside the research center, still wounded by grenade fragments and wearing the receiver on his clothing, bent over in uncontrollable grief.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 차가운 금속 잔해 위로 상체를 엎드린 채 오열하는 슬픈 뒷모습.\n\nLOCATION (lock): At the devastated outdoor aftermath site beside the island research institute, where the robot's remaining fragments lie at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ground outside the institute (Exposed around 이현우 and 찰리's remains) — Seen obliquely downward from behind 이현우; used as Creates negative space around the compact embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination holds the bowed back and metal remains legible without retaining any glow from the vanished beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's powered-off remains consist of only part of his burned metal torso, held against Hyunwoo as Hyunwoo folds his upper body over them in grief. No intact limb arrangement or precise orientation of the surviving torso is specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Night has fallen outside the devastated research center, and the white beam is gone. Only part of Charlie's burned-away torso remains; its remaining power has now shut off. 이현우: He is outside the research center, still wounded by grenade fragments and wearing the receiver on his clothing, bent over in uncontrollable grief.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우가 찰리의 차가운 금속 잔해 위로 상체를 엎드린 채 오열하는 슬픈 뒷모습.\n\nLOCATION (lock): At the devastated outdoor aftermath site beside the island research institute, where the robot's remaining fragments lie at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ground outside the institute (Exposed around 이현우 and 찰리's remains) — Seen obliquely downward from behind 이현우; used as Creates negative space around the compact embrace.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime ambient illumination holds the bowed back and metal remains legible without retaining any glow from the vanished beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's powered-off remains consist of only part of his burned metal torso, held against Hyunwoo as Hyunwoo folds his upper body over them in grief. No intact limb arrangement or precise orientation of the surviving torso is specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Night has fallen outside the devastated research center, and the white beam is gone. Only part of Charlie's burned-away torso remains; its remaining power has now shut off. 이현우: He is outside the research center, still wounded by grenade fragments and wearing the receiver on his clothing, bent over in uncontrollable grief.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S89sh73__bgfirst_bg.png",
     "asset_id": "1ce33cdc-bf13-4ac9-9425-cf756e8b577e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S89sh73.png",
     "asset_id": "38594bd0-bd30-4888-bdd8-80137721ab5d",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 찰리의 몸체 잔해와 분리 부품: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1622352>",
     "asset_id": "c42224c8-15cc-4469-ba02-214fb735bdd8",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L138B03.png",
     "asset_id": "c3cd0288-8fbb-4409-8af7-a3a514a3a34d",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 찰리의 몸체 잔해와 분리 부품: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1622352>",
     "asset_id": "c42224c8-15cc-4469-ba02-214fb735bdd8",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 엎드려 있는 이현우의 등 뒤에서 비스듬히 아래를 향해 잔해를 내려다봄.",
    "built_space": "위치 레퍼런스의 해안선과 연구소 건물이 조명이 켜진 상태까지 매우 정확하게 일치함.",
    "entities": "이현우의 흰색 의상은 일치하나, 찰리의 잔해가 프롭 레퍼런스와 전혀 상관없는 정체불명의 매끄러운 금속 헬멧 형태로 묘사됨. 수신기 미확인.",
    "hard_violations": [],
    "physics": "땅에 꿇은 무릎과 잔해 위에 얹힌 팔 등 무게 중심과 지지 상태가 자연스러움."
   },
   {
    "label": "B",
    "direction": "카메라는 이현우의 뒤에서 엎드린 인물과 로봇 잔해를 향하고 있음.",
    "built_space": "해안가와 연구소 건물이 위치 레퍼런스와 일치하게 배치되어 있으나, 건물의 조명은 꺼진 상태임.",
    "entities": "이현우의 의상이 일치하고 등에 수신기가 부착되어 있음. 찰리의 잔해는 프롭 레퍼런스의 주황색 장갑과 기계적 디테일을 훌륭하게 반영함.",
    "hard_violations": [
     "[gpt-high] 요청하지 않은 수류탄 세 개가 잔해 주변에 추가되어 있습니다.",
     "[gpt-high] 몸통 일부만 남아야 하는 찰리에게 손이 달린 로봇 팔이 추가로 남아 있습니다."
    ],
    "physics": "무릎을 꿇고 로봇 잔해 위로 엎드린 상체가 지면과 잔해에 의해 안정적으로 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롭 레퍼런스(찰리의 로봇 몸통)의 주황색과 회색 디테일을 정확히 구현했으며, 의상에 부착된 수신기 묘사도 훌륭함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리의 잔해가 프롭 레퍼런스와 전혀 다른 단순한 회색 금속 덩어리로 묘사되어 지시사항을 크게 위반함."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 이현우의 뒤에서 엎드린 인물과 로봇 잔해를 향하고 있음.",
        "built_space": "해안가와 연구소 건물이 위치 레퍼런스와 일치하게 배치되어 있으나, 건물의 조명은 꺼진 상태임.",
        "entities": "이현우의 의상이 일치하고 등에 수신기가 부착되어 있음. 찰리의 잔해는 프롭 레퍼런스의 주황색 장갑과 기계적 디테일을 훌륭하게 반영함.",
        "hard_violations": [],
        "physics": "무릎을 꿇고 로봇 잔해 위로 엎드린 상체가 지면과 잔해에 의해 안정적으로 지지됨."
       },
       {
        "label": "A",
        "direction": "카메라는 엎드려 있는 이현우의 등 뒤에서 비스듬히 아래를 향해 잔해를 내려다봄.",
        "built_space": "위치 레퍼런스의 해안선과 연구소 건물이 조명이 켜진 상태까지 매우 정확하게 일치함.",
        "entities": "이현우의 흰색 의상은 일치하나, 찰리의 잔해가 프롭 레퍼런스와 전혀 상관없는 정체불명의 매끄러운 금속 헬멧 형태로 묘사됨. 수신기 미확인.",
        "hard_violations": [],
        "physics": "땅에 꿇은 무릎과 잔해 위에 얹힌 팔 등 무게 중심과 지지 상태가 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롭 레퍼런스(찰리의 로봇 몸통)의 주황색과 회색 디테일을 정확히 구현했으며, 의상에 부착된 수신기 묘사도 훌륭함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리의 잔해가 프롭 레퍼런스와 전혀 다른 단순한 회색 금속 덩어리로 묘사되어 지시사항을 크게 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 이현우의 뒤에서 엎드린 인물과 로봇 잔해를 향하고 있음.",
        "built_space": "해안가와 연구소 건물이 위치 레퍼런스와 일치하게 배치되어 있으나, 건물의 조명은 꺼진 상태임.",
        "entities": "이현우의 의상이 일치하고 등에 수신기가 부착되어 있음. 찰리의 잔해는 프롭 레퍼런스의 주황색 장갑과 기계적 디테일을 훌륭하게 반영함.",
        "hard_violations": [],
        "physics": "무릎을 꿇고 로봇 잔해 위로 엎드린 상체가 지면과 잔해에 의해 안정적으로 지지됨."
       },
       {
        "label": "A",
        "direction": "카메라는 엎드려 있는 이현우의 등 뒤에서 비스듬히 아래를 향해 잔해를 내려다봄.",
        "built_space": "위치 레퍼런스의 해안선과 연구소 건물이 조명이 켜진 상태까지 매우 정확하게 일치함.",
        "entities": "이현우의 흰색 의상은 일치하나, 찰리의 잔해가 프롭 레퍼런스와 전혀 상관없는 정체불명의 매끄러운 금속 헬멧 형태로 묘사됨. 수신기 미확인.",
        "hard_violations": [],
        "physics": "땅에 꿇은 무릎과 잔해 위에 얹힌 팔 등 무게 중심과 지지 상태가 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "현우의 뒤가 아닌 앞에서 촬영했고, 불필요한 수류탄과 로봇 팔까지 추가되어 ‘몸통 일부를 껴안고 엎드린 슬픈 뒷모습’에서 크게 벗어납니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "잔해에 상체를 기댄 뒷모습과 와이드 구도는 충실하지만, 지면을 비스듬히 내려다보는 시점이 약하고 명시적인 낮 시간 고정도 지키지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우의 머리와 얼굴은 아래의 로봇 잔해를 향하고 두 손도 그 주변에 놓여 있습니다. 그러나 카메라는 현우의 앞쪽에서 숙인 정수리와 어깨를 보고 있어, 요구한 뒤쪽 시점이 아닙니다. 무기 조준이나 이동 동작은 없습니다.",
        "built_space": "왼쪽 바다, 오른쪽 뒤 연구소 한 동, 옥상 원형 구조물 하나, 중앙 출입구 하나, 오른쪽 울타리와 능선의 철탑 하나가 보입니다. 해안 흙길과 바위는 장소 참조와 대체로 맞지만, 참조의 건물 조명은 대부분 꺼져 있습니다. 현우는 길 중앙의 잔해 앞에 있으며, 수평선과 하늘을 넓게 포함한 시점은 지면 중심의 후방 하향 구도와 다릅니다.",
        "entities": "사람은 한 명이며 검은 헝클어진 머리, 가는 체격, 흰 긴소매 옷은 현우 참조와 대체로 맞습니다. 얼굴이 가려져 정확한 나이와 얼굴 정체성은 확인할 수 없습니다. 등 위에는 수신기로 보이는 작은 장치가 있습니다. 로봇은 검정·주황색 금속 외장과 배선을 갖지만 몸통 외에 손이 달린 팔도 남아 있습니다. 주변에는 수류탄 형태의 물체 세 개가 보입니다. 읽을 수 있는 글씨나 빛기둥은 없습니다. 조명은 밤이며 명시적인 낮 고정과 불일치합니다.",
        "hard_violations": [
         "요청하지 않은 수류탄 세 개가 잔해 주변에 추가되어 있습니다.",
         "몸통 일부만 남아야 하는 찰리에게 손이 달린 로봇 팔이 추가로 남아 있습니다."
        ],
        "physics": "로봇 몸통과 분리 부품은 지면에 놓여 있고 현우의 상체는 잔해 쪽으로 접혀 있습니다. 하체의 접지는 잔해에 가려져 있지만, 공중에 떠 있다고 볼 근거는 없습니다. 작은 수신기는 옷 위에 밀착되어 있으며 명백히 지지 없이 떠 있는 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "현우는 카메라에 등을 보이고 머리를 오른쪽 앞의 몸통 잔해로 숙입니다. 두 팔과 상체도 같은 잔해를 감싸고 있어 행동의 대상이 분명합니다. 얼굴을 카메라 쪽으로 돌리지 않아 요구한 슬픈 뒷모습에 맞습니다.",
        "built_space": "왼쪽 해안과 바다, 오른쪽 뒤 연구소 한 동, 원형 옥상 구조물 하나, 중앙 출입구 하나, 오른쪽 울타리와 능선의 철탑 하나가 보입니다. 출입구와 외벽의 조명, 연구소로 이어지는 바위 섞인 흙길도 참조와 잘 맞습니다. 현우와 잔해는 화면 아래 중앙에 모여 있고 주변 지면이 노출되지만, 하늘과 수평선의 비중이 커 요청한 비스듬한 하향 시점은 충분히 강조되지 않습니다.",
        "entities": "사람은 현우 한 명뿐이며 헝클어진 검은 머리, 마른 체격, 흰 상하의와 흰 신발이 참조에 부합합니다. 뒷모습이므로 한국계 얼굴과 정확한 나이는 직접 확인할 수 없습니다. 옷의 구멍과 검댕은 부상 이후 상태를 표현하지만 수신기는 식별되지 않습니다. 찰리는 검게 탄 금속 몸통 일부와 주변 외장 파편으로 보이며 온전한 팔다리나 작동광은 없습니다. 읽을 수 있는 글씨도 없습니다. 달빛이 있는 밤으로, 명시적인 낮 시간 고정은 충족하지 못합니다.",
        "hard_violations": [],
        "physics": "현우의 두 무릎과 접힌 정강이가 지면을 받치고, 앞으로 숙인 상체와 팔은 몸통 잔해에 기대어 있습니다. 금속 잔해의 아래쪽과 주변 파편도 땅에 닿아 있습니다. 꺼진 로봇이 스스로 자세를 유지하거나 지지 없이 떠 있는 부분은 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "현우의 뒤가 아닌 앞에서 촬영했고, 불필요한 수류탄과 로봇 팔까지 추가되어 ‘몸통 일부를 껴안고 엎드린 슬픈 뒷모습’에서 크게 벗어납니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "잔해에 상체를 기댄 뒷모습과 와이드 구도는 충실하지만, 지면을 비스듬히 내려다보는 시점이 약하고 명시적인 낮 시간 고정도 지키지 않았습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우의 머리와 얼굴은 아래의 로봇 잔해를 향하고 두 손도 그 주변에 놓여 있습니다. 그러나 카메라는 현우의 앞쪽에서 숙인 정수리와 어깨를 보고 있어, 요구한 뒤쪽 시점이 아닙니다. 무기 조준이나 이동 동작은 없습니다.",
        "built_space": "왼쪽 바다, 오른쪽 뒤 연구소 한 동, 옥상 원형 구조물 하나, 중앙 출입구 하나, 오른쪽 울타리와 능선의 철탑 하나가 보입니다. 해안 흙길과 바위는 장소 참조와 대체로 맞지만, 참조의 건물 조명은 대부분 꺼져 있습니다. 현우는 길 중앙의 잔해 앞에 있으며, 수평선과 하늘을 넓게 포함한 시점은 지면 중심의 후방 하향 구도와 다릅니다.",
        "entities": "사람은 한 명이며 검은 헝클어진 머리, 가는 체격, 흰 긴소매 옷은 현우 참조와 대체로 맞습니다. 얼굴이 가려져 정확한 나이와 얼굴 정체성은 확인할 수 없습니다. 등 위에는 수신기로 보이는 작은 장치가 있습니다. 로봇은 검정·주황색 금속 외장과 배선을 갖지만 몸통 외에 손이 달린 팔도 남아 있습니다. 주변에는 수류탄 형태의 물체 세 개가 보입니다. 읽을 수 있는 글씨나 빛기둥은 없습니다. 조명은 밤이며 명시적인 낮 고정과 불일치합니다.",
        "hard_violations": [
         "요청하지 않은 수류탄 세 개가 잔해 주변에 추가되어 있습니다.",
         "몸통 일부만 남아야 하는 찰리에게 손이 달린 로봇 팔이 추가로 남아 있습니다."
        ],
        "physics": "로봇 몸통과 분리 부품은 지면에 놓여 있고 현우의 상체는 잔해 쪽으로 접혀 있습니다. 하체의 접지는 잔해에 가려져 있지만, 공중에 떠 있다고 볼 근거는 없습니다. 작은 수신기는 옷 위에 밀착되어 있으며 명백히 지지 없이 떠 있는 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "현우는 카메라에 등을 보이고 머리를 오른쪽 앞의 몸통 잔해로 숙입니다. 두 팔과 상체도 같은 잔해를 감싸고 있어 행동의 대상이 분명합니다. 얼굴을 카메라 쪽으로 돌리지 않아 요구한 슬픈 뒷모습에 맞습니다.",
        "built_space": "왼쪽 해안과 바다, 오른쪽 뒤 연구소 한 동, 원형 옥상 구조물 하나, 중앙 출입구 하나, 오른쪽 울타리와 능선의 철탑 하나가 보입니다. 출입구와 외벽의 조명, 연구소로 이어지는 바위 섞인 흙길도 참조와 잘 맞습니다. 현우와 잔해는 화면 아래 중앙에 모여 있고 주변 지면이 노출되지만, 하늘과 수평선의 비중이 커 요청한 비스듬한 하향 시점은 충분히 강조되지 않습니다.",
        "entities": "사람은 현우 한 명뿐이며 헝클어진 검은 머리, 마른 체격, 흰 상하의와 흰 신발이 참조에 부합합니다. 뒷모습이므로 한국계 얼굴과 정확한 나이는 직접 확인할 수 없습니다. 옷의 구멍과 검댕은 부상 이후 상태를 표현하지만 수신기는 식별되지 않습니다. 찰리는 검게 탄 금속 몸통 일부와 주변 외장 파편으로 보이며 온전한 팔다리나 작동광은 없습니다. 읽을 수 있는 글씨도 없습니다. 달빛이 있는 밤으로, 명시적인 낮 시간 고정은 충족하지 못합니다.",
        "hard_violations": [],
        "physics": "현우의 두 무릎과 접힌 정강이가 지면을 받치고, 앞으로 숙인 상체와 팔은 몸통 잔해에 기대어 있습니다. 금속 잔해의 아래쪽과 주변 파편도 땅에 닿아 있습니다. 꺼진 로봇이 스스로 자세를 유지하거나 지지 없이 떠 있는 부분은 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.286
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.036
   },
   "violations": {
    "B": [
     "[gpt-high] 요청하지 않은 수류탄 세 개가 잔해 주변에 추가되어 있습니다.",
     "[gpt-high] 몸통 일부만 남아야 하는 찰리에게 손이 달린 로봇 팔이 추가로 남아 있습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1036,
   "A": 1571
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1036,
    "verdict_ko": "프롭 레퍼런스(찰리의 로봇 몸통)의 주황색과 회색 디테일을 정확히 구현했으며, 의상에 부착된 수신기 묘사도 훌륭함.  ★위반: [gpt-high] 요청하지 않은 수류탄 세 개가 잔해 주변에 추가되어 있습니다. / [gpt-high] 몸통 일부만 남아야 하는 찰리에게 손이 달린 로봇 팔이 추가로 남아 있습니다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "찰리의 잔해가 프롭 레퍼런스와 전혀 다른 단순한 회색 금속 덩어리로 묘사되어 지시사항을 크게 위반함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L138B03.png",
    "asset_id": "c3cd0288-8fbb-4409-8af7-a3a514a3a34d",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 찰리의 몸체 잔해와 분리 부품: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1622352>",
    "asset_id": "c42224c8-15cc-4469-ba02-214fb735bdd8",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf7f-4b7a-7860-abef-d8d3fe3f53af",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S89sh73__bgfirst_bg.png",
   "bg_asset_id": "1ce33cdc-bf13-4ac9-9425-cf756e8b577e",
   "bg_record_key": "S89sh73::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S89sh73::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:37:40.610251+00:00",
  "fingerprint": "7fa23a076bb0d60a5f1b62c3bd36a7457cdee20e5e20c05a2a6e30daf9c99387",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S89sh73_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S89sh73_sel.png",
  "source_sha256": "f1417ea3244aa52e08c21ad2f4a70a061f92a490e9f50005c1cf436e9174d253",
  "file": "S89sh73_cine.png",
  "staged_sha256": "b9172037d7c906deeaf6c8c69bbc1223a42ceed92bb9ce41880dfd09a3e12544",
  "latency_ms": 11422
 },
 "S90sh3::signage": {
  "fp": "2525620079922d28",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::5561f9253f547356": {
  "subjects": [],
  "subject_text": "찰리의 부품 보관과 재조립이 이루어지는 제주도 연구실\n전체적으로 어두운 넓은 연구실. 대형 모니터 아래 제단처럼 높은 테이블이 있으며 수술대와 기계 부품이 배치되어 있다.",
  "identity": "canonical",
  "scope_id": "L143",
  "scope_role": "location_interior",
  "scope_sha": "d475f5e1ac9c8ca9"
 },
 "S90sh3::bgfirst_bg": {
  "input_fingerprint": "d6d849af3ddf1048",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 상처투성이인 이현우가 테이블 앞 의자에 앉아 부서진 금속 부품들을 하염없이 내려다보는 쓸쓸한 뒷모습.\n\nLOCATION (lock): At a chair facing an altar-like parts table inside a large dark research room, with a large display glowing above the fragments.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Altar-like table and disassembled parts (찰리's separated parts rest on the tabletop) — The near edge and top are visible beyond 이현우's shoulder; used as Receives his downward attention and links the foreground figure to the middle distance; Large monitor (Displaying 찰리 units distributed around the world) — Its image-bearing face is visible beyond the table, showing the worldwide units; used as Supplies a distant counterpoint to the disassembled individual on the table; 이현우's chair (Occupied) — Its rear and left side are partially visible beneath him; used as Anchors the seated posture and the camera's left-side approach.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the spacious room generally dark, with restrained tonal separation across 이현우's wounded back, the parts, and the visible monitor content.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 상처투성이인 이현우가 테이블 앞 의자에 앉아 부서진 금속 부품들을 하염없이 내려다보는 쓸쓸한 뒷모습.\n\nLOCATION (lock): At a chair facing an altar-like parts table inside a large dark research room, with a large display glowing above the fragments.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Altar-like table and disassembled parts (찰리's separated parts rest on the tabletop) — The near edge and top are visible beyond 이현우's shoulder; used as Receives his downward attention and links the foreground figure to the middle distance; Large monitor (Displaying 찰리 units distributed around the world) — Its image-bearing face is visible beyond the table, showing the worldwide units; used as Supplies a distant counterpoint to the disassembled individual on the table; 이현우's chair (Occupied) — Its rear and left side are partially visible beneath him; used as Anchors the seated posture and the camera's left-side approach.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the spacious room generally dark, with restrained tonal separation across 이현우's wounded back, the parts, and the visible monitor content.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S90sh3__bgfirst_bg.png",
  "asset_id": "6ee93a75-a171-4dbe-807f-f7cc6d367594",
  "input_asset_ids": [
   "c747d2f6-6600-48e1-a0d5-384bb63f1959",
   "b5b4d60b-4572-4d4c-83bf-d4c6bae59496"
  ]
 },
 "S90sh3": {
  "input_fingerprint": "d0a01b53f676a649",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 상처투성이인 이현우가 테이블 앞 의자에 앉아 부서진 금속 부품들을 하염없이 내려다보는 쓸쓸한 뒷모습.\n\nLOCATION (lock): At a chair facing an altar-like parts table inside a large dark research room, with a large display glowing above the fragments. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Altar-like table and disassembled parts (찰리's separated parts rest on the tabletop) — The near edge and top are visible beyond 이현우's shoulder; used as Receives his downward attention and links the foreground figure to the middle distance; Large monitor (Displaying 찰리 units distributed around the world) — Its image-bearing face is visible beyond the table, showing the worldwide units; used as Supplies a distant counterpoint to the disassembled individual on the table; 이현우's chair (Occupied) — Its rear and left side are partially visible beneath him; used as Anchors the seated posture and the camera's left-side approach.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the spacious room generally dark, with restrained tonal separation across 이현우's wounded back, the parts, and the visible monitor content.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's dismantled metal parts rest separately on the altar-like table in front of Hyunwoo. There is no assembled torso, head, or limb pose, and the exact placement and orientation of the individual parts are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large research room is dim, with disassembled Charlie components arranged on an altar-like table. A large monitor displays Charlie energy-source units distributed across Dubai, Europe, China and Africa. 이현우: He sits behind the display table, still covered in wounds.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 상처투성이인 이현우가 테이블 앞 의자에 앉아 부서진 금속 부품들을 하염없이 내려다보는 쓸쓸한 뒷모습.\n\nLOCATION (lock): At a chair facing an altar-like parts table inside a large dark research room, with a large display glowing above the fragments. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Altar-like table and disassembled parts (찰리's separated parts rest on the tabletop) — The near edge and top are visible beyond 이현우's shoulder; used as Receives his downward attention and links the foreground figure to the middle distance; Large monitor (Displaying 찰리 units distributed around the world) — Its image-bearing face is visible beyond the table, showing the worldwide units; used as Supplies a distant counterpoint to the disassembled individual on the table; 이현우's chair (Occupied) — Its rear and left side are partially visible beneath him; used as Anchors the seated posture and the camera's left-side approach.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the spacious room generally dark, with restrained tonal separation across 이현우's wounded back, the parts, and the visible monitor content.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's dismantled metal parts rest separately on the altar-like table in front of Hyunwoo. There is no assembled torso, head, or limb pose, and the exact placement and orientation of the individual parts are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large research room is dim, with disassembled Charlie components arranged on an altar-like table. A large monitor displays Charlie energy-source units distributed across Dubai, Europe, China and Africa. 이현우: He sits behind the display table, still covered in wounds.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 상처투성이인 이현우가 테이블 앞 의자에 앉아 부서진 금속 부품들을 하염없이 내려다보는 쓸쓸한 뒷모습.\n\nLOCATION (lock): At a chair facing an altar-like parts table inside a large dark research room, with a large display glowing above the fragments. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Altar-like table and disassembled parts (찰리's separated parts rest on the tabletop) — The near edge and top are visible beyond 이현우's shoulder; used as Receives his downward attention and links the foreground figure to the middle distance; Large monitor (Displaying 찰리 units distributed around the world) — Its image-bearing face is visible beyond the table, showing the worldwide units; used as Supplies a distant counterpoint to the disassembled individual on the table; 이현우's chair (Occupied) — Its rear and left side are partially visible beneath him; used as Anchors the seated posture and the camera's left-side approach.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the spacious room generally dark, with restrained tonal separation across 이현우's wounded back, the parts, and the visible monitor content.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's dismantled metal parts rest separately on the altar-like table in front of Hyunwoo. There is no assembled torso, head, or limb pose, and the exact placement and orientation of the individual parts are unspecified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large research room is dim, with disassembled Charlie components arranged on an altar-like table. A large monitor displays Charlie energy-source units distributed across Dubai, Europe, China and Africa. 이현우: He sits behind the display table, still covered in wounds.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 남성 얼굴, 18세의 젊은 얼굴, 검은 머리칼, 헝클어진 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S90sh3__bgfirst_bg.png",
     "asset_id": "6ee93a75-a171-4dbe-807f-f7cc6d367594",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S90sh3.png",
     "asset_id": "c747d2f6-6600-48e1-a0d5-384bb63f1959",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 찰리의 몸체 잔해와 분리 부품: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1622352>",
     "asset_id": "c42224c8-15cc-4469-ba02-214fb735bdd8",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L143B01.png",
     "asset_id": "b5b4d60b-4572-4d4c-83bf-d4c6bae59496",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:811095>",
     "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 찰리의 몸체 잔해와 분리 부품: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1622352>",
     "asset_id": "c42224c8-15cc-4469-ba02-214fb735bdd8",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선이 카메라를 등진 채 테이블 위 금속 부품들을 향해 아래로 고정됨.",
    "built_space": "어두운 연구실 내부에 긴 제단형 금속 테이블이 중앙에 있고, 그 뒤 벽면에 대형 모니터가 위치함. 전경 좌측에 의자가 배치되어 카메라 접근 방향과 일치함.",
    "entities": "상처가 있는 흰색 셔츠 차림의 이현우 뒷모습. 테이블 위에 흩어진 분해된 로봇 부품들, 세계 지도를 띄운 대형 모니터가 모두 지시대로 묘사됨.",
    "hard_violations": [],
    "physics": "이현우는 의자에 체중을 싣고 안정적으로 앉아 있으며, 모든 금속 부품은 테이블 표면에 중력에 맞게 놓여 있음."
   },
   {
    "label": "B",
    "direction": "이현우의 시선이 테이블 위에 놓인 로봇 상체를 향해 아래로 향함.",
    "built_space": "공간이 어두우나, 테이블의 크기가 기준 이미지보다 작고 대형 연구실의 느낌이 축소됨. 모니터도 벽면에 작게 배치됨.",
    "entities": "기준과 다른 어두운 재킷을 입은 이현우의 측면. 테이블 위에는 팔이 붙어있는 등 조립된 상태의 로봇 상체가 있음. 모니터에 영문 지역명이 명확히 표시됨.",
    "hard_violations": [
     "[gemini-pro] leaked markers/diagrams/text (모니터에 Europe, Africa 등 읽을 수 있는 텍스트가 노출됨)",
     "[gpt-high] 원래 모니터 아래의 제단 테이블과 별개로 전경에 부품 테이블을 추가해 지정된 핵심 공간 설비를 중복했습니다.",
     "[gpt-high] 모니터의 지역명이 읽혀 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반합니다.",
     "[gpt-high] 분리된 부품만 놓여 있어야 하는데 중앙에 흉부와 어깨가 연결된 몸통 조립체가 남아 있습니다."
    ],
    "physics": "이현우는 의자에 앉아 한쪽 팔을 다리에 올린 채 지탱하고 있으며, 부품들은 테이블 위에 놓여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 '뒷모습' 앵글과 캐릭터의 흰색 의상을 정확히 반영했으며, 지정된 공간 구성과 텍스트 금지 조건을 모두 잘 준수함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "명시적으로 금지된 읽을 수 있는 텍스트(지도 상의 지역명)가 노출되었으며, 요구된 뒷모습이 아닌 측면 앵글로 묘사되고 의상도 불일치함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 카메라를 등진 채 테이블 위 금속 부품들을 향해 아래로 고정됨.",
        "built_space": "어두운 연구실 내부에 긴 제단형 금속 테이블이 중앙에 있고, 그 뒤 벽면에 대형 모니터가 위치함. 전경 좌측에 의자가 배치되어 카메라 접근 방향과 일치함.",
        "entities": "상처가 있는 흰색 셔츠 차림의 이현우 뒷모습. 테이블 위에 흩어진 분해된 로봇 부품들, 세계 지도를 띄운 대형 모니터가 모두 지시대로 묘사됨.",
        "hard_violations": [],
        "physics": "이현우는 의자에 체중을 싣고 안정적으로 앉아 있으며, 모든 금속 부품은 테이블 표면에 중력에 맞게 놓여 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 테이블 위에 놓인 로봇 상체를 향해 아래로 향함.",
        "built_space": "공간이 어두우나, 테이블의 크기가 기준 이미지보다 작고 대형 연구실의 느낌이 축소됨. 모니터도 벽면에 작게 배치됨.",
        "entities": "기준과 다른 어두운 재킷을 입은 이현우의 측면. 테이블 위에는 팔이 붙어있는 등 조립된 상태의 로봇 상체가 있음. 모니터에 영문 지역명이 명확히 표시됨.",
        "hard_violations": [
         "leaked markers/diagrams/text (모니터에 Europe, Africa 등 읽을 수 있는 텍스트가 노출됨)"
        ],
        "physics": "이현우는 의자에 앉아 한쪽 팔을 다리에 올린 채 지탱하고 있으며, 부품들은 테이블 위에 놓여 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 '뒷모습' 앵글과 캐릭터의 흰색 의상을 정확히 반영했으며, 지정된 공간 구성과 텍스트 금지 조건을 모두 잘 준수함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "명시적으로 금지된 읽을 수 있는 텍스트(지도 상의 지역명)가 노출되었으며, 요구된 뒷모습이 아닌 측면 앵글로 묘사되고 의상도 불일치함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 카메라를 등진 채 테이블 위 금속 부품들을 향해 아래로 고정됨.",
        "built_space": "어두운 연구실 내부에 긴 제단형 금속 테이블이 중앙에 있고, 그 뒤 벽면에 대형 모니터가 위치함. 전경 좌측에 의자가 배치되어 카메라 접근 방향과 일치함.",
        "entities": "상처가 있는 흰색 셔츠 차림의 이현우 뒷모습. 테이블 위에 흩어진 분해된 로봇 부품들, 세계 지도를 띄운 대형 모니터가 모두 지시대로 묘사됨.",
        "hard_violations": [],
        "physics": "이현우는 의자에 체중을 싣고 안정적으로 앉아 있으며, 모든 금속 부품은 테이블 표면에 중력에 맞게 놓여 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 테이블 위에 놓인 로봇 상체를 향해 아래로 향함.",
        "built_space": "공간이 어두우나, 테이블의 크기가 기준 이미지보다 작고 대형 연구실의 느낌이 축소됨. 모니터도 벽면에 작게 배치됨.",
        "entities": "기준과 다른 어두운 재킷을 입은 이현우의 측면. 테이블 위에는 팔이 붙어있는 등 조립된 상태의 로봇 상체가 있음. 모니터에 영문 지역명이 명확히 표시됨.",
        "hard_violations": [
         "leaked markers/diagrams/text (모니터에 Europe, Africa 등 읽을 수 있는 텍스트가 노출됨)"
        ],
        "physics": "이현우는 의자에 앉아 한쪽 팔을 다리에 올린 채 지탱하고 있으며, 부품들은 테이블 위에 놓여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "상처와 내려다보는 동작은 보이지만, 제단 테이블을 중복하고 읽을 수 있는 지명을 노출했으며, 흰옷 대신 회색 재킷과 조립된 몸통을 보여 핵심 조건을 위반합니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "흰옷을 입고 고개를 숙인 이현우의 쓸쓸한 뒷모습과 단일 제단 테이블의 공간 관계가 충실하지만, 상처와 세계 각지의 찰리 유닛 묘사는 다소 불명확합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남성은 화면 오른쪽 아래로 고개를 숙이고 있으며, 보이는 눈의 방향은 가까운 테이블 가장자리와 부품 쪽입니다. 뒤쪽 모니터를 바라보지는 않습니다. 왼쪽 뒤보다는 측면에 가까운 접근으로 얼굴 옆면이 상당히 드러납니다.",
        "built_space": "큰 벽면 모니터 하나, 왼쪽 유리 수납장, 오른쪽 선반과 작업대가 보입니다. 그러나 모니터 아래에 원래의 긴 금속 테이블을 남겨 둔 채 인물 앞에 별도의 부품 테이블을 추가해, 제단 역할의 테이블이 앞뒤 두 개가 되었습니다. 남성이 앉은 의자의 등받이 일부는 왼쪽 아래에 보입니다. 모니터 화면이 인물과 카메라를 향하는 방향 자체는 가능합니다.",
        "entities": "보이는 인물은 한 명이며 검은 헝클어진 머리의 젊은 동아시아계 남성입니다. 얼굴과 목, 손에는 상처가 있지만 참조의 흰색 상의 대신 짙은 회색 재킷을 입었습니다. 부품의 주황색 도장, 어두운 금속, 마모와 회로기판은 참조와 유사합니다. 다만 중앙에는 흉부와 어깨 구조가 연결된 큰 몸통 조립체가 남아 있습니다. 모니터는 세계지도와 발광 지점을 보여 주지만 찰리 유닛 자체는 구체적으로 보이지 않으며, 여러 지역명이 읽힙니다.",
        "hard_violations": [
         "원래 모니터 아래의 제단 테이블과 별개로 전경에 부품 테이블을 추가해 지정된 핵심 공간 설비를 중복했습니다.",
         "모니터의 지역명이 읽혀 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반합니다.",
         "분리된 부품만 놓여 있어야 하는데 중앙에 흉부와 어깨가 연결된 몸통 조립체가 남아 있습니다."
        ],
        "physics": "남성의 몸은 의자 좌판에 지지되고 등은 등받이 앞에 놓여 있습니다. 한 손은 무릎에 얹혀 있고 다른 손은 테이블 가장자리 부근에 놓인 팔에서 아래로 늘어집니다. 금속 부품과 몸통 조립체는 모두 상판에 닿아 있으며 공중에 떠 있는 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "남성은 카메라에 등을 보이고 머리를 앞쪽 아래로 숙입니다. 눈은 가려져 정확한 시선은 확인할 수 없지만, 머리 방향은 앞의 부품 테이블 쪽이며 높은 벽면 모니터 쪽은 아닙니다. 뒤와 왼쪽에서 접근한 구도로 얼굴을 억지로 드러내지 않습니다.",
        "built_space": "긴 금속 제단 테이블 하나와 그 위쪽의 대형 벽면 모니터 하나가 보입니다. 왼쪽 유리 수납장과 이동식 장비, 오른쪽 금속 선반과 작업대, 노출된 천장 배관이 참조 장소와 대응합니다. 인물은 테이블 앞 의자에 앉아 있고 의자 뒤와 측면 일부가 전경에 보입니다. 어깨 너머로 테이블의 가까운 가장자리와 상판, 그 뒤 모니터의 표시 면이 함께 보이며 별도의 제단 테이블은 없습니다.",
        "entities": "인물은 한 명이고, 헝클어진 검은 머리와 마른 젊은 남성의 체격, 흰색 긴소매 옷이 참조와 부합합니다. 얼굴이 보이지 않아 정확한 얼굴 동일성과 나이는 확인할 수 없습니다. 옷의 찢김과 혈흔 같은 얼룩은 보이지만 노출된 피부의 상처는 두드러지지 않습니다. 상판에는 주황색과 회색의 파손 외장, 금속 프레임, 원통형 부품, 회로기판이 분리되어 있습니다. 오른쪽에는 비교적 큰 파손 외장 덩어리가 남아 있으나 완성된 인체형 몸통이나 팔다리 자세는 보이지 않습니다. 모니터에는 세계 연결망과 시설 영상이 있으나 지정된 네 지역의 찰리 에너지원 유닛을 각각 식별하기는 어렵습니다. 명확하게 읽히는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "엉덩이와 허벅지는 의자 좌판에 지지되고 상체는 자연스럽게 앞으로 숙여져 있습니다. 팔은 몸 앞쪽 아래로 내려가며 손은 가려져 있어 접촉 상태를 확인할 수 없습니다. 모든 부품은 상판 위에 놓여 있고, 벽면 모니터와 천장 조명도 구조물에 고정되어 있습니다. 근거 없이 떠 있는 몸이나 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "상처와 내려다보는 동작은 보이지만, 제단 테이블을 중복하고 읽을 수 있는 지명을 노출했으며, 흰옷 대신 회색 재킷과 조립된 몸통을 보여 핵심 조건을 위반합니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "흰옷을 입고 고개를 숙인 이현우의 쓸쓸한 뒷모습과 단일 제단 테이블의 공간 관계가 충실하지만, 상처와 세계 각지의 찰리 유닛 묘사는 다소 불명확합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "남성은 화면 오른쪽 아래로 고개를 숙이고 있으며, 보이는 눈의 방향은 가까운 테이블 가장자리와 부품 쪽입니다. 뒤쪽 모니터를 바라보지는 않습니다. 왼쪽 뒤보다는 측면에 가까운 접근으로 얼굴 옆면이 상당히 드러납니다.",
        "built_space": "큰 벽면 모니터 하나, 왼쪽 유리 수납장, 오른쪽 선반과 작업대가 보입니다. 그러나 모니터 아래에 원래의 긴 금속 테이블을 남겨 둔 채 인물 앞에 별도의 부품 테이블을 추가해, 제단 역할의 테이블이 앞뒤 두 개가 되었습니다. 남성이 앉은 의자의 등받이 일부는 왼쪽 아래에 보입니다. 모니터 화면이 인물과 카메라를 향하는 방향 자체는 가능합니다.",
        "entities": "보이는 인물은 한 명이며 검은 헝클어진 머리의 젊은 동아시아계 남성입니다. 얼굴과 목, 손에는 상처가 있지만 참조의 흰색 상의 대신 짙은 회색 재킷을 입었습니다. 부품의 주황색 도장, 어두운 금속, 마모와 회로기판은 참조와 유사합니다. 다만 중앙에는 흉부와 어깨 구조가 연결된 큰 몸통 조립체가 남아 있습니다. 모니터는 세계지도와 발광 지점을 보여 주지만 찰리 유닛 자체는 구체적으로 보이지 않으며, 여러 지역명이 읽힙니다.",
        "hard_violations": [
         "원래 모니터 아래의 제단 테이블과 별개로 전경에 부품 테이블을 추가해 지정된 핵심 공간 설비를 중복했습니다.",
         "모니터의 지역명이 읽혀 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반합니다.",
         "분리된 부품만 놓여 있어야 하는데 중앙에 흉부와 어깨가 연결된 몸통 조립체가 남아 있습니다."
        ],
        "physics": "남성의 몸은 의자 좌판에 지지되고 등은 등받이 앞에 놓여 있습니다. 한 손은 무릎에 얹혀 있고 다른 손은 테이블 가장자리 부근에 놓인 팔에서 아래로 늘어집니다. 금속 부품과 몸통 조립체는 모두 상판에 닿아 있으며 공중에 떠 있는 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "남성은 카메라에 등을 보이고 머리를 앞쪽 아래로 숙입니다. 눈은 가려져 정확한 시선은 확인할 수 없지만, 머리 방향은 앞의 부품 테이블 쪽이며 높은 벽면 모니터 쪽은 아닙니다. 뒤와 왼쪽에서 접근한 구도로 얼굴을 억지로 드러내지 않습니다.",
        "built_space": "긴 금속 제단 테이블 하나와 그 위쪽의 대형 벽면 모니터 하나가 보입니다. 왼쪽 유리 수납장과 이동식 장비, 오른쪽 금속 선반과 작업대, 노출된 천장 배관이 참조 장소와 대응합니다. 인물은 테이블 앞 의자에 앉아 있고 의자 뒤와 측면 일부가 전경에 보입니다. 어깨 너머로 테이블의 가까운 가장자리와 상판, 그 뒤 모니터의 표시 면이 함께 보이며 별도의 제단 테이블은 없습니다.",
        "entities": "인물은 한 명이고, 헝클어진 검은 머리와 마른 젊은 남성의 체격, 흰색 긴소매 옷이 참조와 부합합니다. 얼굴이 보이지 않아 정확한 얼굴 동일성과 나이는 확인할 수 없습니다. 옷의 찢김과 혈흔 같은 얼룩은 보이지만 노출된 피부의 상처는 두드러지지 않습니다. 상판에는 주황색과 회색의 파손 외장, 금속 프레임, 원통형 부품, 회로기판이 분리되어 있습니다. 오른쪽에는 비교적 큰 파손 외장 덩어리가 남아 있으나 완성된 인체형 몸통이나 팔다리 자세는 보이지 않습니다. 모니터에는 세계 연결망과 시설 영상이 있으나 지정된 네 지역의 찰리 에너지원 유닛을 각각 식별하기는 어렵습니다. 명확하게 읽히는 글자는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "엉덩이와 허벅지는 의자 좌판에 지지되고 상체는 자연스럽게 앞으로 숙여져 있습니다. 팔은 몸 앞쪽 아래로 내려가며 손은 가려져 있어 접촉 상태를 확인할 수 없습니다. 모든 부품은 상판 위에 놓여 있고, 벽면 모니터와 천장 조명도 구조물에 고정되어 있습니다. 근거 없이 떠 있는 몸이나 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.679
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.429
   },
   "violations": {
    "B": [
     "[gemini-pro] leaked markers/diagrams/text (모니터에 Europe, Africa 등 읽을 수 있는 텍스트가 노출됨)",
     "[gpt-high] 원래 모니터 아래의 제단 테이블과 별개로 전경에 부품 테이블을 추가해 지정된 핵심 공간 설비를 중복했습니다.",
     "[gpt-high] 모니터의 지역명이 읽혀 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반합니다.",
     "[gpt-high] 분리된 부품만 놓여 있어야 하는데 중앙에 흉부와 어깨가 연결된 몸통 조립체가 남아 있습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 '뒷모습' 앵글과 캐릭터의 흰색 의상을 정확히 반영했으며, 지정된 공간 구성과 텍스트 금지 조건을 모두 잘 준수함."
   },
   {
    "label": "B",
    "score": 429,
    "verdict_ko": "명시적으로 금지된 읽을 수 있는 텍스트(지도 상의 지역명)가 노출되었으며, 요구된 뒷모습이 아닌 측면 앵글로 묘사되고 의상도 불일치함.  ★위반: [gemini-pro] leaked markers/diagrams/text (모니터에 Europe, Africa 등 읽을 수 있는 텍스트가 노출됨) / [gpt-high] 원래 모니터 아래의 제단 테이블과 별개로 전경에 부품 테이블을 추가해 지정된 핵심 공간 설비를 중복했습니다. / [gpt-high] 모니터의 지역명이 읽혀 이미지 어디에도 읽을 수 있는 글자를 두지 말라는 조건을 위반합니다. / [gpt-high] 분리된 부품만 놓여 있어야 하는데 중앙에 흉부와 어깨가 연결된 몸통 조립체가 남아 있습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/episodes/089c4a91-4d48-420f-9ded-550269221fa0/images/background_chain/L143B01.png",
    "asset_id": "b5b4d60b-4572-4d4c-83bf-d4c6bae59496",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:811095>",
    "asset_id": "cd8664ae-a4d7-4fd8-96f0-53401108d65a",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 찰리의 몸체 잔해와 분리 부품: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1622352>",
    "asset_id": "c42224c8-15cc-4469-ba02-214fb735bdd8",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf86-0711-7654-a56c-8870d414a1a6",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S90sh3__bgfirst_bg.png",
   "bg_asset_id": "6ee93a75-a171-4dbe-807f-f7cc6d367594",
   "bg_record_key": "S90sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S90sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:39:28.285556+00:00",
  "fingerprint": "986d9d88cd5a3de82fd8d893beb13dd93ea7809d00139658c38edd72173e849a",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S90sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S90sh3_sel.png",
  "source_sha256": "b1c6ea6316476c751e00b95796ddcb55557d484476c4da41fc7c092f5da6d068",
  "file": "S90sh3_cine.png",
  "staged_sha256": "8d9366d076971f1fde8f6caad7f8856896dc14bc879a2d7cf23dbdebb6a71950",
  "latency_ms": 11150
 },
 "S90sh11::signage": {
  "fp": "089f8bc87199f709",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S90sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:40:29.661984+00:00",
  "fingerprint": "db6c114ac5f91a34eb3d9cb4d6b17211f9d77d0c33692770f8c3bf8e98f1577f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S90sh11_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S90sh11_sel.png",
  "source_sha256": "57fefa3d6dc7be06c2f9d377a2a4a78bf3a7946ea002c41ed5a89c7a0c0104b8",
  "file": "S90sh11_cine.png",
  "staged_sha256": "287e8cea1098c3ef4a253f0d19b5de78eab25521d72ad35c26f99b7dae1dd8e7",
  "latency_ms": 10052
 },
 "S90sh18::signage": {
  "fp": "48ba5fbd4375abce",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S90sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:41:25.974060+00:00",
  "fingerprint": "29dd0e2260b84a34f435bb78a14d96d510bf4e7aeb90cee405e6d47c6cfeb45f",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S90sh18_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S90sh18_sel.png",
  "source_sha256": "211c623a6e78ab2c1a82e569f6e5da845a2d3fbcbb5ccb072d8c2f5abc974040",
  "file": "S90sh18_cine.png",
  "staged_sha256": "1c719cf9a5f870ef10b796a753061d417878615ce5c05cb64a3334aaf1f97d09",
  "latency_ms": 11313
 },
 "S91sh3::signage": {
  "fp": "3910cf8ad5321e09",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::caee92de16d80d21": {
  "subjects": [],
  "subject_text": "에필로그 섬바위 주변 얕은 바다\n섬바위 둘레로 투명한 바닷물이 펼쳐진 얕은 해역. 수면 아래 물고기가 선명하게 보이며 가까운 바위 쪽에는 식탁이 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L144",
  "scope_role": "location_exterior",
  "scope_sha": "47171111828a2ba6"
 },
 "groupbg::islet_fishing_rocks": {
  "input_fingerprint": "64768149956dc75b",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "islet_fishing_rocks",
    "tags": [
     "S91sh10",
     "S91sh3",
     "S91sh7"
    ]
   },
   "context_sig": "181a5a774602c7e1"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: In clear shallow seawater immediately beside a rocky island outcrop, outdoors in daylight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n에필로그 섬바위 주변 얕은 바다: 바닥이 투명하게 들여다보이는 맑은 물속에 물고기들이 헤엄치고 기계 다리가 물장구를 치는 평화로운 해변. (특징: 속이 비치는 깨끗하고 얕은 바닷물; 물속을 유유히 헤엄치는 자연 물고기들; 수면 위로 튀어 오르는 물보라)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 91. 에필로그 / 어느 바다 섬바위 – D 자막: 한 달 후\n- 앰버와 라울도 함께 고기를 잡는다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved: In clear shallow seawater immediately beside a rocky island outcrop, outdoors in daylight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n에필로그 섬바위 주변 얕은 바다: 바닥이 투명하게 들여다보이는 맑은 물속에 물고기들이 헤엄치고 기계 다리가 물장구를 치는 평화로운 해변. (특징: 속이 비치는 깨끗하고 얕은 바닷물; 물속을 유유히 헤엄치는 자연 물고기들; 수면 위로 튀어 오르는 물보라)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 91. 에필로그 / 어느 바다 섬바위 – D 자막: 한 달 후\n- 앰버와 라울도 함께 고기를 잡는다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_islet_fishing_rocks_cfb6f6.png",
  "asset_id": "32dc4e39-740f-414c-9b0c-efd8c44ae18f",
  "input_asset_ids": [
   "2d7c3ffa-9387-43ec-b144-f1208b52f575"
  ],
  "origin_tag": "S91sh3",
  "place_text": "In clear shallow seawater immediately beside a rocky island outcrop, outdoors in daylight.",
  "origin_inputs": {
   "place_text": "In clear shallow seawater immediately beside a rocky island outcrop, outdoors in daylight.",
   "time_of_day_en": "day",
   "conti_asset_id": "2d7c3ffa-9387-43ec-b144-f1208b52f575"
  }
 },
 "S91sh3::bgfirst_bg": {
  "input_fingerprint": "29d27f60d7229161",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 튀어 오르는 물보라 속에서 앰버와 라울이 양손으로 작은 그물을 쥔 채 활짝 웃는 역동적인 구도.\n\nLOCATION (lock): In clear shallow seawater immediately beside a rocky island outcrop, outdoors in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Small fishing net (Held with both participants' hands during fishing) — Seen obliquely between the two bodies, below their faces; used as Connects their asynchronous gestures without obscuring either pair of hands; Seawater (Clear water disturbed by fishing, with splashes above its surface) — The near surface is visible beneath the hands; movement interrupts the view into the water; used as Supplies foreground motion and establishes the activity's physical setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the smiling faces and splashing seawater clearly separated, giving the fishing moment an open, lively contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 튀어 오르는 물보라 속에서 앰버와 라울이 양손으로 작은 그물을 쥔 채 활짝 웃는 역동적인 구도.\n\nLOCATION (lock): In clear shallow seawater immediately beside a rocky island outcrop, outdoors in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Small fishing net (Held with both participants' hands during fishing) — Seen obliquely between the two bodies, below their faces; used as Connects their asynchronous gestures without obscuring either pair of hands; Seawater (Clear water disturbed by fishing, with splashes above its surface) — The near surface is visible beneath the hands; movement interrupts the view into the water; used as Supplies foreground motion and establishes the activity's physical setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the smiling faces and splashing seawater clearly separated, giving the fishing moment an open, lively contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S91sh3__bgfirst_bg.png",
  "asset_id": "022958ff-3960-4516-93f5-9904bf4de833",
  "input_asset_ids": [
   "2d7c3ffa-9387-43ec-b144-f1208b52f575",
   "32dc4e39-740f-414c-9b0c-efd8c44ae18f"
  ]
 },
 "S91sh3": {
  "input_fingerprint": "c8b0200a2c55aeed",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 튀어 오르는 물보라 속에서 앰버와 라울이 양손으로 작은 그물을 쥔 채 활짝 웃는 역동적인 구도.\n\nLOCATION (lock): In clear shallow seawater immediately beside a rocky island outcrop, outdoors in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Small fishing net (Held with both participants' hands during fishing) — Seen obliquely between the two bodies, below their faces; used as Connects their asynchronous gestures without obscuring either pair of hands; Seawater (Clear water disturbed by fishing, with splashes above its surface) — The near surface is visible beneath the hands; movement interrupts the view into the water; used as Supplies foreground motion and establishes the activity's physical setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the smiling faces and splashing seawater clearly separated, giving the fishing moment an open, lively contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): One month later, clear seawater reveals fish below the surface beside the island rocks, and food is being arranged on a dining table. Charlie is reconstructed and active again, attempting to catch fish. 앰버: She is at the water catching fish. 라울: He is at the water catching fish.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버와 라울 right now, so 앰버와 라울's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버와 라울: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 튀어 오르는 물보라 속에서 앰버와 라울이 양손으로 작은 그물을 쥔 채 활짝 웃는 역동적인 구도.\n\nLOCATION (lock): In clear shallow seawater immediately beside a rocky island outcrop, outdoors in daylight. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Small fishing net (Held with both participants' hands during fishing) — Seen obliquely between the two bodies, below their faces; used as Connects their asynchronous gestures without obscuring either pair of hands; Seawater (Clear water disturbed by fishing, with splashes above its surface) — The near surface is visible beneath the hands; movement interrupts the view into the water; used as Supplies foreground motion and establishes the activity's physical setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the smiling faces and splashing seawater clearly separated, giving the fishing moment an open, lively contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): One month later, clear seawater reveals fish below the surface beside the island rocks, and food is being arranged on a dining table. Charlie is reconstructed and active again, attempting to catch fish. 앰버: She is at the water catching fish. 라울: He is at the water catching fish.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버와 라울 right now, so 앰버와 라울's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버와 라울: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2069; Korean-speaking by default, with multiethnic, multinational refugee communities and explicitly stated ethnic identities preserved. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 튀어 오르는 물보라 속에서 앰버와 라울이 양손으로 작은 그물을 쥔 채 활짝 웃는 역동적인 구도.\n\nLOCATION (lock): In clear shallow seawater immediately beside a rocky island outcrop, outdoors in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Small fishing net (Held with both participants' hands during fishing) — Seen obliquely between the two bodies, below their faces; used as Connects their asynchronous gestures without obscuring either pair of hands; Seawater (Clear water disturbed by fishing, with splashes above its surface) — The near surface is visible beneath the hands; movement interrupts the view into the water; used as Supplies foreground motion and establishes the activity's physical setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the smiling faces and splashing seawater clearly separated, giving the fishing moment an open, lively contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): One month later, clear seawater reveals fish below the surface beside the island rocks, and food is being arranged on a dining table. Charlie is reconstructed and active again, attempting to catch fish. 앰버: She is at the water catching fish. 라울: He is at the water catching fish.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버와 라울 right now, so 앰버와 라울's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버와 라울: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계·백인 혼혈 얼굴, 10세 여자아이, 금발 머리, 커다란 눈); 라울 (라틴계·흑인 혼혈 얼굴, 10세 남자아이, 어린 얼굴 윤곽, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S91sh3__bgfirst_bg.png",
     "asset_id": "022958ff-3960-4516-93f5-9904bf4de833",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/conti/conti_S91sh3.png",
     "asset_id": "2d7c3ffa-9387-43ec-b144-f1208b52f575",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:877537>",
     "asset_id": "18e1f1bb-4262-4a67-9570-9b29761c3527",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_islet_fishing_rocks_cfb6f6.png",
     "asset_id": "32dc4e39-740f-414c-9b0c-efd8c44ae18f",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:877537>",
     "asset_id": "18e1f1bb-4262-4a67-9570-9b29761c3527",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:943979>",
     "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물의 시선이 모두 수면과 그물을 향해 집중되어 있음.",
    "built_space": "바위섬이 있는 얕고 맑은 바닷물 속에 두 인물이 올바르게 위치함.",
    "entities": "앰버는 흰 옷과 외양 참조가 일치함. 라울은 얼굴은 맞으나 상의를 탈의했고 꽁지머리가 아님. 물고기가 든 작은 그물이 존재함.",
    "hard_violations": [],
    "physics": "물 속에 서서 양손으로 그물의 가장자리를 안정적으로 쥐고 있으며, 튀어 오르는 물보라가 자연스럽게 지탱됨."
   },
   {
    "label": "B",
    "direction": "두 인물이 카메라 렌즈를 정면으로 응시하고 있음.",
    "built_space": "얕고 맑은 바닷물 환경이며, 뒷배경에 바위섬이 적절히 배치됨.",
    "entities": "앰버가 지정된 흰 셔츠 대신 웻수트를 입음. 라울은 녹색 셔츠를 입었으나 프롬프트에 없는 두건을 착용함. 나무 프레임의 뜰채형 그물이 있음.",
    "hard_violations": [
     "[gemini-pro] 라울의 오른손(그물 손잡이를 잡은 앞쪽 손) 손가락이 6개로 렌더링된 물리적으로 불가능한 해부학적 오류"
    ],
    "physics": "물 속에 서서 양손으로 그물 손잡이를 쥐고 자세를 유지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "라울의 의상(상의 탈의)과 헤어스타일(꽁지머리 누락)이 참조 이미지와 다르지만, 치명적 오류 없이 지시된 양손 그물잡이 액션과 물보라를 자연스럽게 연출했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "라울의 오른손 손가락이 6개로 묘사되는 치명적인 해부학적 오류가 발생했으며, 앰버가 웻수트를 입어 복장 지침을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물의 시선이 모두 수면과 그물을 향해 집중되어 있음.",
        "built_space": "바위섬이 있는 얕고 맑은 바닷물 속에 두 인물이 올바르게 위치함.",
        "entities": "앰버는 흰 옷과 외양 참조가 일치함. 라울은 얼굴은 맞으나 상의를 탈의했고 꽁지머리가 아님. 물고기가 든 작은 그물이 존재함.",
        "hard_violations": [],
        "physics": "물 속에 서서 양손으로 그물의 가장자리를 안정적으로 쥐고 있으며, 튀어 오르는 물보라가 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "두 인물이 카메라 렌즈를 정면으로 응시하고 있음.",
        "built_space": "얕고 맑은 바닷물 환경이며, 뒷배경에 바위섬이 적절히 배치됨.",
        "entities": "앰버가 지정된 흰 셔츠 대신 웻수트를 입음. 라울은 녹색 셔츠를 입었으나 프롬프트에 없는 두건을 착용함. 나무 프레임의 뜰채형 그물이 있음.",
        "hard_violations": [
         "라울의 오른손(그물 손잡이를 잡은 앞쪽 손) 손가락이 6개로 렌더링된 물리적으로 불가능한 해부학적 오류"
        ],
        "physics": "물 속에 서서 양손으로 그물 손잡이를 쥐고 자세를 유지함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "라울의 의상(상의 탈의)과 헤어스타일(꽁지머리 누락)이 참조 이미지와 다르지만, 치명적 오류 없이 지시된 양손 그물잡이 액션과 물보라를 자연스럽게 연출했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "라울의 오른손 손가락이 6개로 묘사되는 치명적인 해부학적 오류가 발생했으며, 앰버가 웻수트를 입어 복장 지침을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물의 시선이 모두 수면과 그물을 향해 집중되어 있음.",
        "built_space": "바위섬이 있는 얕고 맑은 바닷물 속에 두 인물이 올바르게 위치함.",
        "entities": "앰버는 흰 옷과 외양 참조가 일치함. 라울은 얼굴은 맞으나 상의를 탈의했고 꽁지머리가 아님. 물고기가 든 작은 그물이 존재함.",
        "hard_violations": [],
        "physics": "물 속에 서서 양손으로 그물의 가장자리를 안정적으로 쥐고 있으며, 튀어 오르는 물보라가 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "두 인물이 카메라 렌즈를 정면으로 응시하고 있음.",
        "built_space": "얕고 맑은 바닷물 환경이며, 뒷배경에 바위섬이 적절히 배치됨.",
        "entities": "앰버가 지정된 흰 셔츠 대신 웻수트를 입음. 라울은 녹색 셔츠를 입었으나 프롬프트에 없는 두건을 착용함. 나무 프레임의 뜰채형 그물이 있음.",
        "hard_violations": [
         "라울의 오른손(그물 손잡이를 잡은 앞쪽 손) 손가락이 6개로 렌더링된 물리적으로 불가능한 해부학적 오류"
        ],
        "physics": "물 속에 서서 양손으로 그물 손잡이를 쥐고 자세를 유지함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 아이가 그물을 내려다보며 서로 다른 높이로 양손을 움직이는 미디엄 숏이 핵심 낚시 동작에 더 충실하지만, 라울의 상의와 꽁지머리는 참조와 다릅니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "양손으로 든 그물과 활짝 웃는 표정은 맞지만, 둘 다 카메라를 바라보는 대칭적인 자세라 낚시 중의 비동시적 동작이 약하고 앰버의 의상도 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버와 라울 모두 그물이나 물속이 아니라 카메라를 향해 활짝 웃습니다. 네 손은 가운데 그물의 테두리와 손잡이를 잡고 있지만, 양쪽 팔의 높이와 동작이 거의 대칭이라 서로 다른 순간의 낚시 동작보다는 함께 들어 보이는 자세입니다.",
        "built_space": "인공 구조물이나 고정 설비는 없습니다. 오른쪽의 갈라진 바위 절벽, 왼쪽 먼 작은 섬, 앞쪽 투명한 바닷물이 장소 참조의 주요 배치를 따릅니다. 두 아이의 상반신을 담은 미디엄 숏이며 그물은 얼굴 아래 중앙에 있습니다. 그물 입구가 카메라에 상당히 정면으로 제시됩니다.",
        "entities": "아이 두 명만 보입니다. 왼쪽은 금발의 약 10세 여자아이로 앰버의 얼굴과 체격에 대체로 부합하지만, 흰 단추 상의 대신 카키색 지퍼 상의를 입었습니다. 오른쪽은 짙은 피부와 곱슬머리의 약 10세 남자아이로 라울의 외형에 대체로 부합하며, 낡은 올리브색 상의는 맞지만 참조에 없는 머리띠가 있습니다. 혼혈 배경 자체는 외형만으로 확정할 수 없습니다. 중앙에 나무 테두리의 바구니형 그물 하나와 물보라가 보이며 물고기는 뚜렷하지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "그물은 각 아이의 두 손, 총 네 손이 실제로 잡고 있어 지지가 분명합니다. 몸 아래쪽은 물에 잠겨 있고 수중에 굽힌 다리 일부가 보여 얕은 바닥에 몸을 낮춘 자세로 해석됩니다. 발의 접지는 선명하지 않지만 공중에 떠 있는 몸은 아닙니다. 중앙 물보라는 수면과 그물 부근에서 솟아 올라 낚시 동작으로 발생할 수 있으나, 정지한 듯 대칭적인 팔 자세 때문에 그 발생 동작은 덜 명확합니다."
       },
       {
        "label": "B",
        "direction": "두 아이의 시선이 얼굴 아래의 그물과 잡힌 물고기를 향합니다. 앰버는 왼쪽 그물 끝을 높게, 라울은 오른쪽 부분을 다른 높이로 잡고 있어 비동시적인 손동작이 보입니다. 그물은 두 몸 사이에서 비스듬히 펼쳐져 있고 네 손이 모두 드러납니다.",
        "built_space": "인공 구조물이나 고정 설비는 없습니다. 오른쪽 소나무가 있는 밝은 바위 절벽, 왼쪽 먼 섬, 가운데 흩어진 낮은 암초와 투명한 전경 수면이 장소 참조에 잘 대응합니다. 두 아이의 머리부터 허리 부근까지 담은 미디엄 숏이며, 손 바로 아래 수면을 물보라가 가로막아 지정된 전경 관계를 구현합니다.",
        "entities": "금발의 약 10세 여자아이와 짙은 피부의 약 10세 남자아이 두 명만 있습니다. 앰버의 긴 금발, 어린 얼굴과 흰 단추 상의는 참조에 가깝습니다. 라울의 어린 얼굴과 곱슬머리는 대체로 부합하지만 뒤로 묶은 꽁지머리는 확인되지 않고, 참조의 올리브색 상의 없이 상체를 드러냈습니다. 명시된 혼혈 배경은 외형만으로 확정할 수 없습니다. 작은 유연한 그물 하나, 그물 안의 물고기들과 전경 수중 물고기들, 물보라가 보입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "각 아이가 양손으로 그물 가장자리를 직접 잡고 있으며, 그물 중앙은 물고기의 무게를 받아 아래로 처집니다. 아이들의 몸은 수면 아래로 이어져 얕은 물에 서서 앞으로 기울이는 자세로 읽히고, 발은 수면과 화면 경계에 가려져 있습니다. 그물 속 물고기는 망에 받쳐지고 수중 물고기는 물속에 있습니다. 물보라는 당겨 올리는 그물과 몸 주변 수면에서 발생해 동작과 연결됩니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 아이가 그물을 내려다보며 서로 다른 높이로 양손을 움직이는 미디엄 숏이 핵심 낚시 동작에 더 충실하지만, 라울의 상의와 꽁지머리는 참조와 다릅니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "양손으로 든 그물과 활짝 웃는 표정은 맞지만, 둘 다 카메라를 바라보는 대칭적인 자세라 낚시 중의 비동시적 동작이 약하고 앰버의 의상도 다릅니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버와 라울 모두 그물이나 물속이 아니라 카메라를 향해 활짝 웃습니다. 네 손은 가운데 그물의 테두리와 손잡이를 잡고 있지만, 양쪽 팔의 높이와 동작이 거의 대칭이라 서로 다른 순간의 낚시 동작보다는 함께 들어 보이는 자세입니다.",
        "built_space": "인공 구조물이나 고정 설비는 없습니다. 오른쪽의 갈라진 바위 절벽, 왼쪽 먼 작은 섬, 앞쪽 투명한 바닷물이 장소 참조의 주요 배치를 따릅니다. 두 아이의 상반신을 담은 미디엄 숏이며 그물은 얼굴 아래 중앙에 있습니다. 그물 입구가 카메라에 상당히 정면으로 제시됩니다.",
        "entities": "아이 두 명만 보입니다. 왼쪽은 금발의 약 10세 여자아이로 앰버의 얼굴과 체격에 대체로 부합하지만, 흰 단추 상의 대신 카키색 지퍼 상의를 입었습니다. 오른쪽은 짙은 피부와 곱슬머리의 약 10세 남자아이로 라울의 외형에 대체로 부합하며, 낡은 올리브색 상의는 맞지만 참조에 없는 머리띠가 있습니다. 혼혈 배경 자체는 외형만으로 확정할 수 없습니다. 중앙에 나무 테두리의 바구니형 그물 하나와 물보라가 보이며 물고기는 뚜렷하지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "그물은 각 아이의 두 손, 총 네 손이 실제로 잡고 있어 지지가 분명합니다. 몸 아래쪽은 물에 잠겨 있고 수중에 굽힌 다리 일부가 보여 얕은 바닥에 몸을 낮춘 자세로 해석됩니다. 발의 접지는 선명하지 않지만 공중에 떠 있는 몸은 아닙니다. 중앙 물보라는 수면과 그물 부근에서 솟아 올라 낚시 동작으로 발생할 수 있으나, 정지한 듯 대칭적인 팔 자세 때문에 그 발생 동작은 덜 명확합니다."
       },
       {
        "label": "A",
        "direction": "두 아이의 시선이 얼굴 아래의 그물과 잡힌 물고기를 향합니다. 앰버는 왼쪽 그물 끝을 높게, 라울은 오른쪽 부분을 다른 높이로 잡고 있어 비동시적인 손동작이 보입니다. 그물은 두 몸 사이에서 비스듬히 펼쳐져 있고 네 손이 모두 드러납니다.",
        "built_space": "인공 구조물이나 고정 설비는 없습니다. 오른쪽 소나무가 있는 밝은 바위 절벽, 왼쪽 먼 섬, 가운데 흩어진 낮은 암초와 투명한 전경 수면이 장소 참조에 잘 대응합니다. 두 아이의 머리부터 허리 부근까지 담은 미디엄 숏이며, 손 바로 아래 수면을 물보라가 가로막아 지정된 전경 관계를 구현합니다.",
        "entities": "금발의 약 10세 여자아이와 짙은 피부의 약 10세 남자아이 두 명만 있습니다. 앰버의 긴 금발, 어린 얼굴과 흰 단추 상의는 참조에 가깝습니다. 라울의 어린 얼굴과 곱슬머리는 대체로 부합하지만 뒤로 묶은 꽁지머리는 확인되지 않고, 참조의 올리브색 상의 없이 상체를 드러냈습니다. 명시된 혼혈 배경은 외형만으로 확정할 수 없습니다. 작은 유연한 그물 하나, 그물 안의 물고기들과 전경 수중 물고기들, 물보라가 보입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "각 아이가 양손으로 그물 가장자리를 직접 잡고 있으며, 그물 중앙은 물고기의 무게를 받아 아래로 처집니다. 아이들의 몸은 수면 아래로 이어져 얕은 물에 서서 앞으로 기울이는 자세로 읽히고, 발은 수면과 화면 경계에 가려져 있습니다. 그물 속 물고기는 망에 받쳐지고 수중 물고기는 물속에 있습니다. 물보라는 당겨 올리는 그물과 몸 주변 수면에서 발생해 동작과 연결됩니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.25
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.0
   },
   "violations": {
    "B": [
     "[gemini-pro] 라울의 오른손(그물 손잡이를 잡은 앞쪽 손) 손가락이 6개로 렌더링된 물리적으로 불가능한 해부학적 오류"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "라울의 의상(상의 탈의)과 헤어스타일(꽁지머리 누락)이 참조 이미지와 다르지만, 치명적 오류 없이 지시된 양손 그물잡이 액션과 물보라를 자연스럽게 연출했습니다."
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "라울의 오른손 손가락이 6개로 묘사되는 치명적인 해부학적 오류가 발생했으며, 앰버가 웻수트를 입어 복장 지침을 위반했습니다.  ★위반: [gemini-pro] 라울의 오른손(그물 손잡이를 잡은 앞쪽 손) 손가락이 6개로 렌더링된 물리적으로 불가능한 해부학적 오류"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/groupbg_islet_fishing_rocks_cfb6f6.png",
    "asset_id": "32dc4e39-740f-414c-9b0c-efd8c44ae18f",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:877537>",
    "asset_id": "18e1f1bb-4262-4a67-9570-9b29761c3527",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:943979>",
    "asset_id": "f243f48a-63d7-4ff0-93e1-50a8a167e7fa",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06aacf94-1aec-752a-b921-f0c3bb1e88ff",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/edfaa1fa-6366-4bfc-bca4-4403163b092e/images/089c4a91-4d48-420f-9ded-550269221fa0/scene/recipe/S91sh3__bgfirst_bg.png",
   "bg_asset_id": "022958ff-3960-4516-93f5-9904bf4de833",
   "bg_record_key": "S91sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "islet_fishing_rocks",
   "groupbg_asset_id": "32dc4e39-740f-414c-9b0c-efd8c44ae18f"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S91sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:44:05.464584+00:00",
  "fingerprint": "6faf5686b17a27a416d25d29b4f4d9106a5f389c65a3a17a6b5a5637739c4f89",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S91sh3_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S91sh3_sel.png",
  "source_sha256": "ec1c871ff9785fc512606e4df2f77011b0d7856b9cf4ff754ba5b0eadca117e2",
  "file": "S91sh3_cine.png",
  "staged_sha256": "1d7898fa34374c5416a819b10d12ffd24a0e6613c4e7be479cec6f679b7476ce",
  "latency_ms": 10152
 },
 "S91sh7::signage": {
  "fp": "4da8d0d6a23f5caa",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S91sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:45:03.611523+00:00",
  "fingerprint": "7f1744a80a250b29a0a9027fc3d527e5e463d49983937ab9695ddce075e531b1",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S91sh7_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S91sh7_sel.png",
  "source_sha256": "9e41603d4789d6b4466d41f5514476a7411a45e6365111785fad5b48405e0290",
  "file": "S91sh7_cine.png",
  "staged_sha256": "a6be5c75eb46dafaa87d8b84f1f4732eb220d8ed97948f4c5156d6e9a320f4b3",
  "latency_ms": 10745
 },
 "S91sh10::signage": {
  "fp": "b5601060dc788642",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S91sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-18T08:46:01.553399+00:00",
  "fingerprint": "ae3a94c6bd69eb5116cb90107f0e2914c31b6b52b20a41d93de6970c501271b9",
  "fingerprint_version": 2,
  "provider": "xai",
  "endpoint": "xai/images/edits",
  "model": "grok-imagine-image",
  "multiroll_tag": "still_S91sh10_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S91sh10_sel.png",
  "source_sha256": "baabedac88166674c828b7d5352a3ec79d91b5d15d2e67ff5cdccd06829aadd6",
  "file": "S91sh10_cine.png",
  "staged_sha256": "f2d3568c5c8fce065d7bf4d0910bda59877cd032acafabb6cd593977a044c0ad",
  "latency_ms": 10038
 }
}